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