drivers/md/raid5.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * raid5.c : Multiple Devices driver for Linux
   4  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   5  *         Copyright (C) 1999, 2000 Ingo Molnar
   6  *         Copyright (C) 2002, 2003 H. Peter Anvin
   7  *
   8  * RAID-4/5/6 management functions.
   9  * Thanks to Penguin Computing for making the RAID-6 development possible
  10  * by donating a test server!
  11  */
  12
  13 /*
  14  * BITMAP UNPLUGGING:
  15  *
  16  * The sequencing for updating the bitmap reliably is a little
  17  * subtle (and I got it wrong the first time) so it deserves some
  18  * explanation.
  19  *
  20  * We group bitmap updates into batches.  Each batch has a number.
  21  * We may write out several batches at once, but that isn't very important.
  22  * conf->seq_write is the number of the last batch successfully written.
  23  * conf->seq_flush is the number of the last batch that was closed to
  24  *    new additions.
  25  * When we discover that we will need to write to any block in a stripe
  26  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
  27  * the number of the batch it will be in. This is seq_flush+1.
  28  * When we are ready to do a write, if that batch hasn't been written yet,
  29  *   we plug the array and queue the stripe for later.
  30  * When an unplug happens, we increment bm_flush, thus closing the current
  31  *   batch.
  32  * When we notice that bm_flush > bm_write, we write out all pending updates
  33  * to the bitmap, and advance bm_write to where bm_flush was.
  34  * This may occasionally write a bit out twice, but is sure never to
  35  * miss any bits.
  36  */
  37
  38 #include <linux/blkdev.h>
  39 #include <linux/kthread.h>
  40 #include <linux/raid/pq.h>
  41 #include <linux/async_tx.h>
  42 #include <linux/module.h>
  43 #include <linux/async.h>
  44 #include <linux/seq_file.h>
  45 #include <linux/cpu.h>
  46 #include <linux/slab.h>
  47 #include <linux/ratelimit.h>
  48 #include <linux/nodemask.h>
  49
  50 #include <trace/events/block.h>
  51 #include <linux/list_sort.h>
  52
  53 #include "md.h"
  54 #include "raid5.h"
  55 #include "raid0.h"
  56 #include "md-bitmap.h"
  57 #include "raid5-log.h"
  58
  59 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
  60
  61 #define cpu_to_group(cpu) cpu_to_node(cpu)
  62 #define ANY_GROUP NUMA_NO_NODE
  63
  64 static bool devices_handle_discard_safely = false;
  65 module_param(devices_handle_discard_safely, bool, 0644);
  66 MODULE_PARM_DESC(devices_handle_discard_safely,
  67                  "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
  68 static struct workqueue_struct *raid5_wq;
  69
  70 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
  71 {
  72         int hash = (sect >> RAID5_STRIPE_SHIFT(conf)) & HASH_MASK;
  73         return &conf->stripe_hashtbl[hash];
  74 }
  75
  76 static inline int stripe_hash_locks_hash(struct r5conf *conf, sector_t sect)
  77 {
  78         return (sect >> RAID5_STRIPE_SHIFT(conf)) & STRIPE_HASH_LOCKS_MASK;
  79 }
  80
  81 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
  82 {
  83         spin_lock_irq(conf->hash_locks + hash);
  84         spin_lock(&conf->device_lock);
  85 }
  86
  87 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
  88 {
  89         spin_unlock(&conf->device_lock);
  90         spin_unlock_irq(conf->hash_locks + hash);
  91 }
  92
  93 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
  94 {
  95         int i;
  96         spin_lock_irq(conf->hash_locks);
  97         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
  98                 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
  99         spin_lock(&conf->device_lock);
 100 }
 101
 102 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
 103 {
 104         int i;
 105         spin_unlock(&conf->device_lock);
 106         for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
 107                 spin_unlock(conf->hash_locks + i);
 108         spin_unlock_irq(conf->hash_locks);
 109 }
 110
 111 /* Find first data disk in a raid6 stripe */
 112 static inline int raid6_d0(struct stripe_head *sh)
 113 {
 114         if (sh->ddf_layout)
 115                 /* ddf always start from first device */
 116                 return 0;
 117         /* md starts just after Q block */
 118         if (sh->qd_idx == sh->disks - 1)
 119                 return 0;
 120         else
 121                 return sh->qd_idx + 1;
 122 }
 123 static inline int raid6_next_disk(int disk, int raid_disks)
 124 {
 125         disk++;
 126         return (disk < raid_disks) ? disk : 0;
 127 }
 128
 129 /* When walking through the disks in a raid5, starting at raid6_d0,
 130  * We need to map each disk to a 'slot', where the data disks are slot
 131  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 132  * is raid_disks-1.  This help does that mapping.
 133  */
 134 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 135                              int *count, int syndrome_disks)
 136 {
 137         int slot = *count;
 138
 139         if (sh->ddf_layout)
 140                 (*count)++;
 141         if (idx == sh->pd_idx)
 142                 return syndrome_disks;
 143         if (idx == sh->qd_idx)
 144                 return syndrome_disks + 1;
 145         if (!sh->ddf_layout)
 146                 (*count)++;
 147         return slot;
 148 }
 149
 150 static void print_raid5_conf (struct r5conf *conf);
 151
 152 static int stripe_operations_active(struct stripe_head *sh)
 153 {
 154         return sh->check_state || sh->reconstruct_state ||
 155                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 156                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 157 }
 158
 159 static bool stripe_is_lowprio(struct stripe_head *sh)
 160 {
 161         return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
 162                 test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
 163                !test_bit(STRIPE_R5C_CACHING, &sh->state);
 164 }
 165
 166 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
 167 {
 168         struct r5conf *conf = sh->raid_conf;
 169         struct r5worker_group *group;
 170         int thread_cnt;
 171         int i, cpu = sh->cpu;
 172
 173         if (!cpu_online(cpu)) {
 174                 cpu = cpumask_any(cpu_online_mask);
 175                 sh->cpu = cpu;
 176         }
 177
 178         if (list_empty(&sh->lru)) {
 179                 struct r5worker_group *group;
 180                 group = conf->worker_groups + cpu_to_group(cpu);
 181                 if (stripe_is_lowprio(sh))
 182                         list_add_tail(&sh->lru, &group->loprio_list);
 183                 else
 184                         list_add_tail(&sh->lru, &group->handle_list);
 185                 group->stripes_cnt++;
 186                 sh->group = group;
 187         }
 188
 189         if (conf->worker_cnt_per_group == 0) {
 190                 md_wakeup_thread(conf->mddev->thread);
 191                 return;
 192         }
 193
 194         group = conf->worker_groups + cpu_to_group(sh->cpu);
 195
 196         group->workers[0].working = true;
 197         /* at least one worker should run to avoid race */
 198         queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
 199
 200         thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
 201         /* wakeup more workers */
 202         for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
 203                 if (group->workers[i].working == false) {
 204                         group->workers[i].working = true;
 205                         queue_work_on(sh->cpu, raid5_wq,
 206                                       &group->workers[i].work);
 207                         thread_cnt--;
 208                 }
 209         }
 210 }
 211
 212 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
 213                               struct list_head *temp_inactive_list)
 214 {
 215         int i;
 216         int injournal = 0;      /* number of date pages with R5_InJournal */
 217
 218         BUG_ON(!list_empty(&sh->lru));
 219         BUG_ON(atomic_read(&conf->active_stripes)==0);
 220
 221         if (r5c_is_writeback(conf->log))
 222                 for (i = sh->disks; i--; )
 223                         if (test_bit(R5_InJournal, &sh->dev[i].flags))
 224                                 injournal++;
 225         /*
 226          * In the following cases, the stripe cannot be released to cached
 227          * lists. Therefore, we make the stripe write out and set
 228          * STRIPE_HANDLE:
 229          *   1. when quiesce in r5c write back;
 230          *   2. when resync is requested fot the stripe.
 231          */
 232         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
 233             (conf->quiesce && r5c_is_writeback(conf->log) &&
 234              !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
 235                 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
 236                         r5c_make_stripe_write_out(sh);
 237                 set_bit(STRIPE_HANDLE, &sh->state);
 238         }
 239
 240         if (test_bit(STRIPE_HANDLE, &sh->state)) {
 241                 if (test_bit(STRIPE_DELAYED, &sh->state) &&
 242                     !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 243                         list_add_tail(&sh->lru, &conf->delayed_list);
 244                 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 245                            sh->bm_seq - conf->seq_write > 0)
 246                         list_add_tail(&sh->lru, &conf->bitmap_list);
 247                 else {
 248                         clear_bit(STRIPE_DELAYED, &sh->state);
 249                         clear_bit(STRIPE_BIT_DELAY, &sh->state);
 250                         if (conf->worker_cnt_per_group == 0) {
 251                                 if (stripe_is_lowprio(sh))
 252                                         list_add_tail(&sh->lru,
 253                                                         &conf->loprio_list);
 254                                 else
 255                                         list_add_tail(&sh->lru,
 256                                                         &conf->handle_list);
 257                         } else {
 258                                 raid5_wakeup_stripe_thread(sh);
 259                                 return;
 260                         }
 261                 }
 262                 md_wakeup_thread(conf->mddev->thread);
 263         } else {
 264                 BUG_ON(stripe_operations_active(sh));
 265                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 266                         if (atomic_dec_return(&conf->preread_active_stripes)
 267                             < IO_THRESHOLD)
 268                                 md_wakeup_thread(conf->mddev->thread);
 269                 atomic_dec(&conf->active_stripes);
 270                 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
 271                         if (!r5c_is_writeback(conf->log))
 272                                 list_add_tail(&sh->lru, temp_inactive_list);
 273                         else {
 274                                 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
 275                                 if (injournal == 0)
 276                                         list_add_tail(&sh->lru, temp_inactive_list);
 277                                 else if (injournal == conf->raid_disks - conf->max_degraded) {
 278                                         /* full stripe */
 279                                         if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
 280                                                 atomic_inc(&conf->r5c_cached_full_stripes);
 281                                         if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
 282                                                 atomic_dec(&conf->r5c_cached_partial_stripes);
 283                                         list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
 284                                         r5c_check_cached_full_stripe(conf);
 285                                 } else
 286                                         /*
 287                                          * STRIPE_R5C_PARTIAL_STRIPE is set in
 288                                          * r5c_try_caching_write(). No need to
 289                                          * set it again.
 290                                          */
 291                                         list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
 292                         }
 293                 }
 294         }
 295 }
 296
 297 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
 298                              struct list_head *temp_inactive_list)
 299 {
 300         if (atomic_dec_and_test(&sh->count))
 301                 do_release_stripe(conf, sh, temp_inactive_list);
 302 }
 303
 304 /*
 305  * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
 306  *
 307  * Be careful: Only one task can add/delete stripes from temp_inactive_list at
 308  * given time. Adding stripes only takes device lock, while deleting stripes
 309  * only takes hash lock.
 310  */
 311 static void release_inactive_stripe_list(struct r5conf *conf,
 312                                          struct list_head *temp_inactive_list,
 313                                          int hash)
 314 {
 315         int size;
 316         bool do_wakeup = false;
 317         unsigned long flags;
 318
 319         if (hash == NR_STRIPE_HASH_LOCKS) {
 320                 size = NR_STRIPE_HASH_LOCKS;
 321                 hash = NR_STRIPE_HASH_LOCKS - 1;
 322         } else
 323                 size = 1;
 324         while (size) {
 325                 struct list_head *list = &temp_inactive_list[size - 1];
 326
 327                 /*
 328                  * We don't hold any lock here yet, raid5_get_active_stripe() might
 329                  * remove stripes from the list
 330                  */
 331                 if (!list_empty_careful(list)) {
 332                         spin_lock_irqsave(conf->hash_locks + hash, flags);
 333                         if (list_empty(conf->inactive_list + hash) &&
 334                             !list_empty(list))
 335                                 atomic_dec(&conf->empty_inactive_list_nr);
 336                         list_splice_tail_init(list, conf->inactive_list + hash);
 337                         do_wakeup = true;
 338                         spin_unlock_irqrestore(conf->hash_locks + hash, flags);
 339                 }
 340                 size--;
 341                 hash--;
 342         }
 343
 344         if (do_wakeup) {
 345                 wake_up(&conf->wait_for_stripe);
 346                 if (atomic_read(&conf->active_stripes) == 0)
 347                         wake_up(&conf->wait_for_quiescent);
 348                 if (conf->retry_read_aligned)
 349                         md_wakeup_thread(conf->mddev->thread);
 350         }
 351 }
 352
 353 /* should hold conf->device_lock already */
 354 static int release_stripe_list(struct r5conf *conf,
 355                                struct list_head *temp_inactive_list)
 356 {
 357         struct stripe_head *sh, *t;
 358         int count = 0;
 359         struct llist_node *head;
 360
 361         head = llist_del_all(&conf->released_stripes);
 362         head = llist_reverse_order(head);
 363         llist_for_each_entry_safe(sh, t, head, release_list) {
 364                 int hash;
 365
 366                 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
 367                 smp_mb();
 368                 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
 369                 /*
 370                  * Don't worry the bit is set here, because if the bit is set
 371                  * again, the count is always > 1. This is true for
 372                  * STRIPE_ON_UNPLUG_LIST bit too.
 373                  */
 374                 hash = sh->hash_lock_index;
 375                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
 376                 count++;
 377         }
 378
 379         return count;
 380 }
 381
 382 void raid5_release_stripe(struct stripe_head *sh)
 383 {
 384         struct r5conf *conf = sh->raid_conf;
 385         unsigned long flags;
 386         struct list_head list;
 387         int hash;
 388         bool wakeup;
 389
 390         /* Avoid release_list until the last reference.
 391          */
 392         if (atomic_add_unless(&sh->count, -1, 1))
 393                 return;
 394
 395         if (unlikely(!conf->mddev->thread) ||
 396                 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
 397                 goto slow_path;
 398         wakeup = llist_add(&sh->release_list, &conf->released_stripes);
 399         if (wakeup)
 400                 md_wakeup_thread(conf->mddev->thread);
 401         return;
 402 slow_path:
 403         /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
 404         if (atomic_dec_and_lock_irqsave(&sh->count, &conf->device_lock, flags)) {
 405                 INIT_LIST_HEAD(&list);
 406                 hash = sh->hash_lock_index;
 407                 do_release_stripe(conf, sh, &list);
 408                 spin_unlock_irqrestore(&conf->device_lock, flags);
 409                 release_inactive_stripe_list(conf, &list, hash);
 410         }
 411 }
 412
 413 static inline void remove_hash(struct stripe_head *sh)
 414 {
 415         pr_debug("remove_hash(), stripe %llu\n",
 416                 (unsigned long long)sh->sector);
 417
 418         hlist_del_init(&sh->hash);
 419 }
 420
 421 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 422 {
 423         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 424
 425         pr_debug("insert_hash(), stripe %llu\n",
 426                 (unsigned long long)sh->sector);
 427
 428         hlist_add_head(&sh->hash, hp);
 429 }
 430
 431 /* find an idle stripe, make sure it is unhashed, and return it. */
 432 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
 433 {
 434         struct stripe_head *sh = NULL;
 435         struct list_head *first;
 436
 437         if (list_empty(conf->inactive_list + hash))
 438                 goto out;
 439         first = (conf->inactive_list + hash)->next;
 440         sh = list_entry(first, struct stripe_head, lru);
 441         list_del_init(first);
 442         remove_hash(sh);
 443         atomic_inc(&conf->active_stripes);
 444         BUG_ON(hash != sh->hash_lock_index);
 445         if (list_empty(conf->inactive_list + hash))
 446                 atomic_inc(&conf->empty_inactive_list_nr);
 447 out:
 448         return sh;
 449 }
 450
 451 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
 452 static void free_stripe_pages(struct stripe_head *sh)
 453 {
 454         int i;
 455         struct page *p;
 456
 457         /* Have not allocate page pool */
 458         if (!sh->pages)
 459                 return;
 460
 461         for (i = 0; i < sh->nr_pages; i++) {
 462                 p = sh->pages[i];
 463                 if (p)
 464                         put_page(p);
 465                 sh->pages[i] = NULL;
 466         }
 467 }
 468
 469 static int alloc_stripe_pages(struct stripe_head *sh, gfp_t gfp)
 470 {
 471         int i;
 472         struct page *p;
 473
 474         for (i = 0; i < sh->nr_pages; i++) {
 475                 /* The page have allocated. */
 476                 if (sh->pages[i])
 477                         continue;
 478
 479                 p = alloc_page(gfp);
 480                 if (!p) {
 481                         free_stripe_pages(sh);
 482                         return -ENOMEM;
 483                 }
 484                 sh->pages[i] = p;
 485         }
 486         return 0;
 487 }
 488
 489 static int
 490 init_stripe_shared_pages(struct stripe_head *sh, struct r5conf *conf, int disks)
 491 {
 492         int nr_pages, cnt;
 493
 494         if (sh->pages)
 495                 return 0;
 496
 497         /* Each of the sh->dev[i] need one conf->stripe_size */
 498         cnt = PAGE_SIZE / conf->stripe_size;
 499         nr_pages = (disks + cnt - 1) / cnt;
 500
 501         sh->pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
 502         if (!sh->pages)
 503                 return -ENOMEM;
 504         sh->nr_pages = nr_pages;
 505         sh->stripes_per_page = cnt;
 506         return 0;
 507 }
 508 #endif
 509
 510 static void shrink_buffers(struct stripe_head *sh)
 511 {
 512         int i;
 513         int num = sh->raid_conf->pool_size;
 514
 515 #if PAGE_SIZE == DEFAULT_STRIPE_SIZE
 516         for (i = 0; i < num ; i++) {
 517                 struct page *p;
 518
 519                 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
 520                 p = sh->dev[i].page;
 521                 if (!p)
 522                         continue;
 523                 sh->dev[i].page = NULL;
 524                 put_page(p);
 525         }
 526 #else
 527         for (i = 0; i < num; i++)
 528                 sh->dev[i].page = NULL;
 529         free_stripe_pages(sh); /* Free pages */
 530 #endif
 531 }
 532
 533 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
 534 {
 535         int i;
 536         int num = sh->raid_conf->pool_size;
 537
 538 #if PAGE_SIZE == DEFAULT_STRIPE_SIZE
 539         for (i = 0; i < num; i++) {
 540                 struct page *page;
 541
 542                 if (!(page = alloc_page(gfp))) {
 543                         return 1;
 544                 }
 545                 sh->dev[i].page = page;
 546                 sh->dev[i].orig_page = page;
 547                 sh->dev[i].offset = 0;
 548         }
 549 #else
 550         if (alloc_stripe_pages(sh, gfp))
 551                 return -ENOMEM;
 552
 553         for (i = 0; i < num; i++) {
 554                 sh->dev[i].page = raid5_get_dev_page(sh, i);
 555                 sh->dev[i].orig_page = sh->dev[i].page;
 556                 sh->dev[i].offset = raid5_get_page_offset(sh, i);
 557         }
 558 #endif
 559         return 0;
 560 }
 561
 562 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 563                             struct stripe_head *sh);
 564
 565 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 566 {
 567         struct r5conf *conf = sh->raid_conf;
 568         int i, seq;
 569
 570         BUG_ON(atomic_read(&sh->count) != 0);
 571         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 572         BUG_ON(stripe_operations_active(sh));
 573         BUG_ON(sh->batch_head);
 574
 575         pr_debug("init_stripe called, stripe %llu\n",
 576                 (unsigned long long)sector);
 577 retry:
 578         seq = read_seqcount_begin(&conf->gen_lock);
 579         sh->generation = conf->generation - previous;
 580         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 581         sh->sector = sector;
 582         stripe_set_idx(sector, conf, previous, sh);
 583         sh->state = 0;
 584
 585         for (i = sh->disks; i--; ) {
 586                 struct r5dev *dev = &sh->dev[i];
 587
 588                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 589                     test_bit(R5_LOCKED, &dev->flags)) {
 590                         pr_err("sector=%llx i=%d %p %p %p %p %d\n",
 591                                (unsigned long long)sh->sector, i, dev->toread,
 592                                dev->read, dev->towrite, dev->written,
 593                                test_bit(R5_LOCKED, &dev->flags));
 594                         WARN_ON(1);
 595                 }
 596                 dev->flags = 0;
 597                 dev->sector = raid5_compute_blocknr(sh, i, previous);
 598         }
 599         if (read_seqcount_retry(&conf->gen_lock, seq))
 600                 goto retry;
 601         sh->overwrite_disks = 0;
 602         insert_hash(conf, sh);
 603         sh->cpu = smp_processor_id();
 604         set_bit(STRIPE_BATCH_READY, &sh->state);
 605 }
 606
 607 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 608                                          short generation)
 609 {
 610         struct stripe_head *sh;
 611
 612         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 613         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 614                 if (sh->sector == sector && sh->generation == generation)
 615                         return sh;
 616         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 617         return NULL;
 618 }
 619
 620 /*
 621  * Need to check if array has failed when deciding whether to:
 622  *  - start an array
 623  *  - remove non-faulty devices
 624  *  - add a spare
 625  *  - allow a reshape
 626  * This determination is simple when no reshape is happening.
 627  * However if there is a reshape, we need to carefully check
 628  * both the before and after sections.
 629  * This is because some failed devices may only affect one
 630  * of the two sections, and some non-in_sync devices may
 631  * be insync in the section most affected by failed devices.
 632  */
 633 int raid5_calc_degraded(struct r5conf *conf)
 634 {
 635         int degraded, degraded2;
 636         int i;
 637
 638         rcu_read_lock();
 639         degraded = 0;
 640         for (i = 0; i < conf->previous_raid_disks; i++) {
 641                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 642                 if (rdev && test_bit(Faulty, &rdev->flags))
 643                         rdev = rcu_dereference(conf->disks[i].replacement);
 644                 if (!rdev || test_bit(Faulty, &rdev->flags))
 645                         degraded++;
 646                 else if (test_bit(In_sync, &rdev->flags))
 647                         ;
 648                 else
 649                         /* not in-sync or faulty.
 650                          * If the reshape increases the number of devices,
 651                          * this is being recovered by the reshape, so
 652                          * this 'previous' section is not in_sync.
 653                          * If the number of devices is being reduced however,
 654                          * the device can only be part of the array if
 655                          * we are reverting a reshape, so this section will
 656                          * be in-sync.
 657                          */
 658                         if (conf->raid_disks >= conf->previous_raid_disks)
 659                                 degraded++;
 660         }
 661         rcu_read_unlock();
 662         if (conf->raid_disks == conf->previous_raid_disks)
 663                 return degraded;
 664         rcu_read_lock();
 665         degraded2 = 0;
 666         for (i = 0; i < conf->raid_disks; i++) {
 667                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 668                 if (rdev && test_bit(Faulty, &rdev->flags))
 669                         rdev = rcu_dereference(conf->disks[i].replacement);
 670                 if (!rdev || test_bit(Faulty, &rdev->flags))
 671                         degraded2++;
 672                 else if (test_bit(In_sync, &rdev->flags))
 673                         ;
 674                 else
 675                         /* not in-sync or faulty.
 676                          * If reshape increases the number of devices, this
 677                          * section has already been recovered, else it
 678                          * almost certainly hasn't.
 679                          */
 680                         if (conf->raid_disks <= conf->previous_raid_disks)
 681                                 degraded2++;
 682         }
 683         rcu_read_unlock();
 684         if (degraded2 > degraded)
 685                 return degraded2;
 686         return degraded;
 687 }
 688
 689 static int has_failed(struct r5conf *conf)
 690 {
 691         int degraded;
 692
 693         if (conf->mddev->reshape_position == MaxSector)
 694                 return conf->mddev->degraded > conf->max_degraded;
 695
 696         degraded = raid5_calc_degraded(conf);
 697         if (degraded > conf->max_degraded)
 698                 return 1;
 699         return 0;
 700 }
 701
 702 struct stripe_head *
 703 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
 704                         int previous, int noblock, int noquiesce)
 705 {
 706         struct stripe_head *sh;
 707         int hash = stripe_hash_locks_hash(conf, sector);
 708         int inc_empty_inactive_list_flag;
 709
 710         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 711
 712         spin_lock_irq(conf->hash_locks + hash);
 713
 714         do {
 715                 wait_event_lock_irq(conf->wait_for_quiescent,
 716                                     conf->quiesce == 0 || noquiesce,
 717                                     *(conf->hash_locks + hash));
 718                 sh = __find_stripe(conf, sector, conf->generation - previous);
 719                 if (!sh) {
 720                         if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 721                                 sh = get_free_stripe(conf, hash);
 722                                 if (!sh && !test_bit(R5_DID_ALLOC,
 723                                                      &conf->cache_state))
 724                                         set_bit(R5_ALLOC_MORE,
 725                                                 &conf->cache_state);
 726                         }
 727                         if (noblock && sh == NULL)
 728                                 break;
 729
 730                         r5c_check_stripe_cache_usage(conf);
 731                         if (!sh) {
 732                                 set_bit(R5_INACTIVE_BLOCKED,
 733                                         &conf->cache_state);
 734                                 r5l_wake_reclaim(conf->log, 0);
 735                                 wait_event_lock_irq(
 736                                         conf->wait_for_stripe,
 737                                         !list_empty(conf->inactive_list + hash) &&
 738                                         (atomic_read(&conf->active_stripes)
 739                                          < (conf->max_nr_stripes * 3 / 4)
 740                                          || !test_bit(R5_INACTIVE_BLOCKED,
 741                                                       &conf->cache_state)),
 742                                         *(conf->hash_locks + hash));
 743                                 clear_bit(R5_INACTIVE_BLOCKED,
 744                                           &conf->cache_state);
 745                         } else {
 746                                 init_stripe(sh, sector, previous);
 747                                 atomic_inc(&sh->count);
 748                         }
 749                 } else if (!atomic_inc_not_zero(&sh->count)) {
 750                         spin_lock(&conf->device_lock);
 751                         if (!atomic_read(&sh->count)) {
 752                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 753                                         atomic_inc(&conf->active_stripes);
 754                                 BUG_ON(list_empty(&sh->lru) &&
 755                                        !test_bit(STRIPE_EXPANDING, &sh->state));
 756                                 inc_empty_inactive_list_flag = 0;
 757                                 if (!list_empty(conf->inactive_list + hash))
 758                                         inc_empty_inactive_list_flag = 1;
 759                                 list_del_init(&sh->lru);
 760                                 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 761                                         atomic_inc(&conf->empty_inactive_list_nr);
 762                                 if (sh->group) {
 763                                         sh->group->stripes_cnt--;
 764                                         sh->group = NULL;
 765                                 }
 766                         }
 767                         atomic_inc(&sh->count);
 768                         spin_unlock(&conf->device_lock);
 769                 }
 770         } while (sh == NULL);
 771
 772         spin_unlock_irq(conf->hash_locks + hash);
 773         return sh;
 774 }
 775
 776 static bool is_full_stripe_write(struct stripe_head *sh)
 777 {
 778         BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 779         return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 780 }
 781
 782 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 783                 __acquires(&sh1->stripe_lock)
 784                 __acquires(&sh2->stripe_lock)
 785 {
 786         if (sh1 > sh2) {
 787                 spin_lock_irq(&sh2->stripe_lock);
 788                 spin_lock_nested(&sh1->stripe_lock, 1);
 789         } else {
 790                 spin_lock_irq(&sh1->stripe_lock);
 791                 spin_lock_nested(&sh2->stripe_lock, 1);
 792         }
 793 }
 794
 795 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 796                 __releases(&sh1->stripe_lock)
 797                 __releases(&sh2->stripe_lock)
 798 {
 799         spin_unlock(&sh1->stripe_lock);
 800         spin_unlock_irq(&sh2->stripe_lock);
 801 }
 802
 803 /* Only freshly new full stripe normal write stripe can be added to a batch list */
 804 static bool stripe_can_batch(struct stripe_head *sh)
 805 {
 806         struct r5conf *conf = sh->raid_conf;
 807
 808         if (raid5_has_log(conf) || raid5_has_ppl(conf))
 809                 return false;
 810         return test_bit(STRIPE_BATCH_READY, &sh->state) &&
 811                 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
 812                 is_full_stripe_write(sh);
 813 }
 814
 815 /* we only do back search */
 816 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
 817 {
 818         struct stripe_head *head;
 819         sector_t head_sector, tmp_sec;
 820         int hash;
 821         int dd_idx;
 822         int inc_empty_inactive_list_flag;
 823
 824         /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
 825         tmp_sec = sh->sector;
 826         if (!sector_div(tmp_sec, conf->chunk_sectors))
 827                 return;
 828         head_sector = sh->sector - RAID5_STRIPE_SECTORS(conf);
 829
 830         hash = stripe_hash_locks_hash(conf, head_sector);
 831         spin_lock_irq(conf->hash_locks + hash);
 832         head = __find_stripe(conf, head_sector, conf->generation);
 833         if (head && !atomic_inc_not_zero(&head->count)) {
 834                 spin_lock(&conf->device_lock);
 835                 if (!atomic_read(&head->count)) {
 836                         if (!test_bit(STRIPE_HANDLE, &head->state))
 837                                 atomic_inc(&conf->active_stripes);
 838                         BUG_ON(list_empty(&head->lru) &&
 839                                !test_bit(STRIPE_EXPANDING, &head->state));
 840                         inc_empty_inactive_list_flag = 0;
 841                         if (!list_empty(conf->inactive_list + hash))
 842                                 inc_empty_inactive_list_flag = 1;
 843                         list_del_init(&head->lru);
 844                         if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 845                                 atomic_inc(&conf->empty_inactive_list_nr);
 846                         if (head->group) {
 847                                 head->group->stripes_cnt--;
 848                                 head->group = NULL;
 849                         }
 850                 }
 851                 atomic_inc(&head->count);
 852                 spin_unlock(&conf->device_lock);
 853         }
 854         spin_unlock_irq(conf->hash_locks + hash);
 855
 856         if (!head)
 857                 return;
 858         if (!stripe_can_batch(head))
 859                 goto out;
 860
 861         lock_two_stripes(head, sh);
 862         /* clear_batch_ready clear the flag */
 863         if (!stripe_can_batch(head) || !stripe_can_batch(sh))
 864                 goto unlock_out;
 865
 866         if (sh->batch_head)
 867                 goto unlock_out;
 868
 869         dd_idx = 0;
 870         while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
 871                 dd_idx++;
 872         if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
 873             bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
 874                 goto unlock_out;
 875
 876         if (head->batch_head) {
 877                 spin_lock(&head->batch_head->batch_lock);
 878                 /* This batch list is already running */
 879                 if (!stripe_can_batch(head)) {
 880                         spin_unlock(&head->batch_head->batch_lock);
 881                         goto unlock_out;
 882                 }
 883                 /*
 884                  * We must assign batch_head of this stripe within the
 885                  * batch_lock, otherwise clear_batch_ready of batch head
 886                  * stripe could clear BATCH_READY bit of this stripe and
 887                  * this stripe->batch_head doesn't get assigned, which
 888                  * could confuse clear_batch_ready for this stripe
 889                  */
 890                 sh->batch_head = head->batch_head;
 891
 892                 /*
 893                  * at this point, head's BATCH_READY could be cleared, but we
 894                  * can still add the stripe to batch list
 895                  */
 896                 list_add(&sh->batch_list, &head->batch_list);
 897                 spin_unlock(&head->batch_head->batch_lock);
 898         } else {
 899                 head->batch_head = head;
 900                 sh->batch_head = head->batch_head;
 901                 spin_lock(&head->batch_lock);
 902                 list_add_tail(&sh->batch_list, &head->batch_list);
 903                 spin_unlock(&head->batch_lock);
 904         }
 905
 906         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 907                 if (atomic_dec_return(&conf->preread_active_stripes)
 908                     < IO_THRESHOLD)
 909                         md_wakeup_thread(conf->mddev->thread);
 910
 911         if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
 912                 int seq = sh->bm_seq;
 913                 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
 914                     sh->batch_head->bm_seq > seq)
 915                         seq = sh->batch_head->bm_seq;
 916                 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
 917                 sh->batch_head->bm_seq = seq;
 918         }
 919
 920         atomic_inc(&sh->count);
 921 unlock_out:
 922         unlock_two_stripes(head, sh);
 923 out:
 924         raid5_release_stripe(head);
 925 }
 926
 927 /* Determine if 'data_offset' or 'new_data_offset' should be used
 928  * in this stripe_head.
 929  */
 930 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
 931 {
 932         sector_t progress = conf->reshape_progress;
 933         /* Need a memory barrier to make sure we see the value
 934          * of conf->generation, or ->data_offset that was set before
 935          * reshape_progress was updated.
 936          */
 937         smp_rmb();
 938         if (progress == MaxSector)
 939                 return 0;
 940         if (sh->generation == conf->generation - 1)
 941                 return 0;
 942         /* We are in a reshape, and this is a new-generation stripe,
 943          * so use new_data_offset.
 944          */
 945         return 1;
 946 }
 947
 948 static void dispatch_bio_list(struct bio_list *tmp)
 949 {
 950         struct bio *bio;
 951
 952         while ((bio = bio_list_pop(tmp)))
 953                 submit_bio_noacct(bio);
 954 }
 955
 956 static int cmp_stripe(void *priv, struct list_head *a, struct list_head *b)
 957 {
 958         const struct r5pending_data *da = list_entry(a,
 959                                 struct r5pending_data, sibling);
 960         const struct r5pending_data *db = list_entry(b,
 961                                 struct r5pending_data, sibling);
 962         if (da->sector > db->sector)
 963                 return 1;
 964         if (da->sector < db->sector)
 965                 return -1;
 966         return 0;
 967 }
 968
 969 static void dispatch_defer_bios(struct r5conf *conf, int target,
 970                                 struct bio_list *list)
 971 {
 972         struct r5pending_data *data;
 973         struct list_head *first, *next = NULL;
 974         int cnt = 0;
 975
 976         if (conf->pending_data_cnt == 0)
 977                 return;
 978
 979         list_sort(NULL, &conf->pending_list, cmp_stripe);
 980
 981         first = conf->pending_list.next;
 982
 983         /* temporarily move the head */
 984         if (conf->next_pending_data)
 985                 list_move_tail(&conf->pending_list,
 986                                 &conf->next_pending_data->sibling);
 987
 988         while (!list_empty(&conf->pending_list)) {
 989                 data = list_first_entry(&conf->pending_list,
 990                         struct r5pending_data, sibling);
 991                 if (&data->sibling == first)
 992                         first = data->sibling.next;
 993                 next = data->sibling.next;
 994
 995                 bio_list_merge(list, &data->bios);
 996                 list_move(&data->sibling, &conf->free_list);
 997                 cnt++;
 998                 if (cnt >= target)
 999                         break;
1000         }
1001         conf->pending_data_cnt -= cnt;
1002         BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
1003
1004         if (next != &conf->pending_list)
1005                 conf->next_pending_data = list_entry(next,
1006                                 struct r5pending_data, sibling);
1007         else
1008                 conf->next_pending_data = NULL;
1009         /* list isn't empty */
1010         if (first != &conf->pending_list)
1011                 list_move_tail(&conf->pending_list, first);
1012 }
1013
1014 static void flush_deferred_bios(struct r5conf *conf)
1015 {
1016         struct bio_list tmp = BIO_EMPTY_LIST;
1017
1018         if (conf->pending_data_cnt == 0)
1019                 return;
1020
1021         spin_lock(&conf->pending_bios_lock);
1022         dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
1023         BUG_ON(conf->pending_data_cnt != 0);
1024         spin_unlock(&conf->pending_bios_lock);
1025
1026         dispatch_bio_list(&tmp);
1027 }
1028
1029 static void defer_issue_bios(struct r5conf *conf, sector_t sector,
1030                                 struct bio_list *bios)
1031 {
1032         struct bio_list tmp = BIO_EMPTY_LIST;
1033         struct r5pending_data *ent;
1034
1035         spin_lock(&conf->pending_bios_lock);
1036         ent = list_first_entry(&conf->free_list, struct r5pending_data,
1037                                                         sibling);
1038         list_move_tail(&ent->sibling, &conf->pending_list);
1039         ent->sector = sector;
1040         bio_list_init(&ent->bios);
1041         bio_list_merge(&ent->bios, bios);
1042         conf->pending_data_cnt++;
1043         if (conf->pending_data_cnt >= PENDING_IO_MAX)
1044                 dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
1045
1046         spin_unlock(&conf->pending_bios_lock);
1047
1048         dispatch_bio_list(&tmp);
1049 }
1050
1051 static void
1052 raid5_end_read_request(struct bio *bi);
1053 static void
1054 raid5_end_write_request(struct bio *bi);
1055
1056 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
1057 {
1058         struct r5conf *conf = sh->raid_conf;
1059         int i, disks = sh->disks;
1060         struct stripe_head *head_sh = sh;
1061         struct bio_list pending_bios = BIO_EMPTY_LIST;
1062         bool should_defer;
1063
1064         might_sleep();
1065
1066         if (log_stripe(sh, s) == 0)
1067                 return;
1068
1069         should_defer = conf->batch_bio_dispatch && conf->group_cnt;
1070
1071         for (i = disks; i--; ) {
1072                 int op, op_flags = 0;
1073                 int replace_only = 0;
1074                 struct bio *bi, *rbi;
1075                 struct md_rdev *rdev, *rrdev = NULL;
1076
1077                 sh = head_sh;
1078                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1079                         op = REQ_OP_WRITE;
1080                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1081                                 op_flags = REQ_FUA;
1082                         if (test_bit(R5_Discard, &sh->dev[i].flags))
1083                                 op = REQ_OP_DISCARD;
1084                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1085                         op = REQ_OP_READ;
1086                 else if (test_and_clear_bit(R5_WantReplace,
1087                                             &sh->dev[i].flags)) {
1088                         op = REQ_OP_WRITE;
1089                         replace_only = 1;
1090                 } else
1091                         continue;
1092                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1093                         op_flags |= REQ_SYNC;
1094
1095 again:
1096                 bi = &sh->dev[i].req;
1097                 rbi = &sh->dev[i].rreq; /* For writing to replacement */
1098
1099                 rcu_read_lock();
1100                 rrdev = rcu_dereference(conf->disks[i].replacement);
1101                 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1102                 rdev = rcu_dereference(conf->disks[i].rdev);
1103                 if (!rdev) {
1104                         rdev = rrdev;
1105                         rrdev = NULL;
1106                 }
1107                 if (op_is_write(op)) {
1108                         if (replace_only)
1109                                 rdev = NULL;
1110                         if (rdev == rrdev)
1111                                 /* We raced and saw duplicates */
1112                                 rrdev = NULL;
1113                 } else {
1114                         if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1115                                 rdev = rrdev;
1116                         rrdev = NULL;
1117                 }
1118
1119                 if (rdev && test_bit(Faulty, &rdev->flags))
1120                         rdev = NULL;
1121                 if (rdev)
1122                         atomic_inc(&rdev->nr_pending);
1123                 if (rrdev && test_bit(Faulty, &rrdev->flags))
1124                         rrdev = NULL;
1125                 if (rrdev)
1126                         atomic_inc(&rrdev->nr_pending);
1127                 rcu_read_unlock();
1128
1129                 /* We have already checked bad blocks for reads.  Now
1130                  * need to check for writes.  We never accept write errors
1131                  * on the replacement, so we don't to check rrdev.
1132                  */
1133                 while (op_is_write(op) && rdev &&
1134                        test_bit(WriteErrorSeen, &rdev->flags)) {
1135                         sector_t first_bad;
1136                         int bad_sectors;
1137                         int bad = is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
1138                                               &first_bad, &bad_sectors);
1139                         if (!bad)
1140                                 break;
1141
1142                         if (bad < 0) {
1143                                 set_bit(BlockedBadBlocks, &rdev->flags);
1144                                 if (!conf->mddev->external &&
1145                                     conf->mddev->sb_flags) {
1146                                         /* It is very unlikely, but we might
1147                                          * still need to write out the
1148                                          * bad block log - better give it
1149                                          * a chance*/
1150                                         md_check_recovery(conf->mddev);
1151                                 }
1152                                 /*
1153                                  * Because md_wait_for_blocked_rdev
1154                                  * will dec nr_pending, we must
1155                                  * increment it first.
1156                                  */
1157                                 atomic_inc(&rdev->nr_pending);
1158                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
1159                         } else {
1160                                 /* Acknowledged bad block - skip the write */
1161                                 rdev_dec_pending(rdev, conf->mddev);
1162                                 rdev = NULL;
1163                         }
1164                 }
1165
1166                 if (rdev) {
1167                         if (s->syncing || s->expanding || s->expanded
1168                             || s->replacing)
1169                                 md_sync_acct(rdev->bdev, RAID5_STRIPE_SECTORS(conf));
1170
1171                         set_bit(STRIPE_IO_STARTED, &sh->state);
1172
1173                         bio_set_dev(bi, rdev->bdev);
1174                         bio_set_op_attrs(bi, op, op_flags);
1175                         bi->bi_end_io = op_is_write(op)
1176                                 ? raid5_end_write_request
1177                                 : raid5_end_read_request;
1178                         bi->bi_private = sh;
1179
1180                         pr_debug("%s: for %llu schedule op %d on disc %d\n",
1181                                 __func__, (unsigned long long)sh->sector,
1182                                 bi->bi_opf, i);
1183                         atomic_inc(&sh->count);
1184                         if (sh != head_sh)
1185                                 atomic_inc(&head_sh->count);
1186                         if (use_new_offset(conf, sh))
1187                                 bi->bi_iter.bi_sector = (sh->sector
1188                                                  + rdev->new_data_offset);
1189                         else
1190                                 bi->bi_iter.bi_sector = (sh->sector
1191                                                  + rdev->data_offset);
1192                         if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1193                                 bi->bi_opf |= REQ_NOMERGE;
1194
1195                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1196                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1197
1198                         if (!op_is_write(op) &&
1199                             test_bit(R5_InJournal, &sh->dev[i].flags))
1200                                 /*
1201                                  * issuing read for a page in journal, this
1202                                  * must be preparing for prexor in rmw; read
1203                                  * the data into orig_page
1204                                  */
1205                                 sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1206                         else
1207                                 sh->dev[i].vec.bv_page = sh->dev[i].page;
1208                         bi->bi_vcnt = 1;
1209                         bi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1210                         bi->bi_io_vec[0].bv_offset = sh->dev[i].offset;
1211                         bi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1212                         bi->bi_write_hint = sh->dev[i].write_hint;
1213                         if (!rrdev)
1214                                 sh->dev[i].write_hint = RWH_WRITE_LIFE_NOT_SET;
1215                         /*
1216                          * If this is discard request, set bi_vcnt 0. We don't
1217                          * want to confuse SCSI because SCSI will replace payload
1218                          */
1219                         if (op == REQ_OP_DISCARD)
1220                                 bi->bi_vcnt = 0;
1221                         if (rrdev)
1222                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1223
1224                         if (conf->mddev->gendisk)
1225                                 trace_block_bio_remap(bi,
1226                                                 disk_devt(conf->mddev->gendisk),
1227                                                 sh->dev[i].sector);
1228                         if (should_defer && op_is_write(op))
1229                                 bio_list_add(&pending_bios, bi);
1230                         else
1231                                 submit_bio_noacct(bi);
1232                 }
1233                 if (rrdev) {
1234                         if (s->syncing || s->expanding || s->expanded
1235                             || s->replacing)
1236                                 md_sync_acct(rrdev->bdev, RAID5_STRIPE_SECTORS(conf));
1237
1238                         set_bit(STRIPE_IO_STARTED, &sh->state);
1239
1240                         bio_set_dev(rbi, rrdev->bdev);
1241                         bio_set_op_attrs(rbi, op, op_flags);
1242                         BUG_ON(!op_is_write(op));
1243                         rbi->bi_end_io = raid5_end_write_request;
1244                         rbi->bi_private = sh;
1245
1246                         pr_debug("%s: for %llu schedule op %d on "
1247                                  "replacement disc %d\n",
1248                                 __func__, (unsigned long long)sh->sector,
1249                                 rbi->bi_opf, i);
1250                         atomic_inc(&sh->count);
1251                         if (sh != head_sh)
1252                                 atomic_inc(&head_sh->count);
1253                         if (use_new_offset(conf, sh))
1254                                 rbi->bi_iter.bi_sector = (sh->sector
1255                                                   + rrdev->new_data_offset);
1256                         else
1257                                 rbi->bi_iter.bi_sector = (sh->sector
1258                                                   + rrdev->data_offset);
1259                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1260                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1261                         sh->dev[i].rvec.bv_page = sh->dev[i].page;
1262                         rbi->bi_vcnt = 1;
1263                         rbi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1264                         rbi->bi_io_vec[0].bv_offset = sh->dev[i].offset;
1265                         rbi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1266                         rbi->bi_write_hint = sh->dev[i].write_hint;
1267                         sh->dev[i].write_hint = RWH_WRITE_LIFE_NOT_SET;
1268                         /*
1269                          * If this is discard request, set bi_vcnt 0. We don't
1270                          * want to confuse SCSI because SCSI will replace payload
1271                          */
1272                         if (op == REQ_OP_DISCARD)
1273                                 rbi->bi_vcnt = 0;
1274                         if (conf->mddev->gendisk)
1275                                 trace_block_bio_remap(rbi,
1276                                                 disk_devt(conf->mddev->gendisk),
1277                                                 sh->dev[i].sector);
1278                         if (should_defer && op_is_write(op))
1279                                 bio_list_add(&pending_bios, rbi);
1280                         else
1281                                 submit_bio_noacct(rbi);
1282                 }
1283                 if (!rdev && !rrdev) {
1284                         if (op_is_write(op))
1285                                 set_bit(STRIPE_DEGRADED, &sh->state);
1286                         pr_debug("skip op %d on disc %d for sector %llu\n",
1287                                 bi->bi_opf, i, (unsigned long long)sh->sector);
1288                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1289                         set_bit(STRIPE_HANDLE, &sh->state);
1290                 }
1291
1292                 if (!head_sh->batch_head)
1293                         continue;
1294                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1295                                       batch_list);
1296                 if (sh != head_sh)
1297                         goto again;
1298         }
1299
1300         if (should_defer && !bio_list_empty(&pending_bios))
1301                 defer_issue_bios(conf, head_sh->sector, &pending_bios);
1302 }
1303
1304 static struct dma_async_tx_descriptor *
1305 async_copy_data(int frombio, struct bio *bio, struct page **page,
1306         unsigned int poff, sector_t sector, struct dma_async_tx_descriptor *tx,
1307         struct stripe_head *sh, int no_skipcopy)
1308 {
1309         struct bio_vec bvl;
1310         struct bvec_iter iter;
1311         struct page *bio_page;
1312         int page_offset;
1313         struct async_submit_ctl submit;
1314         enum async_tx_flags flags = 0;
1315         struct r5conf *conf = sh->raid_conf;
1316
1317         if (bio->bi_iter.bi_sector >= sector)
1318                 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1319         else
1320                 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1321
1322         if (frombio)
1323                 flags |= ASYNC_TX_FENCE;
1324         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1325
1326         bio_for_each_segment(bvl, bio, iter) {
1327                 int len = bvl.bv_len;
1328                 int clen;
1329                 int b_offset = 0;
1330
1331                 if (page_offset < 0) {
1332                         b_offset = -page_offset;
1333                         page_offset += b_offset;
1334                         len -= b_offset;
1335                 }
1336
1337                 if (len > 0 && page_offset + len > RAID5_STRIPE_SIZE(conf))
1338                         clen = RAID5_STRIPE_SIZE(conf) - page_offset;
1339                 else
1340                         clen = len;
1341
1342                 if (clen > 0) {
1343                         b_offset += bvl.bv_offset;
1344                         bio_page = bvl.bv_page;
1345                         if (frombio) {
1346                                 if (conf->skip_copy &&
1347                                     b_offset == 0 && page_offset == 0 &&
1348                                     clen == RAID5_STRIPE_SIZE(conf) &&
1349                                     !no_skipcopy)
1350                                         *page = bio_page;
1351                                 else
1352                                         tx = async_memcpy(*page, bio_page, page_offset + poff,
1353                                                   b_offset, clen, &submit);
1354                         } else
1355                                 tx = async_memcpy(bio_page, *page, b_offset,
1356                                                   page_offset + poff, clen, &submit);
1357                 }
1358                 /* chain the operations */
1359                 submit.depend_tx = tx;
1360
1361                 if (clen < len) /* hit end of page */
1362                         break;
1363                 page_offset +=  len;
1364         }
1365
1366         return tx;
1367 }
1368
1369 static void ops_complete_biofill(void *stripe_head_ref)
1370 {
1371         struct stripe_head *sh = stripe_head_ref;
1372         int i;
1373         struct r5conf *conf = sh->raid_conf;
1374
1375         pr_debug("%s: stripe %llu\n", __func__,
1376                 (unsigned long long)sh->sector);
1377
1378         /* clear completed biofills */
1379         for (i = sh->disks; i--; ) {
1380                 struct r5dev *dev = &sh->dev[i];
1381
1382                 /* acknowledge completion of a biofill operation */
1383                 /* and check if we need to reply to a read request,
1384                  * new R5_Wantfill requests are held off until
1385                  * !STRIPE_BIOFILL_RUN
1386                  */
1387                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1388                         struct bio *rbi, *rbi2;
1389
1390                         BUG_ON(!dev->read);
1391                         rbi = dev->read;
1392                         dev->read = NULL;
1393                         while (rbi && rbi->bi_iter.bi_sector <
1394                                 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1395                                 rbi2 = r5_next_bio(conf, rbi, dev->sector);
1396                                 bio_endio(rbi);
1397                                 rbi = rbi2;
1398                         }
1399                 }
1400         }
1401         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1402
1403         set_bit(STRIPE_HANDLE, &sh->state);
1404         raid5_release_stripe(sh);
1405 }
1406
1407 static void ops_run_biofill(struct stripe_head *sh)
1408 {
1409         struct dma_async_tx_descriptor *tx = NULL;
1410         struct async_submit_ctl submit;
1411         int i;
1412         struct r5conf *conf = sh->raid_conf;
1413
1414         BUG_ON(sh->batch_head);
1415         pr_debug("%s: stripe %llu\n", __func__,
1416                 (unsigned long long)sh->sector);
1417
1418         for (i = sh->disks; i--; ) {
1419                 struct r5dev *dev = &sh->dev[i];
1420                 if (test_bit(R5_Wantfill, &dev->flags)) {
1421                         struct bio *rbi;
1422                         spin_lock_irq(&sh->stripe_lock);
1423                         dev->read = rbi = dev->toread;
1424                         dev->toread = NULL;
1425                         spin_unlock_irq(&sh->stripe_lock);
1426                         while (rbi && rbi->bi_iter.bi_sector <
1427                                 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1428                                 tx = async_copy_data(0, rbi, &dev->page,
1429                                                      dev->offset,
1430                                                      dev->sector, tx, sh, 0);
1431                                 rbi = r5_next_bio(conf, rbi, dev->sector);
1432                         }
1433                 }
1434         }
1435
1436         atomic_inc(&sh->count);
1437         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1438         async_trigger_callback(&submit);
1439 }
1440
1441 static void mark_target_uptodate(struct stripe_head *sh, int target)
1442 {
1443         struct r5dev *tgt;
1444
1445         if (target < 0)
1446                 return;
1447
1448         tgt = &sh->dev[target];
1449         set_bit(R5_UPTODATE, &tgt->flags);
1450         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1451         clear_bit(R5_Wantcompute, &tgt->flags);
1452 }
1453
1454 static void ops_complete_compute(void *stripe_head_ref)
1455 {
1456         struct stripe_head *sh = stripe_head_ref;
1457
1458         pr_debug("%s: stripe %llu\n", __func__,
1459                 (unsigned long long)sh->sector);
1460
1461         /* mark the computed target(s) as uptodate */
1462         mark_target_uptodate(sh, sh->ops.target);
1463         mark_target_uptodate(sh, sh->ops.target2);
1464
1465         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1466         if (sh->check_state == check_state_compute_run)
1467                 sh->check_state = check_state_compute_result;
1468         set_bit(STRIPE_HANDLE, &sh->state);
1469         raid5_release_stripe(sh);
1470 }
1471
1472 /* return a pointer to the address conversion region of the scribble buffer */
1473 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1474 {
1475         return percpu->scribble + i * percpu->scribble_obj_size;
1476 }
1477
1478 /* return a pointer to the address conversion region of the scribble buffer */
1479 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1480                                  struct raid5_percpu *percpu, int i)
1481 {
1482         return (void *) (to_addr_page(percpu, i) + sh->disks + 2);
1483 }
1484
1485 /*
1486  * Return a pointer to record offset address.
1487  */
1488 static unsigned int *
1489 to_addr_offs(struct stripe_head *sh, struct raid5_percpu *percpu)
1490 {
1491         return (unsigned int *) (to_addr_conv(sh, percpu, 0) + sh->disks + 2);
1492 }
1493
1494 static struct dma_async_tx_descriptor *
1495 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1496 {
1497         int disks = sh->disks;
1498         struct page **xor_srcs = to_addr_page(percpu, 0);
1499         unsigned int *off_srcs = to_addr_offs(sh, percpu);
1500         int target = sh->ops.target;
1501         struct r5dev *tgt = &sh->dev[target];
1502         struct page *xor_dest = tgt->page;
1503         unsigned int off_dest = tgt->offset;
1504         int count = 0;
1505         struct dma_async_tx_descriptor *tx;
1506         struct async_submit_ctl submit;
1507         int i;
1508
1509         BUG_ON(sh->batch_head);
1510
1511         pr_debug("%s: stripe %llu block: %d\n",
1512                 __func__, (unsigned long long)sh->sector, target);
1513         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1514
1515         for (i = disks; i--; ) {
1516                 if (i != target) {
1517                         off_srcs[count] = sh->dev[i].offset;
1518                         xor_srcs[count++] = sh->dev[i].page;
1519                 }
1520         }
1521
1522         atomic_inc(&sh->count);
1523
1524         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1525                           ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1526         if (unlikely(count == 1))
1527                 tx = async_memcpy(xor_dest, xor_srcs[0], off_dest, off_srcs[0],
1528                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1529         else
1530                 tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
1531                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1532
1533         return tx;
1534 }
1535
1536 /* set_syndrome_sources - populate source buffers for gen_syndrome
1537  * @srcs - (struct page *) array of size sh->disks
1538  * @offs - (unsigned int) array of offset for each page
1539  * @sh - stripe_head to parse
1540  *
1541  * Populates srcs in proper layout order for the stripe and returns the
1542  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1543  * destination buffer is recorded in srcs[count] and the Q destination
1544  * is recorded in srcs[count+1]].
1545  */
1546 static int set_syndrome_sources(struct page **srcs,
1547                                 unsigned int *offs,
1548                                 struct stripe_head *sh,
1549                                 int srctype)
1550 {
1551         int disks = sh->disks;
1552         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1553         int d0_idx = raid6_d0(sh);
1554         int count;
1555         int i;
1556
1557         for (i = 0; i < disks; i++)
1558                 srcs[i] = NULL;
1559
1560         count = 0;
1561         i = d0_idx;
1562         do {
1563                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1564                 struct r5dev *dev = &sh->dev[i];
1565
1566                 if (i == sh->qd_idx || i == sh->pd_idx ||
1567                     (srctype == SYNDROME_SRC_ALL) ||
1568                     (srctype == SYNDROME_SRC_WANT_DRAIN &&
1569                      (test_bit(R5_Wantdrain, &dev->flags) ||
1570                       test_bit(R5_InJournal, &dev->flags))) ||
1571                     (srctype == SYNDROME_SRC_WRITTEN &&
1572                      (dev->written ||
1573                       test_bit(R5_InJournal, &dev->flags)))) {
1574                         if (test_bit(R5_InJournal, &dev->flags))
1575                                 srcs[slot] = sh->dev[i].orig_page;
1576                         else
1577                                 srcs[slot] = sh->dev[i].page;
1578                         /*
1579                          * For R5_InJournal, PAGE_SIZE must be 4KB and will
1580                          * not shared page. In that case, dev[i].offset
1581                          * is 0.
1582                          */
1583                         offs[slot] = sh->dev[i].offset;
1584                 }
1585                 i = raid6_next_disk(i, disks);
1586         } while (i != d0_idx);
1587
1588         return syndrome_disks;
1589 }
1590
1591 static struct dma_async_tx_descriptor *
1592 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1593 {
1594         int disks = sh->disks;
1595         struct page **blocks = to_addr_page(percpu, 0);
1596         unsigned int *offs = to_addr_offs(sh, percpu);
1597         int target;
1598         int qd_idx = sh->qd_idx;
1599         struct dma_async_tx_descriptor *tx;
1600         struct async_submit_ctl submit;
1601         struct r5dev *tgt;
1602         struct page *dest;
1603         unsigned int dest_off;
1604         int i;
1605         int count;
1606
1607         BUG_ON(sh->batch_head);
1608         if (sh->ops.target < 0)
1609                 target = sh->ops.target2;
1610         else if (sh->ops.target2 < 0)
1611                 target = sh->ops.target;
1612         else
1613                 /* we should only have one valid target */
1614                 BUG();
1615         BUG_ON(target < 0);
1616         pr_debug("%s: stripe %llu block: %d\n",
1617                 __func__, (unsigned long long)sh->sector, target);
1618
1619         tgt = &sh->dev[target];
1620         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1621         dest = tgt->page;
1622         dest_off = tgt->offset;
1623
1624         atomic_inc(&sh->count);
1625
1626         if (target == qd_idx) {
1627                 count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_ALL);
1628                 blocks[count] = NULL; /* regenerating p is not necessary */
1629                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1630                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1631                                   ops_complete_compute, sh,
1632                                   to_addr_conv(sh, percpu, 0));
1633                 tx = async_gen_syndrome(blocks, offs, count+2,
1634                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1635         } else {
1636                 /* Compute any data- or p-drive using XOR */
1637                 count = 0;
1638                 for (i = disks; i-- ; ) {
1639                         if (i == target || i == qd_idx)
1640                                 continue;
1641                         offs[count] = sh->dev[i].offset;
1642                         blocks[count++] = sh->dev[i].page;
1643                 }
1644
1645                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1646                                   NULL, ops_complete_compute, sh,
1647                                   to_addr_conv(sh, percpu, 0));
1648                 tx = async_xor_offs(dest, dest_off, blocks, offs, count,
1649                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1650         }
1651
1652         return tx;
1653 }
1654
1655 static struct dma_async_tx_descriptor *
1656 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1657 {
1658         int i, count, disks = sh->disks;
1659         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1660         int d0_idx = raid6_d0(sh);
1661         int faila = -1, failb = -1;
1662         int target = sh->ops.target;
1663         int target2 = sh->ops.target2;
1664         struct r5dev *tgt = &sh->dev[target];
1665         struct r5dev *tgt2 = &sh->dev[target2];
1666         struct dma_async_tx_descriptor *tx;
1667         struct page **blocks = to_addr_page(percpu, 0);
1668         unsigned int *offs = to_addr_offs(sh, percpu);
1669         struct async_submit_ctl submit;
1670
1671         BUG_ON(sh->batch_head);
1672         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1673                  __func__, (unsigned long long)sh->sector, target, target2);
1674         BUG_ON(target < 0 || target2 < 0);
1675         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1676         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1677
1678         /* we need to open-code set_syndrome_sources to handle the
1679          * slot number conversion for 'faila' and 'failb'
1680          */
1681         for (i = 0; i < disks ; i++) {
1682                 offs[i] = 0;
1683                 blocks[i] = NULL;
1684         }
1685         count = 0;
1686         i = d0_idx;
1687         do {
1688                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1689
1690                 offs[slot] = sh->dev[i].offset;
1691                 blocks[slot] = sh->dev[i].page;
1692
1693                 if (i == target)
1694                         faila = slot;
1695                 if (i == target2)
1696                         failb = slot;
1697                 i = raid6_next_disk(i, disks);
1698         } while (i != d0_idx);
1699
1700         BUG_ON(faila == failb);
1701         if (failb < faila)
1702                 swap(faila, failb);
1703         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1704                  __func__, (unsigned long long)sh->sector, faila, failb);
1705
1706         atomic_inc(&sh->count);
1707
1708         if (failb == syndrome_disks+1) {
1709                 /* Q disk is one of the missing disks */
1710                 if (faila == syndrome_disks) {
1711                         /* Missing P+Q, just recompute */
1712                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1713                                           ops_complete_compute, sh,
1714                                           to_addr_conv(sh, percpu, 0));
1715                         return async_gen_syndrome(blocks, offs, syndrome_disks+2,
1716                                                   RAID5_STRIPE_SIZE(sh->raid_conf),
1717                                                   &submit);
1718                 } else {
1719                         struct page *dest;
1720                         unsigned int dest_off;
1721                         int data_target;
1722                         int qd_idx = sh->qd_idx;
1723
1724                         /* Missing D+Q: recompute D from P, then recompute Q */
1725                         if (target == qd_idx)
1726                                 data_target = target2;
1727                         else
1728                                 data_target = target;
1729
1730                         count = 0;
1731                         for (i = disks; i-- ; ) {
1732                                 if (i == data_target || i == qd_idx)
1733                                         continue;
1734                                 offs[count] = sh->dev[i].offset;
1735                                 blocks[count++] = sh->dev[i].page;
1736                         }
1737                         dest = sh->dev[data_target].page;
1738                         dest_off = sh->dev[data_target].offset;
1739                         init_async_submit(&submit,
1740                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1741                                           NULL, NULL, NULL,
1742                                           to_addr_conv(sh, percpu, 0));
1743                         tx = async_xor_offs(dest, dest_off, blocks, offs, count,
1744                                        RAID5_STRIPE_SIZE(sh->raid_conf),
1745                                        &submit);
1746
1747                         count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_ALL);
1748                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1749                                           ops_complete_compute, sh,
1750                                           to_addr_conv(sh, percpu, 0));
1751                         return async_gen_syndrome(blocks, offs, count+2,
1752                                                   RAID5_STRIPE_SIZE(sh->raid_conf),
1753                                                   &submit);
1754                 }
1755         } else {
1756                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1757                                   ops_complete_compute, sh,
1758                                   to_addr_conv(sh, percpu, 0));
1759                 if (failb == syndrome_disks) {
1760                         /* We're missing D+P. */
1761                         return async_raid6_datap_recov(syndrome_disks+2,
1762                                                 RAID5_STRIPE_SIZE(sh->raid_conf),
1763                                                 faila,
1764                                                 blocks, offs, &submit);
1765                 } else {
1766                         /* We're missing D+D. */
1767                         return async_raid6_2data_recov(syndrome_disks+2,
1768                                                 RAID5_STRIPE_SIZE(sh->raid_conf),
1769                                                 faila, failb,
1770                                                 blocks, offs, &submit);
1771                 }
1772         }
1773 }
1774
1775 static void ops_complete_prexor(void *stripe_head_ref)
1776 {
1777         struct stripe_head *sh = stripe_head_ref;
1778
1779         pr_debug("%s: stripe %llu\n", __func__,
1780                 (unsigned long long)sh->sector);
1781
1782         if (r5c_is_writeback(sh->raid_conf->log))
1783                 /*
1784                  * raid5-cache write back uses orig_page during prexor.
1785                  * After prexor, it is time to free orig_page
1786                  */
1787                 r5c_release_extra_page(sh);
1788 }
1789
1790 static struct dma_async_tx_descriptor *
1791 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1792                 struct dma_async_tx_descriptor *tx)
1793 {
1794         int disks = sh->disks;
1795         struct page **xor_srcs = to_addr_page(percpu, 0);
1796         unsigned int *off_srcs = to_addr_offs(sh, percpu);
1797         int count = 0, pd_idx = sh->pd_idx, i;
1798         struct async_submit_ctl submit;
1799
1800         /* existing parity data subtracted */
1801         unsigned int off_dest = off_srcs[count] = sh->dev[pd_idx].offset;
1802         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1803
1804         BUG_ON(sh->batch_head);
1805         pr_debug("%s: stripe %llu\n", __func__,
1806                 (unsigned long long)sh->sector);
1807
1808         for (i = disks; i--; ) {
1809                 struct r5dev *dev = &sh->dev[i];
1810                 /* Only process blocks that are known to be uptodate */
1811                 if (test_bit(R5_InJournal, &dev->flags)) {
1812                         /*
1813                          * For this case, PAGE_SIZE must be equal to 4KB and
1814                          * page offset is zero.
1815                          */
1816                         off_srcs[count] = dev->offset;
1817                         xor_srcs[count++] = dev->orig_page;
1818                 } else if (test_bit(R5_Wantdrain, &dev->flags)) {
1819                         off_srcs[count] = dev->offset;
1820                         xor_srcs[count++] = dev->page;
1821                 }
1822         }
1823
1824         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1825                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1826         tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
1827                         RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1828
1829         return tx;
1830 }
1831
1832 static struct dma_async_tx_descriptor *
1833 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1834                 struct dma_async_tx_descriptor *tx)
1835 {
1836         struct page **blocks = to_addr_page(percpu, 0);
1837         unsigned int *offs = to_addr_offs(sh, percpu);
1838         int count;
1839         struct async_submit_ctl submit;
1840
1841         pr_debug("%s: stripe %llu\n", __func__,
1842                 (unsigned long long)sh->sector);
1843
1844         count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_WANT_DRAIN);
1845
1846         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1847                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1848         tx = async_gen_syndrome(blocks, offs, count+2,
1849                         RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1850
1851         return tx;
1852 }
1853
1854 static struct dma_async_tx_descriptor *
1855 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1856 {
1857         struct r5conf *conf = sh->raid_conf;
1858         int disks = sh->disks;
1859         int i;
1860         struct stripe_head *head_sh = sh;
1861
1862         pr_debug("%s: stripe %llu\n", __func__,
1863                 (unsigned long long)sh->sector);
1864
1865         for (i = disks; i--; ) {
1866                 struct r5dev *dev;
1867                 struct bio *chosen;
1868
1869                 sh = head_sh;
1870                 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1871                         struct bio *wbi;
1872
1873 again:
1874                         dev = &sh->dev[i];
1875                         /*
1876                          * clear R5_InJournal, so when rewriting a page in
1877                          * journal, it is not skipped by r5l_log_stripe()
1878                          */
1879                         clear_bit(R5_InJournal, &dev->flags);
1880                         spin_lock_irq(&sh->stripe_lock);
1881                         chosen = dev->towrite;
1882                         dev->towrite = NULL;
1883                         sh->overwrite_disks = 0;
1884                         BUG_ON(dev->written);
1885                         wbi = dev->written = chosen;
1886                         spin_unlock_irq(&sh->stripe_lock);
1887                         WARN_ON(dev->page != dev->orig_page);
1888
1889                         while (wbi && wbi->bi_iter.bi_sector <
1890                                 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1891                                 if (wbi->bi_opf & REQ_FUA)
1892                                         set_bit(R5_WantFUA, &dev->flags);
1893                                 if (wbi->bi_opf & REQ_SYNC)
1894                                         set_bit(R5_SyncIO, &dev->flags);
1895                                 if (bio_op(wbi) == REQ_OP_DISCARD)
1896                                         set_bit(R5_Discard, &dev->flags);
1897                                 else {
1898                                         tx = async_copy_data(1, wbi, &dev->page,
1899                                                              dev->offset,
1900                                                              dev->sector, tx, sh,
1901                                                              r5c_is_writeback(conf->log));
1902                                         if (dev->page != dev->orig_page &&
1903                                             !r5c_is_writeback(conf->log)) {
1904                                                 set_bit(R5_SkipCopy, &dev->flags);
1905                                                 clear_bit(R5_UPTODATE, &dev->flags);
1906                                                 clear_bit(R5_OVERWRITE, &dev->flags);
1907                                         }
1908                                 }
1909                                 wbi = r5_next_bio(conf, wbi, dev->sector);
1910                         }
1911
1912                         if (head_sh->batch_head) {
1913                                 sh = list_first_entry(&sh->batch_list,
1914                                                       struct stripe_head,
1915                                                       batch_list);
1916                                 if (sh == head_sh)
1917                                         continue;
1918                                 goto again;
1919                         }
1920                 }
1921         }
1922
1923         return tx;
1924 }
1925
1926 static void ops_complete_reconstruct(void *stripe_head_ref)
1927 {
1928         struct stripe_head *sh = stripe_head_ref;
1929         int disks = sh->disks;
1930         int pd_idx = sh->pd_idx;
1931         int qd_idx = sh->qd_idx;
1932         int i;
1933         bool fua = false, sync = false, discard = false;
1934
1935         pr_debug("%s: stripe %llu\n", __func__,
1936                 (unsigned long long)sh->sector);
1937
1938         for (i = disks; i--; ) {
1939                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1940                 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1941                 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1942         }
1943
1944         for (i = disks; i--; ) {
1945                 struct r5dev *dev = &sh->dev[i];
1946
1947                 if (dev->written || i == pd_idx || i == qd_idx) {
1948                         if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
1949                                 set_bit(R5_UPTODATE, &dev->flags);
1950                                 if (test_bit(STRIPE_EXPAND_READY, &sh->state))
1951                                         set_bit(R5_Expanded, &dev->flags);
1952                         }
1953                         if (fua)
1954                                 set_bit(R5_WantFUA, &dev->flags);
1955                         if (sync)
1956                                 set_bit(R5_SyncIO, &dev->flags);
1957                 }
1958         }
1959
1960         if (sh->reconstruct_state == reconstruct_state_drain_run)
1961                 sh->reconstruct_state = reconstruct_state_drain_result;
1962         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1963                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1964         else {
1965                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1966                 sh->reconstruct_state = reconstruct_state_result;
1967         }
1968
1969         set_bit(STRIPE_HANDLE, &sh->state);
1970         raid5_release_stripe(sh);
1971 }
1972
1973 static void
1974 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1975                      struct dma_async_tx_descriptor *tx)
1976 {
1977         int disks = sh->disks;
1978         struct page **xor_srcs;
1979         unsigned int *off_srcs;
1980         struct async_submit_ctl submit;
1981         int count, pd_idx = sh->pd_idx, i;
1982         struct page *xor_dest;
1983         unsigned int off_dest;
1984         int prexor = 0;
1985         unsigned long flags;
1986         int j = 0;
1987         struct stripe_head *head_sh = sh;
1988         int last_stripe;
1989
1990         pr_debug("%s: stripe %llu\n", __func__,
1991                 (unsigned long long)sh->sector);
1992
1993         for (i = 0; i < sh->disks; i++) {
1994                 if (pd_idx == i)
1995                         continue;
1996                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1997                         break;
1998         }
1999         if (i >= sh->disks) {
2000                 atomic_inc(&sh->count);
2001                 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
2002                 ops_complete_reconstruct(sh);
2003                 return;
2004         }
2005 again:
2006         count = 0;
2007         xor_srcs = to_addr_page(percpu, j);
2008         off_srcs = to_addr_offs(sh, percpu);
2009         /* check if prexor is active which means only process blocks
2010          * that are part of a read-modify-write (written)
2011          */
2012         if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
2013                 prexor = 1;
2014                 off_dest = off_srcs[count] = sh->dev[pd_idx].offset;
2015                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
2016                 for (i = disks; i--; ) {
2017                         struct r5dev *dev = &sh->dev[i];
2018                         if (head_sh->dev[i].written ||
2019                             test_bit(R5_InJournal, &head_sh->dev[i].flags)) {
2020                                 off_srcs[count] = dev->offset;
2021                                 xor_srcs[count++] = dev->page;
2022                         }
2023                 }
2024         } else {
2025                 xor_dest = sh->dev[pd_idx].page;
2026                 off_dest = sh->dev[pd_idx].offset;
2027                 for (i = disks; i--; ) {
2028                         struct r5dev *dev = &sh->dev[i];
2029                         if (i != pd_idx) {
2030                                 off_srcs[count] = dev->offset;
2031                                 xor_srcs[count++] = dev->page;
2032                         }
2033                 }
2034         }
2035
2036         /* 1/ if we prexor'd then the dest is reused as a source
2037          * 2/ if we did not prexor then we are redoing the parity
2038          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
2039          * for the synchronous xor case
2040          */
2041         last_stripe = !head_sh->batch_head ||
2042                 list_first_entry(&sh->batch_list,
2043                                  struct stripe_head, batch_list) == head_sh;
2044         if (last_stripe) {
2045                 flags = ASYNC_TX_ACK |
2046                         (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
2047
2048                 atomic_inc(&head_sh->count);
2049                 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
2050                                   to_addr_conv(sh, percpu, j));
2051         } else {
2052                 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
2053                 init_async_submit(&submit, flags, tx, NULL, NULL,
2054                                   to_addr_conv(sh, percpu, j));
2055         }
2056
2057         if (unlikely(count == 1))
2058                 tx = async_memcpy(xor_dest, xor_srcs[0], off_dest, off_srcs[0],
2059                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2060         else
2061                 tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
2062                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2063         if (!last_stripe) {
2064                 j++;
2065                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
2066                                       batch_list);
2067                 goto again;
2068         }
2069 }
2070
2071 static void
2072 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
2073                      struct dma_async_tx_descriptor *tx)
2074 {
2075         struct async_submit_ctl submit;
2076         struct page **blocks;
2077         unsigned int *offs;
2078         int count, i, j = 0;
2079         struct stripe_head *head_sh = sh;
2080         int last_stripe;
2081         int synflags;
2082         unsigned long txflags;
2083
2084         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
2085
2086         for (i = 0; i < sh->disks; i++) {
2087                 if (sh->pd_idx == i || sh->qd_idx == i)
2088                         continue;
2089                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
2090                         break;
2091         }
2092         if (i >= sh->disks) {
2093                 atomic_inc(&sh->count);
2094                 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
2095                 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
2096                 ops_complete_reconstruct(sh);
2097                 return;
2098         }
2099
2100 again:
2101         blocks = to_addr_page(percpu, j);
2102         offs = to_addr_offs(sh, percpu);
2103
2104         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
2105                 synflags = SYNDROME_SRC_WRITTEN;
2106                 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
2107         } else {
2108                 synflags = SYNDROME_SRC_ALL;
2109                 txflags = ASYNC_TX_ACK;
2110         }
2111
2112         count = set_syndrome_sources(blocks, offs, sh, synflags);
2113         last_stripe = !head_sh->batch_head ||
2114                 list_first_entry(&sh->batch_list,
2115                                  struct stripe_head, batch_list) == head_sh;
2116
2117         if (last_stripe) {
2118                 atomic_inc(&head_sh->count);
2119                 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
2120                                   head_sh, to_addr_conv(sh, percpu, j));
2121         } else
2122                 init_async_submit(&submit, 0, tx, NULL, NULL,
2123                                   to_addr_conv(sh, percpu, j));
2124         tx = async_gen_syndrome(blocks, offs, count+2,
2125                         RAID5_STRIPE_SIZE(sh->raid_conf),  &submit);
2126         if (!last_stripe) {
2127                 j++;
2128                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
2129                                       batch_list);
2130                 goto again;
2131         }
2132 }
2133
2134 static void ops_complete_check(void *stripe_head_ref)
2135 {
2136         struct stripe_head *sh = stripe_head_ref;
2137
2138         pr_debug("%s: stripe %llu\n", __func__,
2139                 (unsigned long long)sh->sector);
2140
2141         sh->check_state = check_state_check_result;
2142         set_bit(STRIPE_HANDLE, &sh->state);
2143         raid5_release_stripe(sh);
2144 }
2145
2146 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
2147 {
2148         int disks = sh->disks;
2149         int pd_idx = sh->pd_idx;
2150         int qd_idx = sh->qd_idx;
2151         struct page *xor_dest;
2152         unsigned int off_dest;
2153         struct page **xor_srcs = to_addr_page(percpu, 0);
2154         unsigned int *off_srcs = to_addr_offs(sh, percpu);
2155         struct dma_async_tx_descriptor *tx;
2156         struct async_submit_ctl submit;
2157         int count;
2158         int i;
2159
2160         pr_debug("%s: stripe %llu\n", __func__,
2161                 (unsigned long long)sh->sector);
2162
2163         BUG_ON(sh->batch_head);
2164         count = 0;
2165         xor_dest = sh->dev[pd_idx].page;
2166         off_dest = sh->dev[pd_idx].offset;
2167         off_srcs[count] = off_dest;
2168         xor_srcs[count++] = xor_dest;
2169         for (i = disks; i--; ) {
2170                 if (i == pd_idx || i == qd_idx)
2171                         continue;
2172                 off_srcs[count] = sh->dev[i].offset;
2173                 xor_srcs[count++] = sh->dev[i].page;
2174         }
2175
2176         init_async_submit(&submit, 0, NULL, NULL, NULL,
2177                           to_addr_conv(sh, percpu, 0));
2178         tx = async_xor_val_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
2179                            RAID5_STRIPE_SIZE(sh->raid_conf),
2180                            &sh->ops.zero_sum_result, &submit);
2181
2182         atomic_inc(&sh->count);
2183         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2184         tx = async_trigger_callback(&submit);
2185 }
2186
2187 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2188 {
2189         struct page **srcs = to_addr_page(percpu, 0);
2190         unsigned int *offs = to_addr_offs(sh, percpu);
2191         struct async_submit_ctl submit;
2192         int count;
2193
2194         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2195                 (unsigned long long)sh->sector, checkp);
2196
2197         BUG_ON(sh->batch_head);
2198         count = set_syndrome_sources(srcs, offs, sh, SYNDROME_SRC_ALL);
2199         if (!checkp)
2200                 srcs[count] = NULL;
2201
2202         atomic_inc(&sh->count);
2203         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2204                           sh, to_addr_conv(sh, percpu, 0));
2205         async_syndrome_val(srcs, offs, count+2,
2206                            RAID5_STRIPE_SIZE(sh->raid_conf),
2207                            &sh->ops.zero_sum_result, percpu->spare_page, 0, &submit);
2208 }
2209
2210 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2211 {
2212         int overlap_clear = 0, i, disks = sh->disks;
2213         struct dma_async_tx_descriptor *tx = NULL;
2214         struct r5conf *conf = sh->raid_conf;
2215         int level = conf->level;
2216         struct raid5_percpu *percpu;
2217         unsigned long cpu;
2218
2219         cpu = get_cpu();
2220         percpu = per_cpu_ptr(conf->percpu, cpu);
2221         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2222                 ops_run_biofill(sh);
2223                 overlap_clear++;
2224         }
2225
2226         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2227                 if (level < 6)
2228                         tx = ops_run_compute5(sh, percpu);
2229                 else {
2230                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
2231                                 tx = ops_run_compute6_1(sh, percpu);
2232                         else
2233                                 tx = ops_run_compute6_2(sh, percpu);
2234                 }
2235                 /* terminate the chain if reconstruct is not set to be run */
2236                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2237                         async_tx_ack(tx);
2238         }
2239
2240         if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2241                 if (level < 6)
2242                         tx = ops_run_prexor5(sh, percpu, tx);
2243                 else
2244                         tx = ops_run_prexor6(sh, percpu, tx);
2245         }
2246
2247         if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2248                 tx = ops_run_partial_parity(sh, percpu, tx);
2249
2250         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2251                 tx = ops_run_biodrain(sh, tx);
2252                 overlap_clear++;
2253         }
2254
2255         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2256                 if (level < 6)
2257                         ops_run_reconstruct5(sh, percpu, tx);
2258                 else
2259                         ops_run_reconstruct6(sh, percpu, tx);
2260         }
2261
2262         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2263                 if (sh->check_state == check_state_run)
2264                         ops_run_check_p(sh, percpu);
2265                 else if (sh->check_state == check_state_run_q)
2266                         ops_run_check_pq(sh, percpu, 0);
2267                 else if (sh->check_state == check_state_run_pq)
2268                         ops_run_check_pq(sh, percpu, 1);
2269                 else
2270                         BUG();
2271         }
2272
2273         if (overlap_clear && !sh->batch_head)
2274                 for (i = disks; i--; ) {
2275                         struct r5dev *dev = &sh->dev[i];
2276                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2277                                 wake_up(&sh->raid_conf->wait_for_overlap);
2278                 }
2279         put_cpu();
2280 }
2281
2282 static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2283 {
2284 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2285         kfree(sh->pages);
2286 #endif
2287         if (sh->ppl_page)
2288                 __free_page(sh->ppl_page);
2289         kmem_cache_free(sc, sh);
2290 }
2291
2292 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2293         int disks, struct r5conf *conf)
2294 {
2295         struct stripe_head *sh;
2296         int i;
2297
2298         sh = kmem_cache_zalloc(sc, gfp);
2299         if (sh) {
2300                 spin_lock_init(&sh->stripe_lock);
2301                 spin_lock_init(&sh->batch_lock);
2302                 INIT_LIST_HEAD(&sh->batch_list);
2303                 INIT_LIST_HEAD(&sh->lru);
2304                 INIT_LIST_HEAD(&sh->r5c);
2305                 INIT_LIST_HEAD(&sh->log_list);
2306                 atomic_set(&sh->count, 1);
2307                 sh->raid_conf = conf;
2308                 sh->log_start = MaxSector;
2309                 for (i = 0; i < disks; i++) {
2310                         struct r5dev *dev = &sh->dev[i];
2311
2312                         bio_init(&dev->req, &dev->vec, 1);
2313                         bio_init(&dev->rreq, &dev->rvec, 1);
2314                 }
2315
2316                 if (raid5_has_ppl(conf)) {
2317                         sh->ppl_page = alloc_page(gfp);
2318                         if (!sh->ppl_page) {
2319                                 free_stripe(sc, sh);
2320                                 return NULL;
2321                         }
2322                 }
2323 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2324                 if (init_stripe_shared_pages(sh, conf, disks)) {
2325                         free_stripe(sc, sh);
2326                         return NULL;
2327                 }
2328 #endif
2329         }
2330         return sh;
2331 }
2332 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2333 {
2334         struct stripe_head *sh;
2335
2336         sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2337         if (!sh)
2338                 return 0;
2339
2340         if (grow_buffers(sh, gfp)) {
2341                 shrink_buffers(sh);
2342                 free_stripe(conf->slab_cache, sh);
2343                 return 0;
2344         }
2345         sh->hash_lock_index =
2346                 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2347         /* we just created an active stripe so... */
2348         atomic_inc(&conf->active_stripes);
2349
2350         raid5_release_stripe(sh);
2351         conf->max_nr_stripes++;
2352         return 1;
2353 }
2354
2355 static int grow_stripes(struct r5conf *conf, int num)
2356 {
2357         struct kmem_cache *sc;
2358         size_t namelen = sizeof(conf->cache_name[0]);
2359         int devs = max(conf->raid_disks, conf->previous_raid_disks);
2360
2361         if (conf->mddev->gendisk)
2362                 snprintf(conf->cache_name[0], namelen,
2363                         "raid%d-%s", conf->level, mdname(conf->mddev));
2364         else
2365                 snprintf(conf->cache_name[0], namelen,
2366                         "raid%d-%p", conf->level, conf->mddev);
2367         snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2368
2369         conf->active_name = 0;
2370         sc = kmem_cache_create(conf->cache_name[conf->active_name],
2371                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2372                                0, 0, NULL);
2373         if (!sc)
2374                 return 1;
2375         conf->slab_cache = sc;
2376         conf->pool_size = devs;
2377         while (num--)
2378                 if (!grow_one_stripe(conf, GFP_KERNEL))
2379                         return 1;
2380
2381         return 0;
2382 }
2383
2384 /**
2385  * scribble_alloc - allocate percpu scribble buffer for required size
2386  *                  of the scribble region
2387  * @percpu: from for_each_present_cpu() of the caller
2388  * @num: total number of disks in the array
2389  * @cnt: scribble objs count for required size of the scribble region
2390  *
2391  * The scribble buffer size must be enough to contain:
2392  * 1/ a struct page pointer for each device in the array +2
2393  * 2/ room to convert each entry in (1) to its corresponding dma
2394  *    (dma_map_page()) or page (page_address()) address.
2395  *
2396  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2397  * calculate over all devices (not just the data blocks), using zeros in place
2398  * of the P and Q blocks.
2399  */
2400 static int scribble_alloc(struct raid5_percpu *percpu,
2401                           int num, int cnt)
2402 {
2403         size_t obj_size =
2404                 sizeof(struct page *) * (num + 2) +
2405                 sizeof(addr_conv_t) * (num + 2) +
2406                 sizeof(unsigned int) * (num + 2);
2407         void *scribble;
2408
2409         /*
2410          * If here is in raid array suspend context, it is in memalloc noio
2411          * context as well, there is no potential recursive memory reclaim
2412          * I/Os with the GFP_KERNEL flag.
2413          */
2414         scribble = kvmalloc_array(cnt, obj_size, GFP_KERNEL);
2415         if (!scribble)
2416                 return -ENOMEM;
2417
2418         kvfree(percpu->scribble);
2419
2420         percpu->scribble = scribble;
2421         percpu->scribble_obj_size = obj_size;
2422         return 0;
2423 }
2424
2425 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2426 {
2427         unsigned long cpu;
2428         int err = 0;
2429
2430         /*
2431          * Never shrink. And mddev_suspend() could deadlock if this is called
2432          * from raid5d. In that case, scribble_disks and scribble_sectors
2433          * should equal to new_disks and new_sectors
2434          */
2435         if (conf->scribble_disks >= new_disks &&
2436             conf->scribble_sectors >= new_sectors)
2437                 return 0;
2438         mddev_suspend(conf->mddev);
2439         get_online_cpus();
2440
2441         for_each_present_cpu(cpu) {
2442                 struct raid5_percpu *percpu;
2443
2444                 percpu = per_cpu_ptr(conf->percpu, cpu);
2445                 err = scribble_alloc(percpu, new_disks,
2446                                      new_sectors / RAID5_STRIPE_SECTORS(conf));
2447                 if (err)
2448                         break;
2449         }
2450
2451         put_online_cpus();
2452         mddev_resume(conf->mddev);
2453         if (!err) {
2454                 conf->scribble_disks = new_disks;
2455                 conf->scribble_sectors = new_sectors;
2456         }
2457         return err;
2458 }
2459
2460 static int resize_stripes(struct r5conf *conf, int newsize)
2461 {
2462         /* Make all the stripes able to hold 'newsize' devices.
2463          * New slots in each stripe get 'page' set to a new page.
2464          *
2465          * This happens in stages:
2466          * 1/ create a new kmem_cache and allocate the required number of
2467          *    stripe_heads.
2468          * 2/ gather all the old stripe_heads and transfer the pages across
2469          *    to the new stripe_heads.  This will have the side effect of
2470          *    freezing the array as once all stripe_heads have been collected,
2471          *    no IO will be possible.  Old stripe heads are freed once their
2472          *    pages have been transferred over, and the old kmem_cache is
2473          *    freed when all stripes are done.
2474          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2475          *    we simple return a failure status - no need to clean anything up.
2476          * 4/ allocate new pages for the new slots in the new stripe_heads.
2477          *    If this fails, we don't bother trying the shrink the
2478          *    stripe_heads down again, we just leave them as they are.
2479          *    As each stripe_head is processed the new one is released into
2480          *    active service.
2481          *
2482          * Once step2 is started, we cannot afford to wait for a write,
2483          * so we use GFP_NOIO allocations.
2484          */
2485         struct stripe_head *osh, *nsh;
2486         LIST_HEAD(newstripes);
2487         struct disk_info *ndisks;
2488         int err = 0;
2489         struct kmem_cache *sc;
2490         int i;
2491         int hash, cnt;
2492
2493         md_allow_write(conf->mddev);
2494
2495         /* Step 1 */
2496         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2497                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2498                                0, 0, NULL);
2499         if (!sc)
2500                 return -ENOMEM;
2501
2502         /* Need to ensure auto-resizing doesn't interfere */
2503         mutex_lock(&conf->cache_size_mutex);
2504
2505         for (i = conf->max_nr_stripes; i; i--) {
2506                 nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2507                 if (!nsh)
2508                         break;
2509
2510                 list_add(&nsh->lru, &newstripes);
2511         }
2512         if (i) {
2513                 /* didn't get enough, give up */
2514                 while (!list_empty(&newstripes)) {
2515                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
2516                         list_del(&nsh->lru);
2517                         free_stripe(sc, nsh);
2518                 }
2519                 kmem_cache_destroy(sc);
2520                 mutex_unlock(&conf->cache_size_mutex);
2521                 return -ENOMEM;
2522         }
2523         /* Step 2 - Must use GFP_NOIO now.
2524          * OK, we have enough stripes, start collecting inactive
2525          * stripes and copying them over
2526          */
2527         hash = 0;
2528         cnt = 0;
2529         list_for_each_entry(nsh, &newstripes, lru) {
2530                 lock_device_hash_lock(conf, hash);
2531                 wait_event_cmd(conf->wait_for_stripe,
2532                                     !list_empty(conf->inactive_list + hash),
2533                                     unlock_device_hash_lock(conf, hash),
2534                                     lock_device_hash_lock(conf, hash));
2535                 osh = get_free_stripe(conf, hash);
2536                 unlock_device_hash_lock(conf, hash);
2537
2538 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2539         for (i = 0; i < osh->nr_pages; i++) {
2540                 nsh->pages[i] = osh->pages[i];
2541                 osh->pages[i] = NULL;
2542         }
2543 #endif
2544                 for(i=0; i<conf->pool_size; i++) {
2545                         nsh->dev[i].page = osh->dev[i].page;
2546                         nsh->dev[i].orig_page = osh->dev[i].page;
2547                         nsh->dev[i].offset = osh->dev[i].offset;
2548                 }
2549                 nsh->hash_lock_index = hash;
2550                 free_stripe(conf->slab_cache, osh);
2551                 cnt++;
2552                 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2553                     !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2554                         hash++;
2555                         cnt = 0;
2556                 }
2557         }
2558         kmem_cache_destroy(conf->slab_cache);
2559
2560         /* Step 3.
2561          * At this point, we are holding all the stripes so the array
2562          * is completely stalled, so now is a good time to resize
2563          * conf->disks and the scribble region
2564          */
2565         ndisks = kcalloc(newsize, sizeof(struct disk_info), GFP_NOIO);
2566         if (ndisks) {
2567                 for (i = 0; i < conf->pool_size; i++)
2568                         ndisks[i] = conf->disks[i];
2569
2570                 for (i = conf->pool_size; i < newsize; i++) {
2571                         ndisks[i].extra_page = alloc_page(GFP_NOIO);
2572                         if (!ndisks[i].extra_page)
2573                                 err = -ENOMEM;
2574                 }
2575
2576                 if (err) {
2577                         for (i = conf->pool_size; i < newsize; i++)
2578                                 if (ndisks[i].extra_page)
2579                                         put_page(ndisks[i].extra_page);
2580                         kfree(ndisks);
2581                 } else {
2582                         kfree(conf->disks);
2583                         conf->disks = ndisks;
2584                 }
2585         } else
2586                 err = -ENOMEM;
2587
2588         conf->slab_cache = sc;
2589         conf->active_name = 1-conf->active_name;
2590
2591         /* Step 4, return new stripes to service */
2592         while(!list_empty(&newstripes)) {
2593                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2594                 list_del_init(&nsh->lru);
2595
2596 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2597                 for (i = 0; i < nsh->nr_pages; i++) {
2598                         if (nsh->pages[i])
2599                                 continue;
2600                         nsh->pages[i] = alloc_page(GFP_NOIO);
2601                         if (!nsh->pages[i])
2602                                 err = -ENOMEM;
2603                 }
2604
2605                 for (i = conf->raid_disks; i < newsize; i++) {
2606                         if (nsh->dev[i].page)
2607                                 continue;
2608                         nsh->dev[i].page = raid5_get_dev_page(nsh, i);
2609                         nsh->dev[i].orig_page = nsh->dev[i].page;
2610                         nsh->dev[i].offset = raid5_get_page_offset(nsh, i);
2611                 }
2612 #else
2613                 for (i=conf->raid_disks; i < newsize; i++)
2614                         if (nsh->dev[i].page == NULL) {
2615                                 struct page *p = alloc_page(GFP_NOIO);
2616                                 nsh->dev[i].page = p;
2617                                 nsh->dev[i].orig_page = p;
2618                                 nsh->dev[i].offset = 0;
2619                                 if (!p)
2620                                         err = -ENOMEM;
2621                         }
2622 #endif
2623                 raid5_release_stripe(nsh);
2624         }
2625         /* critical section pass, GFP_NOIO no longer needed */
2626
2627         if (!err)
2628                 conf->pool_size = newsize;
2629         mutex_unlock(&conf->cache_size_mutex);
2630
2631         return err;
2632 }
2633
2634 static int drop_one_stripe(struct r5conf *conf)
2635 {
2636         struct stripe_head *sh;
2637         int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2638
2639         spin_lock_irq(conf->hash_locks + hash);
2640         sh = get_free_stripe(conf, hash);
2641         spin_unlock_irq(conf->hash_locks + hash);
2642         if (!sh)
2643                 return 0;
2644         BUG_ON(atomic_read(&sh->count));
2645         shrink_buffers(sh);
2646         free_stripe(conf->slab_cache, sh);
2647         atomic_dec(&conf->active_stripes);
2648         conf->max_nr_stripes--;
2649         return 1;
2650 }
2651
2652 static void shrink_stripes(struct r5conf *conf)
2653 {
2654         while (conf->max_nr_stripes &&
2655                drop_one_stripe(conf))
2656                 ;
2657
2658         kmem_cache_destroy(conf->slab_cache);
2659         conf->slab_cache = NULL;
2660 }
2661
2662 static void raid5_end_read_request(struct bio * bi)
2663 {
2664         struct stripe_head *sh = bi->bi_private;
2665         struct r5conf *conf = sh->raid_conf;
2666         int disks = sh->disks, i;
2667         char b[BDEVNAME_SIZE];
2668         struct md_rdev *rdev = NULL;
2669         sector_t s;
2670
2671         for (i=0 ; i<disks; i++)
2672                 if (bi == &sh->dev[i].req)
2673                         break;
2674
2675         pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2676                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2677                 bi->bi_status);
2678         if (i == disks) {
2679                 bio_reset(bi);
2680                 BUG();
2681                 return;
2682         }
2683         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2684                 /* If replacement finished while this request was outstanding,
2685                  * 'replacement' might be NULL already.
2686                  * In that case it moved down to 'rdev'.
2687                  * rdev is not removed until all requests are finished.
2688                  */
2689                 rdev = conf->disks[i].replacement;
2690         if (!rdev)
2691                 rdev = conf->disks[i].rdev;
2692
2693         if (use_new_offset(conf, sh))
2694                 s = sh->sector + rdev->new_data_offset;
2695         else
2696                 s = sh->sector + rdev->data_offset;
2697         if (!bi->bi_status) {
2698                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2699                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2700                         /* Note that this cannot happen on a
2701                          * replacement device.  We just fail those on
2702                          * any error
2703                          */
2704                         pr_info_ratelimited(
2705                                 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2706                                 mdname(conf->mddev), RAID5_STRIPE_SECTORS(conf),
2707                                 (unsigned long long)s,
2708                                 bdevname(rdev->bdev, b));
2709                         atomic_add(RAID5_STRIPE_SECTORS(conf), &rdev->corrected_errors);
2710                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2711                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2712                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2713                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2714
2715                 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2716                         /*
2717                          * end read for a page in journal, this
2718                          * must be preparing for prexor in rmw
2719                          */
2720                         set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2721
2722                 if (atomic_read(&rdev->read_errors))
2723                         atomic_set(&rdev->read_errors, 0);
2724         } else {
2725                 const char *bdn = bdevname(rdev->bdev, b);
2726                 int retry = 0;
2727                 int set_bad = 0;
2728
2729                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2730                 if (!(bi->bi_status == BLK_STS_PROTECTION))
2731                         atomic_inc(&rdev->read_errors);
2732                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2733                         pr_warn_ratelimited(
2734                                 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2735                                 mdname(conf->mddev),
2736                                 (unsigned long long)s,
2737                                 bdn);
2738                 else if (conf->mddev->degraded >= conf->max_degraded) {
2739                         set_bad = 1;
2740                         pr_warn_ratelimited(
2741                                 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2742                                 mdname(conf->mddev),
2743                                 (unsigned long long)s,
2744                                 bdn);
2745                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2746                         /* Oh, no!!! */
2747                         set_bad = 1;
2748                         pr_warn_ratelimited(
2749                                 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2750                                 mdname(conf->mddev),
2751                                 (unsigned long long)s,
2752                                 bdn);
2753                 } else if (atomic_read(&rdev->read_errors)
2754                          > conf->max_nr_stripes) {
2755                         if (!test_bit(Faulty, &rdev->flags)) {
2756                                 pr_warn("md/raid:%s: %d read_errors > %d stripes\n",
2757                                     mdname(conf->mddev),
2758                                     atomic_read(&rdev->read_errors),
2759                                     conf->max_nr_stripes);
2760                                 pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2761                                     mdname(conf->mddev), bdn);
2762                         }
2763                 } else
2764                         retry = 1;
2765                 if (set_bad && test_bit(In_sync, &rdev->flags)
2766                     && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2767                         retry = 1;
2768                 if (retry)
2769                         if (sh->qd_idx >= 0 && sh->pd_idx == i)
2770                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2771                         else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2772                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2773                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2774                         } else
2775                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2776                 else {
2777                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2778                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2779                         if (!(set_bad
2780                               && test_bit(In_sync, &rdev->flags)
2781                               && rdev_set_badblocks(
2782                                       rdev, sh->sector, RAID5_STRIPE_SECTORS(conf), 0)))
2783                                 md_error(conf->mddev, rdev);
2784                 }
2785         }
2786         rdev_dec_pending(rdev, conf->mddev);
2787         bio_reset(bi);
2788         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2789         set_bit(STRIPE_HANDLE, &sh->state);
2790         raid5_release_stripe(sh);
2791 }
2792
2793 static void raid5_end_write_request(struct bio *bi)
2794 {
2795         struct stripe_head *sh = bi->bi_private;
2796         struct r5conf *conf = sh->raid_conf;
2797         int disks = sh->disks, i;
2798         struct md_rdev *rdev;
2799         sector_t first_bad;
2800         int bad_sectors;
2801         int replacement = 0;
2802
2803         for (i = 0 ; i < disks; i++) {
2804                 if (bi == &sh->dev[i].req) {
2805                         rdev = conf->disks[i].rdev;
2806                         break;
2807                 }
2808                 if (bi == &sh->dev[i].rreq) {
2809                         rdev = conf->disks[i].replacement;
2810                         if (rdev)
2811                                 replacement = 1;
2812                         else
2813                                 /* rdev was removed and 'replacement'
2814                                  * replaced it.  rdev is not removed
2815                                  * until all requests are finished.
2816                                  */
2817                                 rdev = conf->disks[i].rdev;
2818                         break;
2819                 }
2820         }
2821         pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2822                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2823                 bi->bi_status);
2824         if (i == disks) {
2825                 bio_reset(bi);
2826                 BUG();
2827                 return;
2828         }
2829
2830         if (replacement) {
2831                 if (bi->bi_status)
2832                         md_error(conf->mddev, rdev);
2833                 else if (is_badblock(rdev, sh->sector,
2834                                      RAID5_STRIPE_SECTORS(conf),
2835                                      &first_bad, &bad_sectors))
2836                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2837         } else {
2838                 if (bi->bi_status) {
2839                         set_bit(STRIPE_DEGRADED, &sh->state);
2840                         set_bit(WriteErrorSeen, &rdev->flags);
2841                         set_bit(R5_WriteError, &sh->dev[i].flags);
2842                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2843                                 set_bit(MD_RECOVERY_NEEDED,
2844                                         &rdev->mddev->recovery);
2845                 } else if (is_badblock(rdev, sh->sector,
2846                                        RAID5_STRIPE_SECTORS(conf),
2847                                        &first_bad, &bad_sectors)) {
2848                         set_bit(R5_MadeGood, &sh->dev[i].flags);
2849                         if (test_bit(R5_ReadError, &sh->dev[i].flags))
2850                                 /* That was a successful write so make
2851                                  * sure it looks like we already did
2852                                  * a re-write.
2853                                  */
2854                                 set_bit(R5_ReWrite, &sh->dev[i].flags);
2855                 }
2856         }
2857         rdev_dec_pending(rdev, conf->mddev);
2858
2859         if (sh->batch_head && bi->bi_status && !replacement)
2860                 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2861
2862         bio_reset(bi);
2863         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2864                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2865         set_bit(STRIPE_HANDLE, &sh->state);
2866         raid5_release_stripe(sh);
2867
2868         if (sh->batch_head && sh != sh->batch_head)
2869                 raid5_release_stripe(sh->batch_head);
2870 }
2871
2872 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2873 {
2874         char b[BDEVNAME_SIZE];
2875         struct r5conf *conf = mddev->private;
2876         unsigned long flags;
2877         pr_debug("raid456: error called\n");
2878
2879         spin_lock_irqsave(&conf->device_lock, flags);
2880
2881         if (test_bit(In_sync, &rdev->flags) &&
2882             mddev->degraded == conf->max_degraded) {
2883                 /*
2884                  * Don't allow to achieve failed state
2885                  * Don't try to recover this device
2886                  */
2887                 conf->recovery_disabled = mddev->recovery_disabled;
2888                 spin_unlock_irqrestore(&conf->device_lock, flags);
2889                 return;
2890         }
2891
2892         set_bit(Faulty, &rdev->flags);
2893         clear_bit(In_sync, &rdev->flags);
2894         mddev->degraded = raid5_calc_degraded(conf);
2895         spin_unlock_irqrestore(&conf->device_lock, flags);
2896         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2897
2898         set_bit(Blocked, &rdev->flags);
2899         set_mask_bits(&mddev->sb_flags, 0,
2900                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2901         pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2902                 "md/raid:%s: Operation continuing on %d devices.\n",
2903                 mdname(mddev),
2904                 bdevname(rdev->bdev, b),
2905                 mdname(mddev),
2906                 conf->raid_disks - mddev->degraded);
2907         r5c_update_on_rdev_error(mddev, rdev);
2908 }
2909
2910 /*
2911  * Input: a 'big' sector number,
2912  * Output: index of the data and parity disk, and the sector # in them.
2913  */
2914 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2915                               int previous, int *dd_idx,
2916                               struct stripe_head *sh)
2917 {
2918         sector_t stripe, stripe2;
2919         sector_t chunk_number;
2920         unsigned int chunk_offset;
2921         int pd_idx, qd_idx;
2922         int ddf_layout = 0;
2923         sector_t new_sector;
2924         int algorithm = previous ? conf->prev_algo
2925                                  : conf->algorithm;
2926         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2927                                          : conf->chunk_sectors;
2928         int raid_disks = previous ? conf->previous_raid_disks
2929                                   : conf->raid_disks;
2930         int data_disks = raid_disks - conf->max_degraded;
2931
2932         /* First compute the information on this sector */
2933
2934         /*
2935          * Compute the chunk number and the sector offset inside the chunk
2936          */
2937         chunk_offset = sector_div(r_sector, sectors_per_chunk);
2938         chunk_number = r_sector;
2939
2940         /*
2941          * Compute the stripe number
2942          */
2943         stripe = chunk_number;
2944         *dd_idx = sector_div(stripe, data_disks);
2945         stripe2 = stripe;
2946         /*
2947          * Select the parity disk based on the user selected algorithm.
2948          */
2949         pd_idx = qd_idx = -1;
2950         switch(conf->level) {
2951         case 4:
2952                 pd_idx = data_disks;
2953                 break;
2954         case 5:
2955                 switch (algorithm) {
2956                 case ALGORITHM_LEFT_ASYMMETRIC:
2957                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2958                         if (*dd_idx >= pd_idx)
2959                                 (*dd_idx)++;
2960                         break;
2961                 case ALGORITHM_RIGHT_ASYMMETRIC:
2962                         pd_idx = sector_div(stripe2, raid_disks);
2963                         if (*dd_idx >= pd_idx)
2964                                 (*dd_idx)++;
2965                         break;
2966                 case ALGORITHM_LEFT_SYMMETRIC:
2967                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2968                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2969                         break;
2970                 case ALGORITHM_RIGHT_SYMMETRIC:
2971                         pd_idx = sector_div(stripe2, raid_disks);
2972                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2973                         break;
2974                 case ALGORITHM_PARITY_0:
2975                         pd_idx = 0;
2976                         (*dd_idx)++;
2977                         break;
2978                 case ALGORITHM_PARITY_N:
2979                         pd_idx = data_disks;
2980                         break;
2981                 default:
2982                         BUG();
2983                 }
2984                 break;
2985         case 6:
2986
2987                 switch (algorithm) {
2988                 case ALGORITHM_LEFT_ASYMMETRIC:
2989                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2990                         qd_idx = pd_idx + 1;
2991                         if (pd_idx == raid_disks-1) {
2992                                 (*dd_idx)++;    /* Q D D D P */
2993                                 qd_idx = 0;
2994                         } else if (*dd_idx >= pd_idx)
2995                                 (*dd_idx) += 2; /* D D P Q D */
2996                         break;
2997                 case ALGORITHM_RIGHT_ASYMMETRIC:
2998                         pd_idx = sector_div(stripe2, raid_disks);
2999                         qd_idx = pd_idx + 1;
3000                         if (pd_idx == raid_disks-1) {
3001                                 (*dd_idx)++;    /* Q D D D P */
3002                                 qd_idx = 0;
3003                         } else if (*dd_idx >= pd_idx)
3004                                 (*dd_idx) += 2; /* D D P Q D */
3005                         break;
3006                 case ALGORITHM_LEFT_SYMMETRIC:
3007                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3008                         qd_idx = (pd_idx + 1) % raid_disks;
3009                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
3010                         break;
3011                 case ALGORITHM_RIGHT_SYMMETRIC:
3012                         pd_idx = sector_div(stripe2, raid_disks);
3013                         qd_idx = (pd_idx + 1) % raid_disks;
3014                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
3015                         break;
3016
3017                 case ALGORITHM_PARITY_0:
3018                         pd_idx = 0;
3019                         qd_idx = 1;
3020                         (*dd_idx) += 2;
3021                         break;
3022                 case ALGORITHM_PARITY_N:
3023                         pd_idx = data_disks;
3024                         qd_idx = data_disks + 1;
3025                         break;
3026
3027                 case ALGORITHM_ROTATING_ZERO_RESTART:
3028                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
3029                          * of blocks for computing Q is different.
3030                          */
3031                         pd_idx = sector_div(stripe2, raid_disks);
3032                         qd_idx = pd_idx + 1;
3033                         if (pd_idx == raid_disks-1) {
3034                                 (*dd_idx)++;    /* Q D D D P */
3035                                 qd_idx = 0;
3036                         } else if (*dd_idx >= pd_idx)
3037                                 (*dd_idx) += 2; /* D D P Q D */
3038                         ddf_layout = 1;
3039                         break;
3040
3041                 case ALGORITHM_ROTATING_N_RESTART:
3042                         /* Same a left_asymmetric, by first stripe is
3043                          * D D D P Q  rather than
3044                          * Q D D D P
3045                          */
3046                         stripe2 += 1;
3047                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3048                         qd_idx = pd_idx + 1;
3049                         if (pd_idx == raid_disks-1) {
3050                                 (*dd_idx)++;    /* Q D D D P */
3051                                 qd_idx = 0;
3052                         } else if (*dd_idx >= pd_idx)
3053                                 (*dd_idx) += 2; /* D D P Q D */
3054                         ddf_layout = 1;
3055                         break;
3056
3057                 case ALGORITHM_ROTATING_N_CONTINUE:
3058                         /* Same as left_symmetric but Q is before P */
3059                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3060                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
3061                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
3062                         ddf_layout = 1;
3063                         break;
3064
3065                 case ALGORITHM_LEFT_ASYMMETRIC_6:
3066                         /* RAID5 left_asymmetric, with Q on last device */
3067                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
3068                         if (*dd_idx >= pd_idx)
3069                                 (*dd_idx)++;
3070                         qd_idx = raid_disks - 1;
3071                         break;
3072
3073                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3074                         pd_idx = sector_div(stripe2, raid_disks-1);
3075                         if (*dd_idx >= pd_idx)
3076                                 (*dd_idx)++;
3077                         qd_idx = raid_disks - 1;
3078                         break;
3079
3080                 case ALGORITHM_LEFT_SYMMETRIC_6:
3081                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
3082                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
3083                         qd_idx = raid_disks - 1;
3084                         break;
3085
3086                 case ALGORITHM_RIGHT_SYMMETRIC_6:
3087                         pd_idx = sector_div(stripe2, raid_disks-1);
3088                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
3089                         qd_idx = raid_disks - 1;
3090                         break;
3091
3092                 case ALGORITHM_PARITY_0_6:
3093                         pd_idx = 0;
3094                         (*dd_idx)++;
3095                         qd_idx = raid_disks - 1;
3096                         break;
3097
3098                 default:
3099                         BUG();
3100                 }
3101                 break;
3102         }
3103
3104         if (sh) {
3105                 sh->pd_idx = pd_idx;
3106                 sh->qd_idx = qd_idx;
3107                 sh->ddf_layout = ddf_layout;
3108         }
3109         /*
3110          * Finally, compute the new sector number
3111          */
3112         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
3113         return new_sector;
3114 }
3115
3116 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
3117 {
3118         struct r5conf *conf = sh->raid_conf;
3119         int raid_disks = sh->disks;
3120         int data_disks = raid_disks - conf->max_degraded;
3121         sector_t new_sector = sh->sector, check;
3122         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
3123                                          : conf->chunk_sectors;
3124         int algorithm = previous ? conf->prev_algo
3125                                  : conf->algorithm;
3126         sector_t stripe;
3127         int chunk_offset;
3128         sector_t chunk_number;
3129         int dummy1, dd_idx = i;
3130         sector_t r_sector;
3131         struct stripe_head sh2;
3132
3133         chunk_offset = sector_div(new_sector, sectors_per_chunk);
3134         stripe = new_sector;
3135
3136         if (i == sh->pd_idx)
3137                 return 0;
3138         switch(conf->level) {
3139         case 4: break;
3140         case 5:
3141                 switch (algorithm) {
3142                 case ALGORITHM_LEFT_ASYMMETRIC:
3143                 case ALGORITHM_RIGHT_ASYMMETRIC:
3144                         if (i > sh->pd_idx)
3145                                 i--;
3146                         break;
3147                 case ALGORITHM_LEFT_SYMMETRIC:
3148                 case ALGORITHM_RIGHT_SYMMETRIC:
3149                         if (i < sh->pd_idx)
3150                                 i += raid_disks;
3151                         i -= (sh->pd_idx + 1);
3152                         break;
3153                 case ALGORITHM_PARITY_0:
3154                         i -= 1;
3155                         break;
3156                 case ALGORITHM_PARITY_N:
3157                         break;
3158                 default:
3159                         BUG();
3160                 }
3161                 break;
3162         case 6:
3163                 if (i == sh->qd_idx)
3164                         return 0; /* It is the Q disk */
3165                 switch (algorithm) {
3166                 case ALGORITHM_LEFT_ASYMMETRIC:
3167                 case ALGORITHM_RIGHT_ASYMMETRIC:
3168                 case ALGORITHM_ROTATING_ZERO_RESTART:
3169                 case ALGORITHM_ROTATING_N_RESTART:
3170                         if (sh->pd_idx == raid_disks-1)
3171                                 i--;    /* Q D D D P */
3172                         else if (i > sh->pd_idx)
3173                                 i -= 2; /* D D P Q D */
3174                         break;
3175                 case ALGORITHM_LEFT_SYMMETRIC:
3176                 case ALGORITHM_RIGHT_SYMMETRIC:
3177                         if (sh->pd_idx == raid_disks-1)
3178                                 i--; /* Q D D D P */
3179                         else {
3180                                 /* D D P Q D */
3181                                 if (i < sh->pd_idx)
3182                                         i += raid_disks;
3183                                 i -= (sh->pd_idx + 2);
3184                         }
3185                         break;
3186                 case ALGORITHM_PARITY_0:
3187                         i -= 2;
3188                         break;
3189                 case ALGORITHM_PARITY_N:
3190                         break;
3191                 case ALGORITHM_ROTATING_N_CONTINUE:
3192                         /* Like left_symmetric, but P is before Q */
3193                         if (sh->pd_idx == 0)
3194                                 i--;    /* P D D D Q */
3195                         else {
3196                                 /* D D Q P D */
3197                                 if (i < sh->pd_idx)
3198                                         i += raid_disks;
3199                                 i -= (sh->pd_idx + 1);
3200                         }
3201                         break;
3202                 case ALGORITHM_LEFT_ASYMMETRIC_6:
3203                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3204                         if (i > sh->pd_idx)
3205                                 i--;
3206                         break;
3207                 case ALGORITHM_LEFT_SYMMETRIC_6:
3208                 case ALGORITHM_RIGHT_SYMMETRIC_6:
3209                         if (i < sh->pd_idx)
3210                                 i += data_disks + 1;
3211                         i -= (sh->pd_idx + 1);
3212                         break;
3213                 case ALGORITHM_PARITY_0_6:
3214                         i -= 1;
3215                         break;
3216                 default:
3217                         BUG();
3218                 }
3219                 break;
3220         }
3221
3222         chunk_number = stripe * data_disks + i;
3223         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3224
3225         check = raid5_compute_sector(conf, r_sector,
3226                                      previous, &dummy1, &sh2);
3227         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3228                 || sh2.qd_idx != sh->qd_idx) {
3229                 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3230                         mdname(conf->mddev));
3231                 return 0;
3232         }
3233         return r_sector;
3234 }
3235
3236 /*
3237  * There are cases where we want handle_stripe_dirtying() and
3238  * schedule_reconstruction() to delay towrite to some dev of a stripe.
3239  *
3240  * This function checks whether we want to delay the towrite. Specifically,
3241  * we delay the towrite when:
3242  *
3243  *   1. degraded stripe has a non-overwrite to the missing dev, AND this
3244  *      stripe has data in journal (for other devices).
3245  *
3246  *      In this case, when reading data for the non-overwrite dev, it is
3247  *      necessary to handle complex rmw of write back cache (prexor with
3248  *      orig_page, and xor with page). To keep read path simple, we would
3249  *      like to flush data in journal to RAID disks first, so complex rmw
3250  *      is handled in the write patch (handle_stripe_dirtying).
3251  *
3252  *   2. when journal space is critical (R5C_LOG_CRITICAL=1)
3253  *
3254  *      It is important to be able to flush all stripes in raid5-cache.
3255  *      Therefore, we need reserve some space on the journal device for
3256  *      these flushes. If flush operation includes pending writes to the
3257  *      stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3258  *      for the flush out. If we exclude these pending writes from flush
3259  *      operation, we only need (conf->max_degraded + 1) pages per stripe.
3260  *      Therefore, excluding pending writes in these cases enables more
3261  *      efficient use of the journal device.
3262  *
3263  *      Note: To make sure the stripe makes progress, we only delay
3264  *      towrite for stripes with data already in journal (injournal > 0).
3265  *      When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3266  *      no_space_stripes list.
3267  *
3268  *   3. during journal failure
3269  *      In journal failure, we try to flush all cached data to raid disks
3270  *      based on data in stripe cache. The array is read-only to upper
3271  *      layers, so we would skip all pending writes.
3272  *
3273  */
3274 static inline bool delay_towrite(struct r5conf *conf,
3275                                  struct r5dev *dev,
3276                                  struct stripe_head_state *s)
3277 {
3278         /* case 1 above */
3279         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3280             !test_bit(R5_Insync, &dev->flags) && s->injournal)
3281                 return true;
3282         /* case 2 above */
3283         if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3284             s->injournal > 0)
3285                 return true;
3286         /* case 3 above */
3287         if (s->log_failed && s->injournal)
3288                 return true;
3289         return false;
3290 }
3291
3292 static void
3293 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3294                          int rcw, int expand)
3295 {
3296         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3297         struct r5conf *conf = sh->raid_conf;
3298         int level = conf->level;
3299
3300         if (rcw) {
3301                 /*
3302                  * In some cases, handle_stripe_dirtying initially decided to
3303                  * run rmw and allocates extra page for prexor. However, rcw is
3304                  * cheaper later on. We need to free the extra page now,
3305                  * because we won't be able to do that in ops_complete_prexor().
3306                  */
3307                 r5c_release_extra_page(sh);
3308
3309                 for (i = disks; i--; ) {
3310                         struct r5dev *dev = &sh->dev[i];
3311
3312                         if (dev->towrite && !delay_towrite(conf, dev, s)) {
3313                                 set_bit(R5_LOCKED, &dev->flags);
3314                                 set_bit(R5_Wantdrain, &dev->flags);
3315                                 if (!expand)
3316                                         clear_bit(R5_UPTODATE, &dev->flags);
3317                                 s->locked++;
3318                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3319                                 set_bit(R5_LOCKED, &dev->flags);
3320                                 s->locked++;
3321                         }
3322                 }
3323                 /* if we are not expanding this is a proper write request, and
3324                  * there will be bios with new data to be drained into the
3325                  * stripe cache
3326                  */
3327                 if (!expand) {
3328                         if (!s->locked)
3329                                 /* False alarm, nothing to do */
3330                                 return;
3331                         sh->reconstruct_state = reconstruct_state_drain_run;
3332                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3333                 } else
3334                         sh->reconstruct_state = reconstruct_state_run;
3335
3336                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3337
3338                 if (s->locked + conf->max_degraded == disks)
3339                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3340                                 atomic_inc(&conf->pending_full_writes);
3341         } else {
3342                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3343                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3344                 BUG_ON(level == 6 &&
3345                         (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3346                            test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3347
3348                 for (i = disks; i--; ) {
3349                         struct r5dev *dev = &sh->dev[i];
3350                         if (i == pd_idx || i == qd_idx)
3351                                 continue;
3352
3353                         if (dev->towrite &&
3354                             (test_bit(R5_UPTODATE, &dev->flags) ||
3355                              test_bit(R5_Wantcompute, &dev->flags))) {
3356                                 set_bit(R5_Wantdrain, &dev->flags);
3357                                 set_bit(R5_LOCKED, &dev->flags);
3358                                 clear_bit(R5_UPTODATE, &dev->flags);
3359                                 s->locked++;
3360                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3361                                 set_bit(R5_LOCKED, &dev->flags);
3362                                 s->locked++;
3363                         }
3364                 }
3365                 if (!s->locked)
3366                         /* False alarm - nothing to do */
3367                         return;
3368                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3369                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3370                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3371                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3372         }
3373
3374         /* keep the parity disk(s) locked while asynchronous operations
3375          * are in flight
3376          */
3377         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3378         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3379         s->locked++;
3380
3381         if (level == 6) {
3382                 int qd_idx = sh->qd_idx;
3383                 struct r5dev *dev = &sh->dev[qd_idx];
3384
3385                 set_bit(R5_LOCKED, &dev->flags);
3386                 clear_bit(R5_UPTODATE, &dev->flags);
3387                 s->locked++;
3388         }
3389
3390         if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3391             test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3392             !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3393             test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3394                 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3395
3396         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3397                 __func__, (unsigned long long)sh->sector,
3398                 s->locked, s->ops_request);
3399 }
3400
3401 /*
3402  * Each stripe/dev can have one or more bion attached.
3403  * toread/towrite point to the first in a chain.
3404  * The bi_next chain must be in order.
3405  */
3406 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3407                           int forwrite, int previous)
3408 {
3409         struct bio **bip;
3410         struct r5conf *conf = sh->raid_conf;
3411         int firstwrite=0;
3412
3413         pr_debug("adding bi b#%llu to stripe s#%llu\n",
3414                 (unsigned long long)bi->bi_iter.bi_sector,
3415                 (unsigned long long)sh->sector);
3416
3417         spin_lock_irq(&sh->stripe_lock);
3418         sh->dev[dd_idx].write_hint = bi->bi_write_hint;
3419         /* Don't allow new IO added to stripes in batch list */
3420         if (sh->batch_head)
3421                 goto overlap;
3422         if (forwrite) {
3423                 bip = &sh->dev[dd_idx].towrite;
3424                 if (*bip == NULL)
3425                         firstwrite = 1;
3426         } else
3427                 bip = &sh->dev[dd_idx].toread;
3428         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3429                 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3430                         goto overlap;
3431                 bip = & (*bip)->bi_next;
3432         }
3433         if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3434                 goto overlap;
3435
3436         if (forwrite && raid5_has_ppl(conf)) {
3437                 /*
3438                  * With PPL only writes to consecutive data chunks within a
3439                  * stripe are allowed because for a single stripe_head we can
3440                  * only have one PPL entry at a time, which describes one data
3441                  * range. Not really an overlap, but wait_for_overlap can be
3442                  * used to handle this.
3443                  */
3444                 sector_t sector;
3445                 sector_t first = 0;
3446                 sector_t last = 0;
3447                 int count = 0;
3448                 int i;
3449
3450                 for (i = 0; i < sh->disks; i++) {
3451                         if (i != sh->pd_idx &&
3452                             (i == dd_idx || sh->dev[i].towrite)) {
3453                                 sector = sh->dev[i].sector;
3454                                 if (count == 0 || sector < first)
3455                                         first = sector;
3456                                 if (sector > last)
3457                                         last = sector;
3458                                 count++;
3459                         }
3460                 }
3461
3462                 if (first + conf->chunk_sectors * (count - 1) != last)
3463                         goto overlap;
3464         }
3465
3466         if (!forwrite || previous)
3467                 clear_bit(STRIPE_BATCH_READY, &sh->state);
3468
3469         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3470         if (*bip)
3471                 bi->bi_next = *bip;
3472         *bip = bi;
3473         bio_inc_remaining(bi);
3474         md_write_inc(conf->mddev, bi);
3475
3476         if (forwrite) {
3477                 /* check if page is covered */
3478                 sector_t sector = sh->dev[dd_idx].sector;
3479                 for (bi=sh->dev[dd_idx].towrite;
3480                      sector < sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf) &&
3481                              bi && bi->bi_iter.bi_sector <= sector;
3482                      bi = r5_next_bio(conf, bi, sh->dev[dd_idx].sector)) {
3483                         if (bio_end_sector(bi) >= sector)
3484                                 sector = bio_end_sector(bi);
3485                 }
3486                 if (sector >= sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf))
3487                         if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3488                                 sh->overwrite_disks++;
3489         }
3490
3491         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3492                 (unsigned long long)(*bip)->bi_iter.bi_sector,
3493                 (unsigned long long)sh->sector, dd_idx);
3494
3495         if (conf->mddev->bitmap && firstwrite) {
3496                 /* Cannot hold spinlock over bitmap_startwrite,
3497                  * but must ensure this isn't added to a batch until
3498                  * we have added to the bitmap and set bm_seq.
3499                  * So set STRIPE_BITMAP_PENDING to prevent
3500                  * batching.
3501                  * If multiple add_stripe_bio() calls race here they
3502                  * much all set STRIPE_BITMAP_PENDING.  So only the first one
3503                  * to complete "bitmap_startwrite" gets to set
3504                  * STRIPE_BIT_DELAY.  This is important as once a stripe
3505                  * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3506                  * any more.
3507                  */
3508                 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3509                 spin_unlock_irq(&sh->stripe_lock);
3510                 md_bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3511                                      RAID5_STRIPE_SECTORS(conf), 0);
3512                 spin_lock_irq(&sh->stripe_lock);
3513                 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3514                 if (!sh->batch_head) {
3515                         sh->bm_seq = conf->seq_flush+1;
3516                         set_bit(STRIPE_BIT_DELAY, &sh->state);
3517                 }
3518         }
3519         spin_unlock_irq(&sh->stripe_lock);
3520
3521         if (stripe_can_batch(sh))
3522                 stripe_add_to_batch_list(conf, sh);
3523         return 1;
3524
3525  overlap:
3526         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3527         spin_unlock_irq(&sh->stripe_lock);
3528         return 0;
3529 }
3530
3531 static void end_reshape(struct r5conf *conf);
3532
3533 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3534                             struct stripe_head *sh)
3535 {
3536         int sectors_per_chunk =
3537                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3538         int dd_idx;
3539         int chunk_offset = sector_div(stripe, sectors_per_chunk);
3540         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3541
3542         raid5_compute_sector(conf,
3543                              stripe * (disks - conf->max_degraded)
3544                              *sectors_per_chunk + chunk_offset,
3545                              previous,
3546                              &dd_idx, sh);
3547 }
3548
3549 static void
3550 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3551                      struct stripe_head_state *s, int disks)
3552 {
3553         int i;
3554         BUG_ON(sh->batch_head);
3555         for (i = disks; i--; ) {
3556                 struct bio *bi;
3557                 int bitmap_end = 0;
3558
3559                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3560                         struct md_rdev *rdev;
3561                         rcu_read_lock();
3562                         rdev = rcu_dereference(conf->disks[i].rdev);
3563                         if (rdev && test_bit(In_sync, &rdev->flags) &&
3564                             !test_bit(Faulty, &rdev->flags))
3565                                 atomic_inc(&rdev->nr_pending);
3566                         else
3567                                 rdev = NULL;
3568                         rcu_read_unlock();
3569                         if (rdev) {
3570                                 if (!rdev_set_badblocks(
3571                                             rdev,
3572                                             sh->sector,
3573                                             RAID5_STRIPE_SECTORS(conf), 0))
3574                                         md_error(conf->mddev, rdev);
3575                                 rdev_dec_pending(rdev, conf->mddev);
3576                         }
3577                 }
3578                 spin_lock_irq(&sh->stripe_lock);
3579                 /* fail all writes first */
3580                 bi = sh->dev[i].towrite;
3581                 sh->dev[i].towrite = NULL;
3582                 sh->overwrite_disks = 0;
3583                 spin_unlock_irq(&sh->stripe_lock);
3584                 if (bi)
3585                         bitmap_end = 1;
3586
3587                 log_stripe_write_finished(sh);
3588
3589                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3590                         wake_up(&conf->wait_for_overlap);
3591
3592                 while (bi && bi->bi_iter.bi_sector <
3593                         sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3594                         struct bio *nextbi = r5_next_bio(conf, bi, sh->dev[i].sector);
3595
3596                         md_write_end(conf->mddev);
3597                         bio_io_error(bi);
3598                         bi = nextbi;
3599                 }
3600                 if (bitmap_end)
3601                         md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3602                                            RAID5_STRIPE_SECTORS(conf), 0, 0);
3603                 bitmap_end = 0;
3604                 /* and fail all 'written' */
3605                 bi = sh->dev[i].written;
3606                 sh->dev[i].written = NULL;
3607                 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3608                         WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3609                         sh->dev[i].page = sh->dev[i].orig_page;
3610                 }
3611
3612                 if (bi) bitmap_end = 1;
3613                 while (bi && bi->bi_iter.bi_sector <
3614                        sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3615                         struct bio *bi2 = r5_next_bio(conf, bi, sh->dev[i].sector);
3616
3617                         md_write_end(conf->mddev);
3618                         bio_io_error(bi);
3619                         bi = bi2;
3620                 }
3621
3622                 /* fail any reads if this device is non-operational and
3623                  * the data has not reached the cache yet.
3624                  */
3625                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3626                     s->failed > conf->max_degraded &&
3627                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3628                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
3629                         spin_lock_irq(&sh->stripe_lock);
3630                         bi = sh->dev[i].toread;
3631                         sh->dev[i].toread = NULL;
3632                         spin_unlock_irq(&sh->stripe_lock);
3633                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3634                                 wake_up(&conf->wait_for_overlap);
3635                         if (bi)
3636                                 s->to_read--;
3637                         while (bi && bi->bi_iter.bi_sector <
3638                                sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3639                                 struct bio *nextbi =
3640                                         r5_next_bio(conf, bi, sh->dev[i].sector);
3641
3642                                 bio_io_error(bi);
3643                                 bi = nextbi;
3644                         }
3645                 }
3646                 if (bitmap_end)
3647                         md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3648                                            RAID5_STRIPE_SECTORS(conf), 0, 0);
3649                 /* If we were in the middle of a write the parity block might
3650                  * still be locked - so just clear all R5_LOCKED flags
3651                  */
3652                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3653         }
3654         s->to_write = 0;
3655         s->written = 0;
3656
3657         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3658                 if (atomic_dec_and_test(&conf->pending_full_writes))
3659                         md_wakeup_thread(conf->mddev->thread);
3660 }
3661
3662 static void
3663 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3664                    struct stripe_head_state *s)
3665 {
3666         int abort = 0;
3667         int i;
3668
3669         BUG_ON(sh->batch_head);
3670         clear_bit(STRIPE_SYNCING, &sh->state);
3671         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3672                 wake_up(&conf->wait_for_overlap);
3673         s->syncing = 0;
3674         s->replacing = 0;
3675         /* There is nothing more to do for sync/check/repair.
3676          * Don't even need to abort as that is handled elsewhere
3677          * if needed, and not always wanted e.g. if there is a known
3678          * bad block here.
3679          * For recover/replace we need to record a bad block on all
3680          * non-sync devices, or abort the recovery
3681          */
3682         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3683                 /* During recovery devices cannot be removed, so
3684                  * locking and refcounting of rdevs is not needed
3685                  */
3686                 rcu_read_lock();
3687                 for (i = 0; i < conf->raid_disks; i++) {
3688                         struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3689                         if (rdev
3690                             && !test_bit(Faulty, &rdev->flags)
3691                             && !test_bit(In_sync, &rdev->flags)
3692                             && !rdev_set_badblocks(rdev, sh->sector,
3693                                                    RAID5_STRIPE_SECTORS(conf), 0))
3694                                 abort = 1;
3695                         rdev = rcu_dereference(conf->disks[i].replacement);
3696                         if (rdev
3697                             && !test_bit(Faulty, &rdev->flags)
3698                             && !test_bit(In_sync, &rdev->flags)
3699                             && !rdev_set_badblocks(rdev, sh->sector,
3700                                                    RAID5_STRIPE_SECTORS(conf), 0))
3701                                 abort = 1;
3702                 }
3703                 rcu_read_unlock();
3704                 if (abort)
3705                         conf->recovery_disabled =
3706                                 conf->mddev->recovery_disabled;
3707         }
3708         md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), !abort);
3709 }
3710
3711 static int want_replace(struct stripe_head *sh, int disk_idx)
3712 {
3713         struct md_rdev *rdev;
3714         int rv = 0;
3715
3716         rcu_read_lock();
3717         rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3718         if (rdev
3719             && !test_bit(Faulty, &rdev->flags)
3720             && !test_bit(In_sync, &rdev->flags)
3721             && (rdev->recovery_offset <= sh->sector
3722                 || rdev->mddev->recovery_cp <= sh->sector))
3723                 rv = 1;
3724         rcu_read_unlock();
3725         return rv;
3726 }
3727
3728 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3729                            int disk_idx, int disks)
3730 {
3731         struct r5dev *dev = &sh->dev[disk_idx];
3732         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3733                                   &sh->dev[s->failed_num[1]] };
3734         int i;
3735         bool force_rcw = (sh->raid_conf->rmw_level == PARITY_DISABLE_RMW);
3736
3737
3738         if (test_bit(R5_LOCKED, &dev->flags) ||
3739             test_bit(R5_UPTODATE, &dev->flags))
3740                 /* No point reading this as we already have it or have
3741                  * decided to get it.
3742                  */
3743                 return 0;
3744
3745         if (dev->toread ||
3746             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3747                 /* We need this block to directly satisfy a request */
3748                 return 1;
3749
3750         if (s->syncing || s->expanding ||
3751             (s->replacing && want_replace(sh, disk_idx)))
3752                 /* When syncing, or expanding we read everything.
3753                  * When replacing, we need the replaced block.
3754                  */
3755                 return 1;
3756
3757         if ((s->failed >= 1 && fdev[0]->toread) ||
3758             (s->failed >= 2 && fdev[1]->toread))
3759                 /* If we want to read from a failed device, then
3760                  * we need to actually read every other device.
3761                  */
3762                 return 1;
3763
3764         /* Sometimes neither read-modify-write nor reconstruct-write
3765          * cycles can work.  In those cases we read every block we
3766          * can.  Then the parity-update is certain to have enough to
3767          * work with.
3768          * This can only be a problem when we need to write something,
3769          * and some device has failed.  If either of those tests
3770          * fail we need look no further.
3771          */
3772         if (!s->failed || !s->to_write)
3773                 return 0;
3774
3775         if (test_bit(R5_Insync, &dev->flags) &&
3776             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3777                 /* Pre-reads at not permitted until after short delay
3778                  * to gather multiple requests.  However if this
3779                  * device is no Insync, the block could only be computed
3780                  * and there is no need to delay that.
3781                  */
3782                 return 0;
3783
3784         for (i = 0; i < s->failed && i < 2; i++) {
3785                 if (fdev[i]->towrite &&
3786                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3787                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3788                         /* If we have a partial write to a failed
3789                          * device, then we will need to reconstruct
3790                          * the content of that device, so all other
3791                          * devices must be read.
3792                          */
3793                         return 1;
3794
3795                 if (s->failed >= 2 &&
3796                     (fdev[i]->towrite ||
3797                      s->failed_num[i] == sh->pd_idx ||
3798                      s->failed_num[i] == sh->qd_idx) &&
3799                     !test_bit(R5_UPTODATE, &fdev[i]->flags))
3800                         /* In max degraded raid6, If the failed disk is P, Q,
3801                          * or we want to read the failed disk, we need to do
3802                          * reconstruct-write.
3803                          */
3804                         force_rcw = true;
3805         }
3806
3807         /* If we are forced to do a reconstruct-write, because parity
3808          * cannot be trusted and we are currently recovering it, there
3809          * is extra need to be careful.
3810          * If one of the devices that we would need to read, because
3811          * it is not being overwritten (and maybe not written at all)
3812          * is missing/faulty, then we need to read everything we can.
3813          */
3814         if (!force_rcw &&
3815             sh->sector < sh->raid_conf->mddev->recovery_cp)
3816                 /* reconstruct-write isn't being forced */
3817                 return 0;
3818         for (i = 0; i < s->failed && i < 2; i++) {
3819                 if (s->failed_num[i] != sh->pd_idx &&
3820                     s->failed_num[i] != sh->qd_idx &&
3821                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3822                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3823                         return 1;
3824         }
3825
3826         return 0;
3827 }
3828
3829 /* fetch_block - checks the given member device to see if its data needs
3830  * to be read or computed to satisfy a request.
3831  *
3832  * Returns 1 when no more member devices need to be checked, otherwise returns
3833  * 0 to tell the loop in handle_stripe_fill to continue
3834  */
3835 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3836                        int disk_idx, int disks)
3837 {
3838         struct r5dev *dev = &sh->dev[disk_idx];
3839
3840         /* is the data in this block needed, and can we get it? */
3841         if (need_this_block(sh, s, disk_idx, disks)) {
3842                 /* we would like to get this block, possibly by computing it,
3843                  * otherwise read it if the backing disk is insync
3844                  */
3845                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3846                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3847                 BUG_ON(sh->batch_head);
3848
3849                 /*
3850                  * In the raid6 case if the only non-uptodate disk is P
3851                  * then we already trusted P to compute the other failed
3852                  * drives. It is safe to compute rather than re-read P.
3853                  * In other cases we only compute blocks from failed
3854                  * devices, otherwise check/repair might fail to detect
3855                  * a real inconsistency.
3856                  */
3857
3858                 if ((s->uptodate == disks - 1) &&
3859                     ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3860                     (s->failed && (disk_idx == s->failed_num[0] ||
3861                                    disk_idx == s->failed_num[1])))) {
3862                         /* have disk failed, and we're requested to fetch it;
3863                          * do compute it
3864                          */
3865                         pr_debug("Computing stripe %llu block %d\n",
3866                                (unsigned long long)sh->sector, disk_idx);
3867                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3868                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3869                         set_bit(R5_Wantcompute, &dev->flags);
3870                         sh->ops.target = disk_idx;
3871                         sh->ops.target2 = -1; /* no 2nd target */
3872                         s->req_compute = 1;
3873                         /* Careful: from this point on 'uptodate' is in the eye
3874                          * of raid_run_ops which services 'compute' operations
3875                          * before writes. R5_Wantcompute flags a block that will
3876                          * be R5_UPTODATE by the time it is needed for a
3877                          * subsequent operation.
3878                          */
3879                         s->uptodate++;
3880                         return 1;
3881                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3882                         /* Computing 2-failure is *very* expensive; only
3883                          * do it if failed >= 2
3884                          */
3885                         int other;
3886                         for (other = disks; other--; ) {
3887                                 if (other == disk_idx)
3888                                         continue;
3889                                 if (!test_bit(R5_UPTODATE,
3890                                       &sh->dev[other].flags))
3891                                         break;
3892                         }
3893                         BUG_ON(other < 0);
3894                         pr_debug("Computing stripe %llu blocks %d,%d\n",
3895                                (unsigned long long)sh->sector,
3896                                disk_idx, other);
3897                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3898                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3899                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3900                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
3901                         sh->ops.target = disk_idx;
3902                         sh->ops.target2 = other;
3903                         s->uptodate += 2;
3904                         s->req_compute = 1;
3905                         return 1;
3906                 } else if (test_bit(R5_Insync, &dev->flags)) {
3907                         set_bit(R5_LOCKED, &dev->flags);
3908                         set_bit(R5_Wantread, &dev->flags);
3909                         s->locked++;
3910                         pr_debug("Reading block %d (sync=%d)\n",
3911                                 disk_idx, s->syncing);
3912                 }
3913         }
3914
3915         return 0;
3916 }
3917
3918 /*
3919  * handle_stripe_fill - read or compute data to satisfy pending requests.
3920  */
3921 static void handle_stripe_fill(struct stripe_head *sh,
3922                                struct stripe_head_state *s,
3923                                int disks)
3924 {
3925         int i;
3926
3927         /* look for blocks to read/compute, skip this if a compute
3928          * is already in flight, or if the stripe contents are in the
3929          * midst of changing due to a write
3930          */
3931         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3932             !sh->reconstruct_state) {
3933
3934                 /*
3935                  * For degraded stripe with data in journal, do not handle
3936                  * read requests yet, instead, flush the stripe to raid
3937                  * disks first, this avoids handling complex rmw of write
3938                  * back cache (prexor with orig_page, and then xor with
3939                  * page) in the read path
3940                  */
3941                 if (s->injournal && s->failed) {
3942                         if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3943                                 r5c_make_stripe_write_out(sh);
3944                         goto out;
3945                 }
3946
3947                 for (i = disks; i--; )
3948                         if (fetch_block(sh, s, i, disks))
3949                                 break;
3950         }
3951 out:
3952         set_bit(STRIPE_HANDLE, &sh->state);
3953 }
3954
3955 static void break_stripe_batch_list(struct stripe_head *head_sh,
3956                                     unsigned long handle_flags);
3957 /* handle_stripe_clean_event
3958  * any written block on an uptodate or failed drive can be returned.
3959  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3960  * never LOCKED, so we don't need to test 'failed' directly.
3961  */
3962 static void handle_stripe_clean_event(struct r5conf *conf,
3963         struct stripe_head *sh, int disks)
3964 {
3965         int i;
3966         struct r5dev *dev;
3967         int discard_pending = 0;
3968         struct stripe_head *head_sh = sh;
3969         bool do_endio = false;
3970
3971         for (i = disks; i--; )
3972                 if (sh->dev[i].written) {
3973                         dev = &sh->dev[i];
3974                         if (!test_bit(R5_LOCKED, &dev->flags) &&
3975                             (test_bit(R5_UPTODATE, &dev->flags) ||
3976                              test_bit(R5_Discard, &dev->flags) ||
3977                              test_bit(R5_SkipCopy, &dev->flags))) {
3978                                 /* We can return any write requests */
3979                                 struct bio *wbi, *wbi2;
3980                                 pr_debug("Return write for disc %d\n", i);
3981                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
3982                                         clear_bit(R5_UPTODATE, &dev->flags);
3983                                 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3984                                         WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3985                                 }
3986                                 do_endio = true;
3987
3988 returnbi:
3989                                 dev->page = dev->orig_page;
3990                                 wbi = dev->written;
3991                                 dev->written = NULL;
3992                                 while (wbi && wbi->bi_iter.bi_sector <
3993                                         dev->sector + RAID5_STRIPE_SECTORS(conf)) {
3994                                         wbi2 = r5_next_bio(conf, wbi, dev->sector);
3995                                         md_write_end(conf->mddev);
3996                                         bio_endio(wbi);
3997                                         wbi = wbi2;
3998                                 }
3999                                 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
4000                                                    RAID5_STRIPE_SECTORS(conf),
4001                                                    !test_bit(STRIPE_DEGRADED, &sh->state),
4002                                                    0);
4003                                 if (head_sh->batch_head) {
4004                                         sh = list_first_entry(&sh->batch_list,
4005                                                               struct stripe_head,
4006                                                               batch_list);
4007                                         if (sh != head_sh) {
4008                                                 dev = &sh->dev[i];
4009                                                 goto returnbi;
4010                                         }
4011                                 }
4012                                 sh = head_sh;
4013                                 dev = &sh->dev[i];
4014                         } else if (test_bit(R5_Discard, &dev->flags))
4015                                 discard_pending = 1;
4016                 }
4017
4018         log_stripe_write_finished(sh);
4019
4020         if (!discard_pending &&
4021             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
4022                 int hash;
4023                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
4024                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4025                 if (sh->qd_idx >= 0) {
4026                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
4027                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
4028                 }
4029                 /* now that discard is done we can proceed with any sync */
4030                 clear_bit(STRIPE_DISCARD, &sh->state);
4031                 /*
4032                  * SCSI discard will change some bio fields and the stripe has
4033                  * no updated data, so remove it from hash list and the stripe
4034                  * will be reinitialized
4035                  */
4036 unhash:
4037                 hash = sh->hash_lock_index;
4038                 spin_lock_irq(conf->hash_locks + hash);
4039                 remove_hash(sh);
4040                 spin_unlock_irq(conf->hash_locks + hash);
4041                 if (head_sh->batch_head) {
4042                         sh = list_first_entry(&sh->batch_list,
4043                                               struct stripe_head, batch_list);
4044                         if (sh != head_sh)
4045                                         goto unhash;
4046                 }
4047                 sh = head_sh;
4048
4049                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
4050                         set_bit(STRIPE_HANDLE, &sh->state);
4051
4052         }
4053
4054         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
4055                 if (atomic_dec_and_test(&conf->pending_full_writes))
4056                         md_wakeup_thread(conf->mddev->thread);
4057
4058         if (head_sh->batch_head && do_endio)
4059                 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
4060 }
4061
4062 /*
4063  * For RMW in write back cache, we need extra page in prexor to store the
4064  * old data. This page is stored in dev->orig_page.
4065  *
4066  * This function checks whether we have data for prexor. The exact logic
4067  * is:
4068  *       R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
4069  */
4070 static inline bool uptodate_for_rmw(struct r5dev *dev)
4071 {
4072         return (test_bit(R5_UPTODATE, &dev->flags)) &&
4073                 (!test_bit(R5_InJournal, &dev->flags) ||
4074                  test_bit(R5_OrigPageUPTDODATE, &dev->flags));
4075 }
4076
4077 static int handle_stripe_dirtying(struct r5conf *conf,
4078                                   struct stripe_head *sh,
4079                                   struct stripe_head_state *s,
4080                                   int disks)
4081 {
4082         int rmw = 0, rcw = 0, i;
4083         sector_t recovery_cp = conf->mddev->recovery_cp;
4084
4085         /* Check whether resync is now happening or should start.
4086          * If yes, then the array is dirty (after unclean shutdown or
4087          * initial creation), so parity in some stripes might be inconsistent.
4088          * In this case, we need to always do reconstruct-write, to ensure
4089          * that in case of drive failure or read-error correction, we
4090          * generate correct data from the parity.
4091          */
4092         if (conf->rmw_level == PARITY_DISABLE_RMW ||
4093             (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
4094              s->failed == 0)) {
4095                 /* Calculate the real rcw later - for now make it
4096                  * look like rcw is cheaper
4097                  */
4098                 rcw = 1; rmw = 2;
4099                 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
4100                          conf->rmw_level, (unsigned long long)recovery_cp,
4101                          (unsigned long long)sh->sector);
4102         } else for (i = disks; i--; ) {
4103                 /* would I have to read this buffer for read_modify_write */
4104                 struct r5dev *dev = &sh->dev[i];
4105                 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
4106                      i == sh->pd_idx || i == sh->qd_idx ||
4107                      test_bit(R5_InJournal, &dev->flags)) &&
4108                     !test_bit(R5_LOCKED, &dev->flags) &&
4109                     !(uptodate_for_rmw(dev) ||
4110                       test_bit(R5_Wantcompute, &dev->flags))) {
4111                         if (test_bit(R5_Insync, &dev->flags))
4112                                 rmw++;
4113                         else
4114                                 rmw += 2*disks;  /* cannot read it */
4115                 }
4116                 /* Would I have to read this buffer for reconstruct_write */
4117                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
4118                     i != sh->pd_idx && i != sh->qd_idx &&
4119                     !test_bit(R5_LOCKED, &dev->flags) &&
4120                     !(test_bit(R5_UPTODATE, &dev->flags) ||
4121                       test_bit(R5_Wantcompute, &dev->flags))) {
4122                         if (test_bit(R5_Insync, &dev->flags))
4123                                 rcw++;
4124                         else
4125                                 rcw += 2*disks;
4126                 }
4127         }
4128
4129         pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
4130                  (unsigned long long)sh->sector, sh->state, rmw, rcw);
4131         set_bit(STRIPE_HANDLE, &sh->state);
4132         if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
4133                 /* prefer read-modify-write, but need to get some data */
4134                 if (conf->mddev->queue)
4135                         blk_add_trace_msg(conf->mddev->queue,
4136                                           "raid5 rmw %llu %d",
4137                                           (unsigned long long)sh->sector, rmw);
4138                 for (i = disks; i--; ) {
4139                         struct r5dev *dev = &sh->dev[i];
4140                         if (test_bit(R5_InJournal, &dev->flags) &&
4141                             dev->page == dev->orig_page &&
4142                             !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
4143                                 /* alloc page for prexor */
4144                                 struct page *p = alloc_page(GFP_NOIO);
4145
4146                                 if (p) {
4147                                         dev->orig_page = p;
4148                                         continue;
4149                                 }
4150
4151                                 /*
4152                                  * alloc_page() failed, try use
4153                                  * disk_info->extra_page
4154                                  */
4155                                 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
4156                                                       &conf->cache_state)) {
4157                                         r5c_use_extra_page(sh);
4158                                         break;
4159                                 }
4160
4161                                 /* extra_page in use, add to delayed_list */
4162                                 set_bit(STRIPE_DELAYED, &sh->state);
4163                                 s->waiting_extra_page = 1;
4164                                 return -EAGAIN;
4165                         }
4166                 }
4167
4168                 for (i = disks; i--; ) {
4169                         struct r5dev *dev = &sh->dev[i];
4170                         if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
4171                              i == sh->pd_idx || i == sh->qd_idx ||
4172                              test_bit(R5_InJournal, &dev->flags)) &&
4173                             !test_bit(R5_LOCKED, &dev->flags) &&
4174                             !(uptodate_for_rmw(dev) ||
4175                               test_bit(R5_Wantcompute, &dev->flags)) &&
4176                             test_bit(R5_Insync, &dev->flags)) {
4177                                 if (test_bit(STRIPE_PREREAD_ACTIVE,
4178                                              &sh->state)) {
4179                                         pr_debug("Read_old block %d for r-m-w\n",
4180                                                  i);
4181                                         set_bit(R5_LOCKED, &dev->flags);
4182                                         set_bit(R5_Wantread, &dev->flags);
4183                                         s->locked++;
4184                                 } else
4185                                         set_bit(STRIPE_DELAYED, &sh->state);
4186                         }
4187                 }
4188         }
4189         if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
4190                 /* want reconstruct write, but need to get some data */
4191                 int qread =0;
4192                 rcw = 0;
4193                 for (i = disks; i--; ) {
4194                         struct r5dev *dev = &sh->dev[i];
4195                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
4196                             i != sh->pd_idx && i != sh->qd_idx &&
4197                             !test_bit(R5_LOCKED, &dev->flags) &&
4198                             !(test_bit(R5_UPTODATE, &dev->flags) ||
4199                               test_bit(R5_Wantcompute, &dev->flags))) {
4200                                 rcw++;
4201                                 if (test_bit(R5_Insync, &dev->flags) &&
4202                                     test_bit(STRIPE_PREREAD_ACTIVE,
4203                                              &sh->state)) {
4204                                         pr_debug("Read_old block "
4205                                                 "%d for Reconstruct\n", i);
4206                                         set_bit(R5_LOCKED, &dev->flags);
4207                                         set_bit(R5_Wantread, &dev->flags);
4208                                         s->locked++;
4209                                         qread++;
4210                                 } else
4211                                         set_bit(STRIPE_DELAYED, &sh->state);
4212                         }
4213                 }
4214                 if (rcw && conf->mddev->queue)
4215                         blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
4216                                           (unsigned long long)sh->sector,
4217                                           rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
4218         }
4219
4220         if (rcw > disks && rmw > disks &&
4221             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4222                 set_bit(STRIPE_DELAYED, &sh->state);
4223
4224         /* now if nothing is locked, and if we have enough data,
4225          * we can start a write request
4226          */
4227         /* since handle_stripe can be called at any time we need to handle the
4228          * case where a compute block operation has been submitted and then a
4229          * subsequent call wants to start a write request.  raid_run_ops only
4230          * handles the case where compute block and reconstruct are requested
4231          * simultaneously.  If this is not the case then new writes need to be
4232          * held off until the compute completes.
4233          */
4234         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4235             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4236              !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4237                 schedule_reconstruction(sh, s, rcw == 0, 0);
4238         return 0;
4239 }
4240
4241 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4242                                 struct stripe_head_state *s, int disks)
4243 {
4244         struct r5dev *dev = NULL;
4245
4246         BUG_ON(sh->batch_head);
4247         set_bit(STRIPE_HANDLE, &sh->state);
4248
4249         switch (sh->check_state) {
4250         case check_state_idle:
4251                 /* start a new check operation if there are no failures */
4252                 if (s->failed == 0) {
4253                         BUG_ON(s->uptodate != disks);
4254                         sh->check_state = check_state_run;
4255                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4256                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4257                         s->uptodate--;
4258                         break;
4259                 }
4260                 dev = &sh->dev[s->failed_num[0]];
4261                 fallthrough;
4262         case check_state_compute_result:
4263                 sh->check_state = check_state_idle;
4264                 if (!dev)
4265                         dev = &sh->dev[sh->pd_idx];
4266
4267                 /* check that a write has not made the stripe insync */
4268                 if (test_bit(STRIPE_INSYNC, &sh->state))
4269                         break;
4270
4271                 /* either failed parity check, or recovery is happening */
4272                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4273                 BUG_ON(s->uptodate != disks);
4274
4275                 set_bit(R5_LOCKED, &dev->flags);
4276                 s->locked++;
4277                 set_bit(R5_Wantwrite, &dev->flags);
4278
4279                 clear_bit(STRIPE_DEGRADED, &sh->state);
4280                 set_bit(STRIPE_INSYNC, &sh->state);
4281                 break;
4282         case check_state_run:
4283                 break; /* we will be called again upon completion */
4284         case check_state_check_result:
4285                 sh->check_state = check_state_idle;
4286
4287                 /* if a failure occurred during the check operation, leave
4288                  * STRIPE_INSYNC not set and let the stripe be handled again
4289                  */
4290                 if (s->failed)
4291                         break;
4292
4293                 /* handle a successful check operation, if parity is correct
4294                  * we are done.  Otherwise update the mismatch count and repair
4295                  * parity if !MD_RECOVERY_CHECK
4296                  */
4297                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4298                         /* parity is correct (on disc,
4299                          * not in buffer any more)
4300                          */
4301                         set_bit(STRIPE_INSYNC, &sh->state);
4302                 else {
4303                         atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4304                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4305                                 /* don't try to repair!! */
4306                                 set_bit(STRIPE_INSYNC, &sh->state);
4307                                 pr_warn_ratelimited("%s: mismatch sector in range "
4308                                                     "%llu-%llu\n", mdname(conf->mddev),
4309                                                     (unsigned long long) sh->sector,
4310                                                     (unsigned long long) sh->sector +
4311                                                     RAID5_STRIPE_SECTORS(conf));
4312                         } else {
4313                                 sh->check_state = check_state_compute_run;
4314                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4315                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4316                                 set_bit(R5_Wantcompute,
4317                                         &sh->dev[sh->pd_idx].flags);
4318                                 sh->ops.target = sh->pd_idx;
4319                                 sh->ops.target2 = -1;
4320                                 s->uptodate++;
4321                         }
4322                 }
4323                 break;
4324         case check_state_compute_run:
4325                 break;
4326         default:
4327                 pr_err("%s: unknown check_state: %d sector: %llu\n",
4328                        __func__, sh->check_state,
4329                        (unsigned long long) sh->sector);
4330                 BUG();
4331         }
4332 }
4333
4334 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4335                                   struct stripe_head_state *s,
4336                                   int disks)
4337 {
4338         int pd_idx = sh->pd_idx;
4339         int qd_idx = sh->qd_idx;
4340         struct r5dev *dev;
4341
4342         BUG_ON(sh->batch_head);
4343         set_bit(STRIPE_HANDLE, &sh->state);
4344
4345         BUG_ON(s->failed > 2);
4346
4347         /* Want to check and possibly repair P and Q.
4348          * However there could be one 'failed' device, in which
4349          * case we can only check one of them, possibly using the
4350          * other to generate missing data
4351          */
4352
4353         switch (sh->check_state) {
4354         case check_state_idle:
4355                 /* start a new check operation if there are < 2 failures */
4356                 if (s->failed == s->q_failed) {
4357                         /* The only possible failed device holds Q, so it
4358                          * makes sense to check P (If anything else were failed,
4359                          * we would have used P to recreate it).
4360                          */
4361                         sh->check_state = check_state_run;
4362                 }
4363                 if (!s->q_failed && s->failed < 2) {
4364                         /* Q is not failed, and we didn't use it to generate
4365                          * anything, so it makes sense to check it
4366                          */
4367                         if (sh->check_state == check_state_run)
4368                                 sh->check_state = check_state_run_pq;
4369                         else
4370                                 sh->check_state = check_state_run_q;
4371                 }
4372
4373                 /* discard potentially stale zero_sum_result */
4374                 sh->ops.zero_sum_result = 0;
4375
4376                 if (sh->check_state == check_state_run) {
4377                         /* async_xor_zero_sum destroys the contents of P */
4378                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4379                         s->uptodate--;
4380                 }
4381                 if (sh->check_state >= check_state_run &&
4382                     sh->check_state <= check_state_run_pq) {
4383                         /* async_syndrome_zero_sum preserves P and Q, so
4384                          * no need to mark them !uptodate here
4385                          */
4386                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4387                         break;
4388                 }
4389
4390                 /* we have 2-disk failure */
4391                 BUG_ON(s->failed != 2);
4392                 fallthrough;
4393         case check_state_compute_result:
4394                 sh->check_state = check_state_idle;
4395
4396                 /* check that a write has not made the stripe insync */
4397                 if (test_bit(STRIPE_INSYNC, &sh->state))
4398                         break;
4399
4400                 /* now write out any block on a failed drive,
4401                  * or P or Q if they were recomputed
4402                  */
4403                 dev = NULL;
4404                 if (s->failed == 2) {
4405                         dev = &sh->dev[s->failed_num[1]];
4406                         s->locked++;
4407                         set_bit(R5_LOCKED, &dev->flags);
4408                         set_bit(R5_Wantwrite, &dev->flags);
4409                 }
4410                 if (s->failed >= 1) {
4411                         dev = &sh->dev[s->failed_num[0]];
4412                         s->locked++;
4413                         set_bit(R5_LOCKED, &dev->flags);
4414                         set_bit(R5_Wantwrite, &dev->flags);
4415                 }
4416                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4417                         dev = &sh->dev[pd_idx];
4418                         s->locked++;
4419                         set_bit(R5_LOCKED, &dev->flags);
4420                         set_bit(R5_Wantwrite, &dev->flags);
4421                 }
4422                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4423                         dev = &sh->dev[qd_idx];
4424                         s->locked++;
4425                         set_bit(R5_LOCKED, &dev->flags);
4426                         set_bit(R5_Wantwrite, &dev->flags);
4427                 }
4428                 if (WARN_ONCE(dev && !test_bit(R5_UPTODATE, &dev->flags),
4429                               "%s: disk%td not up to date\n",
4430                               mdname(conf->mddev),
4431                               dev - (struct r5dev *) &sh->dev)) {
4432                         clear_bit(R5_LOCKED, &dev->flags);
4433                         clear_bit(R5_Wantwrite, &dev->flags);
4434                         s->locked--;
4435                 }
4436                 clear_bit(STRIPE_DEGRADED, &sh->state);
4437
4438                 set_bit(STRIPE_INSYNC, &sh->state);
4439                 break;
4440         case check_state_run:
4441         case check_state_run_q:
4442         case check_state_run_pq:
4443                 break; /* we will be called again upon completion */
4444         case check_state_check_result:
4445                 sh->check_state = check_state_idle;
4446
4447                 /* handle a successful check operation, if parity is correct
4448                  * we are done.  Otherwise update the mismatch count and repair
4449                  * parity if !MD_RECOVERY_CHECK
4450                  */
4451                 if (sh->ops.zero_sum_result == 0) {
4452                         /* both parities are correct */
4453                         if (!s->failed)
4454                                 set_bit(STRIPE_INSYNC, &sh->state);
4455                         else {
4456                                 /* in contrast to the raid5 case we can validate
4457                                  * parity, but still have a failure to write
4458                                  * back
4459                                  */
4460                                 sh->check_state = check_state_compute_result;
4461                                 /* Returning at this point means that we may go
4462                                  * off and bring p and/or q uptodate again so
4463                                  * we make sure to check zero_sum_result again
4464                                  * to verify if p or q need writeback
4465                                  */
4466                         }
4467                 } else {
4468                         atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4469                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4470                                 /* don't try to repair!! */
4471                                 set_bit(STRIPE_INSYNC, &sh->state);
4472                                 pr_warn_ratelimited("%s: mismatch sector in range "
4473                                                     "%llu-%llu\n", mdname(conf->mddev),
4474                                                     (unsigned long long) sh->sector,
4475                                                     (unsigned long long) sh->sector +
4476                                                     RAID5_STRIPE_SECTORS(conf));
4477                         } else {
4478                                 int *target = &sh->ops.target;
4479
4480                                 sh->ops.target = -1;
4481                                 sh->ops.target2 = -1;
4482                                 sh->check_state = check_state_compute_run;
4483                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4484                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4485                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4486                                         set_bit(R5_Wantcompute,
4487                                                 &sh->dev[pd_idx].flags);
4488                                         *target = pd_idx;
4489                                         target = &sh->ops.target2;
4490                                         s->uptodate++;
4491                                 }
4492                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4493                                         set_bit(R5_Wantcompute,
4494                                                 &sh->dev[qd_idx].flags);
4495                                         *target = qd_idx;
4496                                         s->uptodate++;
4497                                 }
4498                         }
4499                 }
4500                 break;
4501         case check_state_compute_run:
4502                 break;
4503         default:
4504                 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4505                         __func__, sh->check_state,
4506                         (unsigned long long) sh->sector);
4507                 BUG();
4508         }
4509 }
4510
4511 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4512 {
4513         int i;
4514
4515         /* We have read all the blocks in this stripe and now we need to
4516          * copy some of them into a target stripe for expand.
4517          */
4518         struct dma_async_tx_descriptor *tx = NULL;
4519         BUG_ON(sh->batch_head);
4520         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4521         for (i = 0; i < sh->disks; i++)
4522                 if (i != sh->pd_idx && i != sh->qd_idx) {
4523                         int dd_idx, j;
4524                         struct stripe_head *sh2;
4525                         struct async_submit_ctl submit;
4526
4527                         sector_t bn = raid5_compute_blocknr(sh, i, 1);
4528                         sector_t s = raid5_compute_sector(conf, bn, 0,
4529                                                           &dd_idx, NULL);
4530                         sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4531                         if (sh2 == NULL)
4532                                 /* so far only the early blocks of this stripe
4533                                  * have been requested.  When later blocks
4534                                  * get requested, we will try again
4535                                  */
4536                                 continue;
4537                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4538                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4539                                 /* must have already done this block */
4540                                 raid5_release_stripe(sh2);
4541                                 continue;
4542                         }
4543
4544                         /* place all the copies on one channel */
4545                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4546                         tx = async_memcpy(sh2->dev[dd_idx].page,
4547                                           sh->dev[i].page, sh2->dev[dd_idx].offset,
4548                                           sh->dev[i].offset, RAID5_STRIPE_SIZE(conf),
4549                                           &submit);
4550
4551                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4552                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4553                         for (j = 0; j < conf->raid_disks; j++)
4554                                 if (j != sh2->pd_idx &&
4555                                     j != sh2->qd_idx &&
4556                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
4557                                         break;
4558                         if (j == conf->raid_disks) {
4559                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4560                                 set_bit(STRIPE_HANDLE, &sh2->state);
4561                         }
4562                         raid5_release_stripe(sh2);
4563
4564                 }
4565         /* done submitting copies, wait for them to complete */
4566         async_tx_quiesce(&tx);
4567 }
4568
4569 /*
4570  * handle_stripe - do things to a stripe.
4571  *
4572  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4573  * state of various bits to see what needs to be done.
4574  * Possible results:
4575  *    return some read requests which now have data
4576  *    return some write requests which are safely on storage
4577  *    schedule a read on some buffers
4578  *    schedule a write of some buffers
4579  *    return confirmation of parity correctness
4580  *
4581  */
4582
4583 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4584 {
4585         struct r5conf *conf = sh->raid_conf;
4586         int disks = sh->disks;
4587         struct r5dev *dev;
4588         int i;
4589         int do_recovery = 0;
4590
4591         memset(s, 0, sizeof(*s));
4592
4593         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4594         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4595         s->failed_num[0] = -1;
4596         s->failed_num[1] = -1;
4597         s->log_failed = r5l_log_disk_error(conf);
4598
4599         /* Now to look around and see what can be done */
4600         rcu_read_lock();
4601         for (i=disks; i--; ) {
4602                 struct md_rdev *rdev;
4603                 sector_t first_bad;
4604                 int bad_sectors;
4605                 int is_bad = 0;
4606
4607                 dev = &sh->dev[i];
4608
4609                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4610                          i, dev->flags,
4611                          dev->toread, dev->towrite, dev->written);
4612                 /* maybe we can reply to a read
4613                  *
4614                  * new wantfill requests are only permitted while
4615                  * ops_complete_biofill is guaranteed to be inactive
4616                  */
4617                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4618                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4619                         set_bit(R5_Wantfill, &dev->flags);
4620
4621                 /* now count some things */
4622                 if (test_bit(R5_LOCKED, &dev->flags))
4623                         s->locked++;
4624                 if (test_bit(R5_UPTODATE, &dev->flags))
4625                         s->uptodate++;
4626                 if (test_bit(R5_Wantcompute, &dev->flags)) {
4627                         s->compute++;
4628                         BUG_ON(s->compute > 2);
4629                 }
4630
4631                 if (test_bit(R5_Wantfill, &dev->flags))
4632                         s->to_fill++;
4633                 else if (dev->toread)
4634                         s->to_read++;
4635                 if (dev->towrite) {
4636                         s->to_write++;
4637                         if (!test_bit(R5_OVERWRITE, &dev->flags))
4638                                 s->non_overwrite++;
4639                 }
4640                 if (dev->written)
4641                         s->written++;
4642                 /* Prefer to use the replacement for reads, but only
4643                  * if it is recovered enough and has no bad blocks.
4644                  */
4645                 rdev = rcu_dereference(conf->disks[i].replacement);
4646                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4647                     rdev->recovery_offset >= sh->sector + RAID5_STRIPE_SECTORS(conf) &&
4648                     !is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
4649                                  &first_bad, &bad_sectors))
4650                         set_bit(R5_ReadRepl, &dev->flags);
4651                 else {
4652                         if (rdev && !test_bit(Faulty, &rdev->flags))
4653                                 set_bit(R5_NeedReplace, &dev->flags);
4654                         else
4655                                 clear_bit(R5_NeedReplace, &dev->flags);
4656                         rdev = rcu_dereference(conf->disks[i].rdev);
4657                         clear_bit(R5_ReadRepl, &dev->flags);
4658                 }
4659                 if (rdev && test_bit(Faulty, &rdev->flags))
4660                         rdev = NULL;
4661                 if (rdev) {
4662                         is_bad = is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
4663                                              &first_bad, &bad_sectors);
4664                         if (s->blocked_rdev == NULL
4665                             && (test_bit(Blocked, &rdev->flags)
4666                                 || is_bad < 0)) {
4667                                 if (is_bad < 0)
4668                                         set_bit(BlockedBadBlocks,
4669                                                 &rdev->flags);
4670                                 s->blocked_rdev = rdev;
4671                                 atomic_inc(&rdev->nr_pending);
4672                         }
4673                 }
4674                 clear_bit(R5_Insync, &dev->flags);
4675                 if (!rdev)
4676                         /* Not in-sync */;
4677                 else if (is_bad) {
4678                         /* also not in-sync */
4679                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4680                             test_bit(R5_UPTODATE, &dev->flags)) {
4681                                 /* treat as in-sync, but with a read error
4682                                  * which we can now try to correct
4683                                  */
4684                                 set_bit(R5_Insync, &dev->flags);
4685                                 set_bit(R5_ReadError, &dev->flags);
4686                         }
4687                 } else if (test_bit(In_sync, &rdev->flags))
4688                         set_bit(R5_Insync, &dev->flags);
4689                 else if (sh->sector + RAID5_STRIPE_SECTORS(conf) <= rdev->recovery_offset)
4690                         /* in sync if before recovery_offset */
4691                         set_bit(R5_Insync, &dev->flags);
4692                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4693                          test_bit(R5_Expanded, &dev->flags))
4694                         /* If we've reshaped into here, we assume it is Insync.
4695                          * We will shortly update recovery_offset to make
4696                          * it official.
4697                          */
4698                         set_bit(R5_Insync, &dev->flags);
4699
4700                 if (test_bit(R5_WriteError, &dev->flags)) {
4701                         /* This flag does not apply to '.replacement'
4702                          * only to .rdev, so make sure to check that*/
4703                         struct md_rdev *rdev2 = rcu_dereference(
4704                                 conf->disks[i].rdev);
4705                         if (rdev2 == rdev)
4706                                 clear_bit(R5_Insync, &dev->flags);
4707                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4708                                 s->handle_bad_blocks = 1;
4709                                 atomic_inc(&rdev2->nr_pending);
4710                         } else
4711                                 clear_bit(R5_WriteError, &dev->flags);
4712                 }
4713                 if (test_bit(R5_MadeGood, &dev->flags)) {
4714                         /* This flag does not apply to '.replacement'
4715                          * only to .rdev, so make sure to check that*/
4716                         struct md_rdev *rdev2 = rcu_dereference(
4717                                 conf->disks[i].rdev);
4718                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4719                                 s->handle_bad_blocks = 1;
4720                                 atomic_inc(&rdev2->nr_pending);
4721                         } else
4722                                 clear_bit(R5_MadeGood, &dev->flags);
4723                 }
4724                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4725                         struct md_rdev *rdev2 = rcu_dereference(
4726                                 conf->disks[i].replacement);
4727                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4728                                 s->handle_bad_blocks = 1;
4729                                 atomic_inc(&rdev2->nr_pending);
4730                         } else
4731                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
4732                 }
4733                 if (!test_bit(R5_Insync, &dev->flags)) {
4734                         /* The ReadError flag will just be confusing now */
4735                         clear_bit(R5_ReadError, &dev->flags);
4736                         clear_bit(R5_ReWrite, &dev->flags);
4737                 }
4738                 if (test_bit(R5_ReadError, &dev->flags))
4739                         clear_bit(R5_Insync, &dev->flags);
4740                 if (!test_bit(R5_Insync, &dev->flags)) {
4741                         if (s->failed < 2)
4742                                 s->failed_num[s->failed] = i;
4743                         s->failed++;
4744                         if (rdev && !test_bit(Faulty, &rdev->flags))
4745                                 do_recovery = 1;
4746                         else if (!rdev) {
4747                                 rdev = rcu_dereference(
4748                                     conf->disks[i].replacement);
4749                                 if (rdev && !test_bit(Faulty, &rdev->flags))
4750                                         do_recovery = 1;
4751                         }
4752                 }
4753
4754                 if (test_bit(R5_InJournal, &dev->flags))
4755                         s->injournal++;
4756                 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4757                         s->just_cached++;
4758         }
4759         if (test_bit(STRIPE_SYNCING, &sh->state)) {
4760                 /* If there is a failed device being replaced,
4761                  *     we must be recovering.
4762                  * else if we are after recovery_cp, we must be syncing
4763                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4764                  * else we can only be replacing
4765                  * sync and recovery both need to read all devices, and so
4766                  * use the same flag.
4767                  */
4768                 if (do_recovery ||
4769                     sh->sector >= conf->mddev->recovery_cp ||
4770                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4771                         s->syncing = 1;
4772                 else
4773                         s->replacing = 1;
4774         }
4775         rcu_read_unlock();
4776 }
4777
4778 /*
4779  * Return '1' if this is a member of batch, or '0' if it is a lone stripe or
4780  * a head which can now be handled.
4781  */
4782 static int clear_batch_ready(struct stripe_head *sh)
4783 {
4784         struct stripe_head *tmp;
4785         if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4786                 return (sh->batch_head && sh->batch_head != sh);
4787         spin_lock(&sh->stripe_lock);
4788         if (!sh->batch_head) {
4789                 spin_unlock(&sh->stripe_lock);
4790                 return 0;
4791         }
4792
4793         /*
4794          * this stripe could be added to a batch list before we check
4795          * BATCH_READY, skips it
4796          */
4797         if (sh->batch_head != sh) {
4798                 spin_unlock(&sh->stripe_lock);
4799                 return 1;
4800         }
4801         spin_lock(&sh->batch_lock);
4802         list_for_each_entry(tmp, &sh->batch_list, batch_list)
4803                 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4804         spin_unlock(&sh->batch_lock);
4805         spin_unlock(&sh->stripe_lock);
4806
4807         /*
4808          * BATCH_READY is cleared, no new stripes can be added.
4809          * batch_list can be accessed without lock
4810          */
4811         return 0;
4812 }
4813
4814 static void break_stripe_batch_list(struct stripe_head *head_sh,
4815                                     unsigned long handle_flags)
4816 {
4817         struct stripe_head *sh, *next;
4818         int i;
4819         int do_wakeup = 0;
4820
4821         list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4822
4823                 list_del_init(&sh->batch_list);
4824
4825                 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4826                                           (1 << STRIPE_SYNCING) |
4827                                           (1 << STRIPE_REPLACED) |
4828                                           (1 << STRIPE_DELAYED) |
4829                                           (1 << STRIPE_BIT_DELAY) |
4830                                           (1 << STRIPE_FULL_WRITE) |
4831                                           (1 << STRIPE_BIOFILL_RUN) |
4832                                           (1 << STRIPE_COMPUTE_RUN)  |
4833                                           (1 << STRIPE_DISCARD) |
4834                                           (1 << STRIPE_BATCH_READY) |
4835                                           (1 << STRIPE_BATCH_ERR) |
4836                                           (1 << STRIPE_BITMAP_PENDING)),
4837                         "stripe state: %lx\n", sh->state);
4838                 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4839                                               (1 << STRIPE_REPLACED)),
4840                         "head stripe state: %lx\n", head_sh->state);
4841
4842                 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4843                                             (1 << STRIPE_PREREAD_ACTIVE) |
4844                                             (1 << STRIPE_DEGRADED) |
4845                                             (1 << STRIPE_ON_UNPLUG_LIST)),
4846                               head_sh->state & (1 << STRIPE_INSYNC));
4847
4848                 sh->check_state = head_sh->check_state;
4849                 sh->reconstruct_state = head_sh->reconstruct_state;
4850                 spin_lock_irq(&sh->stripe_lock);
4851                 sh->batch_head = NULL;
4852                 spin_unlock_irq(&sh->stripe_lock);
4853                 for (i = 0; i < sh->disks; i++) {
4854                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4855                                 do_wakeup = 1;
4856                         sh->dev[i].flags = head_sh->dev[i].flags &
4857                                 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4858                 }
4859                 if (handle_flags == 0 ||
4860                     sh->state & handle_flags)
4861                         set_bit(STRIPE_HANDLE, &sh->state);
4862                 raid5_release_stripe(sh);
4863         }
4864         spin_lock_irq(&head_sh->stripe_lock);
4865         head_sh->batch_head = NULL;
4866         spin_unlock_irq(&head_sh->stripe_lock);
4867         for (i = 0; i < head_sh->disks; i++)
4868                 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4869                         do_wakeup = 1;
4870         if (head_sh->state & handle_flags)
4871                 set_bit(STRIPE_HANDLE, &head_sh->state);
4872
4873         if (do_wakeup)
4874                 wake_up(&head_sh->raid_conf->wait_for_overlap);
4875 }
4876
4877 static void handle_stripe(struct stripe_head *sh)
4878 {
4879         struct stripe_head_state s;
4880         struct r5conf *conf = sh->raid_conf;
4881         int i;
4882         int prexor;
4883         int disks = sh->disks;
4884         struct r5dev *pdev, *qdev;
4885
4886         clear_bit(STRIPE_HANDLE, &sh->state);
4887
4888         /*
4889          * handle_stripe should not continue handle the batched stripe, only
4890          * the head of batch list or lone stripe can continue. Otherwise we
4891          * could see break_stripe_batch_list warns about the STRIPE_ACTIVE
4892          * is set for the batched stripe.
4893          */
4894         if (clear_batch_ready(sh))
4895                 return;
4896
4897         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4898                 /* already being handled, ensure it gets handled
4899                  * again when current action finishes */
4900                 set_bit(STRIPE_HANDLE, &sh->state);
4901                 return;
4902         }
4903
4904         if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4905                 break_stripe_batch_list(sh, 0);
4906
4907         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4908                 spin_lock(&sh->stripe_lock);
4909                 /*
4910                  * Cannot process 'sync' concurrently with 'discard'.
4911                  * Flush data in r5cache before 'sync'.
4912                  */
4913                 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4914                     !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4915                     !test_bit(STRIPE_DISCARD, &sh->state) &&
4916                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4917                         set_bit(STRIPE_SYNCING, &sh->state);
4918                         clear_bit(STRIPE_INSYNC, &sh->state);
4919                         clear_bit(STRIPE_REPLACED, &sh->state);
4920                 }
4921                 spin_unlock(&sh->stripe_lock);
4922         }
4923         clear_bit(STRIPE_DELAYED, &sh->state);
4924
4925         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4926                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4927                (unsigned long long)sh->sector, sh->state,
4928                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4929                sh->check_state, sh->reconstruct_state);
4930
4931         analyse_stripe(sh, &s);
4932
4933         if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4934                 goto finish;
4935
4936         if (s.handle_bad_blocks ||
4937             test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4938                 set_bit(STRIPE_HANDLE, &sh->state);
4939                 goto finish;
4940         }
4941
4942         if (unlikely(s.blocked_rdev)) {
4943                 if (s.syncing || s.expanding || s.expanded ||
4944                     s.replacing || s.to_write || s.written) {
4945                         set_bit(STRIPE_HANDLE, &sh->state);
4946                         goto finish;
4947                 }
4948                 /* There is nothing for the blocked_rdev to block */
4949                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4950                 s.blocked_rdev = NULL;
4951         }
4952
4953         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4954                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4955                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4956         }
4957
4958         pr_debug("locked=%d uptodate=%d to_read=%d"
4959                " to_write=%d failed=%d failed_num=%d,%d\n",
4960                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4961                s.failed_num[0], s.failed_num[1]);
4962         /*
4963          * check if the array has lost more than max_degraded devices and,
4964          * if so, some requests might need to be failed.
4965          *
4966          * When journal device failed (log_failed), we will only process
4967          * the stripe if there is data need write to raid disks
4968          */
4969         if (s.failed > conf->max_degraded ||
4970             (s.log_failed && s.injournal == 0)) {
4971                 sh->check_state = 0;
4972                 sh->reconstruct_state = 0;
4973                 break_stripe_batch_list(sh, 0);
4974                 if (s.to_read+s.to_write+s.written)
4975                         handle_failed_stripe(conf, sh, &s, disks);
4976                 if (s.syncing + s.replacing)
4977                         handle_failed_sync(conf, sh, &s);
4978         }
4979
4980         /* Now we check to see if any write operations have recently
4981          * completed
4982          */
4983         prexor = 0;
4984         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4985                 prexor = 1;
4986         if (sh->reconstruct_state == reconstruct_state_drain_result ||
4987             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4988                 sh->reconstruct_state = reconstruct_state_idle;
4989
4990                 /* All the 'written' buffers and the parity block are ready to
4991                  * be written back to disk
4992                  */
4993                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4994                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4995                 BUG_ON(sh->qd_idx >= 0 &&
4996                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4997                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4998                 for (i = disks; i--; ) {
4999                         struct r5dev *dev = &sh->dev[i];
5000                         if (test_bit(R5_LOCKED, &dev->flags) &&
5001                                 (i == sh->pd_idx || i == sh->qd_idx ||
5002                                  dev->written || test_bit(R5_InJournal,
5003                                                           &dev->flags))) {
5004                                 pr_debug("Writing block %d\n", i);
5005                                 set_bit(R5_Wantwrite, &dev->flags);
5006                                 if (prexor)
5007                                         continue;
5008                                 if (s.failed > 1)
5009                                         continue;
5010                                 if (!test_bit(R5_Insync, &dev->flags) ||
5011                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
5012                                      s.failed == 0))
5013                                         set_bit(STRIPE_INSYNC, &sh->state);
5014                         }
5015                 }
5016                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5017                         s.dec_preread_active = 1;
5018         }
5019
5020         /*
5021          * might be able to return some write requests if the parity blocks
5022          * are safe, or on a failed drive
5023          */
5024         pdev = &sh->dev[sh->pd_idx];
5025         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
5026                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
5027         qdev = &sh->dev[sh->qd_idx];
5028         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
5029                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
5030                 || conf->level < 6;
5031
5032         if (s.written &&
5033             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
5034                              && !test_bit(R5_LOCKED, &pdev->flags)
5035                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
5036                                  test_bit(R5_Discard, &pdev->flags))))) &&
5037             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
5038                              && !test_bit(R5_LOCKED, &qdev->flags)
5039                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
5040                                  test_bit(R5_Discard, &qdev->flags))))))
5041                 handle_stripe_clean_event(conf, sh, disks);
5042
5043         if (s.just_cached)
5044                 r5c_handle_cached_data_endio(conf, sh, disks);
5045         log_stripe_write_finished(sh);
5046
5047         /* Now we might consider reading some blocks, either to check/generate
5048          * parity, or to satisfy requests
5049          * or to load a block that is being partially written.
5050          */
5051         if (s.to_read || s.non_overwrite
5052             || (s.to_write && s.failed)
5053             || (s.syncing && (s.uptodate + s.compute < disks))
5054             || s.replacing
5055             || s.expanding)
5056                 handle_stripe_fill(sh, &s, disks);
5057
5058         /*
5059          * When the stripe finishes full journal write cycle (write to journal
5060          * and raid disk), this is the clean up procedure so it is ready for
5061          * next operation.
5062          */
5063         r5c_finish_stripe_write_out(conf, sh, &s);
5064
5065         /*
5066          * Now to consider new write requests, cache write back and what else,
5067          * if anything should be read.  We do not handle new writes when:
5068          * 1/ A 'write' operation (copy+xor) is already in flight.
5069          * 2/ A 'check' operation is in flight, as it may clobber the parity
5070          *    block.
5071          * 3/ A r5c cache log write is in flight.
5072          */
5073
5074         if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
5075                 if (!r5c_is_writeback(conf->log)) {
5076                         if (s.to_write)
5077                                 handle_stripe_dirtying(conf, sh, &s, disks);
5078                 } else { /* write back cache */
5079                         int ret = 0;
5080
5081                         /* First, try handle writes in caching phase */
5082                         if (s.to_write)
5083                                 ret = r5c_try_caching_write(conf, sh, &s,
5084                                                             disks);
5085                         /*
5086                          * If caching phase failed: ret == -EAGAIN
5087                          *    OR
5088                          * stripe under reclaim: !caching && injournal
5089                          *
5090                          * fall back to handle_stripe_dirtying()
5091                          */
5092                         if (ret == -EAGAIN ||
5093                             /* stripe under reclaim: !caching && injournal */
5094                             (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
5095                              s.injournal > 0)) {
5096                                 ret = handle_stripe_dirtying(conf, sh, &s,
5097                                                              disks);
5098                                 if (ret == -EAGAIN)
5099                                         goto finish;
5100                         }
5101                 }
5102         }
5103
5104         /* maybe we need to check and possibly fix the parity for this stripe
5105          * Any reads will already have been scheduled, so we just see if enough
5106          * data is available.  The parity check is held off while parity
5107          * dependent operations are in flight.
5108          */
5109         if (sh->check_state ||
5110             (s.syncing && s.locked == 0 &&
5111              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
5112              !test_bit(STRIPE_INSYNC, &sh->state))) {
5113                 if (conf->level == 6)
5114                         handle_parity_checks6(conf, sh, &s, disks);
5115                 else
5116                         handle_parity_checks5(conf, sh, &s, disks);
5117         }
5118
5119         if ((s.replacing || s.syncing) && s.locked == 0
5120             && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
5121             && !test_bit(STRIPE_REPLACED, &sh->state)) {
5122                 /* Write out to replacement devices where possible */
5123                 for (i = 0; i < conf->raid_disks; i++)
5124                         if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
5125                                 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
5126                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
5127                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
5128                                 s.locked++;
5129                         }
5130                 if (s.replacing)
5131                         set_bit(STRIPE_INSYNC, &sh->state);
5132                 set_bit(STRIPE_REPLACED, &sh->state);
5133         }
5134         if ((s.syncing || s.replacing) && s.locked == 0 &&
5135             !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
5136             test_bit(STRIPE_INSYNC, &sh->state)) {
5137                 md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
5138                 clear_bit(STRIPE_SYNCING, &sh->state);
5139                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
5140                         wake_up(&conf->wait_for_overlap);
5141         }
5142
5143         /* If the failed drives are just a ReadError, then we might need
5144          * to progress the repair/check process
5145          */
5146         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
5147                 for (i = 0; i < s.failed; i++) {
5148                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
5149                         if (test_bit(R5_ReadError, &dev->flags)
5150                             && !test_bit(R5_LOCKED, &dev->flags)
5151                             && test_bit(R5_UPTODATE, &dev->flags)
5152                                 ) {
5153                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
5154                                         set_bit(R5_Wantwrite, &dev->flags);
5155                                         set_bit(R5_ReWrite, &dev->flags);
5156                                 } else
5157                                         /* let's read it back */
5158                                         set_bit(R5_Wantread, &dev->flags);
5159                                 set_bit(R5_LOCKED, &dev->flags);
5160                                 s.locked++;
5161                         }
5162                 }
5163
5164         /* Finish reconstruct operations initiated by the expansion process */
5165         if (sh->reconstruct_state == reconstruct_state_result) {
5166                 struct stripe_head *sh_src
5167                         = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
5168                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
5169                         /* sh cannot be written until sh_src has been read.
5170                          * so arrange for sh to be delayed a little
5171                          */
5172                         set_bit(STRIPE_DELAYED, &sh->state);
5173                         set_bit(STRIPE_HANDLE, &sh->state);
5174                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
5175                                               &sh_src->state))
5176                                 atomic_inc(&conf->preread_active_stripes);
5177                         raid5_release_stripe(sh_src);
5178                         goto finish;
5179                 }
5180                 if (sh_src)
5181                         raid5_release_stripe(sh_src);
5182
5183                 sh->reconstruct_state = reconstruct_state_idle;
5184                 clear_bit(STRIPE_EXPANDING, &sh->state);
5185                 for (i = conf->raid_disks; i--; ) {
5186                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
5187                         set_bit(R5_LOCKED, &sh->dev[i].flags);
5188                         s.locked++;
5189                 }
5190         }
5191
5192         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
5193             !sh->reconstruct_state) {
5194                 /* Need to write out all blocks after computing parity */
5195                 sh->disks = conf->raid_disks;
5196                 stripe_set_idx(sh->sector, conf, 0, sh);
5197                 schedule_reconstruction(sh, &s, 1, 1);
5198         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
5199                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
5200                 atomic_dec(&conf->reshape_stripes);
5201                 wake_up(&conf->wait_for_overlap);
5202                 md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
5203         }
5204
5205         if (s.expanding && s.locked == 0 &&
5206             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
5207                 handle_stripe_expansion(conf, sh);
5208
5209 finish:
5210         /* wait for this device to become unblocked */
5211         if (unlikely(s.blocked_rdev)) {
5212                 if (conf->mddev->external)
5213                         md_wait_for_blocked_rdev(s.blocked_rdev,
5214                                                  conf->mddev);
5215                 else
5216                         /* Internal metadata will immediately
5217                          * be written by raid5d, so we don't
5218                          * need to wait here.
5219                          */
5220                         rdev_dec_pending(s.blocked_rdev,
5221                                          conf->mddev);
5222         }
5223
5224         if (s.handle_bad_blocks)
5225                 for (i = disks; i--; ) {
5226                         struct md_rdev *rdev;
5227                         struct r5dev *dev = &sh->dev[i];
5228                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
5229                                 /* We own a safe reference to the rdev */
5230                                 rdev = conf->disks[i].rdev;
5231                                 if (!rdev_set_badblocks(rdev, sh->sector,
5232                                                         RAID5_STRIPE_SECTORS(conf), 0))
5233                                         md_error(conf->mddev, rdev);
5234                                 rdev_dec_pending(rdev, conf->mddev);
5235                         }
5236                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5237                                 rdev = conf->disks[i].rdev;
5238                                 rdev_clear_badblocks(rdev, sh->sector,
5239                                                      RAID5_STRIPE_SECTORS(conf), 0);
5240                                 rdev_dec_pending(rdev, conf->mddev);
5241                         }
5242                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5243                                 rdev = conf->disks[i].replacement;
5244                                 if (!rdev)
5245                                         /* rdev have been moved down */
5246                                         rdev = conf->disks[i].rdev;
5247                                 rdev_clear_badblocks(rdev, sh->sector,
5248                                                      RAID5_STRIPE_SECTORS(conf), 0);
5249                                 rdev_dec_pending(rdev, conf->mddev);
5250                         }
5251                 }
5252
5253         if (s.ops_request)
5254                 raid_run_ops(sh, s.ops_request);
5255
5256         ops_run_io(sh, &s);
5257
5258         if (s.dec_preread_active) {
5259                 /* We delay this until after ops_run_io so that if make_request
5260                  * is waiting on a flush, it won't continue until the writes
5261                  * have actually been submitted.
5262                  */
5263                 atomic_dec(&conf->preread_active_stripes);
5264                 if (atomic_read(&conf->preread_active_stripes) <
5265                     IO_THRESHOLD)
5266                         md_wakeup_thread(conf->mddev->thread);
5267         }
5268
5269         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5270 }
5271
5272 static void raid5_activate_delayed(struct r5conf *conf)
5273 {
5274         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5275                 while (!list_empty(&conf->delayed_list)) {
5276                         struct list_head *l = conf->delayed_list.next;
5277                         struct stripe_head *sh;
5278                         sh = list_entry(l, struct stripe_head, lru);
5279                         list_del_init(l);
5280                         clear_bit(STRIPE_DELAYED, &sh->state);
5281                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5282                                 atomic_inc(&conf->preread_active_stripes);
5283                         list_add_tail(&sh->lru, &conf->hold_list);
5284                         raid5_wakeup_stripe_thread(sh);
5285                 }
5286         }
5287 }
5288
5289 static void activate_bit_delay(struct r5conf *conf,
5290         struct list_head *temp_inactive_list)
5291 {
5292         /* device_lock is held */
5293         struct list_head head;
5294         list_add(&head, &conf->bitmap_list);
5295         list_del_init(&conf->bitmap_list);
5296         while (!list_empty(&head)) {
5297                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5298                 int hash;
5299                 list_del_init(&sh->lru);
5300                 atomic_inc(&sh->count);
5301                 hash = sh->hash_lock_index;
5302                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5303         }
5304 }
5305
5306 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5307 {
5308         struct r5conf *conf = mddev->private;
5309         sector_t sector = bio->bi_iter.bi_sector;
5310         unsigned int chunk_sectors;
5311         unsigned int bio_sectors = bio_sectors(bio);
5312
5313         WARN_ON_ONCE(bio->bi_bdev->bd_partno);
5314
5315         chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5316         return  chunk_sectors >=
5317                 ((sector & (chunk_sectors - 1)) + bio_sectors);
5318 }
5319
5320 /*
5321  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
5322  *  later sampled by raid5d.
5323  */
5324 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5325 {
5326         unsigned long flags;
5327
5328         spin_lock_irqsave(&conf->device_lock, flags);
5329
5330         bi->bi_next = conf->retry_read_aligned_list;
5331         conf->retry_read_aligned_list = bi;
5332
5333         spin_unlock_irqrestore(&conf->device_lock, flags);
5334         md_wakeup_thread(conf->mddev->thread);
5335 }
5336
5337 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5338                                          unsigned int *offset)
5339 {
5340         struct bio *bi;
5341
5342         bi = conf->retry_read_aligned;
5343         if (bi) {
5344                 *offset = conf->retry_read_offset;
5345                 conf->retry_read_aligned = NULL;
5346                 return bi;
5347         }
5348         bi = conf->retry_read_aligned_list;
5349         if(bi) {
5350                 conf->retry_read_aligned_list = bi->bi_next;
5351                 bi->bi_next = NULL;
5352                 *offset = 0;
5353         }
5354
5355         return bi;
5356 }
5357
5358 /*
5359  *  The "raid5_align_endio" should check if the read succeeded and if it
5360  *  did, call bio_endio on the original bio (having bio_put the new bio
5361  *  first).
5362  *  If the read failed..
5363  */
5364 static void raid5_align_endio(struct bio *bi)
5365 {
5366         struct bio* raid_bi  = bi->bi_private;
5367         struct mddev *mddev;
5368         struct r5conf *conf;
5369         struct md_rdev *rdev;
5370         blk_status_t error = bi->bi_status;
5371
5372         bio_put(bi);
5373
5374         rdev = (void*)raid_bi->bi_next;
5375         raid_bi->bi_next = NULL;
5376         mddev = rdev->mddev;
5377         conf = mddev->private;
5378
5379         rdev_dec_pending(rdev, conf->mddev);
5380
5381         if (!error) {
5382                 bio_endio(raid_bi);
5383                 if (atomic_dec_and_test(&conf->active_aligned_reads))
5384                         wake_up(&conf->wait_for_quiescent);
5385                 return;
5386         }
5387
5388         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5389
5390         add_bio_to_retry(raid_bi, conf);
5391 }
5392
5393 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5394 {
5395         struct r5conf *conf = mddev->private;
5396         struct bio *align_bio;
5397         struct md_rdev *rdev;
5398         sector_t sector, end_sector, first_bad;
5399         int bad_sectors, dd_idx;
5400
5401         if (!in_chunk_boundary(mddev, raid_bio)) {
5402                 pr_debug("%s: non aligned\n", __func__);
5403                 return 0;
5404         }
5405
5406         sector = raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector, 0,
5407                                       &dd_idx, NULL);
5408         end_sector = bio_end_sector(raid_bio);
5409
5410         rcu_read_lock();
5411         if (r5c_big_stripe_cached(conf, sector))
5412                 goto out_rcu_unlock;
5413
5414         rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5415         if (!rdev || test_bit(Faulty, &rdev->flags) ||
5416             rdev->recovery_offset < end_sector) {
5417                 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5418                 if (!rdev)
5419                         goto out_rcu_unlock;
5420                 if (test_bit(Faulty, &rdev->flags) ||
5421                     !(test_bit(In_sync, &rdev->flags) ||
5422                       rdev->recovery_offset >= end_sector))
5423                         goto out_rcu_unlock;
5424         }
5425
5426         atomic_inc(&rdev->nr_pending);
5427         rcu_read_unlock();
5428
5429         align_bio = bio_clone_fast(raid_bio, GFP_NOIO, &mddev->bio_set);
5430         bio_set_dev(align_bio, rdev->bdev);
5431         align_bio->bi_end_io = raid5_align_endio;
5432         align_bio->bi_private = raid_bio;
5433         align_bio->bi_iter.bi_sector = sector;
5434
5435         raid_bio->bi_next = (void *)rdev;
5436
5437         if (is_badblock(rdev, sector, bio_sectors(align_bio), &first_bad,
5438                         &bad_sectors)) {
5439                 bio_put(align_bio);
5440                 rdev_dec_pending(rdev, mddev);
5441                 return 0;
5442         }
5443
5444         /* No reshape active, so we can trust rdev->data_offset */
5445         align_bio->bi_iter.bi_sector += rdev->data_offset;
5446
5447         spin_lock_irq(&conf->device_lock);
5448         wait_event_lock_irq(conf->wait_for_quiescent, conf->quiesce == 0,
5449                             conf->device_lock);
5450         atomic_inc(&conf->active_aligned_reads);
5451         spin_unlock_irq(&conf->device_lock);
5452
5453         if (mddev->gendisk)
5454                 trace_block_bio_remap(align_bio, disk_devt(mddev->gendisk),
5455                                       raid_bio->bi_iter.bi_sector);
5456         submit_bio_noacct(align_bio);
5457         return 1;
5458
5459 out_rcu_unlock:
5460         rcu_read_unlock();
5461         return 0;
5462 }
5463
5464 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5465 {
5466         struct bio *split;
5467         sector_t sector = raid_bio->bi_iter.bi_sector;
5468         unsigned chunk_sects = mddev->chunk_sectors;
5469         unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5470
5471         if (sectors < bio_sectors(raid_bio)) {
5472                 struct r5conf *conf = mddev->private;
5473                 split = bio_split(raid_bio, sectors, GFP_NOIO, &conf->bio_split);
5474                 bio_chain(split, raid_bio);
5475                 submit_bio_noacct(raid_bio);
5476                 raid_bio = split;
5477         }
5478
5479         if (!raid5_read_one_chunk(mddev, raid_bio))
5480                 return raid_bio;
5481
5482         return NULL;
5483 }
5484
5485 /* __get_priority_stripe - get the next stripe to process
5486  *
5487  * Full stripe writes are allowed to pass preread active stripes up until
5488  * the bypass_threshold is exceeded.  In general the bypass_count
5489  * increments when the handle_list is handled before the hold_list; however, it
5490  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5491  * stripe with in flight i/o.  The bypass_count will be reset when the
5492  * head of the hold_list has changed, i.e. the head was promoted to the
5493  * handle_list.
5494  */
5495 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5496 {
5497         struct stripe_head *sh, *tmp;
5498         struct list_head *handle_list = NULL;
5499         struct r5worker_group *wg;
5500         bool second_try = !r5c_is_writeback(conf->log) &&
5501                 !r5l_log_disk_error(conf);
5502         bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5503                 r5l_log_disk_error(conf);
5504
5505 again:
5506         wg = NULL;
5507         sh = NULL;
5508         if (conf->worker_cnt_per_group == 0) {
5509                 handle_list = try_loprio ? &conf->loprio_list :
5510                                         &conf->handle_list;
5511         } else if (group != ANY_GROUP) {
5512                 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5513                                 &conf->worker_groups[group].handle_list;
5514                 wg = &conf->worker_groups[group];
5515         } else {
5516                 int i;
5517                 for (i = 0; i < conf->group_cnt; i++) {
5518                         handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5519                                 &conf->worker_groups[i].handle_list;
5520                         wg = &conf->worker_groups[i];
5521                         if (!list_empty(handle_list))
5522                                 break;
5523                 }
5524         }
5525
5526         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5527                   __func__,
5528                   list_empty(handle_list) ? "empty" : "busy",
5529                   list_empty(&conf->hold_list) ? "empty" : "busy",
5530                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
5531
5532         if (!list_empty(handle_list)) {
5533                 sh = list_entry(handle_list->next, typeof(*sh), lru);
5534
5535                 if (list_empty(&conf->hold_list))
5536                         conf->bypass_count = 0;
5537                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5538                         if (conf->hold_list.next == conf->last_hold)
5539                                 conf->bypass_count++;
5540                         else {
5541                                 conf->last_hold = conf->hold_list.next;
5542                                 conf->bypass_count -= conf->bypass_threshold;
5543                                 if (conf->bypass_count < 0)
5544                                         conf->bypass_count = 0;
5545                         }
5546                 }
5547         } else if (!list_empty(&conf->hold_list) &&
5548                    ((conf->bypass_threshold &&
5549                      conf->bypass_count > conf->bypass_threshold) ||
5550                     atomic_read(&conf->pending_full_writes) == 0)) {
5551
5552                 list_for_each_entry(tmp, &conf->hold_list,  lru) {
5553                         if (conf->worker_cnt_per_group == 0 ||
5554                             group == ANY_GROUP ||
5555                             !cpu_online(tmp->cpu) ||
5556                             cpu_to_group(tmp->cpu) == group) {
5557                                 sh = tmp;
5558                                 break;
5559                         }
5560                 }
5561
5562                 if (sh) {
5563                         conf->bypass_count -= conf->bypass_threshold;
5564                         if (conf->bypass_count < 0)
5565                                 conf->bypass_count = 0;
5566                 }
5567                 wg = NULL;
5568         }
5569
5570         if (!sh) {
5571                 if (second_try)
5572                         return NULL;
5573                 second_try = true;
5574                 try_loprio = !try_loprio;
5575                 goto again;
5576         }
5577
5578         if (wg) {
5579                 wg->stripes_cnt--;
5580                 sh->group = NULL;
5581         }
5582         list_del_init(&sh->lru);
5583         BUG_ON(atomic_inc_return(&sh->count) != 1);
5584         return sh;
5585 }
5586
5587 struct raid5_plug_cb {
5588         struct blk_plug_cb      cb;
5589         struct list_head        list;
5590         struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5591 };
5592
5593 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5594 {
5595         struct raid5_plug_cb *cb = container_of(
5596                 blk_cb, struct raid5_plug_cb, cb);
5597         struct stripe_head *sh;
5598         struct mddev *mddev = cb->cb.data;
5599         struct r5conf *conf = mddev->private;
5600         int cnt = 0;
5601         int hash;
5602
5603         if (cb->list.next && !list_empty(&cb->list)) {
5604                 spin_lock_irq(&conf->device_lock);
5605                 while (!list_empty(&cb->list)) {
5606                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
5607                         list_del_init(&sh->lru);
5608                         /*
5609                          * avoid race release_stripe_plug() sees
5610                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
5611                          * is still in our list
5612                          */
5613                         smp_mb__before_atomic();
5614                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5615                         /*
5616                          * STRIPE_ON_RELEASE_LIST could be set here. In that
5617                          * case, the count is always > 1 here
5618                          */
5619                         hash = sh->hash_lock_index;
5620                         __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5621                         cnt++;
5622                 }
5623                 spin_unlock_irq(&conf->device_lock);
5624         }
5625         release_inactive_stripe_list(conf, cb->temp_inactive_list,
5626                                      NR_STRIPE_HASH_LOCKS);
5627         if (mddev->queue)
5628                 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5629         kfree(cb);
5630 }
5631
5632 static void release_stripe_plug(struct mddev *mddev,
5633                                 struct stripe_head *sh)
5634 {
5635         struct blk_plug_cb *blk_cb = blk_check_plugged(
5636                 raid5_unplug, mddev,
5637                 sizeof(struct raid5_plug_cb));
5638         struct raid5_plug_cb *cb;
5639
5640         if (!blk_cb) {
5641                 raid5_release_stripe(sh);
5642                 return;
5643         }
5644
5645         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5646
5647         if (cb->list.next == NULL) {
5648                 int i;
5649                 INIT_LIST_HEAD(&cb->list);
5650                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5651                         INIT_LIST_HEAD(cb->temp_inactive_list + i);
5652         }
5653
5654         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5655                 list_add_tail(&sh->lru, &cb->list);
5656         else
5657                 raid5_release_stripe(sh);
5658 }
5659
5660 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5661 {
5662         struct r5conf *conf = mddev->private;
5663         sector_t logical_sector, last_sector;
5664         struct stripe_head *sh;
5665         int stripe_sectors;
5666
5667         if (mddev->reshape_position != MaxSector)
5668                 /* Skip discard while reshape is happening */
5669                 return;
5670
5671         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
5672         last_sector = bio_end_sector(bi);
5673
5674         bi->bi_next = NULL;
5675
5676         stripe_sectors = conf->chunk_sectors *
5677                 (conf->raid_disks - conf->max_degraded);
5678         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5679                                                stripe_sectors);
5680         sector_div(last_sector, stripe_sectors);
5681
5682         logical_sector *= conf->chunk_sectors;
5683         last_sector *= conf->chunk_sectors;
5684
5685         for (; logical_sector < last_sector;
5686              logical_sector += RAID5_STRIPE_SECTORS(conf)) {
5687                 DEFINE_WAIT(w);
5688                 int d;
5689         again:
5690                 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5691                 prepare_to_wait(&conf->wait_for_overlap, &w,
5692                                 TASK_UNINTERRUPTIBLE);
5693                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5694                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5695                         raid5_release_stripe(sh);
5696                         schedule();
5697                         goto again;
5698                 }
5699                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5700                 spin_lock_irq(&sh->stripe_lock);
5701                 for (d = 0; d < conf->raid_disks; d++) {
5702                         if (d == sh->pd_idx || d == sh->qd_idx)
5703                                 continue;
5704                         if (sh->dev[d].towrite || sh->dev[d].toread) {
5705                                 set_bit(R5_Overlap, &sh->dev[d].flags);
5706                                 spin_unlock_irq(&sh->stripe_lock);
5707                                 raid5_release_stripe(sh);
5708                                 schedule();
5709                                 goto again;
5710                         }
5711                 }
5712                 set_bit(STRIPE_DISCARD, &sh->state);
5713                 finish_wait(&conf->wait_for_overlap, &w);
5714                 sh->overwrite_disks = 0;
5715                 for (d = 0; d < conf->raid_disks; d++) {
5716                         if (d == sh->pd_idx || d == sh->qd_idx)
5717                                 continue;
5718                         sh->dev[d].towrite = bi;
5719                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5720                         bio_inc_remaining(bi);
5721                         md_write_inc(mddev, bi);
5722                         sh->overwrite_disks++;
5723                 }
5724                 spin_unlock_irq(&sh->stripe_lock);
5725                 if (conf->mddev->bitmap) {
5726                         for (d = 0;
5727                              d < conf->raid_disks - conf->max_degraded;
5728                              d++)
5729                                 md_bitmap_startwrite(mddev->bitmap,
5730                                                      sh->sector,
5731                                                      RAID5_STRIPE_SECTORS(conf),
5732                                                      0);
5733                         sh->bm_seq = conf->seq_flush + 1;
5734                         set_bit(STRIPE_BIT_DELAY, &sh->state);
5735                 }
5736
5737                 set_bit(STRIPE_HANDLE, &sh->state);
5738                 clear_bit(STRIPE_DELAYED, &sh->state);
5739                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5740                         atomic_inc(&conf->preread_active_stripes);
5741                 release_stripe_plug(mddev, sh);
5742         }
5743
5744         bio_endio(bi);
5745 }
5746
5747 static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
5748 {
5749         struct r5conf *conf = mddev->private;
5750         int dd_idx;
5751         sector_t new_sector;
5752         sector_t logical_sector, last_sector;
5753         struct stripe_head *sh;
5754         const int rw = bio_data_dir(bi);
5755         DEFINE_WAIT(w);
5756         bool do_prepare;
5757         bool do_flush = false;
5758
5759         if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5760                 int ret = log_handle_flush_request(conf, bi);
5761
5762                 if (ret == 0)
5763                         return true;
5764                 if (ret == -ENODEV) {
5765                         if (md_flush_request(mddev, bi))
5766                                 return true;
5767                 }
5768                 /* ret == -EAGAIN, fallback */
5769                 /*
5770                  * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5771                  * we need to flush journal device
5772                  */
5773                 do_flush = bi->bi_opf & REQ_PREFLUSH;
5774         }
5775
5776         if (!md_write_start(mddev, bi))
5777                 return false;
5778         /*
5779          * If array is degraded, better not do chunk aligned read because
5780          * later we might have to read it again in order to reconstruct
5781          * data on failed drives.
5782          */
5783         if (rw == READ && mddev->degraded == 0 &&
5784             mddev->reshape_position == MaxSector) {
5785                 bi = chunk_aligned_read(mddev, bi);
5786                 if (!bi)
5787                         return true;
5788         }
5789
5790         if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5791                 make_discard_request(mddev, bi);
5792                 md_write_end(mddev);
5793                 return true;
5794         }
5795
5796         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
5797         last_sector = bio_end_sector(bi);
5798         bi->bi_next = NULL;
5799
5800         prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5801         for (; logical_sector < last_sector; logical_sector += RAID5_STRIPE_SECTORS(conf)) {
5802                 int previous;
5803                 int seq;
5804
5805                 do_prepare = false;
5806         retry:
5807                 seq = read_seqcount_begin(&conf->gen_lock);
5808                 previous = 0;
5809                 if (do_prepare)
5810                         prepare_to_wait(&conf->wait_for_overlap, &w,
5811                                 TASK_UNINTERRUPTIBLE);
5812                 if (unlikely(conf->reshape_progress != MaxSector)) {
5813                         /* spinlock is needed as reshape_progress may be
5814                          * 64bit on a 32bit platform, and so it might be
5815                          * possible to see a half-updated value
5816                          * Of course reshape_progress could change after
5817                          * the lock is dropped, so once we get a reference
5818                          * to the stripe that we think it is, we will have
5819                          * to check again.
5820                          */
5821                         spin_lock_irq(&conf->device_lock);
5822                         if (mddev->reshape_backwards
5823                             ? logical_sector < conf->reshape_progress
5824                             : logical_sector >= conf->reshape_progress) {
5825                                 previous = 1;
5826                         } else {
5827                                 if (mddev->reshape_backwards
5828                                     ? logical_sector < conf->reshape_safe
5829                                     : logical_sector >= conf->reshape_safe) {
5830                                         spin_unlock_irq(&conf->device_lock);
5831                                         schedule();
5832                                         do_prepare = true;
5833                                         goto retry;
5834                                 }
5835                         }
5836                         spin_unlock_irq(&conf->device_lock);
5837                 }
5838
5839                 new_sector = raid5_compute_sector(conf, logical_sector,
5840                                                   previous,
5841                                                   &dd_idx, NULL);
5842                 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5843                         (unsigned long long)new_sector,
5844                         (unsigned long long)logical_sector);
5845
5846                 sh = raid5_get_active_stripe(conf, new_sector, previous,
5847                                        (bi->bi_opf & REQ_RAHEAD), 0);
5848                 if (sh) {
5849                         if (unlikely(previous)) {
5850                                 /* expansion might have moved on while waiting for a
5851                                  * stripe, so we must do the range check again.
5852                                  * Expansion could still move past after this
5853                                  * test, but as we are holding a reference to
5854                                  * 'sh', we know that if that happens,
5855                                  *  STRIPE_EXPANDING will get set and the expansion
5856                                  * won't proceed until we finish with the stripe.
5857                                  */
5858                                 int must_retry = 0;
5859                                 spin_lock_irq(&conf->device_lock);
5860                                 if (mddev->reshape_backwards
5861                                     ? logical_sector >= conf->reshape_progress
5862                                     : logical_sector < conf->reshape_progress)
5863                                         /* mismatch, need to try again */
5864                                         must_retry = 1;
5865                                 spin_unlock_irq(&conf->device_lock);
5866                                 if (must_retry) {
5867                                         raid5_release_stripe(sh);
5868                                         schedule();
5869                                         do_prepare = true;
5870                                         goto retry;
5871                                 }
5872                         }
5873                         if (read_seqcount_retry(&conf->gen_lock, seq)) {
5874                                 /* Might have got the wrong stripe_head
5875                                  * by accident
5876                                  */
5877                                 raid5_release_stripe(sh);
5878                                 goto retry;
5879                         }
5880
5881                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5882                             !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5883                                 /* Stripe is busy expanding or
5884                                  * add failed due to overlap.  Flush everything
5885                                  * and wait a while
5886                                  */
5887                                 md_wakeup_thread(mddev->thread);
5888                                 raid5_release_stripe(sh);
5889                                 schedule();
5890                                 do_prepare = true;
5891                                 goto retry;
5892                         }
5893                         if (do_flush) {
5894                                 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5895                                 /* we only need flush for one stripe */
5896                                 do_flush = false;
5897                         }
5898
5899                         set_bit(STRIPE_HANDLE, &sh->state);
5900                         clear_bit(STRIPE_DELAYED, &sh->state);
5901                         if ((!sh->batch_head || sh == sh->batch_head) &&
5902                             (bi->bi_opf & REQ_SYNC) &&
5903                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5904                                 atomic_inc(&conf->preread_active_stripes);
5905                         release_stripe_plug(mddev, sh);
5906                 } else {
5907                         /* cannot get stripe for read-ahead, just give-up */
5908                         bi->bi_status = BLK_STS_IOERR;
5909                         break;
5910                 }
5911         }
5912         finish_wait(&conf->wait_for_overlap, &w);
5913
5914         if (rw == WRITE)
5915                 md_write_end(mddev);
5916         bio_endio(bi);
5917         return true;
5918 }
5919
5920 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5921
5922 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5923 {
5924         /* reshaping is quite different to recovery/resync so it is
5925          * handled quite separately ... here.
5926          *
5927          * On each call to sync_request, we gather one chunk worth of
5928          * destination stripes and flag them as expanding.
5929          * Then we find all the source stripes and request reads.
5930          * As the reads complete, handle_stripe will copy the data
5931          * into the destination stripe and release that stripe.
5932          */
5933         struct r5conf *conf = mddev->private;
5934         struct stripe_head *sh;
5935         struct md_rdev *rdev;
5936         sector_t first_sector, last_sector;
5937         int raid_disks = conf->previous_raid_disks;
5938         int data_disks = raid_disks - conf->max_degraded;
5939         int new_data_disks = conf->raid_disks - conf->max_degraded;
5940         int i;
5941         int dd_idx;
5942         sector_t writepos, readpos, safepos;
5943         sector_t stripe_addr;
5944         int reshape_sectors;
5945         struct list_head stripes;
5946         sector_t retn;
5947
5948         if (sector_nr == 0) {
5949                 /* If restarting in the middle, skip the initial sectors */
5950                 if (mddev->reshape_backwards &&
5951                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5952                         sector_nr = raid5_size(mddev, 0, 0)
5953                                 - conf->reshape_progress;
5954                 } else if (mddev->reshape_backwards &&
5955                            conf->reshape_progress == MaxSector) {
5956                         /* shouldn't happen, but just in case, finish up.*/
5957                         sector_nr = MaxSector;
5958                 } else if (!mddev->reshape_backwards &&
5959                            conf->reshape_progress > 0)
5960                         sector_nr = conf->reshape_progress;
5961                 sector_div(sector_nr, new_data_disks);
5962                 if (sector_nr) {
5963                         mddev->curr_resync_completed = sector_nr;
5964                         sysfs_notify_dirent_safe(mddev->sysfs_completed);
5965                         *skipped = 1;
5966                         retn = sector_nr;
5967                         goto finish;
5968                 }
5969         }
5970
5971         /* We need to process a full chunk at a time.
5972          * If old and new chunk sizes differ, we need to process the
5973          * largest of these
5974          */
5975
5976         reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5977
5978         /* We update the metadata at least every 10 seconds, or when
5979          * the data about to be copied would over-write the source of
5980          * the data at the front of the range.  i.e. one new_stripe
5981          * along from reshape_progress new_maps to after where
5982          * reshape_safe old_maps to
5983          */
5984         writepos = conf->reshape_progress;
5985         sector_div(writepos, new_data_disks);
5986         readpos = conf->reshape_progress;
5987         sector_div(readpos, data_disks);
5988         safepos = conf->reshape_safe;
5989         sector_div(safepos, data_disks);
5990         if (mddev->reshape_backwards) {
5991                 BUG_ON(writepos < reshape_sectors);
5992                 writepos -= reshape_sectors;
5993                 readpos += reshape_sectors;
5994                 safepos += reshape_sectors;
5995         } else {
5996                 writepos += reshape_sectors;
5997                 /* readpos and safepos are worst-case calculations.
5998                  * A negative number is overly pessimistic, and causes
5999                  * obvious problems for unsigned storage.  So clip to 0.
6000                  */
6001                 readpos -= min_t(sector_t, reshape_sectors, readpos);
6002                 safepos -= min_t(sector_t, reshape_sectors, safepos);
6003         }
6004
6005         /* Having calculated the 'writepos' possibly use it
6006          * to set 'stripe_addr' which is where we will write to.
6007          */
6008         if (mddev->reshape_backwards) {
6009                 BUG_ON(conf->reshape_progress == 0);
6010                 stripe_addr = writepos;
6011                 BUG_ON((mddev->dev_sectors &
6012                         ~((sector_t)reshape_sectors - 1))
6013                        - reshape_sectors - stripe_addr
6014                        != sector_nr);
6015         } else {
6016                 BUG_ON(writepos != sector_nr + reshape_sectors);
6017                 stripe_addr = sector_nr;
6018         }
6019
6020         /* 'writepos' is the most advanced device address we might write.
6021          * 'readpos' is the least advanced device address we might read.
6022          * 'safepos' is the least address recorded in the metadata as having
6023          *     been reshaped.
6024          * If there is a min_offset_diff, these are adjusted either by
6025          * increasing the safepos/readpos if diff is negative, or
6026          * increasing writepos if diff is positive.
6027          * If 'readpos' is then behind 'writepos', there is no way that we can
6028          * ensure safety in the face of a crash - that must be done by userspace
6029          * making a backup of the data.  So in that case there is no particular
6030          * rush to update metadata.
6031          * Otherwise if 'safepos' is behind 'writepos', then we really need to
6032          * update the metadata to advance 'safepos' to match 'readpos' so that
6033          * we can be safe in the event of a crash.
6034          * So we insist on updating metadata if safepos is behind writepos and
6035          * readpos is beyond writepos.
6036          * In any case, update the metadata every 10 seconds.
6037          * Maybe that number should be configurable, but I'm not sure it is
6038          * worth it.... maybe it could be a multiple of safemode_delay???
6039          */
6040         if (conf->min_offset_diff < 0) {
6041                 safepos += -conf->min_offset_diff;
6042                 readpos += -conf->min_offset_diff;
6043         } else
6044                 writepos += conf->min_offset_diff;
6045
6046         if ((mddev->reshape_backwards
6047              ? (safepos > writepos && readpos < writepos)
6048              : (safepos < writepos && readpos > writepos)) ||
6049             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
6050                 /* Cannot proceed until we've updated the superblock... */
6051                 wait_event(conf->wait_for_overlap,
6052                            atomic_read(&conf->reshape_stripes)==0
6053                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6054                 if (atomic_read(&conf->reshape_stripes) != 0)
6055                         return 0;
6056                 mddev->reshape_position = conf->reshape_progress;
6057                 mddev->curr_resync_completed = sector_nr;
6058                 if (!mddev->reshape_backwards)
6059                         /* Can update recovery_offset */
6060                         rdev_for_each(rdev, mddev)
6061                                 if (rdev->raid_disk >= 0 &&
6062                                     !test_bit(Journal, &rdev->flags) &&
6063                                     !test_bit(In_sync, &rdev->flags) &&
6064                                     rdev->recovery_offset < sector_nr)
6065                                         rdev->recovery_offset = sector_nr;
6066
6067                 conf->reshape_checkpoint = jiffies;
6068                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6069                 md_wakeup_thread(mddev->thread);
6070                 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
6071                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6072                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6073                         return 0;
6074                 spin_lock_irq(&conf->device_lock);
6075                 conf->reshape_safe = mddev->reshape_position;
6076                 spin_unlock_irq(&conf->device_lock);
6077                 wake_up(&conf->wait_for_overlap);
6078                 sysfs_notify_dirent_safe(mddev->sysfs_completed);
6079         }
6080
6081         INIT_LIST_HEAD(&stripes);
6082         for (i = 0; i < reshape_sectors; i += RAID5_STRIPE_SECTORS(conf)) {
6083                 int j;
6084                 int skipped_disk = 0;
6085                 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
6086                 set_bit(STRIPE_EXPANDING, &sh->state);
6087                 atomic_inc(&conf->reshape_stripes);
6088                 /* If any of this stripe is beyond the end of the old
6089                  * array, then we need to zero those blocks
6090                  */
6091                 for (j=sh->disks; j--;) {
6092                         sector_t s;
6093                         if (j == sh->pd_idx)
6094                                 continue;
6095                         if (conf->level == 6 &&
6096                             j == sh->qd_idx)
6097                                 continue;
6098                         s = raid5_compute_blocknr(sh, j, 0);
6099                         if (s < raid5_size(mddev, 0, 0)) {
6100                                 skipped_disk = 1;
6101                                 continue;
6102                         }
6103                         memset(page_address(sh->dev[j].page), 0, RAID5_STRIPE_SIZE(conf));
6104                         set_bit(R5_Expanded, &sh->dev[j].flags);
6105                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
6106                 }
6107                 if (!skipped_disk) {
6108                         set_bit(STRIPE_EXPAND_READY, &sh->state);
6109                         set_bit(STRIPE_HANDLE, &sh->state);
6110                 }
6111                 list_add(&sh->lru, &stripes);
6112         }
6113         spin_lock_irq(&conf->device_lock);
6114         if (mddev->reshape_backwards)
6115                 conf->reshape_progress -= reshape_sectors * new_data_disks;
6116         else
6117                 conf->reshape_progress += reshape_sectors * new_data_disks;
6118         spin_unlock_irq(&conf->device_lock);
6119         /* Ok, those stripe are ready. We can start scheduling
6120          * reads on the source stripes.
6121          * The source stripes are determined by mapping the first and last
6122          * block on the destination stripes.
6123          */
6124         first_sector =
6125                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
6126                                      1, &dd_idx, NULL);
6127         last_sector =
6128                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
6129                                             * new_data_disks - 1),
6130                                      1, &dd_idx, NULL);
6131         if (last_sector >= mddev->dev_sectors)
6132                 last_sector = mddev->dev_sectors - 1;
6133         while (first_sector <= last_sector) {
6134                 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
6135                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
6136                 set_bit(STRIPE_HANDLE, &sh->state);
6137                 raid5_release_stripe(sh);
6138                 first_sector += RAID5_STRIPE_SECTORS(conf);
6139         }
6140         /* Now that the sources are clearly marked, we can release
6141          * the destination stripes
6142          */
6143         while (!list_empty(&stripes)) {
6144                 sh = list_entry(stripes.next, struct stripe_head, lru);
6145                 list_del_init(&sh->lru);
6146                 raid5_release_stripe(sh);
6147         }
6148         /* If this takes us to the resync_max point where we have to pause,
6149          * then we need to write out the superblock.
6150          */
6151         sector_nr += reshape_sectors;
6152         retn = reshape_sectors;
6153 finish:
6154         if (mddev->curr_resync_completed > mddev->resync_max ||
6155             (sector_nr - mddev->curr_resync_completed) * 2
6156             >= mddev->resync_max - mddev->curr_resync_completed) {
6157                 /* Cannot proceed until we've updated the superblock... */
6158                 wait_event(conf->wait_for_overlap,
6159                            atomic_read(&conf->reshape_stripes) == 0
6160                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6161                 if (atomic_read(&conf->reshape_stripes) != 0)
6162                         goto ret;
6163                 mddev->reshape_position = conf->reshape_progress;
6164                 mddev->curr_resync_completed = sector_nr;
6165                 if (!mddev->reshape_backwards)
6166                         /* Can update recovery_offset */
6167                         rdev_for_each(rdev, mddev)
6168                                 if (rdev->raid_disk >= 0 &&
6169                                     !test_bit(Journal, &rdev->flags) &&
6170                                     !test_bit(In_sync, &rdev->flags) &&
6171                                     rdev->recovery_offset < sector_nr)
6172                                         rdev->recovery_offset = sector_nr;
6173                 conf->reshape_checkpoint = jiffies;
6174                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6175                 md_wakeup_thread(mddev->thread);
6176                 wait_event(mddev->sb_wait,
6177                            !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
6178                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6179                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6180                         goto ret;
6181                 spin_lock_irq(&conf->device_lock);
6182                 conf->reshape_safe = mddev->reshape_position;
6183                 spin_unlock_irq(&conf->device_lock);
6184                 wake_up(&conf->wait_for_overlap);
6185                 sysfs_notify_dirent_safe(mddev->sysfs_completed);
6186         }
6187 ret:
6188         return retn;
6189 }
6190
6191 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6192                                           int *skipped)
6193 {
6194         struct r5conf *conf = mddev->private;
6195         struct stripe_head *sh;
6196         sector_t max_sector = mddev->dev_sectors;
6197         sector_t sync_blocks;
6198         int still_degraded = 0;
6199         int i;
6200
6201         if (sector_nr >= max_sector) {
6202                 /* just being told to finish up .. nothing much to do */
6203
6204                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6205                         end_reshape(conf);
6206                         return 0;
6207                 }
6208
6209                 if (mddev->curr_resync < max_sector) /* aborted */
6210                         md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6211                                            &sync_blocks, 1);
6212                 else /* completed sync */
6213                         conf->fullsync = 0;
6214                 md_bitmap_close_sync(mddev->bitmap);
6215
6216                 return 0;
6217         }
6218
6219         /* Allow raid5_quiesce to complete */
6220         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6221
6222         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6223                 return reshape_request(mddev, sector_nr, skipped);
6224
6225         /* No need to check resync_max as we never do more than one
6226          * stripe, and as resync_max will always be on a chunk boundary,
6227          * if the check in md_do_sync didn't fire, there is no chance
6228          * of overstepping resync_max here
6229          */
6230
6231         /* if there is too many failed drives and we are trying
6232          * to resync, then assert that we are finished, because there is
6233          * nothing we can do.
6234          */
6235         if (mddev->degraded >= conf->max_degraded &&
6236             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6237                 sector_t rv = mddev->dev_sectors - sector_nr;
6238                 *skipped = 1;
6239                 return rv;
6240         }
6241         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6242             !conf->fullsync &&
6243             !md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6244             sync_blocks >= RAID5_STRIPE_SECTORS(conf)) {
6245                 /* we can skip this block, and probably more */
6246                 do_div(sync_blocks, RAID5_STRIPE_SECTORS(conf));
6247                 *skipped = 1;
6248                 /* keep things rounded to whole stripes */
6249                 return sync_blocks * RAID5_STRIPE_SECTORS(conf);
6250         }
6251
6252         md_bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6253
6254         sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
6255         if (sh == NULL) {
6256                 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
6257                 /* make sure we don't swamp the stripe cache if someone else
6258                  * is trying to get access
6259                  */
6260                 schedule_timeout_uninterruptible(1);
6261         }
6262         /* Need to check if array will still be degraded after recovery/resync
6263          * Note in case of > 1 drive failures it's possible we're rebuilding
6264          * one drive while leaving another faulty drive in array.
6265          */
6266         rcu_read_lock();
6267         for (i = 0; i < conf->raid_disks; i++) {
6268                 struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
6269
6270                 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6271                         still_degraded = 1;
6272         }
6273         rcu_read_unlock();
6274
6275         md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6276
6277         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6278         set_bit(STRIPE_HANDLE, &sh->state);
6279
6280         raid5_release_stripe(sh);
6281
6282         return RAID5_STRIPE_SECTORS(conf);
6283 }
6284
6285 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6286                                unsigned int offset)
6287 {
6288         /* We may not be able to submit a whole bio at once as there
6289          * may not be enough stripe_heads available.
6290          * We cannot pre-allocate enough stripe_heads as we may need
6291          * more than exist in the cache (if we allow ever large chunks).
6292          * So we do one stripe head at a time and record in
6293          * ->bi_hw_segments how many have been done.
6294          *
6295          * We *know* that this entire raid_bio is in one chunk, so
6296          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6297          */
6298         struct stripe_head *sh;
6299         int dd_idx;
6300         sector_t sector, logical_sector, last_sector;
6301         int scnt = 0;
6302         int handled = 0;
6303
6304         logical_sector = raid_bio->bi_iter.bi_sector &
6305                 ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
6306         sector = raid5_compute_sector(conf, logical_sector,
6307                                       0, &dd_idx, NULL);
6308         last_sector = bio_end_sector(raid_bio);
6309
6310         for (; logical_sector < last_sector;
6311              logical_sector += RAID5_STRIPE_SECTORS(conf),
6312                      sector += RAID5_STRIPE_SECTORS(conf),
6313                      scnt++) {
6314
6315                 if (scnt < offset)
6316                         /* already done this stripe */
6317                         continue;
6318
6319                 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
6320
6321                 if (!sh) {
6322                         /* failed to get a stripe - must wait */
6323                         conf->retry_read_aligned = raid_bio;
6324                         conf->retry_read_offset = scnt;
6325                         return handled;
6326                 }
6327
6328                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6329                         raid5_release_stripe(sh);
6330                         conf->retry_read_aligned = raid_bio;
6331                         conf->retry_read_offset = scnt;
6332                         return handled;
6333                 }
6334
6335                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6336                 handle_stripe(sh);
6337                 raid5_release_stripe(sh);
6338                 handled++;
6339         }
6340
6341         bio_endio(raid_bio);
6342
6343         if (atomic_dec_and_test(&conf->active_aligned_reads))
6344                 wake_up(&conf->wait_for_quiescent);
6345         return handled;
6346 }
6347
6348 static int handle_active_stripes(struct r5conf *conf, int group,
6349                                  struct r5worker *worker,
6350                                  struct list_head *temp_inactive_list)
6351                 __releases(&conf->device_lock)
6352                 __acquires(&conf->device_lock)
6353 {
6354         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6355         int i, batch_size = 0, hash;
6356         bool release_inactive = false;
6357
6358         while (batch_size < MAX_STRIPE_BATCH &&
6359                         (sh = __get_priority_stripe(conf, group)) != NULL)
6360                 batch[batch_size++] = sh;
6361
6362         if (batch_size == 0) {
6363                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6364                         if (!list_empty(temp_inactive_list + i))
6365                                 break;
6366                 if (i == NR_STRIPE_HASH_LOCKS) {
6367                         spin_unlock_irq(&conf->device_lock);
6368                         log_flush_stripe_to_raid(conf);
6369                         spin_lock_irq(&conf->device_lock);
6370                         return batch_size;
6371                 }
6372                 release_inactive = true;
6373         }
6374         spin_unlock_irq(&conf->device_lock);
6375
6376         release_inactive_stripe_list(conf, temp_inactive_list,
6377                                      NR_STRIPE_HASH_LOCKS);
6378
6379         r5l_flush_stripe_to_raid(conf->log);
6380         if (release_inactive) {
6381                 spin_lock_irq(&conf->device_lock);
6382                 return 0;
6383         }
6384
6385         for (i = 0; i < batch_size; i++)
6386                 handle_stripe(batch[i]);
6387         log_write_stripe_run(conf);
6388
6389         cond_resched();
6390
6391         spin_lock_irq(&conf->device_lock);
6392         for (i = 0; i < batch_size; i++) {
6393                 hash = batch[i]->hash_lock_index;
6394                 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6395         }
6396         return batch_size;
6397 }
6398
6399 static void raid5_do_work(struct work_struct *work)
6400 {
6401         struct r5worker *worker = container_of(work, struct r5worker, work);
6402         struct r5worker_group *group = worker->group;
6403         struct r5conf *conf = group->conf;
6404         struct mddev *mddev = conf->mddev;
6405         int group_id = group - conf->worker_groups;
6406         int handled;
6407         struct blk_plug plug;
6408
6409         pr_debug("+++ raid5worker active\n");
6410
6411         blk_start_plug(&plug);
6412         handled = 0;
6413         spin_lock_irq(&conf->device_lock);
6414         while (1) {
6415                 int batch_size, released;
6416
6417                 released = release_stripe_list(conf, worker->temp_inactive_list);
6418
6419                 batch_size = handle_active_stripes(conf, group_id, worker,
6420                                                    worker->temp_inactive_list);
6421                 worker->working = false;
6422                 if (!batch_size && !released)
6423                         break;
6424                 handled += batch_size;
6425                 wait_event_lock_irq(mddev->sb_wait,
6426                         !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6427                         conf->device_lock);
6428         }
6429         pr_debug("%d stripes handled\n", handled);
6430
6431         spin_unlock_irq(&conf->device_lock);
6432
6433         flush_deferred_bios(conf);
6434
6435         r5l_flush_stripe_to_raid(conf->log);
6436
6437         async_tx_issue_pending_all();
6438         blk_finish_plug(&plug);
6439
6440         pr_debug("--- raid5worker inactive\n");
6441 }
6442
6443 /*
6444  * This is our raid5 kernel thread.
6445  *
6446  * We scan the hash table for stripes which can be handled now.
6447  * During the scan, completed stripes are saved for us by the interrupt
6448  * handler, so that they will not have to wait for our next wakeup.
6449  */
6450 static void raid5d(struct md_thread *thread)
6451 {
6452         struct mddev *mddev = thread->mddev;
6453         struct r5conf *conf = mddev->private;
6454         int handled;
6455         struct blk_plug plug;
6456
6457         pr_debug("+++ raid5d active\n");
6458
6459         md_check_recovery(mddev);
6460
6461         blk_start_plug(&plug);
6462         handled = 0;
6463         spin_lock_irq(&conf->device_lock);
6464         while (1) {
6465                 struct bio *bio;
6466                 int batch_size, released;
6467                 unsigned int offset;
6468
6469                 released = release_stripe_list(conf, conf->temp_inactive_list);
6470                 if (released)
6471                         clear_bit(R5_DID_ALLOC, &conf->cache_state);
6472
6473                 if (
6474                     !list_empty(&conf->bitmap_list)) {
6475                         /* Now is a good time to flush some bitmap updates */
6476                         conf->seq_flush++;
6477                         spin_unlock_irq(&conf->device_lock);
6478                         md_bitmap_unplug(mddev->bitmap);
6479                         spin_lock_irq(&conf->device_lock);
6480                         conf->seq_write = conf->seq_flush;
6481                         activate_bit_delay(conf, conf->temp_inactive_list);
6482                 }
6483                 raid5_activate_delayed(conf);
6484
6485                 while ((bio = remove_bio_from_retry(conf, &offset))) {
6486                         int ok;
6487                         spin_unlock_irq(&conf->device_lock);
6488                         ok = retry_aligned_read(conf, bio, offset);
6489                         spin_lock_irq(&conf->device_lock);
6490                         if (!ok)
6491                                 break;
6492                         handled++;
6493                 }
6494
6495                 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6496                                                    conf->temp_inactive_list);
6497                 if (!batch_size && !released)
6498                         break;
6499                 handled += batch_size;
6500
6501                 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6502                         spin_unlock_irq(&conf->device_lock);
6503                         md_check_recovery(mddev);
6504                         spin_lock_irq(&conf->device_lock);
6505                 }
6506         }
6507         pr_debug("%d stripes handled\n", handled);
6508
6509         spin_unlock_irq(&conf->device_lock);
6510         if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6511             mutex_trylock(&conf->cache_size_mutex)) {
6512                 grow_one_stripe(conf, __GFP_NOWARN);
6513                 /* Set flag even if allocation failed.  This helps
6514                  * slow down allocation requests when mem is short
6515                  */
6516                 set_bit(R5_DID_ALLOC, &conf->cache_state);
6517                 mutex_unlock(&conf->cache_size_mutex);
6518         }
6519
6520         flush_deferred_bios(conf);
6521
6522         r5l_flush_stripe_to_raid(conf->log);
6523
6524         async_tx_issue_pending_all();
6525         blk_finish_plug(&plug);
6526
6527         pr_debug("--- raid5d inactive\n");
6528 }
6529
6530 static ssize_t
6531 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6532 {
6533         struct r5conf *conf;
6534         int ret = 0;
6535         spin_lock(&mddev->lock);
6536         conf = mddev->private;
6537         if (conf)
6538                 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6539         spin_unlock(&mddev->lock);
6540         return ret;
6541 }
6542
6543 int
6544 raid5_set_cache_size(struct mddev *mddev, int size)
6545 {
6546         int result = 0;
6547         struct r5conf *conf = mddev->private;
6548
6549         if (size <= 16 || size > 32768)
6550                 return -EINVAL;
6551
6552         conf->min_nr_stripes = size;
6553         mutex_lock(&conf->cache_size_mutex);
6554         while (size < conf->max_nr_stripes &&
6555                drop_one_stripe(conf))
6556                 ;
6557         mutex_unlock(&conf->cache_size_mutex);
6558
6559         md_allow_write(mddev);
6560
6561         mutex_lock(&conf->cache_size_mutex);
6562         while (size > conf->max_nr_stripes)
6563                 if (!grow_one_stripe(conf, GFP_KERNEL)) {
6564                         conf->min_nr_stripes = conf->max_nr_stripes;
6565                         result = -ENOMEM;
6566                         break;
6567                 }
6568         mutex_unlock(&conf->cache_size_mutex);
6569
6570         return result;
6571 }
6572 EXPORT_SYMBOL(raid5_set_cache_size);
6573
6574 static ssize_t
6575 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6576 {
6577         struct r5conf *conf;
6578         unsigned long new;
6579         int err;
6580
6581         if (len >= PAGE_SIZE)
6582                 return -EINVAL;
6583         if (kstrtoul(page, 10, &new))
6584                 return -EINVAL;
6585         err = mddev_lock(mddev);
6586         if (err)
6587                 return err;
6588         conf = mddev->private;
6589         if (!conf)
6590                 err = -ENODEV;
6591         else
6592                 err = raid5_set_cache_size(mddev, new);
6593         mddev_unlock(mddev);
6594
6595         return err ?: len;
6596 }
6597
6598 static struct md_sysfs_entry
6599 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6600                                 raid5_show_stripe_cache_size,
6601                                 raid5_store_stripe_cache_size);
6602
6603 static ssize_t
6604 raid5_show_rmw_level(struct mddev  *mddev, char *page)
6605 {
6606         struct r5conf *conf = mddev->private;
6607         if (conf)
6608                 return sprintf(page, "%d\n", conf->rmw_level);
6609         else
6610                 return 0;
6611 }
6612
6613 static ssize_t
6614 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6615 {
6616         struct r5conf *conf = mddev->private;
6617         unsigned long new;
6618
6619         if (!conf)
6620                 return -ENODEV;
6621
6622         if (len >= PAGE_SIZE)
6623                 return -EINVAL;
6624
6625         if (kstrtoul(page, 10, &new))
6626                 return -EINVAL;
6627
6628         if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6629                 return -EINVAL;
6630
6631         if (new != PARITY_DISABLE_RMW &&
6632             new != PARITY_ENABLE_RMW &&
6633             new != PARITY_PREFER_RMW)
6634                 return -EINVAL;
6635
6636         conf->rmw_level = new;
6637         return len;
6638 }
6639
6640 static struct md_sysfs_entry
6641 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6642                          raid5_show_rmw_level,
6643                          raid5_store_rmw_level);
6644
6645 static ssize_t
6646 raid5_show_stripe_size(struct mddev  *mddev, char *page)
6647 {
6648         struct r5conf *conf;
6649         int ret = 0;
6650
6651         spin_lock(&mddev->lock);
6652         conf = mddev->private;
6653         if (conf)
6654                 ret = sprintf(page, "%lu\n", RAID5_STRIPE_SIZE(conf));
6655         spin_unlock(&mddev->lock);
6656         return ret;
6657 }
6658
6659 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
6660 static ssize_t
6661 raid5_store_stripe_size(struct mddev  *mddev, const char *page, size_t len)
6662 {
6663         struct r5conf *conf;
6664         unsigned long new;
6665         int err;
6666         int size;
6667
6668         if (len >= PAGE_SIZE)
6669                 return -EINVAL;
6670         if (kstrtoul(page, 10, &new))
6671                 return -EINVAL;
6672
6673         /*
6674          * The value should not be bigger than PAGE_SIZE. It requires to
6675          * be multiple of DEFAULT_STRIPE_SIZE and the value should be power
6676          * of two.
6677          */
6678         if (new % DEFAULT_STRIPE_SIZE != 0 ||
6679                         new > PAGE_SIZE || new == 0 ||
6680                         new != roundup_pow_of_two(new))
6681                 return -EINVAL;
6682
6683         err = mddev_lock(mddev);
6684         if (err)
6685                 return err;
6686
6687         conf = mddev->private;
6688         if (!conf) {
6689                 err = -ENODEV;
6690                 goto out_unlock;
6691         }
6692
6693         if (new == conf->stripe_size)
6694                 goto out_unlock;
6695
6696         pr_debug("md/raid: change stripe_size from %lu to %lu\n",
6697                         conf->stripe_size, new);
6698
6699         if (mddev->sync_thread ||
6700                 test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) ||
6701                 mddev->reshape_position != MaxSector ||
6702                 mddev->sysfs_active) {
6703                 err = -EBUSY;
6704                 goto out_unlock;
6705         }
6706
6707         mddev_suspend(mddev);
6708         mutex_lock(&conf->cache_size_mutex);
6709         size = conf->max_nr_stripes;
6710
6711         shrink_stripes(conf);
6712
6713         conf->stripe_size = new;
6714         conf->stripe_shift = ilog2(new) - 9;
6715         conf->stripe_sectors = new >> 9;
6716         if (grow_stripes(conf, size)) {
6717                 pr_warn("md/raid:%s: couldn't allocate buffers\n",
6718                                 mdname(mddev));
6719                 err = -ENOMEM;
6720         }
6721         mutex_unlock(&conf->cache_size_mutex);
6722         mddev_resume(mddev);
6723
6724 out_unlock:
6725         mddev_unlock(mddev);
6726         return err ?: len;
6727 }
6728
6729 static struct md_sysfs_entry
6730 raid5_stripe_size = __ATTR(stripe_size, 0644,
6731                          raid5_show_stripe_size,
6732                          raid5_store_stripe_size);
6733 #else
6734 static struct md_sysfs_entry
6735 raid5_stripe_size = __ATTR(stripe_size, 0444,
6736                          raid5_show_stripe_size,
6737                          NULL);
6738 #endif
6739
6740 static ssize_t
6741 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6742 {
6743         struct r5conf *conf;
6744         int ret = 0;
6745         spin_lock(&mddev->lock);
6746         conf = mddev->private;
6747         if (conf)
6748                 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6749         spin_unlock(&mddev->lock);
6750         return ret;
6751 }
6752
6753 static ssize_t
6754 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6755 {
6756         struct r5conf *conf;
6757         unsigned long new;
6758         int err;
6759
6760         if (len >= PAGE_SIZE)
6761                 return -EINVAL;
6762         if (kstrtoul(page, 10, &new))
6763                 return -EINVAL;
6764
6765         err = mddev_lock(mddev);
6766         if (err)
6767                 return err;
6768         conf = mddev->private;
6769         if (!conf)
6770                 err = -ENODEV;
6771         else if (new > conf->min_nr_stripes)
6772                 err = -EINVAL;
6773         else
6774                 conf->bypass_threshold = new;
6775         mddev_unlock(mddev);
6776         return err ?: len;
6777 }
6778
6779 static struct md_sysfs_entry
6780 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6781                                         S_IRUGO | S_IWUSR,
6782                                         raid5_show_preread_threshold,
6783                                         raid5_store_preread_threshold);
6784
6785 static ssize_t
6786 raid5_show_skip_copy(struct mddev *mddev, char *page)
6787 {
6788         struct r5conf *conf;
6789         int ret = 0;
6790         spin_lock(&mddev->lock);
6791         conf = mddev->private;
6792         if (conf)
6793                 ret = sprintf(page, "%d\n", conf->skip_copy);
6794         spin_unlock(&mddev->lock);
6795         return ret;
6796 }
6797
6798 static ssize_t
6799 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6800 {
6801         struct r5conf *conf;
6802         unsigned long new;
6803         int err;
6804
6805         if (len >= PAGE_SIZE)
6806                 return -EINVAL;
6807         if (kstrtoul(page, 10, &new))
6808                 return -EINVAL;
6809         new = !!new;
6810
6811         err = mddev_lock(mddev);
6812         if (err)
6813                 return err;
6814         conf = mddev->private;
6815         if (!conf)
6816                 err = -ENODEV;
6817         else if (new != conf->skip_copy) {
6818                 struct request_queue *q = mddev->queue;
6819
6820                 mddev_suspend(mddev);
6821                 conf->skip_copy = new;
6822                 if (new)
6823                         blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES, q);
6824                 else
6825                         blk_queue_flag_clear(QUEUE_FLAG_STABLE_WRITES, q);
6826                 mddev_resume(mddev);
6827         }
6828         mddev_unlock(mddev);
6829         return err ?: len;
6830 }
6831
6832 static struct md_sysfs_entry
6833 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6834                                         raid5_show_skip_copy,
6835                                         raid5_store_skip_copy);
6836
6837 static ssize_t
6838 stripe_cache_active_show(struct mddev *mddev, char *page)
6839 {
6840         struct r5conf *conf = mddev->private;
6841         if (conf)
6842                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6843         else
6844                 return 0;
6845 }
6846
6847 static struct md_sysfs_entry
6848 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6849
6850 static ssize_t
6851 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6852 {
6853         struct r5conf *conf;
6854         int ret = 0;
6855         spin_lock(&mddev->lock);
6856         conf = mddev->private;
6857         if (conf)
6858                 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6859         spin_unlock(&mddev->lock);
6860         return ret;
6861 }
6862
6863 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6864                                int *group_cnt,
6865                                struct r5worker_group **worker_groups);
6866 static ssize_t
6867 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6868 {
6869         struct r5conf *conf;
6870         unsigned int new;
6871         int err;
6872         struct r5worker_group *new_groups, *old_groups;
6873         int group_cnt;
6874
6875         if (len >= PAGE_SIZE)
6876                 return -EINVAL;
6877         if (kstrtouint(page, 10, &new))
6878                 return -EINVAL;
6879         /* 8192 should be big enough */
6880         if (new > 8192)
6881                 return -EINVAL;
6882
6883         err = mddev_lock(mddev);
6884         if (err)
6885                 return err;
6886         conf = mddev->private;
6887         if (!conf)
6888                 err = -ENODEV;
6889         else if (new != conf->worker_cnt_per_group) {
6890                 mddev_suspend(mddev);
6891
6892                 old_groups = conf->worker_groups;
6893                 if (old_groups)
6894                         flush_workqueue(raid5_wq);
6895
6896                 err = alloc_thread_groups(conf, new, &group_cnt, &new_groups);
6897                 if (!err) {
6898                         spin_lock_irq(&conf->device_lock);
6899                         conf->group_cnt = group_cnt;
6900                         conf->worker_cnt_per_group = new;
6901                         conf->worker_groups = new_groups;
6902                         spin_unlock_irq(&conf->device_lock);
6903
6904                         if (old_groups)
6905                                 kfree(old_groups[0].workers);
6906                         kfree(old_groups);
6907                 }
6908                 mddev_resume(mddev);
6909         }
6910         mddev_unlock(mddev);
6911
6912         return err ?: len;
6913 }
6914
6915 static struct md_sysfs_entry
6916 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6917                                 raid5_show_group_thread_cnt,
6918                                 raid5_store_group_thread_cnt);
6919
6920 static struct attribute *raid5_attrs[] =  {
6921         &raid5_stripecache_size.attr,
6922         &raid5_stripecache_active.attr,
6923         &raid5_preread_bypass_threshold.attr,
6924         &raid5_group_thread_cnt.attr,
6925         &raid5_skip_copy.attr,
6926         &raid5_rmw_level.attr,
6927         &raid5_stripe_size.attr,
6928         &r5c_journal_mode.attr,
6929         &ppl_write_hint.attr,
6930         NULL,
6931 };
6932 static struct attribute_group raid5_attrs_group = {
6933         .name = NULL,
6934         .attrs = raid5_attrs,
6935 };
6936
6937 static int alloc_thread_groups(struct r5conf *conf, int cnt, int *group_cnt,
6938                                struct r5worker_group **worker_groups)
6939 {
6940         int i, j, k;
6941         ssize_t size;
6942         struct r5worker *workers;
6943
6944         if (cnt == 0) {
6945                 *group_cnt = 0;
6946                 *worker_groups = NULL;
6947                 return 0;
6948         }
6949         *group_cnt = num_possible_nodes();
6950         size = sizeof(struct r5worker) * cnt;
6951         workers = kcalloc(size, *group_cnt, GFP_NOIO);
6952         *worker_groups = kcalloc(*group_cnt, sizeof(struct r5worker_group),
6953                                  GFP_NOIO);
6954         if (!*worker_groups || !workers) {
6955                 kfree(workers);
6956                 kfree(*worker_groups);
6957                 return -ENOMEM;
6958         }
6959
6960         for (i = 0; i < *group_cnt; i++) {
6961                 struct r5worker_group *group;
6962
6963                 group = &(*worker_groups)[i];
6964                 INIT_LIST_HEAD(&group->handle_list);
6965                 INIT_LIST_HEAD(&group->loprio_list);
6966                 group->conf = conf;
6967                 group->workers = workers + i * cnt;
6968
6969                 for (j = 0; j < cnt; j++) {
6970                         struct r5worker *worker = group->workers + j;
6971                         worker->group = group;
6972                         INIT_WORK(&worker->work, raid5_do_work);
6973
6974                         for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6975                                 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6976                 }
6977         }
6978
6979         return 0;
6980 }
6981
6982 static void free_thread_groups(struct r5conf *conf)
6983 {
6984         if (conf->worker_groups)
6985                 kfree(conf->worker_groups[0].workers);
6986         kfree(conf->worker_groups);
6987         conf->worker_groups = NULL;
6988 }
6989
6990 static sector_t
6991 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6992 {
6993         struct r5conf *conf = mddev->private;
6994
6995         if (!sectors)
6996                 sectors = mddev->dev_sectors;
6997         if (!raid_disks)
6998                 /* size is defined by the smallest of previous and new size */
6999                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
7000
7001         sectors &= ~((sector_t)conf->chunk_sectors - 1);
7002         sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
7003         return sectors * (raid_disks - conf->max_degraded);
7004 }
7005
7006 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
7007 {
7008         safe_put_page(percpu->spare_page);
7009         percpu->spare_page = NULL;
7010         kvfree(percpu->scribble);
7011         percpu->scribble = NULL;
7012 }
7013
7014 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
7015 {
7016         if (conf->level == 6 && !percpu->spare_page) {
7017                 percpu->spare_page = alloc_page(GFP_KERNEL);
7018                 if (!percpu->spare_page)
7019                         return -ENOMEM;
7020         }
7021
7022         if (scribble_alloc(percpu,
7023                            max(conf->raid_disks,
7024                                conf->previous_raid_disks),
7025                            max(conf->chunk_sectors,
7026                                conf->prev_chunk_sectors)
7027                            / RAID5_STRIPE_SECTORS(conf))) {
7028                 free_scratch_buffer(conf, percpu);
7029                 return -ENOMEM;
7030         }
7031
7032         return 0;
7033 }
7034
7035 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
7036 {
7037         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
7038
7039         free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
7040         return 0;
7041 }
7042
7043 static void raid5_free_percpu(struct r5conf *conf)
7044 {
7045         if (!conf->percpu)
7046                 return;
7047
7048         cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
7049         free_percpu(conf->percpu);
7050 }
7051
7052 static void free_conf(struct r5conf *conf)
7053 {
7054         int i;
7055
7056         log_exit(conf);
7057
7058         unregister_shrinker(&conf->shrinker);
7059         free_thread_groups(conf);
7060         shrink_stripes(conf);
7061         raid5_free_percpu(conf);
7062         for (i = 0; i < conf->pool_size; i++)
7063                 if (conf->disks[i].extra_page)
7064                         put_page(conf->disks[i].extra_page);
7065         kfree(conf->disks);
7066         bioset_exit(&conf->bio_split);
7067         kfree(conf->stripe_hashtbl);
7068         kfree(conf->pending_data);
7069         kfree(conf);
7070 }
7071
7072 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
7073 {
7074         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
7075         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
7076
7077         if (alloc_scratch_buffer(conf, percpu)) {
7078                 pr_warn("%s: failed memory allocation for cpu%u\n",
7079                         __func__, cpu);
7080                 return -ENOMEM;
7081         }
7082         return 0;
7083 }
7084
7085 static int raid5_alloc_percpu(struct r5conf *conf)
7086 {
7087         int err = 0;
7088
7089         conf->percpu = alloc_percpu(struct raid5_percpu);
7090         if (!conf->percpu)
7091                 return -ENOMEM;
7092
7093         err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
7094         if (!err) {
7095                 conf->scribble_disks = max(conf->raid_disks,
7096                         conf->previous_raid_disks);
7097                 conf->scribble_sectors = max(conf->chunk_sectors,
7098                         conf->prev_chunk_sectors);
7099         }
7100         return err;
7101 }
7102
7103 static unsigned long raid5_cache_scan(struct shrinker *shrink,
7104                                       struct shrink_control *sc)
7105 {
7106         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
7107         unsigned long ret = SHRINK_STOP;
7108
7109         if (mutex_trylock(&conf->cache_size_mutex)) {
7110                 ret= 0;
7111                 while (ret < sc->nr_to_scan &&
7112                        conf->max_nr_stripes > conf->min_nr_stripes) {
7113                         if (drop_one_stripe(conf) == 0) {
7114                                 ret = SHRINK_STOP;
7115                                 break;
7116                         }
7117                         ret++;
7118                 }
7119                 mutex_unlock(&conf->cache_size_mutex);
7120         }
7121         return ret;
7122 }
7123
7124 static unsigned long raid5_cache_count(struct shrinker *shrink,
7125                                        struct shrink_control *sc)
7126 {
7127         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
7128
7129         if (conf->max_nr_stripes < conf->min_nr_stripes)
7130                 /* unlikely, but not impossible */
7131                 return 0;
7132         return conf->max_nr_stripes - conf->min_nr_stripes;
7133 }
7134
7135 static struct r5conf *setup_conf(struct mddev *mddev)
7136 {
7137         struct r5conf *conf;
7138         int raid_disk, memory, max_disks;
7139         struct md_rdev *rdev;
7140         struct disk_info *disk;
7141         char pers_name[6];
7142         int i;
7143         int group_cnt;
7144         struct r5worker_group *new_group;
7145         int ret;
7146
7147         if (mddev->new_level != 5
7148             && mddev->new_level != 4
7149             && mddev->new_level != 6) {
7150                 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
7151                         mdname(mddev), mddev->new_level);
7152                 return ERR_PTR(-EIO);
7153         }
7154         if ((mddev->new_level == 5
7155              && !algorithm_valid_raid5(mddev->new_layout)) ||
7156             (mddev->new_level == 6
7157              && !algorithm_valid_raid6(mddev->new_layout))) {
7158                 pr_warn("md/raid:%s: layout %d not supported\n",
7159                         mdname(mddev), mddev->new_layout);
7160                 return ERR_PTR(-EIO);
7161         }
7162         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
7163                 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
7164                         mdname(mddev), mddev->raid_disks);
7165                 return ERR_PTR(-EINVAL);
7166         }
7167
7168         if (!mddev->new_chunk_sectors ||
7169             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
7170             !is_power_of_2(mddev->new_chunk_sectors)) {
7171                 pr_warn("md/raid:%s: invalid chunk size %d\n",
7172                         mdname(mddev), mddev->new_chunk_sectors << 9);
7173                 return ERR_PTR(-EINVAL);
7174         }
7175
7176         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
7177         if (conf == NULL)
7178                 goto abort;
7179
7180 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
7181         conf->stripe_size = DEFAULT_STRIPE_SIZE;
7182         conf->stripe_shift = ilog2(DEFAULT_STRIPE_SIZE) - 9;
7183         conf->stripe_sectors = DEFAULT_STRIPE_SIZE >> 9;
7184 #endif
7185         INIT_LIST_HEAD(&conf->free_list);
7186         INIT_LIST_HEAD(&conf->pending_list);
7187         conf->pending_data = kcalloc(PENDING_IO_MAX,
7188                                      sizeof(struct r5pending_data),
7189                                      GFP_KERNEL);
7190         if (!conf->pending_data)
7191                 goto abort;
7192         for (i = 0; i < PENDING_IO_MAX; i++)
7193                 list_add(&conf->pending_data[i].sibling, &conf->free_list);
7194         /* Don't enable multi-threading by default*/
7195         if (!alloc_thread_groups(conf, 0, &group_cnt, &new_group)) {
7196                 conf->group_cnt = group_cnt;
7197                 conf->worker_cnt_per_group = 0;
7198                 conf->worker_groups = new_group;
7199         } else
7200                 goto abort;
7201         spin_lock_init(&conf->device_lock);
7202         seqcount_spinlock_init(&conf->gen_lock, &conf->device_lock);
7203         mutex_init(&conf->cache_size_mutex);
7204         init_waitqueue_head(&conf->wait_for_quiescent);
7205         init_waitqueue_head(&conf->wait_for_stripe);
7206         init_waitqueue_head(&conf->wait_for_overlap);
7207         INIT_LIST_HEAD(&conf->handle_list);
7208         INIT_LIST_HEAD(&conf->loprio_list);
7209         INIT_LIST_HEAD(&conf->hold_list);
7210         INIT_LIST_HEAD(&conf->delayed_list);
7211         INIT_LIST_HEAD(&conf->bitmap_list);
7212         init_llist_head(&conf->released_stripes);
7213         atomic_set(&conf->active_stripes, 0);
7214         atomic_set(&conf->preread_active_stripes, 0);
7215         atomic_set(&conf->active_aligned_reads, 0);
7216         spin_lock_init(&conf->pending_bios_lock);
7217         conf->batch_bio_dispatch = true;
7218         rdev_for_each(rdev, mddev) {
7219                 if (test_bit(Journal, &rdev->flags))
7220                         continue;
7221                 if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
7222                         conf->batch_bio_dispatch = false;
7223                         break;
7224                 }
7225         }
7226
7227         conf->bypass_threshold = BYPASS_THRESHOLD;
7228         conf->recovery_disabled = mddev->recovery_disabled - 1;
7229
7230         conf->raid_disks = mddev->raid_disks;
7231         if (mddev->reshape_position == MaxSector)
7232                 conf->previous_raid_disks = mddev->raid_disks;
7233         else
7234                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
7235         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
7236
7237         conf->disks = kcalloc(max_disks, sizeof(struct disk_info),
7238                               GFP_KERNEL);
7239
7240         if (!conf->disks)
7241                 goto abort;
7242
7243         for (i = 0; i < max_disks; i++) {
7244                 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
7245                 if (!conf->disks[i].extra_page)
7246                         goto abort;
7247         }
7248
7249         ret = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
7250         if (ret)
7251                 goto abort;
7252         conf->mddev = mddev;
7253
7254         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
7255                 goto abort;
7256
7257         /* We init hash_locks[0] separately to that it can be used
7258          * as the reference lock in the spin_lock_nest_lock() call
7259          * in lock_all_device_hash_locks_irq in order to convince
7260          * lockdep that we know what we are doing.
7261          */
7262         spin_lock_init(conf->hash_locks);
7263         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
7264                 spin_lock_init(conf->hash_locks + i);
7265
7266         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7267                 INIT_LIST_HEAD(conf->inactive_list + i);
7268
7269         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7270                 INIT_LIST_HEAD(conf->temp_inactive_list + i);
7271
7272         atomic_set(&conf->r5c_cached_full_stripes, 0);
7273         INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
7274         atomic_set(&conf->r5c_cached_partial_stripes, 0);
7275         INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
7276         atomic_set(&conf->r5c_flushing_full_stripes, 0);
7277         atomic_set(&conf->r5c_flushing_partial_stripes, 0);
7278
7279         conf->level = mddev->new_level;
7280         conf->chunk_sectors = mddev->new_chunk_sectors;
7281         if (raid5_alloc_percpu(conf) != 0)
7282                 goto abort;
7283
7284         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
7285
7286         rdev_for_each(rdev, mddev) {
7287                 raid_disk = rdev->raid_disk;
7288                 if (raid_disk >= max_disks
7289                     || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7290                         continue;
7291                 disk = conf->disks + raid_disk;
7292
7293                 if (test_bit(Replacement, &rdev->flags)) {
7294                         if (disk->replacement)
7295                                 goto abort;
7296                         disk->replacement = rdev;
7297                 } else {
7298                         if (disk->rdev)
7299                                 goto abort;
7300                         disk->rdev = rdev;
7301                 }
7302
7303                 if (test_bit(In_sync, &rdev->flags)) {
7304                         char b[BDEVNAME_SIZE];
7305                         pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7306                                 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
7307                 } else if (rdev->saved_raid_disk != raid_disk)
7308                         /* Cannot rely on bitmap to complete recovery */
7309                         conf->fullsync = 1;
7310         }
7311
7312         conf->level = mddev->new_level;
7313         if (conf->level == 6) {
7314                 conf->max_degraded = 2;
7315                 if (raid6_call.xor_syndrome)
7316                         conf->rmw_level = PARITY_ENABLE_RMW;
7317                 else
7318                         conf->rmw_level = PARITY_DISABLE_RMW;
7319         } else {
7320                 conf->max_degraded = 1;
7321                 conf->rmw_level = PARITY_ENABLE_RMW;
7322         }
7323         conf->algorithm = mddev->new_layout;
7324         conf->reshape_progress = mddev->reshape_position;
7325         if (conf->reshape_progress != MaxSector) {
7326                 conf->prev_chunk_sectors = mddev->chunk_sectors;
7327                 conf->prev_algo = mddev->layout;
7328         } else {
7329                 conf->prev_chunk_sectors = conf->chunk_sectors;
7330                 conf->prev_algo = conf->algorithm;
7331         }
7332
7333         conf->min_nr_stripes = NR_STRIPES;
7334         if (mddev->reshape_position != MaxSector) {
7335                 int stripes = max_t(int,
7336                         ((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4,
7337                         ((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4);
7338                 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7339                 if (conf->min_nr_stripes != NR_STRIPES)
7340                         pr_info("md/raid:%s: force stripe size %d for reshape\n",
7341                                 mdname(mddev), conf->min_nr_stripes);
7342         }
7343         memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7344                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7345         atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7346         if (grow_stripes(conf, conf->min_nr_stripes)) {
7347                 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7348                         mdname(mddev), memory);
7349                 goto abort;
7350         } else
7351                 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7352         /*
7353          * Losing a stripe head costs more than the time to refill it,
7354          * it reduces the queue depth and so can hurt throughput.
7355          * So set it rather large, scaled by number of devices.
7356          */
7357         conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7358         conf->shrinker.scan_objects = raid5_cache_scan;
7359         conf->shrinker.count_objects = raid5_cache_count;
7360         conf->shrinker.batch = 128;
7361         conf->shrinker.flags = 0;
7362         if (register_shrinker(&conf->shrinker)) {
7363                 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7364                         mdname(mddev));
7365                 goto abort;
7366         }
7367
7368         sprintf(pers_name, "raid%d", mddev->new_level);
7369         conf->thread = md_register_thread(raid5d, mddev, pers_name);
7370         if (!conf->thread) {
7371                 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7372                         mdname(mddev));
7373                 goto abort;
7374         }
7375
7376         return conf;
7377
7378  abort:
7379         if (conf) {
7380                 free_conf(conf);
7381                 return ERR_PTR(-EIO);
7382         } else
7383                 return ERR_PTR(-ENOMEM);
7384 }
7385
7386 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7387 {
7388         switch (algo) {
7389         case ALGORITHM_PARITY_0:
7390                 if (raid_disk < max_degraded)
7391                         return 1;
7392                 break;
7393         case ALGORITHM_PARITY_N:
7394                 if (raid_disk >= raid_disks - max_degraded)
7395                         return 1;
7396                 break;
7397         case ALGORITHM_PARITY_0_6:
7398                 if (raid_disk == 0 ||
7399                     raid_disk == raid_disks - 1)
7400                         return 1;
7401                 break;
7402         case ALGORITHM_LEFT_ASYMMETRIC_6:
7403         case ALGORITHM_RIGHT_ASYMMETRIC_6:
7404         case ALGORITHM_LEFT_SYMMETRIC_6:
7405         case ALGORITHM_RIGHT_SYMMETRIC_6:
7406                 if (raid_disk == raid_disks - 1)
7407                         return 1;
7408         }
7409         return 0;
7410 }
7411
7412 static void raid5_set_io_opt(struct r5conf *conf)
7413 {
7414         blk_queue_io_opt(conf->mddev->queue, (conf->chunk_sectors << 9) *
7415                          (conf->raid_disks - conf->max_degraded));
7416 }
7417
7418 static int raid5_run(struct mddev *mddev)
7419 {
7420         struct r5conf *conf;
7421         int working_disks = 0;
7422         int dirty_parity_disks = 0;
7423         struct md_rdev *rdev;
7424         struct md_rdev *journal_dev = NULL;
7425         sector_t reshape_offset = 0;
7426         int i;
7427         long long min_offset_diff = 0;
7428         int first = 1;
7429
7430         if (mddev_init_writes_pending(mddev) < 0)
7431                 return -ENOMEM;
7432
7433         if (mddev->recovery_cp != MaxSector)
7434                 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7435                           mdname(mddev));
7436
7437         rdev_for_each(rdev, mddev) {
7438                 long long diff;
7439
7440                 if (test_bit(Journal, &rdev->flags)) {
7441                         journal_dev = rdev;
7442                         continue;
7443                 }
7444                 if (rdev->raid_disk < 0)
7445                         continue;
7446                 diff = (rdev->new_data_offset - rdev->data_offset);
7447                 if (first) {
7448                         min_offset_diff = diff;
7449                         first = 0;
7450                 } else if (mddev->reshape_backwards &&
7451                          diff < min_offset_diff)
7452                         min_offset_diff = diff;
7453                 else if (!mddev->reshape_backwards &&
7454                          diff > min_offset_diff)
7455                         min_offset_diff = diff;
7456         }
7457
7458         if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7459             (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7460                 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7461                           mdname(mddev));
7462                 return -EINVAL;
7463         }
7464
7465         if (mddev->reshape_position != MaxSector) {
7466                 /* Check that we can continue the reshape.
7467                  * Difficulties arise if the stripe we would write to
7468                  * next is at or after the stripe we would read from next.
7469                  * For a reshape that changes the number of devices, this
7470                  * is only possible for a very short time, and mdadm makes
7471                  * sure that time appears to have past before assembling
7472                  * the array.  So we fail if that time hasn't passed.
7473                  * For a reshape that keeps the number of devices the same
7474                  * mdadm must be monitoring the reshape can keeping the
7475                  * critical areas read-only and backed up.  It will start
7476                  * the array in read-only mode, so we check for that.
7477                  */
7478                 sector_t here_new, here_old;
7479                 int old_disks;
7480                 int max_degraded = (mddev->level == 6 ? 2 : 1);
7481                 int chunk_sectors;
7482                 int new_data_disks;
7483
7484                 if (journal_dev) {
7485                         pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7486                                 mdname(mddev));
7487                         return -EINVAL;
7488                 }
7489
7490                 if (mddev->new_level != mddev->level) {
7491                         pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7492                                 mdname(mddev));
7493                         return -EINVAL;
7494                 }
7495                 old_disks = mddev->raid_disks - mddev->delta_disks;
7496                 /* reshape_position must be on a new-stripe boundary, and one
7497                  * further up in new geometry must map after here in old
7498                  * geometry.
7499                  * If the chunk sizes are different, then as we perform reshape
7500                  * in units of the largest of the two, reshape_position needs
7501                  * be a multiple of the largest chunk size times new data disks.
7502                  */
7503                 here_new = mddev->reshape_position;
7504                 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7505                 new_data_disks = mddev->raid_disks - max_degraded;
7506                 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7507                         pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7508                                 mdname(mddev));
7509                         return -EINVAL;
7510                 }
7511                 reshape_offset = here_new * chunk_sectors;
7512                 /* here_new is the stripe we will write to */
7513                 here_old = mddev->reshape_position;
7514                 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7515                 /* here_old is the first stripe that we might need to read
7516                  * from */
7517                 if (mddev->delta_disks == 0) {
7518                         /* We cannot be sure it is safe to start an in-place
7519                          * reshape.  It is only safe if user-space is monitoring
7520                          * and taking constant backups.
7521                          * mdadm always starts a situation like this in
7522                          * readonly mode so it can take control before
7523                          * allowing any writes.  So just check for that.
7524                          */
7525                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7526                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
7527                                 /* not really in-place - so OK */;
7528                         else if (mddev->ro == 0) {
7529                                 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7530                                         mdname(mddev));
7531                                 return -EINVAL;
7532                         }
7533                 } else if (mddev->reshape_backwards
7534                     ? (here_new * chunk_sectors + min_offset_diff <=
7535                        here_old * chunk_sectors)
7536                     : (here_new * chunk_sectors >=
7537                        here_old * chunk_sectors + (-min_offset_diff))) {
7538                         /* Reading from the same stripe as writing to - bad */
7539                         pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7540                                 mdname(mddev));
7541                         return -EINVAL;
7542                 }
7543                 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7544                 /* OK, we should be able to continue; */
7545         } else {
7546                 BUG_ON(mddev->level != mddev->new_level);
7547                 BUG_ON(mddev->layout != mddev->new_layout);
7548                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7549                 BUG_ON(mddev->delta_disks != 0);
7550         }
7551
7552         if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7553             test_bit(MD_HAS_PPL, &mddev->flags)) {
7554                 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7555                         mdname(mddev));
7556                 clear_bit(MD_HAS_PPL, &mddev->flags);
7557                 clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7558         }
7559
7560         if (mddev->private == NULL)
7561                 conf = setup_conf(mddev);
7562         else
7563                 conf = mddev->private;
7564
7565         if (IS_ERR(conf))
7566                 return PTR_ERR(conf);
7567
7568         if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7569                 if (!journal_dev) {
7570                         pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7571                                 mdname(mddev));
7572                         mddev->ro = 1;
7573                         set_disk_ro(mddev->gendisk, 1);
7574                 } else if (mddev->recovery_cp == MaxSector)
7575                         set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7576         }
7577
7578         conf->min_offset_diff = min_offset_diff;
7579         mddev->thread = conf->thread;
7580         conf->thread = NULL;
7581         mddev->private = conf;
7582
7583         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7584              i++) {
7585                 rdev = conf->disks[i].rdev;
7586                 if (!rdev && conf->disks[i].replacement) {
7587                         /* The replacement is all we have yet */
7588                         rdev = conf->disks[i].replacement;
7589                         conf->disks[i].replacement = NULL;
7590                         clear_bit(Replacement, &rdev->flags);
7591                         conf->disks[i].rdev = rdev;
7592                 }
7593                 if (!rdev)
7594                         continue;
7595                 if (conf->disks[i].replacement &&
7596                     conf->reshape_progress != MaxSector) {
7597                         /* replacements and reshape simply do not mix. */
7598                         pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7599                         goto abort;
7600                 }
7601                 if (test_bit(In_sync, &rdev->flags)) {
7602                         working_disks++;
7603                         continue;
7604                 }
7605                 /* This disc is not fully in-sync.  However if it
7606                  * just stored parity (beyond the recovery_offset),
7607                  * when we don't need to be concerned about the
7608                  * array being dirty.
7609                  * When reshape goes 'backwards', we never have
7610                  * partially completed devices, so we only need
7611                  * to worry about reshape going forwards.
7612                  */
7613                 /* Hack because v0.91 doesn't store recovery_offset properly. */
7614                 if (mddev->major_version == 0 &&
7615                     mddev->minor_version > 90)
7616                         rdev->recovery_offset = reshape_offset;
7617
7618                 if (rdev->recovery_offset < reshape_offset) {
7619                         /* We need to check old and new layout */
7620                         if (!only_parity(rdev->raid_disk,
7621                                          conf->algorithm,
7622                                          conf->raid_disks,
7623                                          conf->max_degraded))
7624                                 continue;
7625                 }
7626                 if (!only_parity(rdev->raid_disk,
7627                                  conf->prev_algo,
7628                                  conf->previous_raid_disks,
7629                                  conf->max_degraded))
7630                         continue;
7631                 dirty_parity_disks++;
7632         }
7633
7634         /*
7635          * 0 for a fully functional array, 1 or 2 for a degraded array.
7636          */
7637         mddev->degraded = raid5_calc_degraded(conf);
7638
7639         if (has_failed(conf)) {
7640                 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7641                         mdname(mddev), mddev->degraded, conf->raid_disks);
7642                 goto abort;
7643         }
7644
7645         /* device size must be a multiple of chunk size */
7646         mddev->dev_sectors &= ~((sector_t)mddev->chunk_sectors - 1);
7647         mddev->resync_max_sectors = mddev->dev_sectors;
7648
7649         if (mddev->degraded > dirty_parity_disks &&
7650             mddev->recovery_cp != MaxSector) {
7651                 if (test_bit(MD_HAS_PPL, &mddev->flags))
7652                         pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7653                                 mdname(mddev));
7654                 else if (mddev->ok_start_degraded)
7655                         pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7656                                 mdname(mddev));
7657                 else {
7658                         pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7659                                 mdname(mddev));
7660                         goto abort;
7661                 }
7662         }
7663
7664         pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7665                 mdname(mddev), conf->level,
7666                 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7667                 mddev->new_layout);
7668
7669         print_raid5_conf(conf);
7670
7671         if (conf->reshape_progress != MaxSector) {
7672                 conf->reshape_safe = conf->reshape_progress;
7673                 atomic_set(&conf->reshape_stripes, 0);
7674                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7675                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7676                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7677                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7678                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7679                                                         "reshape");
7680                 if (!mddev->sync_thread)
7681                         goto abort;
7682         }
7683
7684         /* Ok, everything is just fine now */
7685         if (mddev->to_remove == &raid5_attrs_group)
7686                 mddev->to_remove = NULL;
7687         else if (mddev->kobj.sd &&
7688             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7689                 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7690                         mdname(mddev));
7691         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7692
7693         if (mddev->queue) {
7694                 int chunk_size;
7695                 /* read-ahead size must cover two whole stripes, which
7696                  * is 2 * (datadisks) * chunksize where 'n' is the
7697                  * number of raid devices
7698                  */
7699                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7700                 int stripe = data_disks *
7701                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7702
7703                 chunk_size = mddev->chunk_sectors << 9;
7704                 blk_queue_io_min(mddev->queue, chunk_size);
7705                 raid5_set_io_opt(conf);
7706                 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7707                 /*
7708                  * We can only discard a whole stripe. It doesn't make sense to
7709                  * discard data disk but write parity disk
7710                  */
7711                 stripe = stripe * PAGE_SIZE;
7712                 /* Round up to power of 2, as discard handling
7713                  * currently assumes that */
7714                 while ((stripe-1) & stripe)
7715                         stripe = (stripe | (stripe-1)) + 1;
7716                 mddev->queue->limits.discard_alignment = stripe;
7717                 mddev->queue->limits.discard_granularity = stripe;
7718
7719                 blk_queue_max_write_same_sectors(mddev->queue, 0);
7720                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
7721
7722                 rdev_for_each(rdev, mddev) {
7723                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7724                                           rdev->data_offset << 9);
7725                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7726                                           rdev->new_data_offset << 9);
7727                 }
7728
7729                 /*
7730                  * zeroing is required, otherwise data
7731                  * could be lost. Consider a scenario: discard a stripe
7732                  * (the stripe could be inconsistent if
7733                  * discard_zeroes_data is 0); write one disk of the
7734                  * stripe (the stripe could be inconsistent again
7735                  * depending on which disks are used to calculate
7736                  * parity); the disk is broken; The stripe data of this
7737                  * disk is lost.
7738                  *
7739                  * We only allow DISCARD if the sysadmin has confirmed that
7740                  * only safe devices are in use by setting a module parameter.
7741                  * A better idea might be to turn DISCARD into WRITE_ZEROES
7742                  * requests, as that is required to be safe.
7743                  */
7744                 if (devices_handle_discard_safely &&
7745                     mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7746                     mddev->queue->limits.discard_granularity >= stripe)
7747                         blk_queue_flag_set(QUEUE_FLAG_DISCARD,
7748                                                 mddev->queue);
7749                 else
7750                         blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
7751                                                 mddev->queue);
7752
7753                 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7754         }
7755
7756         if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
7757                 goto abort;
7758
7759         return 0;
7760 abort:
7761         md_unregister_thread(&mddev->thread);
7762         print_raid5_conf(conf);
7763         free_conf(conf);
7764         mddev->private = NULL;
7765         pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7766         return -EIO;
7767 }
7768
7769 static void raid5_free(struct mddev *mddev, void *priv)
7770 {
7771         struct r5conf *conf = priv;
7772
7773         free_conf(conf);
7774         mddev->to_remove = &raid5_attrs_group;
7775 }
7776
7777 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7778 {
7779         struct r5conf *conf = mddev->private;
7780         int i;
7781
7782         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7783                 conf->chunk_sectors / 2, mddev->layout);
7784         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7785         rcu_read_lock();
7786         for (i = 0; i < conf->raid_disks; i++) {
7787                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7788                 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7789         }
7790         rcu_read_unlock();
7791         seq_printf (seq, "]");
7792 }
7793
7794 static void print_raid5_conf (struct r5conf *conf)
7795 {
7796         int i;
7797         struct disk_info *tmp;
7798
7799         pr_debug("RAID conf printout:\n");
7800         if (!conf) {
7801                 pr_debug("(conf==NULL)\n");
7802                 return;
7803         }
7804         pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7805                conf->raid_disks,
7806                conf->raid_disks - conf->mddev->degraded);
7807
7808         for (i = 0; i < conf->raid_disks; i++) {
7809                 char b[BDEVNAME_SIZE];
7810                 tmp = conf->disks + i;
7811                 if (tmp->rdev)
7812                         pr_debug(" disk %d, o:%d, dev:%s\n",
7813                                i, !test_bit(Faulty, &tmp->rdev->flags),
7814                                bdevname(tmp->rdev->bdev, b));
7815         }
7816 }
7817
7818 static int raid5_spare_active(struct mddev *mddev)
7819 {
7820         int i;
7821         struct r5conf *conf = mddev->private;
7822         struct disk_info *tmp;
7823         int count = 0;
7824         unsigned long flags;
7825
7826         for (i = 0; i < conf->raid_disks; i++) {
7827                 tmp = conf->disks + i;
7828                 if (tmp->replacement
7829                     && tmp->replacement->recovery_offset == MaxSector
7830                     && !test_bit(Faulty, &tmp->replacement->flags)
7831                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7832                         /* Replacement has just become active. */
7833                         if (!tmp->rdev
7834                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7835                                 count++;
7836                         if (tmp->rdev) {
7837                                 /* Replaced device not technically faulty,
7838                                  * but we need to be sure it gets removed
7839                                  * and never re-added.
7840                                  */
7841                                 set_bit(Faulty, &tmp->rdev->flags);
7842                                 sysfs_notify_dirent_safe(
7843                                         tmp->rdev->sysfs_state);
7844                         }
7845                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7846                 } else if (tmp->rdev
7847                     && tmp->rdev->recovery_offset == MaxSector
7848                     && !test_bit(Faulty, &tmp->rdev->flags)
7849                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7850                         count++;
7851                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7852                 }
7853         }
7854         spin_lock_irqsave(&conf->device_lock, flags);
7855         mddev->degraded = raid5_calc_degraded(conf);
7856         spin_unlock_irqrestore(&conf->device_lock, flags);
7857         print_raid5_conf(conf);
7858         return count;
7859 }
7860
7861 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7862 {
7863         struct r5conf *conf = mddev->private;
7864         int err = 0;
7865         int number = rdev->raid_disk;
7866         struct md_rdev **rdevp;
7867         struct disk_info *p = conf->disks + number;
7868
7869         print_raid5_conf(conf);
7870         if (test_bit(Journal, &rdev->flags) && conf->log) {
7871                 /*
7872                  * we can't wait pending write here, as this is called in
7873                  * raid5d, wait will deadlock.
7874                  * neilb: there is no locking about new writes here,
7875                  * so this cannot be safe.
7876                  */
7877                 if (atomic_read(&conf->active_stripes) ||
7878                     atomic_read(&conf->r5c_cached_full_stripes) ||
7879                     atomic_read(&conf->r5c_cached_partial_stripes)) {
7880                         return -EBUSY;
7881                 }
7882                 log_exit(conf);
7883                 return 0;
7884         }
7885         if (rdev == p->rdev)
7886                 rdevp = &p->rdev;
7887         else if (rdev == p->replacement)
7888                 rdevp = &p->replacement;
7889         else
7890                 return 0;
7891
7892         if (number >= conf->raid_disks &&
7893             conf->reshape_progress == MaxSector)
7894                 clear_bit(In_sync, &rdev->flags);
7895
7896         if (test_bit(In_sync, &rdev->flags) ||
7897             atomic_read(&rdev->nr_pending)) {
7898                 err = -EBUSY;
7899                 goto abort;
7900         }
7901         /* Only remove non-faulty devices if recovery
7902          * isn't possible.
7903          */
7904         if (!test_bit(Faulty, &rdev->flags) &&
7905             mddev->recovery_disabled != conf->recovery_disabled &&
7906             !has_failed(conf) &&
7907             (!p->replacement || p->replacement == rdev) &&
7908             number < conf->raid_disks) {
7909                 err = -EBUSY;
7910                 goto abort;
7911         }
7912         *rdevp = NULL;
7913         if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7914                 synchronize_rcu();
7915                 if (atomic_read(&rdev->nr_pending)) {
7916                         /* lost the race, try later */
7917                         err = -EBUSY;
7918                         *rdevp = rdev;
7919                 }
7920         }
7921         if (!err) {
7922                 err = log_modify(conf, rdev, false);
7923                 if (err)
7924                         goto abort;
7925         }
7926         if (p->replacement) {
7927                 /* We must have just cleared 'rdev' */
7928                 p->rdev = p->replacement;
7929                 clear_bit(Replacement, &p->replacement->flags);
7930                 smp_mb(); /* Make sure other CPUs may see both as identical
7931                            * but will never see neither - if they are careful
7932                            */
7933                 p->replacement = NULL;
7934
7935                 if (!err)
7936                         err = log_modify(conf, p->rdev, true);
7937         }
7938
7939         clear_bit(WantReplacement, &rdev->flags);
7940 abort:
7941
7942         print_raid5_conf(conf);
7943         return err;
7944 }
7945
7946 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7947 {
7948         struct r5conf *conf = mddev->private;
7949         int ret, err = -EEXIST;
7950         int disk;
7951         struct disk_info *p;
7952         int first = 0;
7953         int last = conf->raid_disks - 1;
7954
7955         if (test_bit(Journal, &rdev->flags)) {
7956                 if (conf->log)
7957                         return -EBUSY;
7958
7959                 rdev->raid_disk = 0;
7960                 /*
7961                  * The array is in readonly mode if journal is missing, so no
7962                  * write requests running. We should be safe
7963                  */
7964                 ret = log_init(conf, rdev, false);
7965                 if (ret)
7966                         return ret;
7967
7968                 ret = r5l_start(conf->log);
7969                 if (ret)
7970                         return ret;
7971
7972                 return 0;
7973         }
7974         if (mddev->recovery_disabled == conf->recovery_disabled)
7975                 return -EBUSY;
7976
7977         if (rdev->saved_raid_disk < 0 && has_failed(conf))
7978                 /* no point adding a device */
7979                 return -EINVAL;
7980
7981         if (rdev->raid_disk >= 0)
7982                 first = last = rdev->raid_disk;
7983
7984         /*
7985          * find the disk ... but prefer rdev->saved_raid_disk
7986          * if possible.
7987          */
7988         if (rdev->saved_raid_disk >= 0 &&
7989             rdev->saved_raid_disk >= first &&
7990             conf->disks[rdev->saved_raid_disk].rdev == NULL)
7991                 first = rdev->saved_raid_disk;
7992
7993         for (disk = first; disk <= last; disk++) {
7994                 p = conf->disks + disk;
7995                 if (p->rdev == NULL) {
7996                         clear_bit(In_sync, &rdev->flags);
7997                         rdev->raid_disk = disk;
7998                         if (rdev->saved_raid_disk != disk)
7999                                 conf->fullsync = 1;
8000                         rcu_assign_pointer(p->rdev, rdev);
8001
8002                         err = log_modify(conf, rdev, true);
8003
8004                         goto out;
8005                 }
8006         }
8007         for (disk = first; disk <= last; disk++) {
8008                 p = conf->disks + disk;
8009                 if (test_bit(WantReplacement, &p->rdev->flags) &&
8010                     p->replacement == NULL) {
8011                         clear_bit(In_sync, &rdev->flags);
8012                         set_bit(Replacement, &rdev->flags);
8013                         rdev->raid_disk = disk;
8014                         err = 0;
8015                         conf->fullsync = 1;
8016                         rcu_assign_pointer(p->replacement, rdev);
8017                         break;
8018                 }
8019         }
8020 out:
8021         print_raid5_conf(conf);
8022         return err;
8023 }
8024
8025 static int raid5_resize(struct mddev *mddev, sector_t sectors)
8026 {
8027         /* no resync is happening, and there is enough space
8028          * on all devices, so we can resize.
8029          * We need to make sure resync covers any new space.
8030          * If the array is shrinking we should possibly wait until
8031          * any io in the removed space completes, but it hardly seems
8032          * worth it.
8033          */
8034         sector_t newsize;
8035         struct r5conf *conf = mddev->private;
8036
8037         if (raid5_has_log(conf) || raid5_has_ppl(conf))
8038                 return -EINVAL;
8039         sectors &= ~((sector_t)conf->chunk_sectors - 1);
8040         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
8041         if (mddev->external_size &&
8042             mddev->array_sectors > newsize)
8043                 return -EINVAL;
8044         if (mddev->bitmap) {
8045                 int ret = md_bitmap_resize(mddev->bitmap, sectors, 0, 0);
8046                 if (ret)
8047                         return ret;
8048         }
8049         md_set_array_sectors(mddev, newsize);
8050         if (sectors > mddev->dev_sectors &&
8051             mddev->recovery_cp > mddev->dev_sectors) {
8052                 mddev->recovery_cp = mddev->dev_sectors;
8053                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
8054         }
8055         mddev->dev_sectors = sectors;
8056         mddev->resync_max_sectors = sectors;
8057         return 0;
8058 }
8059
8060 static int check_stripe_cache(struct mddev *mddev)
8061 {
8062         /* Can only proceed if there are plenty of stripe_heads.
8063          * We need a minimum of one full stripe,, and for sensible progress
8064          * it is best to have about 4 times that.
8065          * If we require 4 times, then the default 256 4K stripe_heads will
8066          * allow for chunk sizes up to 256K, which is probably OK.
8067          * If the chunk size is greater, user-space should request more
8068          * stripe_heads first.
8069          */
8070         struct r5conf *conf = mddev->private;
8071         if (((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
8072             > conf->min_nr_stripes ||
8073             ((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
8074             > conf->min_nr_stripes) {
8075                 pr_warn("md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
8076                         mdname(mddev),
8077                         ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
8078                          / RAID5_STRIPE_SIZE(conf))*4);
8079                 return 0;
8080         }
8081         return 1;
8082 }
8083
8084 static int check_reshape(struct mddev *mddev)
8085 {
8086         struct r5conf *conf = mddev->private;
8087
8088         if (raid5_has_log(conf) || raid5_has_ppl(conf))
8089                 return -EINVAL;
8090         if (mddev->delta_disks == 0 &&
8091             mddev->new_layout == mddev->layout &&
8092             mddev->new_chunk_sectors == mddev->chunk_sectors)
8093                 return 0; /* nothing to do */
8094         if (has_failed(conf))
8095                 return -EINVAL;
8096         if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
8097                 /* We might be able to shrink, but the devices must
8098                  * be made bigger first.
8099                  * For raid6, 4 is the minimum size.
8100                  * Otherwise 2 is the minimum
8101                  */
8102                 int min = 2;
8103                 if (mddev->level == 6)
8104                         min = 4;
8105                 if (mddev->raid_disks + mddev->delta_disks < min)
8106                         return -EINVAL;
8107         }
8108
8109         if (!check_stripe_cache(mddev))
8110                 return -ENOSPC;
8111
8112         if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
8113             mddev->delta_disks > 0)
8114                 if (resize_chunks(conf,
8115                                   conf->previous_raid_disks
8116                                   + max(0, mddev->delta_disks),
8117                                   max(mddev->new_chunk_sectors,
8118                                       mddev->chunk_sectors)
8119                             ) < 0)
8120                         return -ENOMEM;
8121
8122         if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
8123                 return 0; /* never bother to shrink */
8124         return resize_stripes(conf, (conf->previous_raid_disks
8125                                      + mddev->delta_disks));
8126 }
8127
8128 static int raid5_start_reshape(struct mddev *mddev)
8129 {
8130         struct r5conf *conf = mddev->private;
8131         struct md_rdev *rdev;
8132         int spares = 0;
8133         unsigned long flags;
8134
8135         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
8136                 return -EBUSY;
8137
8138         if (!check_stripe_cache(mddev))
8139                 return -ENOSPC;
8140
8141         if (has_failed(conf))
8142                 return -EINVAL;
8143
8144         rdev_for_each(rdev, mddev) {
8145                 if (!test_bit(In_sync, &rdev->flags)
8146                     && !test_bit(Faulty, &rdev->flags))
8147                         spares++;
8148         }
8149
8150         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
8151                 /* Not enough devices even to make a degraded array
8152                  * of that size
8153                  */
8154                 return -EINVAL;
8155
8156         /* Refuse to reduce size of the array.  Any reductions in
8157          * array size must be through explicit setting of array_size
8158          * attribute.
8159          */
8160         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
8161             < mddev->array_sectors) {
8162                 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
8163                         mdname(mddev));
8164                 return -EINVAL;
8165         }
8166
8167         atomic_set(&conf->reshape_stripes, 0);
8168         spin_lock_irq(&conf->device_lock);
8169         write_seqcount_begin(&conf->gen_lock);
8170         conf->previous_raid_disks = conf->raid_disks;
8171         conf->raid_disks += mddev->delta_disks;
8172         conf->prev_chunk_sectors = conf->chunk_sectors;
8173         conf->chunk_sectors = mddev->new_chunk_sectors;
8174         conf->prev_algo = conf->algorithm;
8175         conf->algorithm = mddev->new_layout;
8176         conf->generation++;
8177         /* Code that selects data_offset needs to see the generation update
8178          * if reshape_progress has been set - so a memory barrier needed.
8179          */
8180         smp_mb();
8181         if (mddev->reshape_backwards)
8182                 conf->reshape_progress = raid5_size(mddev, 0, 0);
8183         else
8184                 conf->reshape_progress = 0;
8185         conf->reshape_safe = conf->reshape_progress;
8186         write_seqcount_end(&conf->gen_lock);
8187         spin_unlock_irq(&conf->device_lock);
8188
8189         /* Now make sure any requests that proceeded on the assumption
8190          * the reshape wasn't running - like Discard or Read - have
8191          * completed.
8192          */
8193         mddev_suspend(mddev);
8194         mddev_resume(mddev);
8195
8196         /* Add some new drives, as many as will fit.
8197          * We know there are enough to make the newly sized array work.
8198          * Don't add devices if we are reducing the number of
8199          * devices in the array.  This is because it is not possible
8200          * to correctly record the "partially reconstructed" state of
8201          * such devices during the reshape and confusion could result.
8202          */
8203         if (mddev->delta_disks >= 0) {
8204                 rdev_for_each(rdev, mddev)
8205                         if (rdev->raid_disk < 0 &&
8206                             !test_bit(Faulty, &rdev->flags)) {
8207                                 if (raid5_add_disk(mddev, rdev) == 0) {
8208                                         if (rdev->raid_disk
8209                                             >= conf->previous_raid_disks)
8210                                                 set_bit(In_sync, &rdev->flags);
8211                                         else
8212                                                 rdev->recovery_offset = 0;
8213
8214                                         /* Failure here is OK */
8215                                         sysfs_link_rdev(mddev, rdev);
8216                                 }
8217                         } else if (rdev->raid_disk >= conf->previous_raid_disks
8218                                    && !test_bit(Faulty, &rdev->flags)) {
8219                                 /* This is a spare that was manually added */
8220                                 set_bit(In_sync, &rdev->flags);
8221                         }
8222
8223                 /* When a reshape changes the number of devices,
8224                  * ->degraded is measured against the larger of the
8225                  * pre and post number of devices.
8226                  */
8227                 spin_lock_irqsave(&conf->device_lock, flags);
8228                 mddev->degraded = raid5_calc_degraded(conf);
8229                 spin_unlock_irqrestore(&conf->device_lock, flags);
8230         }
8231         mddev->raid_disks = conf->raid_disks;
8232         mddev->reshape_position = conf->reshape_progress;
8233         set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8234
8235         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
8236         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
8237         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
8238         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
8239         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
8240         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
8241                                                 "reshape");
8242         if (!mddev->sync_thread) {
8243                 mddev->recovery = 0;
8244                 spin_lock_irq(&conf->device_lock);
8245                 write_seqcount_begin(&conf->gen_lock);
8246                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
8247                 mddev->new_chunk_sectors =
8248                         conf->chunk_sectors = conf->prev_chunk_sectors;
8249                 mddev->new_layout = conf->algorithm = conf->prev_algo;
8250                 rdev_for_each(rdev, mddev)
8251                         rdev->new_data_offset = rdev->data_offset;
8252                 smp_wmb();
8253                 conf->generation --;
8254                 conf->reshape_progress = MaxSector;
8255                 mddev->reshape_position = MaxSector;
8256                 write_seqcount_end(&conf->gen_lock);
8257                 spin_unlock_irq(&conf->device_lock);
8258                 return -EAGAIN;
8259         }
8260         conf->reshape_checkpoint = jiffies;
8261         md_wakeup_thread(mddev->sync_thread);
8262         md_new_event(mddev);
8263         return 0;
8264 }
8265
8266 /* This is called from the reshape thread and should make any
8267  * changes needed in 'conf'
8268  */
8269 static void end_reshape(struct r5conf *conf)
8270 {
8271
8272         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
8273                 struct md_rdev *rdev;
8274
8275                 spin_lock_irq(&conf->device_lock);
8276                 conf->previous_raid_disks = conf->raid_disks;
8277                 md_finish_reshape(conf->mddev);
8278                 smp_wmb();
8279                 conf->reshape_progress = MaxSector;
8280                 conf->mddev->reshape_position = MaxSector;
8281                 rdev_for_each(rdev, conf->mddev)
8282                         if (rdev->raid_disk >= 0 &&
8283                             !test_bit(Journal, &rdev->flags) &&
8284                             !test_bit(In_sync, &rdev->flags))
8285                                 rdev->recovery_offset = MaxSector;
8286                 spin_unlock_irq(&conf->device_lock);
8287                 wake_up(&conf->wait_for_overlap);
8288
8289                 if (conf->mddev->queue)
8290                         raid5_set_io_opt(conf);
8291         }
8292 }
8293
8294 /* This is called from the raid5d thread with mddev_lock held.
8295  * It makes config changes to the device.
8296  */
8297 static void raid5_finish_reshape(struct mddev *mddev)
8298 {
8299         struct r5conf *conf = mddev->private;
8300
8301         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8302
8303                 if (mddev->delta_disks <= 0) {
8304                         int d;
8305                         spin_lock_irq(&conf->device_lock);
8306                         mddev->degraded = raid5_calc_degraded(conf);
8307                         spin_unlock_irq(&conf->device_lock);
8308                         for (d = conf->raid_disks ;
8309                              d < conf->raid_disks - mddev->delta_disks;
8310                              d++) {
8311                                 struct md_rdev *rdev = conf->disks[d].rdev;
8312                                 if (rdev)
8313                                         clear_bit(In_sync, &rdev->flags);
8314                                 rdev = conf->disks[d].replacement;
8315                                 if (rdev)
8316                                         clear_bit(In_sync, &rdev->flags);
8317                         }
8318                 }
8319                 mddev->layout = conf->algorithm;
8320                 mddev->chunk_sectors = conf->chunk_sectors;
8321                 mddev->reshape_position = MaxSector;
8322                 mddev->delta_disks = 0;
8323                 mddev->reshape_backwards = 0;
8324         }
8325 }
8326
8327 static void raid5_quiesce(struct mddev *mddev, int quiesce)
8328 {
8329         struct r5conf *conf = mddev->private;
8330
8331         if (quiesce) {
8332                 /* stop all writes */
8333                 lock_all_device_hash_locks_irq(conf);
8334                 /* '2' tells resync/reshape to pause so that all
8335                  * active stripes can drain
8336                  */
8337                 r5c_flush_cache(conf, INT_MAX);
8338                 conf->quiesce = 2;
8339                 wait_event_cmd(conf->wait_for_quiescent,
8340                                     atomic_read(&conf->active_stripes) == 0 &&
8341                                     atomic_read(&conf->active_aligned_reads) == 0,
8342                                     unlock_all_device_hash_locks_irq(conf),
8343                                     lock_all_device_hash_locks_irq(conf));
8344                 conf->quiesce = 1;
8345                 unlock_all_device_hash_locks_irq(conf);
8346                 /* allow reshape to continue */
8347                 wake_up(&conf->wait_for_overlap);
8348         } else {
8349                 /* re-enable writes */
8350                 lock_all_device_hash_locks_irq(conf);
8351                 conf->quiesce = 0;
8352                 wake_up(&conf->wait_for_quiescent);
8353                 wake_up(&conf->wait_for_overlap);
8354                 unlock_all_device_hash_locks_irq(conf);
8355         }
8356         log_quiesce(conf, quiesce);
8357 }
8358
8359 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8360 {
8361         struct r0conf *raid0_conf = mddev->private;
8362         sector_t sectors;
8363
8364         /* for raid0 takeover only one zone is supported */
8365         if (raid0_conf->nr_strip_zones > 1) {
8366                 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8367                         mdname(mddev));
8368                 return ERR_PTR(-EINVAL);
8369         }
8370
8371         sectors = raid0_conf->strip_zone[0].zone_end;
8372         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8373         mddev->dev_sectors = sectors;
8374         mddev->new_level = level;
8375         mddev->new_layout = ALGORITHM_PARITY_N;
8376         mddev->new_chunk_sectors = mddev->chunk_sectors;
8377         mddev->raid_disks += 1;
8378         mddev->delta_disks = 1;
8379         /* make sure it will be not marked as dirty */
8380         mddev->recovery_cp = MaxSector;
8381
8382         return setup_conf(mddev);
8383 }
8384
8385 static void *raid5_takeover_raid1(struct mddev *mddev)
8386 {
8387         int chunksect;
8388         void *ret;
8389
8390         if (mddev->raid_disks != 2 ||
8391             mddev->degraded > 1)
8392                 return ERR_PTR(-EINVAL);
8393
8394         /* Should check if there are write-behind devices? */
8395
8396         chunksect = 64*2; /* 64K by default */
8397
8398         /* The array must be an exact multiple of chunksize */
8399         while (chunksect && (mddev->array_sectors & (chunksect-1)))
8400                 chunksect >>= 1;
8401
8402         if ((chunksect<<9) < RAID5_STRIPE_SIZE((struct r5conf *)mddev->private))
8403                 /* array size does not allow a suitable chunk size */
8404                 return ERR_PTR(-EINVAL);
8405
8406         mddev->new_level = 5;
8407         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8408         mddev->new_chunk_sectors = chunksect;
8409
8410         ret = setup_conf(mddev);
8411         if (!IS_ERR(ret))
8412                 mddev_clear_unsupported_flags(mddev,
8413                         UNSUPPORTED_MDDEV_FLAGS);
8414         return ret;
8415 }
8416
8417 static void *raid5_takeover_raid6(struct mddev *mddev)
8418 {
8419         int new_layout;
8420
8421         switch (mddev->layout) {
8422         case ALGORITHM_LEFT_ASYMMETRIC_6:
8423                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8424                 break;
8425         case ALGORITHM_RIGHT_ASYMMETRIC_6:
8426                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8427                 break;
8428         case ALGORITHM_LEFT_SYMMETRIC_6:
8429                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8430                 break;
8431         case ALGORITHM_RIGHT_SYMMETRIC_6:
8432                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8433                 break;
8434         case ALGORITHM_PARITY_0_6:
8435                 new_layout = ALGORITHM_PARITY_0;
8436                 break;
8437         case ALGORITHM_PARITY_N:
8438                 new_layout = ALGORITHM_PARITY_N;
8439                 break;
8440         default:
8441                 return ERR_PTR(-EINVAL);
8442         }
8443         mddev->new_level = 5;
8444         mddev->new_layout = new_layout;
8445         mddev->delta_disks = -1;
8446         mddev->raid_disks -= 1;
8447         return setup_conf(mddev);
8448 }
8449
8450 static int raid5_check_reshape(struct mddev *mddev)
8451 {
8452         /* For a 2-drive array, the layout and chunk size can be changed
8453          * immediately as not restriping is needed.
8454          * For larger arrays we record the new value - after validation
8455          * to be used by a reshape pass.
8456          */
8457         struct r5conf *conf = mddev->private;
8458         int new_chunk = mddev->new_chunk_sectors;
8459
8460         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8461                 return -EINVAL;
8462         if (new_chunk > 0) {
8463                 if (!is_power_of_2(new_chunk))
8464                         return -EINVAL;
8465                 if (new_chunk < (PAGE_SIZE>>9))
8466                         return -EINVAL;
8467                 if (mddev->array_sectors & (new_chunk-1))
8468                         /* not factor of array size */
8469                         return -EINVAL;
8470         }
8471
8472         /* They look valid */
8473
8474         if (mddev->raid_disks == 2) {
8475                 /* can make the change immediately */
8476                 if (mddev->new_layout >= 0) {
8477                         conf->algorithm = mddev->new_layout;
8478                         mddev->layout = mddev->new_layout;
8479                 }
8480                 if (new_chunk > 0) {
8481                         conf->chunk_sectors = new_chunk ;
8482                         mddev->chunk_sectors = new_chunk;
8483                 }
8484                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8485                 md_wakeup_thread(mddev->thread);
8486         }
8487         return check_reshape(mddev);
8488 }
8489
8490 static int raid6_check_reshape(struct mddev *mddev)
8491 {
8492         int new_chunk = mddev->new_chunk_sectors;
8493
8494         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8495                 return -EINVAL;
8496         if (new_chunk > 0) {
8497                 if (!is_power_of_2(new_chunk))
8498                         return -EINVAL;
8499                 if (new_chunk < (PAGE_SIZE >> 9))
8500                         return -EINVAL;
8501                 if (mddev->array_sectors & (new_chunk-1))
8502                         /* not factor of array size */
8503                         return -EINVAL;
8504         }
8505
8506         /* They look valid */
8507         return check_reshape(mddev);
8508 }
8509
8510 static void *raid5_takeover(struct mddev *mddev)
8511 {
8512         /* raid5 can take over:
8513          *  raid0 - if there is only one strip zone - make it a raid4 layout
8514          *  raid1 - if there are two drives.  We need to know the chunk size
8515          *  raid4 - trivial - just use a raid4 layout.
8516          *  raid6 - Providing it is a *_6 layout
8517          */
8518         if (mddev->level == 0)
8519                 return raid45_takeover_raid0(mddev, 5);
8520         if (mddev->level == 1)
8521                 return raid5_takeover_raid1(mddev);
8522         if (mddev->level == 4) {
8523                 mddev->new_layout = ALGORITHM_PARITY_N;
8524                 mddev->new_level = 5;
8525                 return setup_conf(mddev);
8526         }
8527         if (mddev->level == 6)
8528                 return raid5_takeover_raid6(mddev);
8529
8530         return ERR_PTR(-EINVAL);
8531 }
8532
8533 static void *raid4_takeover(struct mddev *mddev)
8534 {
8535         /* raid4 can take over:
8536          *  raid0 - if there is only one strip zone
8537          *  raid5 - if layout is right
8538          */
8539         if (mddev->level == 0)
8540                 return raid45_takeover_raid0(mddev, 4);
8541         if (mddev->level == 5 &&
8542             mddev->layout == ALGORITHM_PARITY_N) {
8543                 mddev->new_layout = 0;
8544                 mddev->new_level = 4;
8545                 return setup_conf(mddev);
8546         }
8547         return ERR_PTR(-EINVAL);
8548 }
8549
8550 static struct md_personality raid5_personality;
8551
8552 static void *raid6_takeover(struct mddev *mddev)
8553 {
8554         /* Currently can only take over a raid5.  We map the
8555          * personality to an equivalent raid6 personality
8556          * with the Q block at the end.
8557          */
8558         int new_layout;
8559
8560         if (mddev->pers != &raid5_personality)
8561                 return ERR_PTR(-EINVAL);
8562         if (mddev->degraded > 1)
8563                 return ERR_PTR(-EINVAL);
8564         if (mddev->raid_disks > 253)
8565                 return ERR_PTR(-EINVAL);
8566         if (mddev->raid_disks < 3)
8567                 return ERR_PTR(-EINVAL);
8568
8569         switch (mddev->layout) {
8570         case ALGORITHM_LEFT_ASYMMETRIC:
8571                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8572                 break;
8573         case ALGORITHM_RIGHT_ASYMMETRIC:
8574                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8575                 break;
8576         case ALGORITHM_LEFT_SYMMETRIC:
8577                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8578                 break;
8579         case ALGORITHM_RIGHT_SYMMETRIC:
8580                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8581                 break;
8582         case ALGORITHM_PARITY_0:
8583                 new_layout = ALGORITHM_PARITY_0_6;
8584                 break;
8585         case ALGORITHM_PARITY_N:
8586                 new_layout = ALGORITHM_PARITY_N;
8587                 break;
8588         default:
8589                 return ERR_PTR(-EINVAL);
8590         }
8591         mddev->new_level = 6;
8592         mddev->new_layout = new_layout;
8593         mddev->delta_disks = 1;
8594         mddev->raid_disks += 1;
8595         return setup_conf(mddev);
8596 }
8597
8598 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8599 {
8600         struct r5conf *conf;
8601         int err;
8602
8603         err = mddev_lock(mddev);
8604         if (err)
8605                 return err;
8606         conf = mddev->private;
8607         if (!conf) {
8608                 mddev_unlock(mddev);
8609                 return -ENODEV;
8610         }
8611
8612         if (strncmp(buf, "ppl", 3) == 0) {
8613                 /* ppl only works with RAID 5 */
8614                 if (!raid5_has_ppl(conf) && conf->level == 5) {
8615                         err = log_init(conf, NULL, true);
8616                         if (!err) {
8617                                 err = resize_stripes(conf, conf->pool_size);
8618                                 if (err)
8619                                         log_exit(conf);
8620                         }
8621                 } else
8622                         err = -EINVAL;
8623         } else if (strncmp(buf, "resync", 6) == 0) {
8624                 if (raid5_has_ppl(conf)) {
8625                         mddev_suspend(mddev);
8626                         log_exit(conf);
8627                         mddev_resume(mddev);
8628                         err = resize_stripes(conf, conf->pool_size);
8629                 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8630                            r5l_log_disk_error(conf)) {
8631                         bool journal_dev_exists = false;
8632                         struct md_rdev *rdev;
8633
8634                         rdev_for_each(rdev, mddev)
8635                                 if (test_bit(Journal, &rdev->flags)) {
8636                                         journal_dev_exists = true;
8637                                         break;
8638                                 }
8639
8640                         if (!journal_dev_exists) {
8641                                 mddev_suspend(mddev);
8642                                 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8643                                 mddev_resume(mddev);
8644                         } else  /* need remove journal device first */
8645                                 err = -EBUSY;
8646                 } else
8647                         err = -EINVAL;
8648         } else {
8649                 err = -EINVAL;
8650         }
8651
8652         if (!err)
8653                 md_update_sb(mddev, 1);
8654
8655         mddev_unlock(mddev);
8656
8657         return err;
8658 }
8659
8660 static int raid5_start(struct mddev *mddev)
8661 {
8662         struct r5conf *conf = mddev->private;
8663
8664         return r5l_start(conf->log);
8665 }
8666
8667 static struct md_personality raid6_personality =
8668 {
8669         .name           = "raid6",
8670         .level          = 6,
8671         .owner          = THIS_MODULE,
8672         .make_request   = raid5_make_request,
8673         .run            = raid5_run,
8674         .start          = raid5_start,
8675         .free           = raid5_free,
8676         .status         = raid5_status,
8677         .error_handler  = raid5_error,
8678         .hot_add_disk   = raid5_add_disk,
8679         .hot_remove_disk= raid5_remove_disk,
8680         .spare_active   = raid5_spare_active,
8681         .sync_request   = raid5_sync_request,
8682         .resize         = raid5_resize,
8683         .size           = raid5_size,
8684         .check_reshape  = raid6_check_reshape,
8685         .start_reshape  = raid5_start_reshape,
8686         .finish_reshape = raid5_finish_reshape,
8687         .quiesce        = raid5_quiesce,
8688         .takeover       = raid6_takeover,
8689         .change_consistency_policy = raid5_change_consistency_policy,
8690 };
8691 static struct md_personality raid5_personality =
8692 {
8693         .name           = "raid5",
8694         .level          = 5,
8695         .owner          = THIS_MODULE,
8696         .make_request   = raid5_make_request,
8697         .run            = raid5_run,
8698         .start          = raid5_start,
8699         .free           = raid5_free,
8700         .status         = raid5_status,
8701         .error_handler  = raid5_error,
8702         .hot_add_disk   = raid5_add_disk,
8703         .hot_remove_disk= raid5_remove_disk,
8704         .spare_active   = raid5_spare_active,
8705         .sync_request   = raid5_sync_request,
8706         .resize         = raid5_resize,
8707         .size           = raid5_size,
8708         .check_reshape  = raid5_check_reshape,
8709         .start_reshape  = raid5_start_reshape,
8710         .finish_reshape = raid5_finish_reshape,
8711         .quiesce        = raid5_quiesce,
8712         .takeover       = raid5_takeover,
8713         .change_consistency_policy = raid5_change_consistency_policy,
8714 };
8715
8716 static struct md_personality raid4_personality =
8717 {
8718         .name           = "raid4",
8719         .level          = 4,
8720         .owner          = THIS_MODULE,
8721         .make_request   = raid5_make_request,
8722         .run            = raid5_run,
8723         .start          = raid5_start,
8724         .free           = raid5_free,
8725         .status         = raid5_status,
8726         .error_handler  = raid5_error,
8727         .hot_add_disk   = raid5_add_disk,
8728         .hot_remove_disk= raid5_remove_disk,
8729         .spare_active   = raid5_spare_active,
8730         .sync_request   = raid5_sync_request,
8731         .resize         = raid5_resize,
8732         .size           = raid5_size,
8733         .check_reshape  = raid5_check_reshape,
8734         .start_reshape  = raid5_start_reshape,
8735         .finish_reshape = raid5_finish_reshape,
8736         .quiesce        = raid5_quiesce,
8737         .takeover       = raid4_takeover,
8738         .change_consistency_policy = raid5_change_consistency_policy,
8739 };
8740
8741 static int __init raid5_init(void)
8742 {
8743         int ret;
8744
8745         raid5_wq = alloc_workqueue("raid5wq",
8746                 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8747         if (!raid5_wq)
8748                 return -ENOMEM;
8749
8750         ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8751                                       "md/raid5:prepare",
8752                                       raid456_cpu_up_prepare,
8753                                       raid456_cpu_dead);
8754         if (ret) {
8755                 destroy_workqueue(raid5_wq);
8756                 return ret;
8757         }
8758         register_md_personality(&raid6_personality);
8759         register_md_personality(&raid5_personality);
8760         register_md_personality(&raid4_personality);
8761         return 0;
8762 }
8763
8764 static void raid5_exit(void)
8765 {
8766         unregister_md_personality(&raid6_personality);
8767         unregister_md_personality(&raid5_personality);
8768         unregister_md_personality(&raid4_personality);
8769         cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8770         destroy_workqueue(raid5_wq);
8771 }
8772
8773 module_init(raid5_init);
8774 module_exit(raid5_exit);
8775 MODULE_LICENSE("GPL");
8776 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8777 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8778 MODULE_ALIAS("md-raid5");
8779 MODULE_ALIAS("md-raid4");
8780 MODULE_ALIAS("md-level-5");
8781 MODULE_ALIAS("md-level-4");
8782 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8783 MODULE_ALIAS("md-raid6");
8784 MODULE_ALIAS("md-level-6");
8785
8786 /* This used to be two separate modules, they were: */
8787 MODULE_ALIAS("raid5");
8788 MODULE_ALIAS("raid6");