drivers/md/raid5.c

   1 /*
   2  * raid5.c : Multiple Devices driver for Linux
   3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   4  *         Copyright (C) 1999, 2000 Ingo Molnar
   5  *         Copyright (C) 2002, 2003 H. Peter Anvin
   6  *
   7  * RAID-4/5/6 management functions.
   8  * Thanks to Penguin Computing for making the RAID-6 development possible
   9  * by donating a test server!
  10  *
  11  * This program is free software; you can redistribute it and/or modify
  12  * it under the terms of the GNU General Public License as published by
  13  * the Free Software Foundation; either version 2, or (at your option)
  14  * any later version.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * (for example /usr/src/linux/COPYING); if not, write to the Free
  18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19  */
  20
  21 /*
  22  * BITMAP UNPLUGGING:
  23  *
  24  * The sequencing for updating the bitmap reliably is a little
  25  * subtle (and I got it wrong the first time) so it deserves some
  26  * explanation.
  27  *
  28  * We group bitmap updates into batches.  Each batch has a number.
  29  * We may write out several batches at once, but that isn't very important.
  30  * conf->seq_write is the number of the last batch successfully written.
  31  * conf->seq_flush is the number of the last batch that was closed to
  32  *    new additions.
  33  * When we discover that we will need to write to any block in a stripe
  34  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
  35  * the number of the batch it will be in. This is seq_flush+1.
  36  * When we are ready to do a write, if that batch hasn't been written yet,
  37  *   we plug the array and queue the stripe for later.
  38  * When an unplug happens, we increment bm_flush, thus closing the current
  39  *   batch.
  40  * When we notice that bm_flush > bm_write, we write out all pending updates
  41  * to the bitmap, and advance bm_write to where bm_flush was.
  42  * This may occasionally write a bit out twice, but is sure never to
  43  * miss any bits.
  44  */
  45
  46 #include <linux/blkdev.h>
  47 #include <linux/kthread.h>
  48 #include <linux/raid/pq.h>
  49 #include <linux/async_tx.h>
  50 #include <linux/module.h>
  51 #include <linux/async.h>
  52 #include <linux/seq_file.h>
  53 #include <linux/cpu.h>
  54 #include <linux/slab.h>
  55 #include <linux/ratelimit.h>
  56 #include <linux/nodemask.h>
  57
  58 #include <trace/events/block.h>
  59 #include <linux/list_sort.h>
  60
  61 #include "md.h"
  62 #include "raid5.h"
  63 #include "raid0.h"
  64 #include "md-bitmap.h"
  65 #include "raid5-log.h"
  66
  67 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
  68
  69 #define cpu_to_group(cpu) cpu_to_node(cpu)
  70 #define ANY_GROUP NUMA_NO_NODE
  71
  72 static bool devices_handle_discard_safely = false;
  73 module_param(devices_handle_discard_safely, bool, 0644);
  74 MODULE_PARM_DESC(devices_handle_discard_safely,
  75                  "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
  76 static struct workqueue_struct *raid5_wq;
  77
  78 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
  79 {
  80         int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
  81         return &conf->stripe_hashtbl[hash];
  82 }
  83
  84 static inline int stripe_hash_locks_hash(sector_t sect)
  85 {
  86         return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
  87 }
  88
  89 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
  90 {
  91         spin_lock_irq(conf->hash_locks + hash);
  92         spin_lock(&conf->device_lock);
  93 }
  94
  95 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
  96 {
  97         spin_unlock(&conf->device_lock);
  98         spin_unlock_irq(conf->hash_locks + hash);
  99 }
 100
 101 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
 102 {
 103         int i;
 104         spin_lock_irq(conf->hash_locks);
 105         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
 106                 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
 107         spin_lock(&conf->device_lock);
 108 }
 109
 110 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
 111 {
 112         int i;
 113         spin_unlock(&conf->device_lock);
 114         for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
 115                 spin_unlock(conf->hash_locks + i);
 116         spin_unlock_irq(conf->hash_locks);
 117 }
 118
 119 /* Find first data disk in a raid6 stripe */
 120 static inline int raid6_d0(struct stripe_head *sh)
 121 {
 122         if (sh->ddf_layout)
 123                 /* ddf always start from first device */
 124                 return 0;
 125         /* md starts just after Q block */
 126         if (sh->qd_idx == sh->disks - 1)
 127                 return 0;
 128         else
 129                 return sh->qd_idx + 1;
 130 }
 131 static inline int raid6_next_disk(int disk, int raid_disks)
 132 {
 133         disk++;
 134         return (disk < raid_disks) ? disk : 0;
 135 }
 136
 137 /* When walking through the disks in a raid5, starting at raid6_d0,
 138  * We need to map each disk to a 'slot', where the data disks are slot
 139  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 140  * is raid_disks-1.  This help does that mapping.
 141  */
 142 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 143                              int *count, int syndrome_disks)
 144 {
 145         int slot = *count;
 146
 147         if (sh->ddf_layout)
 148                 (*count)++;
 149         if (idx == sh->pd_idx)
 150                 return syndrome_disks;
 151         if (idx == sh->qd_idx)
 152                 return syndrome_disks + 1;
 153         if (!sh->ddf_layout)
 154                 (*count)++;
 155         return slot;
 156 }
 157
 158 static void print_raid5_conf (struct r5conf *conf);
 159
 160 static int stripe_operations_active(struct stripe_head *sh)
 161 {
 162         return sh->check_state || sh->reconstruct_state ||
 163                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 164                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 165 }
 166
 167 static bool stripe_is_lowprio(struct stripe_head *sh)
 168 {
 169         return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
 170                 test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
 171                !test_bit(STRIPE_R5C_CACHING, &sh->state);
 172 }
 173
 174 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
 175 {
 176         struct r5conf *conf = sh->raid_conf;
 177         struct r5worker_group *group;
 178         int thread_cnt;
 179         int i, cpu = sh->cpu;
 180
 181         if (!cpu_online(cpu)) {
 182                 cpu = cpumask_any(cpu_online_mask);
 183                 sh->cpu = cpu;
 184         }
 185
 186         if (list_empty(&sh->lru)) {
 187                 struct r5worker_group *group;
 188                 group = conf->worker_groups + cpu_to_group(cpu);
 189                 if (stripe_is_lowprio(sh))
 190                         list_add_tail(&sh->lru, &group->loprio_list);
 191                 else
 192                         list_add_tail(&sh->lru, &group->handle_list);
 193                 group->stripes_cnt++;
 194                 sh->group = group;
 195         }
 196
 197         if (conf->worker_cnt_per_group == 0) {
 198                 md_wakeup_thread(conf->mddev->thread);
 199                 return;
 200         }
 201
 202         group = conf->worker_groups + cpu_to_group(sh->cpu);
 203
 204         group->workers[0].working = true;
 205         /* at least one worker should run to avoid race */
 206         queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
 207
 208         thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
 209         /* wakeup more workers */
 210         for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
 211                 if (group->workers[i].working == false) {
 212                         group->workers[i].working = true;
 213                         queue_work_on(sh->cpu, raid5_wq,
 214                                       &group->workers[i].work);
 215                         thread_cnt--;
 216                 }
 217         }
 218 }
 219
 220 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
 221                               struct list_head *temp_inactive_list)
 222 {
 223         int i;
 224         int injournal = 0;      /* number of date pages with R5_InJournal */
 225
 226         BUG_ON(!list_empty(&sh->lru));
 227         BUG_ON(atomic_read(&conf->active_stripes)==0);
 228
 229         if (r5c_is_writeback(conf->log))
 230                 for (i = sh->disks; i--; )
 231                         if (test_bit(R5_InJournal, &sh->dev[i].flags))
 232                                 injournal++;
 233         /*
 234          * In the following cases, the stripe cannot be released to cached
 235          * lists. Therefore, we make the stripe write out and set
 236          * STRIPE_HANDLE:
 237          *   1. when quiesce in r5c write back;
 238          *   2. when resync is requested fot the stripe.
 239          */
 240         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
 241             (conf->quiesce && r5c_is_writeback(conf->log) &&
 242              !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
 243                 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
 244                         r5c_make_stripe_write_out(sh);
 245                 set_bit(STRIPE_HANDLE, &sh->state);
 246         }
 247
 248         if (test_bit(STRIPE_HANDLE, &sh->state)) {
 249                 if (test_bit(STRIPE_DELAYED, &sh->state) &&
 250                     !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 251                         list_add_tail(&sh->lru, &conf->delayed_list);
 252                 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 253                            sh->bm_seq - conf->seq_write > 0)
 254                         list_add_tail(&sh->lru, &conf->bitmap_list);
 255                 else {
 256                         clear_bit(STRIPE_DELAYED, &sh->state);
 257                         clear_bit(STRIPE_BIT_DELAY, &sh->state);
 258                         if (conf->worker_cnt_per_group == 0) {
 259                                 if (stripe_is_lowprio(sh))
 260                                         list_add_tail(&sh->lru,
 261                                                         &conf->loprio_list);
 262                                 else
 263                                         list_add_tail(&sh->lru,
 264                                                         &conf->handle_list);
 265                         } else {
 266                                 raid5_wakeup_stripe_thread(sh);
 267                                 return;
 268                         }
 269                 }
 270                 md_wakeup_thread(conf->mddev->thread);
 271         } else {
 272                 BUG_ON(stripe_operations_active(sh));
 273                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 274                         if (atomic_dec_return(&conf->preread_active_stripes)
 275                             < IO_THRESHOLD)
 276                                 md_wakeup_thread(conf->mddev->thread);
 277                 atomic_dec(&conf->active_stripes);
 278                 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
 279                         if (!r5c_is_writeback(conf->log))
 280                                 list_add_tail(&sh->lru, temp_inactive_list);
 281                         else {
 282                                 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
 283                                 if (injournal == 0)
 284                                         list_add_tail(&sh->lru, temp_inactive_list);
 285                                 else if (injournal == conf->raid_disks - conf->max_degraded) {
 286                                         /* full stripe */
 287                                         if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
 288                                                 atomic_inc(&conf->r5c_cached_full_stripes);
 289                                         if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
 290                                                 atomic_dec(&conf->r5c_cached_partial_stripes);
 291                                         list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
 292                                         r5c_check_cached_full_stripe(conf);
 293                                 } else
 294                                         /*
 295                                          * STRIPE_R5C_PARTIAL_STRIPE is set in
 296                                          * r5c_try_caching_write(). No need to
 297                                          * set it again.
 298                                          */
 299                                         list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
 300                         }
 301                 }
 302         }
 303 }
 304
 305 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
 306                              struct list_head *temp_inactive_list)
 307 {
 308         if (atomic_dec_and_test(&sh->count))
 309                 do_release_stripe(conf, sh, temp_inactive_list);
 310 }
 311
 312 /*
 313  * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
 314  *
 315  * Be careful: Only one task can add/delete stripes from temp_inactive_list at
 316  * given time. Adding stripes only takes device lock, while deleting stripes
 317  * only takes hash lock.
 318  */
 319 static void release_inactive_stripe_list(struct r5conf *conf,
 320                                          struct list_head *temp_inactive_list,
 321                                          int hash)
 322 {
 323         int size;
 324         bool do_wakeup = false;
 325         unsigned long flags;
 326
 327         if (hash == NR_STRIPE_HASH_LOCKS) {
 328                 size = NR_STRIPE_HASH_LOCKS;
 329                 hash = NR_STRIPE_HASH_LOCKS - 1;
 330         } else
 331                 size = 1;
 332         while (size) {
 333                 struct list_head *list = &temp_inactive_list[size - 1];
 334
 335                 /*
 336                  * We don't hold any lock here yet, raid5_get_active_stripe() might
 337                  * remove stripes from the list
 338                  */
 339                 if (!list_empty_careful(list)) {
 340                         spin_lock_irqsave(conf->hash_locks + hash, flags);
 341                         if (list_empty(conf->inactive_list + hash) &&
 342                             !list_empty(list))
 343                                 atomic_dec(&conf->empty_inactive_list_nr);
 344                         list_splice_tail_init(list, conf->inactive_list + hash);
 345                         do_wakeup = true;
 346                         spin_unlock_irqrestore(conf->hash_locks + hash, flags);
 347                 }
 348                 size--;
 349                 hash--;
 350         }
 351
 352         if (do_wakeup) {
 353                 wake_up(&conf->wait_for_stripe);
 354                 if (atomic_read(&conf->active_stripes) == 0)
 355                         wake_up(&conf->wait_for_quiescent);
 356                 if (conf->retry_read_aligned)
 357                         md_wakeup_thread(conf->mddev->thread);
 358         }
 359 }
 360
 361 /* should hold conf->device_lock already */
 362 static int release_stripe_list(struct r5conf *conf,
 363                                struct list_head *temp_inactive_list)
 364 {
 365         struct stripe_head *sh, *t;
 366         int count = 0;
 367         struct llist_node *head;
 368
 369         head = llist_del_all(&conf->released_stripes);
 370         head = llist_reverse_order(head);
 371         llist_for_each_entry_safe(sh, t, head, release_list) {
 372                 int hash;
 373
 374                 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
 375                 smp_mb();
 376                 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
 377                 /*
 378                  * Don't worry the bit is set here, because if the bit is set
 379                  * again, the count is always > 1. This is true for
 380                  * STRIPE_ON_UNPLUG_LIST bit too.
 381                  */
 382                 hash = sh->hash_lock_index;
 383                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
 384                 count++;
 385         }
 386
 387         return count;
 388 }
 389
 390 void raid5_release_stripe(struct stripe_head *sh)
 391 {
 392         struct r5conf *conf = sh->raid_conf;
 393         unsigned long flags;
 394         struct list_head list;
 395         int hash;
 396         bool wakeup;
 397
 398         /* Avoid release_list until the last reference.
 399          */
 400         if (atomic_add_unless(&sh->count, -1, 1))
 401                 return;
 402
 403         if (unlikely(!conf->mddev->thread) ||
 404                 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
 405                 goto slow_path;
 406         wakeup = llist_add(&sh->release_list, &conf->released_stripes);
 407         if (wakeup)
 408                 md_wakeup_thread(conf->mddev->thread);
 409         return;
 410 slow_path:
 411         /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
 412         if (atomic_dec_and_lock_irqsave(&sh->count, &conf->device_lock, flags)) {
 413                 INIT_LIST_HEAD(&list);
 414                 hash = sh->hash_lock_index;
 415                 do_release_stripe(conf, sh, &list);
 416                 spin_unlock_irqrestore(&conf->device_lock, flags);
 417                 release_inactive_stripe_list(conf, &list, hash);
 418         }
 419 }
 420
 421 static inline void remove_hash(struct stripe_head *sh)
 422 {
 423         pr_debug("remove_hash(), stripe %llu\n",
 424                 (unsigned long long)sh->sector);
 425
 426         hlist_del_init(&sh->hash);
 427 }
 428
 429 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 430 {
 431         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 432
 433         pr_debug("insert_hash(), stripe %llu\n",
 434                 (unsigned long long)sh->sector);
 435
 436         hlist_add_head(&sh->hash, hp);
 437 }
 438
 439 /* find an idle stripe, make sure it is unhashed, and return it. */
 440 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
 441 {
 442         struct stripe_head *sh = NULL;
 443         struct list_head *first;
 444
 445         if (list_empty(conf->inactive_list + hash))
 446                 goto out;
 447         first = (conf->inactive_list + hash)->next;
 448         sh = list_entry(first, struct stripe_head, lru);
 449         list_del_init(first);
 450         remove_hash(sh);
 451         atomic_inc(&conf->active_stripes);
 452         BUG_ON(hash != sh->hash_lock_index);
 453         if (list_empty(conf->inactive_list + hash))
 454                 atomic_inc(&conf->empty_inactive_list_nr);
 455 out:
 456         return sh;
 457 }
 458
 459 static void shrink_buffers(struct stripe_head *sh)
 460 {
 461         struct page *p;
 462         int i;
 463         int num = sh->raid_conf->pool_size;
 464
 465         for (i = 0; i < num ; i++) {
 466                 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
 467                 p = sh->dev[i].page;
 468                 if (!p)
 469                         continue;
 470                 sh->dev[i].page = NULL;
 471                 put_page(p);
 472         }
 473 }
 474
 475 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
 476 {
 477         int i;
 478         int num = sh->raid_conf->pool_size;
 479
 480         for (i = 0; i < num; i++) {
 481                 struct page *page;
 482
 483                 if (!(page = alloc_page(gfp))) {
 484                         return 1;
 485                 }
 486                 sh->dev[i].page = page;
 487                 sh->dev[i].orig_page = page;
 488         }
 489
 490         return 0;
 491 }
 492
 493 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 494                             struct stripe_head *sh);
 495
 496 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 497 {
 498         struct r5conf *conf = sh->raid_conf;
 499         int i, seq;
 500
 501         BUG_ON(atomic_read(&sh->count) != 0);
 502         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 503         BUG_ON(stripe_operations_active(sh));
 504         BUG_ON(sh->batch_head);
 505
 506         pr_debug("init_stripe called, stripe %llu\n",
 507                 (unsigned long long)sector);
 508 retry:
 509         seq = read_seqcount_begin(&conf->gen_lock);
 510         sh->generation = conf->generation - previous;
 511         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 512         sh->sector = sector;
 513         stripe_set_idx(sector, conf, previous, sh);
 514         sh->state = 0;
 515
 516         for (i = sh->disks; i--; ) {
 517                 struct r5dev *dev = &sh->dev[i];
 518
 519                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 520                     test_bit(R5_LOCKED, &dev->flags)) {
 521                         pr_err("sector=%llx i=%d %p %p %p %p %d\n",
 522                                (unsigned long long)sh->sector, i, dev->toread,
 523                                dev->read, dev->towrite, dev->written,
 524                                test_bit(R5_LOCKED, &dev->flags));
 525                         WARN_ON(1);
 526                 }
 527                 dev->flags = 0;
 528                 dev->sector = raid5_compute_blocknr(sh, i, previous);
 529         }
 530         if (read_seqcount_retry(&conf->gen_lock, seq))
 531                 goto retry;
 532         sh->overwrite_disks = 0;
 533         insert_hash(conf, sh);
 534         sh->cpu = smp_processor_id();
 535         set_bit(STRIPE_BATCH_READY, &sh->state);
 536 }
 537
 538 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 539                                          short generation)
 540 {
 541         struct stripe_head *sh;
 542
 543         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 544         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 545                 if (sh->sector == sector && sh->generation == generation)
 546                         return sh;
 547         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 548         return NULL;
 549 }
 550
 551 /*
 552  * Need to check if array has failed when deciding whether to:
 553  *  - start an array
 554  *  - remove non-faulty devices
 555  *  - add a spare
 556  *  - allow a reshape
 557  * This determination is simple when no reshape is happening.
 558  * However if there is a reshape, we need to carefully check
 559  * both the before and after sections.
 560  * This is because some failed devices may only affect one
 561  * of the two sections, and some non-in_sync devices may
 562  * be insync in the section most affected by failed devices.
 563  */
 564 int raid5_calc_degraded(struct r5conf *conf)
 565 {
 566         int degraded, degraded2;
 567         int i;
 568
 569         rcu_read_lock();
 570         degraded = 0;
 571         for (i = 0; i < conf->previous_raid_disks; i++) {
 572                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 573                 if (rdev && test_bit(Faulty, &rdev->flags))
 574                         rdev = rcu_dereference(conf->disks[i].replacement);
 575                 if (!rdev || test_bit(Faulty, &rdev->flags))
 576                         degraded++;
 577                 else if (test_bit(In_sync, &rdev->flags))
 578                         ;
 579                 else
 580                         /* not in-sync or faulty.
 581                          * If the reshape increases the number of devices,
 582                          * this is being recovered by the reshape, so
 583                          * this 'previous' section is not in_sync.
 584                          * If the number of devices is being reduced however,
 585                          * the device can only be part of the array if
 586                          * we are reverting a reshape, so this section will
 587                          * be in-sync.
 588                          */
 589                         if (conf->raid_disks >= conf->previous_raid_disks)
 590                                 degraded++;
 591         }
 592         rcu_read_unlock();
 593         if (conf->raid_disks == conf->previous_raid_disks)
 594                 return degraded;
 595         rcu_read_lock();
 596         degraded2 = 0;
 597         for (i = 0; i < conf->raid_disks; i++) {
 598                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 599                 if (rdev && test_bit(Faulty, &rdev->flags))
 600                         rdev = rcu_dereference(conf->disks[i].replacement);
 601                 if (!rdev || test_bit(Faulty, &rdev->flags))
 602                         degraded2++;
 603                 else if (test_bit(In_sync, &rdev->flags))
 604                         ;
 605                 else
 606                         /* not in-sync or faulty.
 607                          * If reshape increases the number of devices, this
 608                          * section has already been recovered, else it
 609                          * almost certainly hasn't.
 610                          */
 611                         if (conf->raid_disks <= conf->previous_raid_disks)
 612                                 degraded2++;
 613         }
 614         rcu_read_unlock();
 615         if (degraded2 > degraded)
 616                 return degraded2;
 617         return degraded;
 618 }
 619
 620 static int has_failed(struct r5conf *conf)
 621 {
 622         int degraded;
 623
 624         if (conf->mddev->reshape_position == MaxSector)
 625                 return conf->mddev->degraded > conf->max_degraded;
 626
 627         degraded = raid5_calc_degraded(conf);
 628         if (degraded > conf->max_degraded)
 629                 return 1;
 630         return 0;
 631 }
 632
 633 struct stripe_head *
 634 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
 635                         int previous, int noblock, int noquiesce)
 636 {
 637         struct stripe_head *sh;
 638         int hash = stripe_hash_locks_hash(sector);
 639         int inc_empty_inactive_list_flag;
 640
 641         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 642
 643         spin_lock_irq(conf->hash_locks + hash);
 644
 645         do {
 646                 wait_event_lock_irq(conf->wait_for_quiescent,
 647                                     conf->quiesce == 0 || noquiesce,
 648                                     *(conf->hash_locks + hash));
 649                 sh = __find_stripe(conf, sector, conf->generation - previous);
 650                 if (!sh) {
 651                         if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 652                                 sh = get_free_stripe(conf, hash);
 653                                 if (!sh && !test_bit(R5_DID_ALLOC,
 654                                                      &conf->cache_state))
 655                                         set_bit(R5_ALLOC_MORE,
 656                                                 &conf->cache_state);
 657                         }
 658                         if (noblock && sh == NULL)
 659                                 break;
 660
 661                         r5c_check_stripe_cache_usage(conf);
 662                         if (!sh) {
 663                                 set_bit(R5_INACTIVE_BLOCKED,
 664                                         &conf->cache_state);
 665                                 r5l_wake_reclaim(conf->log, 0);
 666                                 wait_event_lock_irq(
 667                                         conf->wait_for_stripe,
 668                                         !list_empty(conf->inactive_list + hash) &&
 669                                         (atomic_read(&conf->active_stripes)
 670                                          < (conf->max_nr_stripes * 3 / 4)
 671                                          || !test_bit(R5_INACTIVE_BLOCKED,
 672                                                       &conf->cache_state)),
 673                                         *(conf->hash_locks + hash));
 674                                 clear_bit(R5_INACTIVE_BLOCKED,
 675                                           &conf->cache_state);
 676                         } else {
 677                                 init_stripe(sh, sector, previous);
 678                                 atomic_inc(&sh->count);
 679                         }
 680                 } else if (!atomic_inc_not_zero(&sh->count)) {
 681                         spin_lock(&conf->device_lock);
 682                         if (!atomic_read(&sh->count)) {
 683                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 684                                         atomic_inc(&conf->active_stripes);
 685                                 BUG_ON(list_empty(&sh->lru) &&
 686                                        !test_bit(STRIPE_EXPANDING, &sh->state));
 687                                 inc_empty_inactive_list_flag = 0;
 688                                 if (!list_empty(conf->inactive_list + hash))
 689                                         inc_empty_inactive_list_flag = 1;
 690                                 list_del_init(&sh->lru);
 691                                 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 692                                         atomic_inc(&conf->empty_inactive_list_nr);
 693                                 if (sh->group) {
 694                                         sh->group->stripes_cnt--;
 695                                         sh->group = NULL;
 696                                 }
 697                         }
 698                         atomic_inc(&sh->count);
 699                         spin_unlock(&conf->device_lock);
 700                 }
 701         } while (sh == NULL);
 702
 703         spin_unlock_irq(conf->hash_locks + hash);
 704         return sh;
 705 }
 706
 707 static bool is_full_stripe_write(struct stripe_head *sh)
 708 {
 709         BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 710         return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 711 }
 712
 713 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 714 {
 715         if (sh1 > sh2) {
 716                 spin_lock_irq(&sh2->stripe_lock);
 717                 spin_lock_nested(&sh1->stripe_lock, 1);
 718         } else {
 719                 spin_lock_irq(&sh1->stripe_lock);
 720                 spin_lock_nested(&sh2->stripe_lock, 1);
 721         }
 722 }
 723
 724 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 725 {
 726         spin_unlock(&sh1->stripe_lock);
 727         spin_unlock_irq(&sh2->stripe_lock);
 728 }
 729
 730 /* Only freshly new full stripe normal write stripe can be added to a batch list */
 731 static bool stripe_can_batch(struct stripe_head *sh)
 732 {
 733         struct r5conf *conf = sh->raid_conf;
 734
 735         if (raid5_has_log(conf) || raid5_has_ppl(conf))
 736                 return false;
 737         return test_bit(STRIPE_BATCH_READY, &sh->state) &&
 738                 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
 739                 is_full_stripe_write(sh);
 740 }
 741
 742 /* we only do back search */
 743 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
 744 {
 745         struct stripe_head *head;
 746         sector_t head_sector, tmp_sec;
 747         int hash;
 748         int dd_idx;
 749         int inc_empty_inactive_list_flag;
 750
 751         /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
 752         tmp_sec = sh->sector;
 753         if (!sector_div(tmp_sec, conf->chunk_sectors))
 754                 return;
 755         head_sector = sh->sector - STRIPE_SECTORS;
 756
 757         hash = stripe_hash_locks_hash(head_sector);
 758         spin_lock_irq(conf->hash_locks + hash);
 759         head = __find_stripe(conf, head_sector, conf->generation);
 760         if (head && !atomic_inc_not_zero(&head->count)) {
 761                 spin_lock(&conf->device_lock);
 762                 if (!atomic_read(&head->count)) {
 763                         if (!test_bit(STRIPE_HANDLE, &head->state))
 764                                 atomic_inc(&conf->active_stripes);
 765                         BUG_ON(list_empty(&head->lru) &&
 766                                !test_bit(STRIPE_EXPANDING, &head->state));
 767                         inc_empty_inactive_list_flag = 0;
 768                         if (!list_empty(conf->inactive_list + hash))
 769                                 inc_empty_inactive_list_flag = 1;
 770                         list_del_init(&head->lru);
 771                         if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 772                                 atomic_inc(&conf->empty_inactive_list_nr);
 773                         if (head->group) {
 774                                 head->group->stripes_cnt--;
 775                                 head->group = NULL;
 776                         }
 777                 }
 778                 atomic_inc(&head->count);
 779                 spin_unlock(&conf->device_lock);
 780         }
 781         spin_unlock_irq(conf->hash_locks + hash);
 782
 783         if (!head)
 784                 return;
 785         if (!stripe_can_batch(head))
 786                 goto out;
 787
 788         lock_two_stripes(head, sh);
 789         /* clear_batch_ready clear the flag */
 790         if (!stripe_can_batch(head) || !stripe_can_batch(sh))
 791                 goto unlock_out;
 792
 793         if (sh->batch_head)
 794                 goto unlock_out;
 795
 796         dd_idx = 0;
 797         while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
 798                 dd_idx++;
 799         if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
 800             bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
 801                 goto unlock_out;
 802
 803         if (head->batch_head) {
 804                 spin_lock(&head->batch_head->batch_lock);
 805                 /* This batch list is already running */
 806                 if (!stripe_can_batch(head)) {
 807                         spin_unlock(&head->batch_head->batch_lock);
 808                         goto unlock_out;
 809                 }
 810                 /*
 811                  * We must assign batch_head of this stripe within the
 812                  * batch_lock, otherwise clear_batch_ready of batch head
 813                  * stripe could clear BATCH_READY bit of this stripe and
 814                  * this stripe->batch_head doesn't get assigned, which
 815                  * could confuse clear_batch_ready for this stripe
 816                  */
 817                 sh->batch_head = head->batch_head;
 818
 819                 /*
 820                  * at this point, head's BATCH_READY could be cleared, but we
 821                  * can still add the stripe to batch list
 822                  */
 823                 list_add(&sh->batch_list, &head->batch_list);
 824                 spin_unlock(&head->batch_head->batch_lock);
 825         } else {
 826                 head->batch_head = head;
 827                 sh->batch_head = head->batch_head;
 828                 spin_lock(&head->batch_lock);
 829                 list_add_tail(&sh->batch_list, &head->batch_list);
 830                 spin_unlock(&head->batch_lock);
 831         }
 832
 833         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 834                 if (atomic_dec_return(&conf->preread_active_stripes)
 835                     < IO_THRESHOLD)
 836                         md_wakeup_thread(conf->mddev->thread);
 837
 838         if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
 839                 int seq = sh->bm_seq;
 840                 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
 841                     sh->batch_head->bm_seq > seq)
 842                         seq = sh->batch_head->bm_seq;
 843                 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
 844                 sh->batch_head->bm_seq = seq;
 845         }
 846
 847         atomic_inc(&sh->count);
 848 unlock_out:
 849         unlock_two_stripes(head, sh);
 850 out:
 851         raid5_release_stripe(head);
 852 }
 853
 854 /* Determine if 'data_offset' or 'new_data_offset' should be used
 855  * in this stripe_head.
 856  */
 857 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
 858 {
 859         sector_t progress = conf->reshape_progress;
 860         /* Need a memory barrier to make sure we see the value
 861          * of conf->generation, or ->data_offset that was set before
 862          * reshape_progress was updated.
 863          */
 864         smp_rmb();
 865         if (progress == MaxSector)
 866                 return 0;
 867         if (sh->generation == conf->generation - 1)
 868                 return 0;
 869         /* We are in a reshape, and this is a new-generation stripe,
 870          * so use new_data_offset.
 871          */
 872         return 1;
 873 }
 874
 875 static void dispatch_bio_list(struct bio_list *tmp)
 876 {
 877         struct bio *bio;
 878
 879         while ((bio = bio_list_pop(tmp)))
 880                 generic_make_request(bio);
 881 }
 882
 883 static int cmp_stripe(void *priv, struct list_head *a, struct list_head *b)
 884 {
 885         const struct r5pending_data *da = list_entry(a,
 886                                 struct r5pending_data, sibling);
 887         const struct r5pending_data *db = list_entry(b,
 888                                 struct r5pending_data, sibling);
 889         if (da->sector > db->sector)
 890                 return 1;
 891         if (da->sector < db->sector)
 892                 return -1;
 893         return 0;
 894 }
 895
 896 static void dispatch_defer_bios(struct r5conf *conf, int target,
 897                                 struct bio_list *list)
 898 {
 899         struct r5pending_data *data;
 900         struct list_head *first, *next = NULL;
 901         int cnt = 0;
 902
 903         if (conf->pending_data_cnt == 0)
 904                 return;
 905
 906         list_sort(NULL, &conf->pending_list, cmp_stripe);
 907
 908         first = conf->pending_list.next;
 909
 910         /* temporarily move the head */
 911         if (conf->next_pending_data)
 912                 list_move_tail(&conf->pending_list,
 913                                 &conf->next_pending_data->sibling);
 914
 915         while (!list_empty(&conf->pending_list)) {
 916                 data = list_first_entry(&conf->pending_list,
 917                         struct r5pending_data, sibling);
 918                 if (&data->sibling == first)
 919                         first = data->sibling.next;
 920                 next = data->sibling.next;
 921
 922                 bio_list_merge(list, &data->bios);
 923                 list_move(&data->sibling, &conf->free_list);
 924                 cnt++;
 925                 if (cnt >= target)
 926                         break;
 927         }
 928         conf->pending_data_cnt -= cnt;
 929         BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
 930
 931         if (next != &conf->pending_list)
 932                 conf->next_pending_data = list_entry(next,
 933                                 struct r5pending_data, sibling);
 934         else
 935                 conf->next_pending_data = NULL;
 936         /* list isn't empty */
 937         if (first != &conf->pending_list)
 938                 list_move_tail(&conf->pending_list, first);
 939 }
 940
 941 static void flush_deferred_bios(struct r5conf *conf)
 942 {
 943         struct bio_list tmp = BIO_EMPTY_LIST;
 944
 945         if (conf->pending_data_cnt == 0)
 946                 return;
 947
 948         spin_lock(&conf->pending_bios_lock);
 949         dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
 950         BUG_ON(conf->pending_data_cnt != 0);
 951         spin_unlock(&conf->pending_bios_lock);
 952
 953         dispatch_bio_list(&tmp);
 954 }
 955
 956 static void defer_issue_bios(struct r5conf *conf, sector_t sector,
 957                                 struct bio_list *bios)
 958 {
 959         struct bio_list tmp = BIO_EMPTY_LIST;
 960         struct r5pending_data *ent;
 961
 962         spin_lock(&conf->pending_bios_lock);
 963         ent = list_first_entry(&conf->free_list, struct r5pending_data,
 964                                                         sibling);
 965         list_move_tail(&ent->sibling, &conf->pending_list);
 966         ent->sector = sector;
 967         bio_list_init(&ent->bios);
 968         bio_list_merge(&ent->bios, bios);
 969         conf->pending_data_cnt++;
 970         if (conf->pending_data_cnt >= PENDING_IO_MAX)
 971                 dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
 972
 973         spin_unlock(&conf->pending_bios_lock);
 974
 975         dispatch_bio_list(&tmp);
 976 }
 977
 978 static void
 979 raid5_end_read_request(struct bio *bi);
 980 static void
 981 raid5_end_write_request(struct bio *bi);
 982
 983 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
 984 {
 985         struct r5conf *conf = sh->raid_conf;
 986         int i, disks = sh->disks;
 987         struct stripe_head *head_sh = sh;
 988         struct bio_list pending_bios = BIO_EMPTY_LIST;
 989         bool should_defer;
 990
 991         might_sleep();
 992
 993         if (log_stripe(sh, s) == 0)
 994                 return;
 995
 996         should_defer = conf->batch_bio_dispatch && conf->group_cnt;
 997
 998         for (i = disks; i--; ) {
 999                 int op, op_flags = 0;
1000                 int replace_only = 0;
1001                 struct bio *bi, *rbi;
1002                 struct md_rdev *rdev, *rrdev = NULL;
1003
1004                 sh = head_sh;
1005                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1006                         op = REQ_OP_WRITE;
1007                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1008                                 op_flags = REQ_FUA;
1009                         if (test_bit(R5_Discard, &sh->dev[i].flags))
1010                                 op = REQ_OP_DISCARD;
1011                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1012                         op = REQ_OP_READ;
1013                 else if (test_and_clear_bit(R5_WantReplace,
1014                                             &sh->dev[i].flags)) {
1015                         op = REQ_OP_WRITE;
1016                         replace_only = 1;
1017                 } else
1018                         continue;
1019                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1020                         op_flags |= REQ_SYNC;
1021
1022 again:
1023                 bi = &sh->dev[i].req;
1024                 rbi = &sh->dev[i].rreq; /* For writing to replacement */
1025
1026                 rcu_read_lock();
1027                 rrdev = rcu_dereference(conf->disks[i].replacement);
1028                 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1029                 rdev = rcu_dereference(conf->disks[i].rdev);
1030                 if (!rdev) {
1031                         rdev = rrdev;
1032                         rrdev = NULL;
1033                 }
1034                 if (op_is_write(op)) {
1035                         if (replace_only)
1036                                 rdev = NULL;
1037                         if (rdev == rrdev)
1038                                 /* We raced and saw duplicates */
1039                                 rrdev = NULL;
1040                 } else {
1041                         if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1042                                 rdev = rrdev;
1043                         rrdev = NULL;
1044                 }
1045
1046                 if (rdev && test_bit(Faulty, &rdev->flags))
1047                         rdev = NULL;
1048                 if (rdev)
1049                         atomic_inc(&rdev->nr_pending);
1050                 if (rrdev && test_bit(Faulty, &rrdev->flags))
1051                         rrdev = NULL;
1052                 if (rrdev)
1053                         atomic_inc(&rrdev->nr_pending);
1054                 rcu_read_unlock();
1055
1056                 /* We have already checked bad blocks for reads.  Now
1057                  * need to check for writes.  We never accept write errors
1058                  * on the replacement, so we don't to check rrdev.
1059                  */
1060                 while (op_is_write(op) && rdev &&
1061                        test_bit(WriteErrorSeen, &rdev->flags)) {
1062                         sector_t first_bad;
1063                         int bad_sectors;
1064                         int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
1065                                               &first_bad, &bad_sectors);
1066                         if (!bad)
1067                                 break;
1068
1069                         if (bad < 0) {
1070                                 set_bit(BlockedBadBlocks, &rdev->flags);
1071                                 if (!conf->mddev->external &&
1072                                     conf->mddev->sb_flags) {
1073                                         /* It is very unlikely, but we might
1074                                          * still need to write out the
1075                                          * bad block log - better give it
1076                                          * a chance*/
1077                                         md_check_recovery(conf->mddev);
1078                                 }
1079                                 /*
1080                                  * Because md_wait_for_blocked_rdev
1081                                  * will dec nr_pending, we must
1082                                  * increment it first.
1083                                  */
1084                                 atomic_inc(&rdev->nr_pending);
1085                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
1086                         } else {
1087                                 /* Acknowledged bad block - skip the write */
1088                                 rdev_dec_pending(rdev, conf->mddev);
1089                                 rdev = NULL;
1090                         }
1091                 }
1092
1093                 if (rdev) {
1094                         if (s->syncing || s->expanding || s->expanded
1095                             || s->replacing)
1096                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1097
1098                         set_bit(STRIPE_IO_STARTED, &sh->state);
1099
1100                         bio_set_dev(bi, rdev->bdev);
1101                         bio_set_op_attrs(bi, op, op_flags);
1102                         bi->bi_end_io = op_is_write(op)
1103                                 ? raid5_end_write_request
1104                                 : raid5_end_read_request;
1105                         bi->bi_private = sh;
1106
1107                         pr_debug("%s: for %llu schedule op %d on disc %d\n",
1108                                 __func__, (unsigned long long)sh->sector,
1109                                 bi->bi_opf, i);
1110                         atomic_inc(&sh->count);
1111                         if (sh != head_sh)
1112                                 atomic_inc(&head_sh->count);
1113                         if (use_new_offset(conf, sh))
1114                                 bi->bi_iter.bi_sector = (sh->sector
1115                                                  + rdev->new_data_offset);
1116                         else
1117                                 bi->bi_iter.bi_sector = (sh->sector
1118                                                  + rdev->data_offset);
1119                         if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1120                                 bi->bi_opf |= REQ_NOMERGE;
1121
1122                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1123                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1124
1125                         if (!op_is_write(op) &&
1126                             test_bit(R5_InJournal, &sh->dev[i].flags))
1127                                 /*
1128                                  * issuing read for a page in journal, this
1129                                  * must be preparing for prexor in rmw; read
1130                                  * the data into orig_page
1131                                  */
1132                                 sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1133                         else
1134                                 sh->dev[i].vec.bv_page = sh->dev[i].page;
1135                         bi->bi_vcnt = 1;
1136                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1137                         bi->bi_io_vec[0].bv_offset = 0;
1138                         bi->bi_iter.bi_size = STRIPE_SIZE;
1139                         bi->bi_write_hint = sh->dev[i].write_hint;
1140                         if (!rrdev)
1141                                 sh->dev[i].write_hint = RWF_WRITE_LIFE_NOT_SET;
1142                         /*
1143                          * If this is discard request, set bi_vcnt 0. We don't
1144                          * want to confuse SCSI because SCSI will replace payload
1145                          */
1146                         if (op == REQ_OP_DISCARD)
1147                                 bi->bi_vcnt = 0;
1148                         if (rrdev)
1149                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1150
1151                         if (conf->mddev->gendisk)
1152                                 trace_block_bio_remap(bi->bi_disk->queue,
1153                                                       bi, disk_devt(conf->mddev->gendisk),
1154                                                       sh->dev[i].sector);
1155                         if (should_defer && op_is_write(op))
1156                                 bio_list_add(&pending_bios, bi);
1157                         else
1158                                 generic_make_request(bi);
1159                 }
1160                 if (rrdev) {
1161                         if (s->syncing || s->expanding || s->expanded
1162                             || s->replacing)
1163                                 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1164
1165                         set_bit(STRIPE_IO_STARTED, &sh->state);
1166
1167                         bio_set_dev(rbi, rrdev->bdev);
1168                         bio_set_op_attrs(rbi, op, op_flags);
1169                         BUG_ON(!op_is_write(op));
1170                         rbi->bi_end_io = raid5_end_write_request;
1171                         rbi->bi_private = sh;
1172
1173                         pr_debug("%s: for %llu schedule op %d on "
1174                                  "replacement disc %d\n",
1175                                 __func__, (unsigned long long)sh->sector,
1176                                 rbi->bi_opf, i);
1177                         atomic_inc(&sh->count);
1178                         if (sh != head_sh)
1179                                 atomic_inc(&head_sh->count);
1180                         if (use_new_offset(conf, sh))
1181                                 rbi->bi_iter.bi_sector = (sh->sector
1182                                                   + rrdev->new_data_offset);
1183                         else
1184                                 rbi->bi_iter.bi_sector = (sh->sector
1185                                                   + rrdev->data_offset);
1186                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1187                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1188                         sh->dev[i].rvec.bv_page = sh->dev[i].page;
1189                         rbi->bi_vcnt = 1;
1190                         rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1191                         rbi->bi_io_vec[0].bv_offset = 0;
1192                         rbi->bi_iter.bi_size = STRIPE_SIZE;
1193                         rbi->bi_write_hint = sh->dev[i].write_hint;
1194                         sh->dev[i].write_hint = RWF_WRITE_LIFE_NOT_SET;
1195                         /*
1196                          * If this is discard request, set bi_vcnt 0. We don't
1197                          * want to confuse SCSI because SCSI will replace payload
1198                          */
1199                         if (op == REQ_OP_DISCARD)
1200                                 rbi->bi_vcnt = 0;
1201                         if (conf->mddev->gendisk)
1202                                 trace_block_bio_remap(rbi->bi_disk->queue,
1203                                                       rbi, disk_devt(conf->mddev->gendisk),
1204                                                       sh->dev[i].sector);
1205                         if (should_defer && op_is_write(op))
1206                                 bio_list_add(&pending_bios, rbi);
1207                         else
1208                                 generic_make_request(rbi);
1209                 }
1210                 if (!rdev && !rrdev) {
1211                         if (op_is_write(op))
1212                                 set_bit(STRIPE_DEGRADED, &sh->state);
1213                         pr_debug("skip op %d on disc %d for sector %llu\n",
1214                                 bi->bi_opf, i, (unsigned long long)sh->sector);
1215                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1216                         set_bit(STRIPE_HANDLE, &sh->state);
1217                 }
1218
1219                 if (!head_sh->batch_head)
1220                         continue;
1221                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1222                                       batch_list);
1223                 if (sh != head_sh)
1224                         goto again;
1225         }
1226
1227         if (should_defer && !bio_list_empty(&pending_bios))
1228                 defer_issue_bios(conf, head_sh->sector, &pending_bios);
1229 }
1230
1231 static struct dma_async_tx_descriptor *
1232 async_copy_data(int frombio, struct bio *bio, struct page **page,
1233         sector_t sector, struct dma_async_tx_descriptor *tx,
1234         struct stripe_head *sh, int no_skipcopy)
1235 {
1236         struct bio_vec bvl;
1237         struct bvec_iter iter;
1238         struct page *bio_page;
1239         int page_offset;
1240         struct async_submit_ctl submit;
1241         enum async_tx_flags flags = 0;
1242
1243         if (bio->bi_iter.bi_sector >= sector)
1244                 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1245         else
1246                 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1247
1248         if (frombio)
1249                 flags |= ASYNC_TX_FENCE;
1250         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1251
1252         bio_for_each_segment(bvl, bio, iter) {
1253                 int len = bvl.bv_len;
1254                 int clen;
1255                 int b_offset = 0;
1256
1257                 if (page_offset < 0) {
1258                         b_offset = -page_offset;
1259                         page_offset += b_offset;
1260                         len -= b_offset;
1261                 }
1262
1263                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1264                         clen = STRIPE_SIZE - page_offset;
1265                 else
1266                         clen = len;
1267
1268                 if (clen > 0) {
1269                         b_offset += bvl.bv_offset;
1270                         bio_page = bvl.bv_page;
1271                         if (frombio) {
1272                                 if (sh->raid_conf->skip_copy &&
1273                                     b_offset == 0 && page_offset == 0 &&
1274                                     clen == STRIPE_SIZE &&
1275                                     !no_skipcopy)
1276                                         *page = bio_page;
1277                                 else
1278                                         tx = async_memcpy(*page, bio_page, page_offset,
1279                                                   b_offset, clen, &submit);
1280                         } else
1281                                 tx = async_memcpy(bio_page, *page, b_offset,
1282                                                   page_offset, clen, &submit);
1283                 }
1284                 /* chain the operations */
1285                 submit.depend_tx = tx;
1286
1287                 if (clen < len) /* hit end of page */
1288                         break;
1289                 page_offset +=  len;
1290         }
1291
1292         return tx;
1293 }
1294
1295 static void ops_complete_biofill(void *stripe_head_ref)
1296 {
1297         struct stripe_head *sh = stripe_head_ref;
1298         int i;
1299
1300         pr_debug("%s: stripe %llu\n", __func__,
1301                 (unsigned long long)sh->sector);
1302
1303         /* clear completed biofills */
1304         for (i = sh->disks; i--; ) {
1305                 struct r5dev *dev = &sh->dev[i];
1306
1307                 /* acknowledge completion of a biofill operation */
1308                 /* and check if we need to reply to a read request,
1309                  * new R5_Wantfill requests are held off until
1310                  * !STRIPE_BIOFILL_RUN
1311                  */
1312                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1313                         struct bio *rbi, *rbi2;
1314
1315                         BUG_ON(!dev->read);
1316                         rbi = dev->read;
1317                         dev->read = NULL;
1318                         while (rbi && rbi->bi_iter.bi_sector <
1319                                 dev->sector + STRIPE_SECTORS) {
1320                                 rbi2 = r5_next_bio(rbi, dev->sector);
1321                                 bio_endio(rbi);
1322                                 rbi = rbi2;
1323                         }
1324                 }
1325         }
1326         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1327
1328         set_bit(STRIPE_HANDLE, &sh->state);
1329         raid5_release_stripe(sh);
1330 }
1331
1332 static void ops_run_biofill(struct stripe_head *sh)
1333 {
1334         struct dma_async_tx_descriptor *tx = NULL;
1335         struct async_submit_ctl submit;
1336         int i;
1337
1338         BUG_ON(sh->batch_head);
1339         pr_debug("%s: stripe %llu\n", __func__,
1340                 (unsigned long long)sh->sector);
1341
1342         for (i = sh->disks; i--; ) {
1343                 struct r5dev *dev = &sh->dev[i];
1344                 if (test_bit(R5_Wantfill, &dev->flags)) {
1345                         struct bio *rbi;
1346                         spin_lock_irq(&sh->stripe_lock);
1347                         dev->read = rbi = dev->toread;
1348                         dev->toread = NULL;
1349                         spin_unlock_irq(&sh->stripe_lock);
1350                         while (rbi && rbi->bi_iter.bi_sector <
1351                                 dev->sector + STRIPE_SECTORS) {
1352                                 tx = async_copy_data(0, rbi, &dev->page,
1353                                                      dev->sector, tx, sh, 0);
1354                                 rbi = r5_next_bio(rbi, dev->sector);
1355                         }
1356                 }
1357         }
1358
1359         atomic_inc(&sh->count);
1360         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1361         async_trigger_callback(&submit);
1362 }
1363
1364 static void mark_target_uptodate(struct stripe_head *sh, int target)
1365 {
1366         struct r5dev *tgt;
1367
1368         if (target < 0)
1369                 return;
1370
1371         tgt = &sh->dev[target];
1372         set_bit(R5_UPTODATE, &tgt->flags);
1373         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1374         clear_bit(R5_Wantcompute, &tgt->flags);
1375 }
1376
1377 static void ops_complete_compute(void *stripe_head_ref)
1378 {
1379         struct stripe_head *sh = stripe_head_ref;
1380
1381         pr_debug("%s: stripe %llu\n", __func__,
1382                 (unsigned long long)sh->sector);
1383
1384         /* mark the computed target(s) as uptodate */
1385         mark_target_uptodate(sh, sh->ops.target);
1386         mark_target_uptodate(sh, sh->ops.target2);
1387
1388         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1389         if (sh->check_state == check_state_compute_run)
1390                 sh->check_state = check_state_compute_result;
1391         set_bit(STRIPE_HANDLE, &sh->state);
1392         raid5_release_stripe(sh);
1393 }
1394
1395 /* return a pointer to the address conversion region of the scribble buffer */
1396 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1397 {
1398         return percpu->scribble + i * percpu->scribble_obj_size;
1399 }
1400
1401 /* return a pointer to the address conversion region of the scribble buffer */
1402 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1403                                  struct raid5_percpu *percpu, int i)
1404 {
1405         return (void *) (to_addr_page(percpu, i) + sh->disks + 2);
1406 }
1407
1408 static struct dma_async_tx_descriptor *
1409 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1410 {
1411         int disks = sh->disks;
1412         struct page **xor_srcs = to_addr_page(percpu, 0);
1413         int target = sh->ops.target;
1414         struct r5dev *tgt = &sh->dev[target];
1415         struct page *xor_dest = tgt->page;
1416         int count = 0;
1417         struct dma_async_tx_descriptor *tx;
1418         struct async_submit_ctl submit;
1419         int i;
1420
1421         BUG_ON(sh->batch_head);
1422
1423         pr_debug("%s: stripe %llu block: %d\n",
1424                 __func__, (unsigned long long)sh->sector, target);
1425         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1426
1427         for (i = disks; i--; )
1428                 if (i != target)
1429                         xor_srcs[count++] = sh->dev[i].page;
1430
1431         atomic_inc(&sh->count);
1432
1433         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1434                           ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1435         if (unlikely(count == 1))
1436                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1437         else
1438                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1439
1440         return tx;
1441 }
1442
1443 /* set_syndrome_sources - populate source buffers for gen_syndrome
1444  * @srcs - (struct page *) array of size sh->disks
1445  * @sh - stripe_head to parse
1446  *
1447  * Populates srcs in proper layout order for the stripe and returns the
1448  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1449  * destination buffer is recorded in srcs[count] and the Q destination
1450  * is recorded in srcs[count+1]].
1451  */
1452 static int set_syndrome_sources(struct page **srcs,
1453                                 struct stripe_head *sh,
1454                                 int srctype)
1455 {
1456         int disks = sh->disks;
1457         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1458         int d0_idx = raid6_d0(sh);
1459         int count;
1460         int i;
1461
1462         for (i = 0; i < disks; i++)
1463                 srcs[i] = NULL;
1464
1465         count = 0;
1466         i = d0_idx;
1467         do {
1468                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1469                 struct r5dev *dev = &sh->dev[i];
1470
1471                 if (i == sh->qd_idx || i == sh->pd_idx ||
1472                     (srctype == SYNDROME_SRC_ALL) ||
1473                     (srctype == SYNDROME_SRC_WANT_DRAIN &&
1474                      (test_bit(R5_Wantdrain, &dev->flags) ||
1475                       test_bit(R5_InJournal, &dev->flags))) ||
1476                     (srctype == SYNDROME_SRC_WRITTEN &&
1477                      (dev->written ||
1478                       test_bit(R5_InJournal, &dev->flags)))) {
1479                         if (test_bit(R5_InJournal, &dev->flags))
1480                                 srcs[slot] = sh->dev[i].orig_page;
1481                         else
1482                                 srcs[slot] = sh->dev[i].page;
1483                 }
1484                 i = raid6_next_disk(i, disks);
1485         } while (i != d0_idx);
1486
1487         return syndrome_disks;
1488 }
1489
1490 static struct dma_async_tx_descriptor *
1491 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1492 {
1493         int disks = sh->disks;
1494         struct page **blocks = to_addr_page(percpu, 0);
1495         int target;
1496         int qd_idx = sh->qd_idx;
1497         struct dma_async_tx_descriptor *tx;
1498         struct async_submit_ctl submit;
1499         struct r5dev *tgt;
1500         struct page *dest;
1501         int i;
1502         int count;
1503
1504         BUG_ON(sh->batch_head);
1505         if (sh->ops.target < 0)
1506                 target = sh->ops.target2;
1507         else if (sh->ops.target2 < 0)
1508                 target = sh->ops.target;
1509         else
1510                 /* we should only have one valid target */
1511                 BUG();
1512         BUG_ON(target < 0);
1513         pr_debug("%s: stripe %llu block: %d\n",
1514                 __func__, (unsigned long long)sh->sector, target);
1515
1516         tgt = &sh->dev[target];
1517         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1518         dest = tgt->page;
1519
1520         atomic_inc(&sh->count);
1521
1522         if (target == qd_idx) {
1523                 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1524                 blocks[count] = NULL; /* regenerating p is not necessary */
1525                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1526                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1527                                   ops_complete_compute, sh,
1528                                   to_addr_conv(sh, percpu, 0));
1529                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1530         } else {
1531                 /* Compute any data- or p-drive using XOR */
1532                 count = 0;
1533                 for (i = disks; i-- ; ) {
1534                         if (i == target || i == qd_idx)
1535                                 continue;
1536                         blocks[count++] = sh->dev[i].page;
1537                 }
1538
1539                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1540                                   NULL, ops_complete_compute, sh,
1541                                   to_addr_conv(sh, percpu, 0));
1542                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1543         }
1544
1545         return tx;
1546 }
1547
1548 static struct dma_async_tx_descriptor *
1549 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1550 {
1551         int i, count, disks = sh->disks;
1552         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1553         int d0_idx = raid6_d0(sh);
1554         int faila = -1, failb = -1;
1555         int target = sh->ops.target;
1556         int target2 = sh->ops.target2;
1557         struct r5dev *tgt = &sh->dev[target];
1558         struct r5dev *tgt2 = &sh->dev[target2];
1559         struct dma_async_tx_descriptor *tx;
1560         struct page **blocks = to_addr_page(percpu, 0);
1561         struct async_submit_ctl submit;
1562
1563         BUG_ON(sh->batch_head);
1564         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1565                  __func__, (unsigned long long)sh->sector, target, target2);
1566         BUG_ON(target < 0 || target2 < 0);
1567         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1568         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1569
1570         /* we need to open-code set_syndrome_sources to handle the
1571          * slot number conversion for 'faila' and 'failb'
1572          */
1573         for (i = 0; i < disks ; i++)
1574                 blocks[i] = NULL;
1575         count = 0;
1576         i = d0_idx;
1577         do {
1578                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1579
1580                 blocks[slot] = sh->dev[i].page;
1581
1582                 if (i == target)
1583                         faila = slot;
1584                 if (i == target2)
1585                         failb = slot;
1586                 i = raid6_next_disk(i, disks);
1587         } while (i != d0_idx);
1588
1589         BUG_ON(faila == failb);
1590         if (failb < faila)
1591                 swap(faila, failb);
1592         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1593                  __func__, (unsigned long long)sh->sector, faila, failb);
1594
1595         atomic_inc(&sh->count);
1596
1597         if (failb == syndrome_disks+1) {
1598                 /* Q disk is one of the missing disks */
1599                 if (faila == syndrome_disks) {
1600                         /* Missing P+Q, just recompute */
1601                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1602                                           ops_complete_compute, sh,
1603                                           to_addr_conv(sh, percpu, 0));
1604                         return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1605                                                   STRIPE_SIZE, &submit);
1606                 } else {
1607                         struct page *dest;
1608                         int data_target;
1609                         int qd_idx = sh->qd_idx;
1610
1611                         /* Missing D+Q: recompute D from P, then recompute Q */
1612                         if (target == qd_idx)
1613                                 data_target = target2;
1614                         else
1615                                 data_target = target;
1616
1617                         count = 0;
1618                         for (i = disks; i-- ; ) {
1619                                 if (i == data_target || i == qd_idx)
1620                                         continue;
1621                                 blocks[count++] = sh->dev[i].page;
1622                         }
1623                         dest = sh->dev[data_target].page;
1624                         init_async_submit(&submit,
1625                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1626                                           NULL, NULL, NULL,
1627                                           to_addr_conv(sh, percpu, 0));
1628                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1629                                        &submit);
1630
1631                         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1632                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1633                                           ops_complete_compute, sh,
1634                                           to_addr_conv(sh, percpu, 0));
1635                         return async_gen_syndrome(blocks, 0, count+2,
1636                                                   STRIPE_SIZE, &submit);
1637                 }
1638         } else {
1639                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1640                                   ops_complete_compute, sh,
1641                                   to_addr_conv(sh, percpu, 0));
1642                 if (failb == syndrome_disks) {
1643                         /* We're missing D+P. */
1644                         return async_raid6_datap_recov(syndrome_disks+2,
1645                                                        STRIPE_SIZE, faila,
1646                                                        blocks, &submit);
1647                 } else {
1648                         /* We're missing D+D. */
1649                         return async_raid6_2data_recov(syndrome_disks+2,
1650                                                        STRIPE_SIZE, faila, failb,
1651                                                        blocks, &submit);
1652                 }
1653         }
1654 }
1655
1656 static void ops_complete_prexor(void *stripe_head_ref)
1657 {
1658         struct stripe_head *sh = stripe_head_ref;
1659
1660         pr_debug("%s: stripe %llu\n", __func__,
1661                 (unsigned long long)sh->sector);
1662
1663         if (r5c_is_writeback(sh->raid_conf->log))
1664                 /*
1665                  * raid5-cache write back uses orig_page during prexor.
1666                  * After prexor, it is time to free orig_page
1667                  */
1668                 r5c_release_extra_page(sh);
1669 }
1670
1671 static struct dma_async_tx_descriptor *
1672 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1673                 struct dma_async_tx_descriptor *tx)
1674 {
1675         int disks = sh->disks;
1676         struct page **xor_srcs = to_addr_page(percpu, 0);
1677         int count = 0, pd_idx = sh->pd_idx, i;
1678         struct async_submit_ctl submit;
1679
1680         /* existing parity data subtracted */
1681         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1682
1683         BUG_ON(sh->batch_head);
1684         pr_debug("%s: stripe %llu\n", __func__,
1685                 (unsigned long long)sh->sector);
1686
1687         for (i = disks; i--; ) {
1688                 struct r5dev *dev = &sh->dev[i];
1689                 /* Only process blocks that are known to be uptodate */
1690                 if (test_bit(R5_InJournal, &dev->flags))
1691                         xor_srcs[count++] = dev->orig_page;
1692                 else if (test_bit(R5_Wantdrain, &dev->flags))
1693                         xor_srcs[count++] = dev->page;
1694         }
1695
1696         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1697                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1698         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1699
1700         return tx;
1701 }
1702
1703 static struct dma_async_tx_descriptor *
1704 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1705                 struct dma_async_tx_descriptor *tx)
1706 {
1707         struct page **blocks = to_addr_page(percpu, 0);
1708         int count;
1709         struct async_submit_ctl submit;
1710
1711         pr_debug("%s: stripe %llu\n", __func__,
1712                 (unsigned long long)sh->sector);
1713
1714         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1715
1716         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1717                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1718         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1719
1720         return tx;
1721 }
1722
1723 static struct dma_async_tx_descriptor *
1724 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1725 {
1726         struct r5conf *conf = sh->raid_conf;
1727         int disks = sh->disks;
1728         int i;
1729         struct stripe_head *head_sh = sh;
1730
1731         pr_debug("%s: stripe %llu\n", __func__,
1732                 (unsigned long long)sh->sector);
1733
1734         for (i = disks; i--; ) {
1735                 struct r5dev *dev;
1736                 struct bio *chosen;
1737
1738                 sh = head_sh;
1739                 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1740                         struct bio *wbi;
1741
1742 again:
1743                         dev = &sh->dev[i];
1744                         /*
1745                          * clear R5_InJournal, so when rewriting a page in
1746                          * journal, it is not skipped by r5l_log_stripe()
1747                          */
1748                         clear_bit(R5_InJournal, &dev->flags);
1749                         spin_lock_irq(&sh->stripe_lock);
1750                         chosen = dev->towrite;
1751                         dev->towrite = NULL;
1752                         sh->overwrite_disks = 0;
1753                         BUG_ON(dev->written);
1754                         wbi = dev->written = chosen;
1755                         spin_unlock_irq(&sh->stripe_lock);
1756                         WARN_ON(dev->page != dev->orig_page);
1757
1758                         while (wbi && wbi->bi_iter.bi_sector <
1759                                 dev->sector + STRIPE_SECTORS) {
1760                                 if (wbi->bi_opf & REQ_FUA)
1761                                         set_bit(R5_WantFUA, &dev->flags);
1762                                 if (wbi->bi_opf & REQ_SYNC)
1763                                         set_bit(R5_SyncIO, &dev->flags);
1764                                 if (bio_op(wbi) == REQ_OP_DISCARD)
1765                                         set_bit(R5_Discard, &dev->flags);
1766                                 else {
1767                                         tx = async_copy_data(1, wbi, &dev->page,
1768                                                              dev->sector, tx, sh,
1769                                                              r5c_is_writeback(conf->log));
1770                                         if (dev->page != dev->orig_page &&
1771                                             !r5c_is_writeback(conf->log)) {
1772                                                 set_bit(R5_SkipCopy, &dev->flags);
1773                                                 clear_bit(R5_UPTODATE, &dev->flags);
1774                                                 clear_bit(R5_OVERWRITE, &dev->flags);
1775                                         }
1776                                 }
1777                                 wbi = r5_next_bio(wbi, dev->sector);
1778                         }
1779
1780                         if (head_sh->batch_head) {
1781                                 sh = list_first_entry(&sh->batch_list,
1782                                                       struct stripe_head,
1783                                                       batch_list);
1784                                 if (sh == head_sh)
1785                                         continue;
1786                                 goto again;
1787                         }
1788                 }
1789         }
1790
1791         return tx;
1792 }
1793
1794 static void ops_complete_reconstruct(void *stripe_head_ref)
1795 {
1796         struct stripe_head *sh = stripe_head_ref;
1797         int disks = sh->disks;
1798         int pd_idx = sh->pd_idx;
1799         int qd_idx = sh->qd_idx;
1800         int i;
1801         bool fua = false, sync = false, discard = false;
1802
1803         pr_debug("%s: stripe %llu\n", __func__,
1804                 (unsigned long long)sh->sector);
1805
1806         for (i = disks; i--; ) {
1807                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1808                 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1809                 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1810         }
1811
1812         for (i = disks; i--; ) {
1813                 struct r5dev *dev = &sh->dev[i];
1814
1815                 if (dev->written || i == pd_idx || i == qd_idx) {
1816                         if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
1817                                 set_bit(R5_UPTODATE, &dev->flags);
1818                                 if (test_bit(STRIPE_EXPAND_READY, &sh->state))
1819                                         set_bit(R5_Expanded, &dev->flags);
1820                         }
1821                         if (fua)
1822                                 set_bit(R5_WantFUA, &dev->flags);
1823                         if (sync)
1824                                 set_bit(R5_SyncIO, &dev->flags);
1825                 }
1826         }
1827
1828         if (sh->reconstruct_state == reconstruct_state_drain_run)
1829                 sh->reconstruct_state = reconstruct_state_drain_result;
1830         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1831                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1832         else {
1833                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1834                 sh->reconstruct_state = reconstruct_state_result;
1835         }
1836
1837         set_bit(STRIPE_HANDLE, &sh->state);
1838         raid5_release_stripe(sh);
1839 }
1840
1841 static void
1842 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1843                      struct dma_async_tx_descriptor *tx)
1844 {
1845         int disks = sh->disks;
1846         struct page **xor_srcs;
1847         struct async_submit_ctl submit;
1848         int count, pd_idx = sh->pd_idx, i;
1849         struct page *xor_dest;
1850         int prexor = 0;
1851         unsigned long flags;
1852         int j = 0;
1853         struct stripe_head *head_sh = sh;
1854         int last_stripe;
1855
1856         pr_debug("%s: stripe %llu\n", __func__,
1857                 (unsigned long long)sh->sector);
1858
1859         for (i = 0; i < sh->disks; i++) {
1860                 if (pd_idx == i)
1861                         continue;
1862                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1863                         break;
1864         }
1865         if (i >= sh->disks) {
1866                 atomic_inc(&sh->count);
1867                 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1868                 ops_complete_reconstruct(sh);
1869                 return;
1870         }
1871 again:
1872         count = 0;
1873         xor_srcs = to_addr_page(percpu, j);
1874         /* check if prexor is active which means only process blocks
1875          * that are part of a read-modify-write (written)
1876          */
1877         if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1878                 prexor = 1;
1879                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1880                 for (i = disks; i--; ) {
1881                         struct r5dev *dev = &sh->dev[i];
1882                         if (head_sh->dev[i].written ||
1883                             test_bit(R5_InJournal, &head_sh->dev[i].flags))
1884                                 xor_srcs[count++] = dev->page;
1885                 }
1886         } else {
1887                 xor_dest = sh->dev[pd_idx].page;
1888                 for (i = disks; i--; ) {
1889                         struct r5dev *dev = &sh->dev[i];
1890                         if (i != pd_idx)
1891                                 xor_srcs[count++] = dev->page;
1892                 }
1893         }
1894
1895         /* 1/ if we prexor'd then the dest is reused as a source
1896          * 2/ if we did not prexor then we are redoing the parity
1897          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1898          * for the synchronous xor case
1899          */
1900         last_stripe = !head_sh->batch_head ||
1901                 list_first_entry(&sh->batch_list,
1902                                  struct stripe_head, batch_list) == head_sh;
1903         if (last_stripe) {
1904                 flags = ASYNC_TX_ACK |
1905                         (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1906
1907                 atomic_inc(&head_sh->count);
1908                 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1909                                   to_addr_conv(sh, percpu, j));
1910         } else {
1911                 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1912                 init_async_submit(&submit, flags, tx, NULL, NULL,
1913                                   to_addr_conv(sh, percpu, j));
1914         }
1915
1916         if (unlikely(count == 1))
1917                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1918         else
1919                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1920         if (!last_stripe) {
1921                 j++;
1922                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1923                                       batch_list);
1924                 goto again;
1925         }
1926 }
1927
1928 static void
1929 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1930                      struct dma_async_tx_descriptor *tx)
1931 {
1932         struct async_submit_ctl submit;
1933         struct page **blocks;
1934         int count, i, j = 0;
1935         struct stripe_head *head_sh = sh;
1936         int last_stripe;
1937         int synflags;
1938         unsigned long txflags;
1939
1940         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1941
1942         for (i = 0; i < sh->disks; i++) {
1943                 if (sh->pd_idx == i || sh->qd_idx == i)
1944                         continue;
1945                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1946                         break;
1947         }
1948         if (i >= sh->disks) {
1949                 atomic_inc(&sh->count);
1950                 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1951                 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1952                 ops_complete_reconstruct(sh);
1953                 return;
1954         }
1955
1956 again:
1957         blocks = to_addr_page(percpu, j);
1958
1959         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1960                 synflags = SYNDROME_SRC_WRITTEN;
1961                 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1962         } else {
1963                 synflags = SYNDROME_SRC_ALL;
1964                 txflags = ASYNC_TX_ACK;
1965         }
1966
1967         count = set_syndrome_sources(blocks, sh, synflags);
1968         last_stripe = !head_sh->batch_head ||
1969                 list_first_entry(&sh->batch_list,
1970                                  struct stripe_head, batch_list) == head_sh;
1971
1972         if (last_stripe) {
1973                 atomic_inc(&head_sh->count);
1974                 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1975                                   head_sh, to_addr_conv(sh, percpu, j));
1976         } else
1977                 init_async_submit(&submit, 0, tx, NULL, NULL,
1978                                   to_addr_conv(sh, percpu, j));
1979         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1980         if (!last_stripe) {
1981                 j++;
1982                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1983                                       batch_list);
1984                 goto again;
1985         }
1986 }
1987
1988 static void ops_complete_check(void *stripe_head_ref)
1989 {
1990         struct stripe_head *sh = stripe_head_ref;
1991
1992         pr_debug("%s: stripe %llu\n", __func__,
1993                 (unsigned long long)sh->sector);
1994
1995         sh->check_state = check_state_check_result;
1996         set_bit(STRIPE_HANDLE, &sh->state);
1997         raid5_release_stripe(sh);
1998 }
1999
2000 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
2001 {
2002         int disks = sh->disks;
2003         int pd_idx = sh->pd_idx;
2004         int qd_idx = sh->qd_idx;
2005         struct page *xor_dest;
2006         struct page **xor_srcs = to_addr_page(percpu, 0);
2007         struct dma_async_tx_descriptor *tx;
2008         struct async_submit_ctl submit;
2009         int count;
2010         int i;
2011
2012         pr_debug("%s: stripe %llu\n", __func__,
2013                 (unsigned long long)sh->sector);
2014
2015         BUG_ON(sh->batch_head);
2016         count = 0;
2017         xor_dest = sh->dev[pd_idx].page;
2018         xor_srcs[count++] = xor_dest;
2019         for (i = disks; i--; ) {
2020                 if (i == pd_idx || i == qd_idx)
2021                         continue;
2022                 xor_srcs[count++] = sh->dev[i].page;
2023         }
2024
2025         init_async_submit(&submit, 0, NULL, NULL, NULL,
2026                           to_addr_conv(sh, percpu, 0));
2027         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
2028                            &sh->ops.zero_sum_result, &submit);
2029
2030         atomic_inc(&sh->count);
2031         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2032         tx = async_trigger_callback(&submit);
2033 }
2034
2035 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2036 {
2037         struct page **srcs = to_addr_page(percpu, 0);
2038         struct async_submit_ctl submit;
2039         int count;
2040
2041         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2042                 (unsigned long long)sh->sector, checkp);
2043
2044         BUG_ON(sh->batch_head);
2045         count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
2046         if (!checkp)
2047                 srcs[count] = NULL;
2048
2049         atomic_inc(&sh->count);
2050         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2051                           sh, to_addr_conv(sh, percpu, 0));
2052         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
2053                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
2054 }
2055
2056 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2057 {
2058         int overlap_clear = 0, i, disks = sh->disks;
2059         struct dma_async_tx_descriptor *tx = NULL;
2060         struct r5conf *conf = sh->raid_conf;
2061         int level = conf->level;
2062         struct raid5_percpu *percpu;
2063         unsigned long cpu;
2064
2065         cpu = get_cpu();
2066         percpu = per_cpu_ptr(conf->percpu, cpu);
2067         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2068                 ops_run_biofill(sh);
2069                 overlap_clear++;
2070         }
2071
2072         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2073                 if (level < 6)
2074                         tx = ops_run_compute5(sh, percpu);
2075                 else {
2076                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
2077                                 tx = ops_run_compute6_1(sh, percpu);
2078                         else
2079                                 tx = ops_run_compute6_2(sh, percpu);
2080                 }
2081                 /* terminate the chain if reconstruct is not set to be run */
2082                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2083                         async_tx_ack(tx);
2084         }
2085
2086         if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2087                 if (level < 6)
2088                         tx = ops_run_prexor5(sh, percpu, tx);
2089                 else
2090                         tx = ops_run_prexor6(sh, percpu, tx);
2091         }
2092
2093         if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2094                 tx = ops_run_partial_parity(sh, percpu, tx);
2095
2096         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2097                 tx = ops_run_biodrain(sh, tx);
2098                 overlap_clear++;
2099         }
2100
2101         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2102                 if (level < 6)
2103                         ops_run_reconstruct5(sh, percpu, tx);
2104                 else
2105                         ops_run_reconstruct6(sh, percpu, tx);
2106         }
2107
2108         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2109                 if (sh->check_state == check_state_run)
2110                         ops_run_check_p(sh, percpu);
2111                 else if (sh->check_state == check_state_run_q)
2112                         ops_run_check_pq(sh, percpu, 0);
2113                 else if (sh->check_state == check_state_run_pq)
2114                         ops_run_check_pq(sh, percpu, 1);
2115                 else
2116                         BUG();
2117         }
2118
2119         if (overlap_clear && !sh->batch_head)
2120                 for (i = disks; i--; ) {
2121                         struct r5dev *dev = &sh->dev[i];
2122                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2123                                 wake_up(&sh->raid_conf->wait_for_overlap);
2124                 }
2125         put_cpu();
2126 }
2127
2128 static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2129 {
2130         if (sh->ppl_page)
2131                 __free_page(sh->ppl_page);
2132         kmem_cache_free(sc, sh);
2133 }
2134
2135 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2136         int disks, struct r5conf *conf)
2137 {
2138         struct stripe_head *sh;
2139         int i;
2140
2141         sh = kmem_cache_zalloc(sc, gfp);
2142         if (sh) {
2143                 spin_lock_init(&sh->stripe_lock);
2144                 spin_lock_init(&sh->batch_lock);
2145                 INIT_LIST_HEAD(&sh->batch_list);
2146                 INIT_LIST_HEAD(&sh->lru);
2147                 INIT_LIST_HEAD(&sh->r5c);
2148                 INIT_LIST_HEAD(&sh->log_list);
2149                 atomic_set(&sh->count, 1);
2150                 sh->raid_conf = conf;
2151                 sh->log_start = MaxSector;
2152                 for (i = 0; i < disks; i++) {
2153                         struct r5dev *dev = &sh->dev[i];
2154
2155                         bio_init(&dev->req, &dev->vec, 1);
2156                         bio_init(&dev->rreq, &dev->rvec, 1);
2157                 }
2158
2159                 if (raid5_has_ppl(conf)) {
2160                         sh->ppl_page = alloc_page(gfp);
2161                         if (!sh->ppl_page) {
2162                                 free_stripe(sc, sh);
2163                                 sh = NULL;
2164                         }
2165                 }
2166         }
2167         return sh;
2168 }
2169 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2170 {
2171         struct stripe_head *sh;
2172
2173         sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2174         if (!sh)
2175                 return 0;
2176
2177         if (grow_buffers(sh, gfp)) {
2178                 shrink_buffers(sh);
2179                 free_stripe(conf->slab_cache, sh);
2180                 return 0;
2181         }
2182         sh->hash_lock_index =
2183                 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2184         /* we just created an active stripe so... */
2185         atomic_inc(&conf->active_stripes);
2186
2187         raid5_release_stripe(sh);
2188         conf->max_nr_stripes++;
2189         return 1;
2190 }
2191
2192 static int grow_stripes(struct r5conf *conf, int num)
2193 {
2194         struct kmem_cache *sc;
2195         size_t namelen = sizeof(conf->cache_name[0]);
2196         int devs = max(conf->raid_disks, conf->previous_raid_disks);
2197
2198         if (conf->mddev->gendisk)
2199                 snprintf(conf->cache_name[0], namelen,
2200                         "raid%d-%s", conf->level, mdname(conf->mddev));
2201         else
2202                 snprintf(conf->cache_name[0], namelen,
2203                         "raid%d-%p", conf->level, conf->mddev);
2204         snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2205
2206         conf->active_name = 0;
2207         sc = kmem_cache_create(conf->cache_name[conf->active_name],
2208                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2209                                0, 0, NULL);
2210         if (!sc)
2211                 return 1;
2212         conf->slab_cache = sc;
2213         conf->pool_size = devs;
2214         while (num--)
2215                 if (!grow_one_stripe(conf, GFP_KERNEL))
2216                         return 1;
2217
2218         return 0;
2219 }
2220
2221 /**
2222  * scribble_len - return the required size of the scribble region
2223  * @num - total number of disks in the array
2224  *
2225  * The size must be enough to contain:
2226  * 1/ a struct page pointer for each device in the array +2
2227  * 2/ room to convert each entry in (1) to its corresponding dma
2228  *    (dma_map_page()) or page (page_address()) address.
2229  *
2230  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2231  * calculate over all devices (not just the data blocks), using zeros in place
2232  * of the P and Q blocks.
2233  */
2234 static int scribble_alloc(struct raid5_percpu *percpu,
2235                           int num, int cnt, gfp_t flags)
2236 {
2237         size_t obj_size =
2238                 sizeof(struct page *) * (num+2) +
2239                 sizeof(addr_conv_t) * (num+2);
2240         void *scribble;
2241
2242         scribble = kvmalloc_array(cnt, obj_size, flags);
2243         if (!scribble)
2244                 return -ENOMEM;
2245
2246         kvfree(percpu->scribble);
2247
2248         percpu->scribble = scribble;
2249         percpu->scribble_obj_size = obj_size;
2250         return 0;
2251 }
2252
2253 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2254 {
2255         unsigned long cpu;
2256         int err = 0;
2257
2258         /*
2259          * Never shrink. And mddev_suspend() could deadlock if this is called
2260          * from raid5d. In that case, scribble_disks and scribble_sectors
2261          * should equal to new_disks and new_sectors
2262          */
2263         if (conf->scribble_disks >= new_disks &&
2264             conf->scribble_sectors >= new_sectors)
2265                 return 0;
2266         mddev_suspend(conf->mddev);
2267         get_online_cpus();
2268
2269         for_each_present_cpu(cpu) {
2270                 struct raid5_percpu *percpu;
2271
2272                 percpu = per_cpu_ptr(conf->percpu, cpu);
2273                 err = scribble_alloc(percpu, new_disks,
2274                                      new_sectors / STRIPE_SECTORS,
2275                                      GFP_NOIO);
2276                 if (err)
2277                         break;
2278         }
2279
2280         put_online_cpus();
2281         mddev_resume(conf->mddev);
2282         if (!err) {
2283                 conf->scribble_disks = new_disks;
2284                 conf->scribble_sectors = new_sectors;
2285         }
2286         return err;
2287 }
2288
2289 static int resize_stripes(struct r5conf *conf, int newsize)
2290 {
2291         /* Make all the stripes able to hold 'newsize' devices.
2292          * New slots in each stripe get 'page' set to a new page.
2293          *
2294          * This happens in stages:
2295          * 1/ create a new kmem_cache and allocate the required number of
2296          *    stripe_heads.
2297          * 2/ gather all the old stripe_heads and transfer the pages across
2298          *    to the new stripe_heads.  This will have the side effect of
2299          *    freezing the array as once all stripe_heads have been collected,
2300          *    no IO will be possible.  Old stripe heads are freed once their
2301          *    pages have been transferred over, and the old kmem_cache is
2302          *    freed when all stripes are done.
2303          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2304          *    we simple return a failure status - no need to clean anything up.
2305          * 4/ allocate new pages for the new slots in the new stripe_heads.
2306          *    If this fails, we don't bother trying the shrink the
2307          *    stripe_heads down again, we just leave them as they are.
2308          *    As each stripe_head is processed the new one is released into
2309          *    active service.
2310          *
2311          * Once step2 is started, we cannot afford to wait for a write,
2312          * so we use GFP_NOIO allocations.
2313          */
2314         struct stripe_head *osh, *nsh;
2315         LIST_HEAD(newstripes);
2316         struct disk_info *ndisks;
2317         int err = 0;
2318         struct kmem_cache *sc;
2319         int i;
2320         int hash, cnt;
2321
2322         md_allow_write(conf->mddev);
2323
2324         /* Step 1 */
2325         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2326                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2327                                0, 0, NULL);
2328         if (!sc)
2329                 return -ENOMEM;
2330
2331         /* Need to ensure auto-resizing doesn't interfere */
2332         mutex_lock(&conf->cache_size_mutex);
2333
2334         for (i = conf->max_nr_stripes; i; i--) {
2335                 nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2336                 if (!nsh)
2337                         break;
2338
2339                 list_add(&nsh->lru, &newstripes);
2340         }
2341         if (i) {
2342                 /* didn't get enough, give up */
2343                 while (!list_empty(&newstripes)) {
2344                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
2345                         list_del(&nsh->lru);
2346                         free_stripe(sc, nsh);
2347                 }
2348                 kmem_cache_destroy(sc);
2349                 mutex_unlock(&conf->cache_size_mutex);
2350                 return -ENOMEM;
2351         }
2352         /* Step 2 - Must use GFP_NOIO now.
2353          * OK, we have enough stripes, start collecting inactive
2354          * stripes and copying them over
2355          */
2356         hash = 0;
2357         cnt = 0;
2358         list_for_each_entry(nsh, &newstripes, lru) {
2359                 lock_device_hash_lock(conf, hash);
2360                 wait_event_cmd(conf->wait_for_stripe,
2361                                     !list_empty(conf->inactive_list + hash),
2362                                     unlock_device_hash_lock(conf, hash),
2363                                     lock_device_hash_lock(conf, hash));
2364                 osh = get_free_stripe(conf, hash);
2365                 unlock_device_hash_lock(conf, hash);
2366
2367                 for(i=0; i<conf->pool_size; i++) {
2368                         nsh->dev[i].page = osh->dev[i].page;
2369                         nsh->dev[i].orig_page = osh->dev[i].page;
2370                 }
2371                 nsh->hash_lock_index = hash;
2372                 free_stripe(conf->slab_cache, osh);
2373                 cnt++;
2374                 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2375                     !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2376                         hash++;
2377                         cnt = 0;
2378                 }
2379         }
2380         kmem_cache_destroy(conf->slab_cache);
2381
2382         /* Step 3.
2383          * At this point, we are holding all the stripes so the array
2384          * is completely stalled, so now is a good time to resize
2385          * conf->disks and the scribble region
2386          */
2387         ndisks = kcalloc(newsize, sizeof(struct disk_info), GFP_NOIO);
2388         if (ndisks) {
2389                 for (i = 0; i < conf->pool_size; i++)
2390                         ndisks[i] = conf->disks[i];
2391
2392                 for (i = conf->pool_size; i < newsize; i++) {
2393                         ndisks[i].extra_page = alloc_page(GFP_NOIO);
2394                         if (!ndisks[i].extra_page)
2395                                 err = -ENOMEM;
2396                 }
2397
2398                 if (err) {
2399                         for (i = conf->pool_size; i < newsize; i++)
2400                                 if (ndisks[i].extra_page)
2401                                         put_page(ndisks[i].extra_page);
2402                         kfree(ndisks);
2403                 } else {
2404                         kfree(conf->disks);
2405                         conf->disks = ndisks;
2406                 }
2407         } else
2408                 err = -ENOMEM;
2409
2410         mutex_unlock(&conf->cache_size_mutex);
2411
2412         conf->slab_cache = sc;
2413         conf->active_name = 1-conf->active_name;
2414
2415         /* Step 4, return new stripes to service */
2416         while(!list_empty(&newstripes)) {
2417                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2418                 list_del_init(&nsh->lru);
2419
2420                 for (i=conf->raid_disks; i < newsize; i++)
2421                         if (nsh->dev[i].page == NULL) {
2422                                 struct page *p = alloc_page(GFP_NOIO);
2423                                 nsh->dev[i].page = p;
2424                                 nsh->dev[i].orig_page = p;
2425                                 if (!p)
2426                                         err = -ENOMEM;
2427                         }
2428                 raid5_release_stripe(nsh);
2429         }
2430         /* critical section pass, GFP_NOIO no longer needed */
2431
2432         if (!err)
2433                 conf->pool_size = newsize;
2434         return err;
2435 }
2436
2437 static int drop_one_stripe(struct r5conf *conf)
2438 {
2439         struct stripe_head *sh;
2440         int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2441
2442         spin_lock_irq(conf->hash_locks + hash);
2443         sh = get_free_stripe(conf, hash);
2444         spin_unlock_irq(conf->hash_locks + hash);
2445         if (!sh)
2446                 return 0;
2447         BUG_ON(atomic_read(&sh->count));
2448         shrink_buffers(sh);
2449         free_stripe(conf->slab_cache, sh);
2450         atomic_dec(&conf->active_stripes);
2451         conf->max_nr_stripes--;
2452         return 1;
2453 }
2454
2455 static void shrink_stripes(struct r5conf *conf)
2456 {
2457         while (conf->max_nr_stripes &&
2458                drop_one_stripe(conf))
2459                 ;
2460
2461         kmem_cache_destroy(conf->slab_cache);
2462         conf->slab_cache = NULL;
2463 }
2464
2465 static void raid5_end_read_request(struct bio * bi)
2466 {
2467         struct stripe_head *sh = bi->bi_private;
2468         struct r5conf *conf = sh->raid_conf;
2469         int disks = sh->disks, i;
2470         char b[BDEVNAME_SIZE];
2471         struct md_rdev *rdev = NULL;
2472         sector_t s;
2473
2474         for (i=0 ; i<disks; i++)
2475                 if (bi == &sh->dev[i].req)
2476                         break;
2477
2478         pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2479                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2480                 bi->bi_status);
2481         if (i == disks) {
2482                 bio_reset(bi);
2483                 BUG();
2484                 return;
2485         }
2486         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2487                 /* If replacement finished while this request was outstanding,
2488                  * 'replacement' might be NULL already.
2489                  * In that case it moved down to 'rdev'.
2490                  * rdev is not removed until all requests are finished.
2491                  */
2492                 rdev = conf->disks[i].replacement;
2493         if (!rdev)
2494                 rdev = conf->disks[i].rdev;
2495
2496         if (use_new_offset(conf, sh))
2497                 s = sh->sector + rdev->new_data_offset;
2498         else
2499                 s = sh->sector + rdev->data_offset;
2500         if (!bi->bi_status) {
2501                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2502                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2503                         /* Note that this cannot happen on a
2504                          * replacement device.  We just fail those on
2505                          * any error
2506                          */
2507                         pr_info_ratelimited(
2508                                 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2509                                 mdname(conf->mddev), STRIPE_SECTORS,
2510                                 (unsigned long long)s,
2511                                 bdevname(rdev->bdev, b));
2512                         atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2513                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2514                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2515                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2516                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2517
2518                 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2519                         /*
2520                          * end read for a page in journal, this
2521                          * must be preparing for prexor in rmw
2522                          */
2523                         set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2524
2525                 if (atomic_read(&rdev->read_errors))
2526                         atomic_set(&rdev->read_errors, 0);
2527         } else {
2528                 const char *bdn = bdevname(rdev->bdev, b);
2529                 int retry = 0;
2530                 int set_bad = 0;
2531
2532                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2533                 atomic_inc(&rdev->read_errors);
2534                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2535                         pr_warn_ratelimited(
2536                                 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2537                                 mdname(conf->mddev),
2538                                 (unsigned long long)s,
2539                                 bdn);
2540                 else if (conf->mddev->degraded >= conf->max_degraded) {
2541                         set_bad = 1;
2542                         pr_warn_ratelimited(
2543                                 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2544                                 mdname(conf->mddev),
2545                                 (unsigned long long)s,
2546                                 bdn);
2547                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2548                         /* Oh, no!!! */
2549                         set_bad = 1;
2550                         pr_warn_ratelimited(
2551                                 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2552                                 mdname(conf->mddev),
2553                                 (unsigned long long)s,
2554                                 bdn);
2555                 } else if (atomic_read(&rdev->read_errors)
2556                          > conf->max_nr_stripes)
2557                         pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2558                                mdname(conf->mddev), bdn);
2559                 else
2560                         retry = 1;
2561                 if (set_bad && test_bit(In_sync, &rdev->flags)
2562                     && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2563                         retry = 1;
2564                 if (retry)
2565                         if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2566                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2567                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2568                         } else
2569                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2570                 else {
2571                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2572                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2573                         if (!(set_bad
2574                               && test_bit(In_sync, &rdev->flags)
2575                               && rdev_set_badblocks(
2576                                       rdev, sh->sector, STRIPE_SECTORS, 0)))
2577                                 md_error(conf->mddev, rdev);
2578                 }
2579         }
2580         rdev_dec_pending(rdev, conf->mddev);
2581         bio_reset(bi);
2582         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2583         set_bit(STRIPE_HANDLE, &sh->state);
2584         raid5_release_stripe(sh);
2585 }
2586
2587 static void raid5_end_write_request(struct bio *bi)
2588 {
2589         struct stripe_head *sh = bi->bi_private;
2590         struct r5conf *conf = sh->raid_conf;
2591         int disks = sh->disks, i;
2592         struct md_rdev *uninitialized_var(rdev);
2593         sector_t first_bad;
2594         int bad_sectors;
2595         int replacement = 0;
2596
2597         for (i = 0 ; i < disks; i++) {
2598                 if (bi == &sh->dev[i].req) {
2599                         rdev = conf->disks[i].rdev;
2600                         break;
2601                 }
2602                 if (bi == &sh->dev[i].rreq) {
2603                         rdev = conf->disks[i].replacement;
2604                         if (rdev)
2605                                 replacement = 1;
2606                         else
2607                                 /* rdev was removed and 'replacement'
2608                                  * replaced it.  rdev is not removed
2609                                  * until all requests are finished.
2610                                  */
2611                                 rdev = conf->disks[i].rdev;
2612                         break;
2613                 }
2614         }
2615         pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2616                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2617                 bi->bi_status);
2618         if (i == disks) {
2619                 bio_reset(bi);
2620                 BUG();
2621                 return;
2622         }
2623
2624         if (replacement) {
2625                 if (bi->bi_status)
2626                         md_error(conf->mddev, rdev);
2627                 else if (is_badblock(rdev, sh->sector,
2628                                      STRIPE_SECTORS,
2629                                      &first_bad, &bad_sectors))
2630                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2631         } else {
2632                 if (bi->bi_status) {
2633                         set_bit(STRIPE_DEGRADED, &sh->state);
2634                         set_bit(WriteErrorSeen, &rdev->flags);
2635                         set_bit(R5_WriteError, &sh->dev[i].flags);
2636                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2637                                 set_bit(MD_RECOVERY_NEEDED,
2638                                         &rdev->mddev->recovery);
2639                 } else if (is_badblock(rdev, sh->sector,
2640                                        STRIPE_SECTORS,
2641                                        &first_bad, &bad_sectors)) {
2642                         set_bit(R5_MadeGood, &sh->dev[i].flags);
2643                         if (test_bit(R5_ReadError, &sh->dev[i].flags))
2644                                 /* That was a successful write so make
2645                                  * sure it looks like we already did
2646                                  * a re-write.
2647                                  */
2648                                 set_bit(R5_ReWrite, &sh->dev[i].flags);
2649                 }
2650         }
2651         rdev_dec_pending(rdev, conf->mddev);
2652
2653         if (sh->batch_head && bi->bi_status && !replacement)
2654                 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2655
2656         bio_reset(bi);
2657         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2658                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2659         set_bit(STRIPE_HANDLE, &sh->state);
2660         raid5_release_stripe(sh);
2661
2662         if (sh->batch_head && sh != sh->batch_head)
2663                 raid5_release_stripe(sh->batch_head);
2664 }
2665
2666 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2667 {
2668         char b[BDEVNAME_SIZE];
2669         struct r5conf *conf = mddev->private;
2670         unsigned long flags;
2671         pr_debug("raid456: error called\n");
2672
2673         spin_lock_irqsave(&conf->device_lock, flags);
2674
2675         if (test_bit(In_sync, &rdev->flags) &&
2676             mddev->degraded == conf->max_degraded) {
2677                 /*
2678                  * Don't allow to achieve failed state
2679                  * Don't try to recover this device
2680                  */
2681                 conf->recovery_disabled = mddev->recovery_disabled;
2682                 spin_unlock_irqrestore(&conf->device_lock, flags);
2683                 return;
2684         }
2685
2686         set_bit(Faulty, &rdev->flags);
2687         clear_bit(In_sync, &rdev->flags);
2688         mddev->degraded = raid5_calc_degraded(conf);
2689         spin_unlock_irqrestore(&conf->device_lock, flags);
2690         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2691
2692         set_bit(Blocked, &rdev->flags);
2693         set_mask_bits(&mddev->sb_flags, 0,
2694                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2695         pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2696                 "md/raid:%s: Operation continuing on %d devices.\n",
2697                 mdname(mddev),
2698                 bdevname(rdev->bdev, b),
2699                 mdname(mddev),
2700                 conf->raid_disks - mddev->degraded);
2701         r5c_update_on_rdev_error(mddev, rdev);
2702 }
2703
2704 /*
2705  * Input: a 'big' sector number,
2706  * Output: index of the data and parity disk, and the sector # in them.
2707  */
2708 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2709                               int previous, int *dd_idx,
2710                               struct stripe_head *sh)
2711 {
2712         sector_t stripe, stripe2;
2713         sector_t chunk_number;
2714         unsigned int chunk_offset;
2715         int pd_idx, qd_idx;
2716         int ddf_layout = 0;
2717         sector_t new_sector;
2718         int algorithm = previous ? conf->prev_algo
2719                                  : conf->algorithm;
2720         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2721                                          : conf->chunk_sectors;
2722         int raid_disks = previous ? conf->previous_raid_disks
2723                                   : conf->raid_disks;
2724         int data_disks = raid_disks - conf->max_degraded;
2725
2726         /* First compute the information on this sector */
2727
2728         /*
2729          * Compute the chunk number and the sector offset inside the chunk
2730          */
2731         chunk_offset = sector_div(r_sector, sectors_per_chunk);
2732         chunk_number = r_sector;
2733
2734         /*
2735          * Compute the stripe number
2736          */
2737         stripe = chunk_number;
2738         *dd_idx = sector_div(stripe, data_disks);
2739         stripe2 = stripe;
2740         /*
2741          * Select the parity disk based on the user selected algorithm.
2742          */
2743         pd_idx = qd_idx = -1;
2744         switch(conf->level) {
2745         case 4:
2746                 pd_idx = data_disks;
2747                 break;
2748         case 5:
2749                 switch (algorithm) {
2750                 case ALGORITHM_LEFT_ASYMMETRIC:
2751                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2752                         if (*dd_idx >= pd_idx)
2753                                 (*dd_idx)++;
2754                         break;
2755                 case ALGORITHM_RIGHT_ASYMMETRIC:
2756                         pd_idx = sector_div(stripe2, raid_disks);
2757                         if (*dd_idx >= pd_idx)
2758                                 (*dd_idx)++;
2759                         break;
2760                 case ALGORITHM_LEFT_SYMMETRIC:
2761                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2762                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2763                         break;
2764                 case ALGORITHM_RIGHT_SYMMETRIC:
2765                         pd_idx = sector_div(stripe2, raid_disks);
2766                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2767                         break;
2768                 case ALGORITHM_PARITY_0:
2769                         pd_idx = 0;
2770                         (*dd_idx)++;
2771                         break;
2772                 case ALGORITHM_PARITY_N:
2773                         pd_idx = data_disks;
2774                         break;
2775                 default:
2776                         BUG();
2777                 }
2778                 break;
2779         case 6:
2780
2781                 switch (algorithm) {
2782                 case ALGORITHM_LEFT_ASYMMETRIC:
2783                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2784                         qd_idx = pd_idx + 1;
2785                         if (pd_idx == raid_disks-1) {
2786                                 (*dd_idx)++;    /* Q D D D P */
2787                                 qd_idx = 0;
2788                         } else if (*dd_idx >= pd_idx)
2789                                 (*dd_idx) += 2; /* D D P Q D */
2790                         break;
2791                 case ALGORITHM_RIGHT_ASYMMETRIC:
2792                         pd_idx = sector_div(stripe2, raid_disks);
2793                         qd_idx = pd_idx + 1;
2794                         if (pd_idx == raid_disks-1) {
2795                                 (*dd_idx)++;    /* Q D D D P */
2796                                 qd_idx = 0;
2797                         } else if (*dd_idx >= pd_idx)
2798                                 (*dd_idx) += 2; /* D D P Q D */
2799                         break;
2800                 case ALGORITHM_LEFT_SYMMETRIC:
2801                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2802                         qd_idx = (pd_idx + 1) % raid_disks;
2803                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2804                         break;
2805                 case ALGORITHM_RIGHT_SYMMETRIC:
2806                         pd_idx = sector_div(stripe2, raid_disks);
2807                         qd_idx = (pd_idx + 1) % raid_disks;
2808                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2809                         break;
2810
2811                 case ALGORITHM_PARITY_0:
2812                         pd_idx = 0;
2813                         qd_idx = 1;
2814                         (*dd_idx) += 2;
2815                         break;
2816                 case ALGORITHM_PARITY_N:
2817                         pd_idx = data_disks;
2818                         qd_idx = data_disks + 1;
2819                         break;
2820
2821                 case ALGORITHM_ROTATING_ZERO_RESTART:
2822                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
2823                          * of blocks for computing Q is different.
2824                          */
2825                         pd_idx = sector_div(stripe2, raid_disks);
2826                         qd_idx = pd_idx + 1;
2827                         if (pd_idx == raid_disks-1) {
2828                                 (*dd_idx)++;    /* Q D D D P */
2829                                 qd_idx = 0;
2830                         } else if (*dd_idx >= pd_idx)
2831                                 (*dd_idx) += 2; /* D D P Q D */
2832                         ddf_layout = 1;
2833                         break;
2834
2835                 case ALGORITHM_ROTATING_N_RESTART:
2836                         /* Same a left_asymmetric, by first stripe is
2837                          * D D D P Q  rather than
2838                          * Q D D D P
2839                          */
2840                         stripe2 += 1;
2841                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2842                         qd_idx = pd_idx + 1;
2843                         if (pd_idx == raid_disks-1) {
2844                                 (*dd_idx)++;    /* Q D D D P */
2845                                 qd_idx = 0;
2846                         } else if (*dd_idx >= pd_idx)
2847                                 (*dd_idx) += 2; /* D D P Q D */
2848                         ddf_layout = 1;
2849                         break;
2850
2851                 case ALGORITHM_ROTATING_N_CONTINUE:
2852                         /* Same as left_symmetric but Q is before P */
2853                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2854                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2855                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2856                         ddf_layout = 1;
2857                         break;
2858
2859                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2860                         /* RAID5 left_asymmetric, with Q on last device */
2861                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2862                         if (*dd_idx >= pd_idx)
2863                                 (*dd_idx)++;
2864                         qd_idx = raid_disks - 1;
2865                         break;
2866
2867                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2868                         pd_idx = sector_div(stripe2, raid_disks-1);
2869                         if (*dd_idx >= pd_idx)
2870                                 (*dd_idx)++;
2871                         qd_idx = raid_disks - 1;
2872                         break;
2873
2874                 case ALGORITHM_LEFT_SYMMETRIC_6:
2875                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2876                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2877                         qd_idx = raid_disks - 1;
2878                         break;
2879
2880                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2881                         pd_idx = sector_div(stripe2, raid_disks-1);
2882                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2883                         qd_idx = raid_disks - 1;
2884                         break;
2885
2886                 case ALGORITHM_PARITY_0_6:
2887                         pd_idx = 0;
2888                         (*dd_idx)++;
2889                         qd_idx = raid_disks - 1;
2890                         break;
2891
2892                 default:
2893                         BUG();
2894                 }
2895                 break;
2896         }
2897
2898         if (sh) {
2899                 sh->pd_idx = pd_idx;
2900                 sh->qd_idx = qd_idx;
2901                 sh->ddf_layout = ddf_layout;
2902         }
2903         /*
2904          * Finally, compute the new sector number
2905          */
2906         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2907         return new_sector;
2908 }
2909
2910 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2911 {
2912         struct r5conf *conf = sh->raid_conf;
2913         int raid_disks = sh->disks;
2914         int data_disks = raid_disks - conf->max_degraded;
2915         sector_t new_sector = sh->sector, check;
2916         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2917                                          : conf->chunk_sectors;
2918         int algorithm = previous ? conf->prev_algo
2919                                  : conf->algorithm;
2920         sector_t stripe;
2921         int chunk_offset;
2922         sector_t chunk_number;
2923         int dummy1, dd_idx = i;
2924         sector_t r_sector;
2925         struct stripe_head sh2;
2926
2927         chunk_offset = sector_div(new_sector, sectors_per_chunk);
2928         stripe = new_sector;
2929
2930         if (i == sh->pd_idx)
2931                 return 0;
2932         switch(conf->level) {
2933         case 4: break;
2934         case 5:
2935                 switch (algorithm) {
2936                 case ALGORITHM_LEFT_ASYMMETRIC:
2937                 case ALGORITHM_RIGHT_ASYMMETRIC:
2938                         if (i > sh->pd_idx)
2939                                 i--;
2940                         break;
2941                 case ALGORITHM_LEFT_SYMMETRIC:
2942                 case ALGORITHM_RIGHT_SYMMETRIC:
2943                         if (i < sh->pd_idx)
2944                                 i += raid_disks;
2945                         i -= (sh->pd_idx + 1);
2946                         break;
2947                 case ALGORITHM_PARITY_0:
2948                         i -= 1;
2949                         break;
2950                 case ALGORITHM_PARITY_N:
2951                         break;
2952                 default:
2953                         BUG();
2954                 }
2955                 break;
2956         case 6:
2957                 if (i == sh->qd_idx)
2958                         return 0; /* It is the Q disk */
2959                 switch (algorithm) {
2960                 case ALGORITHM_LEFT_ASYMMETRIC:
2961                 case ALGORITHM_RIGHT_ASYMMETRIC:
2962                 case ALGORITHM_ROTATING_ZERO_RESTART:
2963                 case ALGORITHM_ROTATING_N_RESTART:
2964                         if (sh->pd_idx == raid_disks-1)
2965                                 i--;    /* Q D D D P */
2966                         else if (i > sh->pd_idx)
2967                                 i -= 2; /* D D P Q D */
2968                         break;
2969                 case ALGORITHM_LEFT_SYMMETRIC:
2970                 case ALGORITHM_RIGHT_SYMMETRIC:
2971                         if (sh->pd_idx == raid_disks-1)
2972                                 i--; /* Q D D D P */
2973                         else {
2974                                 /* D D P Q D */
2975                                 if (i < sh->pd_idx)
2976                                         i += raid_disks;
2977                                 i -= (sh->pd_idx + 2);
2978                         }
2979                         break;
2980                 case ALGORITHM_PARITY_0:
2981                         i -= 2;
2982                         break;
2983                 case ALGORITHM_PARITY_N:
2984                         break;
2985                 case ALGORITHM_ROTATING_N_CONTINUE:
2986                         /* Like left_symmetric, but P is before Q */
2987                         if (sh->pd_idx == 0)
2988                                 i--;    /* P D D D Q */
2989                         else {
2990                                 /* D D Q P D */
2991                                 if (i < sh->pd_idx)
2992                                         i += raid_disks;
2993                                 i -= (sh->pd_idx + 1);
2994                         }
2995                         break;
2996                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2997                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2998                         if (i > sh->pd_idx)
2999                                 i--;
3000                         break;
3001                 case ALGORITHM_LEFT_SYMMETRIC_6:
3002                 case ALGORITHM_RIGHT_SYMMETRIC_6:
3003                         if (i < sh->pd_idx)
3004                                 i += data_disks + 1;
3005                         i -= (sh->pd_idx + 1);
3006                         break;
3007                 case ALGORITHM_PARITY_0_6:
3008                         i -= 1;
3009                         break;
3010                 default:
3011                         BUG();
3012                 }
3013                 break;
3014         }
3015
3016         chunk_number = stripe * data_disks + i;
3017         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3018
3019         check = raid5_compute_sector(conf, r_sector,
3020                                      previous, &dummy1, &sh2);
3021         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3022                 || sh2.qd_idx != sh->qd_idx) {
3023                 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3024                         mdname(conf->mddev));
3025                 return 0;
3026         }
3027         return r_sector;
3028 }
3029
3030 /*
3031  * There are cases where we want handle_stripe_dirtying() and
3032  * schedule_reconstruction() to delay towrite to some dev of a stripe.
3033  *
3034  * This function checks whether we want to delay the towrite. Specifically,
3035  * we delay the towrite when:
3036  *
3037  *   1. degraded stripe has a non-overwrite to the missing dev, AND this
3038  *      stripe has data in journal (for other devices).
3039  *
3040  *      In this case, when reading data for the non-overwrite dev, it is
3041  *      necessary to handle complex rmw of write back cache (prexor with
3042  *      orig_page, and xor with page). To keep read path simple, we would
3043  *      like to flush data in journal to RAID disks first, so complex rmw
3044  *      is handled in the write patch (handle_stripe_dirtying).
3045  *
3046  *   2. when journal space is critical (R5C_LOG_CRITICAL=1)
3047  *
3048  *      It is important to be able to flush all stripes in raid5-cache.
3049  *      Therefore, we need reserve some space on the journal device for
3050  *      these flushes. If flush operation includes pending writes to the
3051  *      stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3052  *      for the flush out. If we exclude these pending writes from flush
3053  *      operation, we only need (conf->max_degraded + 1) pages per stripe.
3054  *      Therefore, excluding pending writes in these cases enables more
3055  *      efficient use of the journal device.
3056  *
3057  *      Note: To make sure the stripe makes progress, we only delay
3058  *      towrite for stripes with data already in journal (injournal > 0).
3059  *      When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3060  *      no_space_stripes list.
3061  *
3062  *   3. during journal failure
3063  *      In journal failure, we try to flush all cached data to raid disks
3064  *      based on data in stripe cache. The array is read-only to upper
3065  *      layers, so we would skip all pending writes.
3066  *
3067  */
3068 static inline bool delay_towrite(struct r5conf *conf,
3069                                  struct r5dev *dev,
3070                                  struct stripe_head_state *s)
3071 {
3072         /* case 1 above */
3073         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3074             !test_bit(R5_Insync, &dev->flags) && s->injournal)
3075                 return true;
3076         /* case 2 above */
3077         if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3078             s->injournal > 0)
3079                 return true;
3080         /* case 3 above */
3081         if (s->log_failed && s->injournal)
3082                 return true;
3083         return false;
3084 }
3085
3086 static void
3087 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3088                          int rcw, int expand)
3089 {
3090         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3091         struct r5conf *conf = sh->raid_conf;
3092         int level = conf->level;
3093
3094         if (rcw) {
3095                 /*
3096                  * In some cases, handle_stripe_dirtying initially decided to
3097                  * run rmw and allocates extra page for prexor. However, rcw is
3098                  * cheaper later on. We need to free the extra page now,
3099                  * because we won't be able to do that in ops_complete_prexor().
3100                  */
3101                 r5c_release_extra_page(sh);
3102
3103                 for (i = disks; i--; ) {
3104                         struct r5dev *dev = &sh->dev[i];
3105
3106                         if (dev->towrite && !delay_towrite(conf, dev, s)) {
3107                                 set_bit(R5_LOCKED, &dev->flags);
3108                                 set_bit(R5_Wantdrain, &dev->flags);
3109                                 if (!expand)
3110                                         clear_bit(R5_UPTODATE, &dev->flags);
3111                                 s->locked++;
3112                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3113                                 set_bit(R5_LOCKED, &dev->flags);
3114                                 s->locked++;
3115                         }
3116                 }
3117                 /* if we are not expanding this is a proper write request, and
3118                  * there will be bios with new data to be drained into the
3119                  * stripe cache
3120                  */
3121                 if (!expand) {
3122                         if (!s->locked)
3123                                 /* False alarm, nothing to do */
3124                                 return;
3125                         sh->reconstruct_state = reconstruct_state_drain_run;
3126                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3127                 } else
3128                         sh->reconstruct_state = reconstruct_state_run;
3129
3130                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3131
3132                 if (s->locked + conf->max_degraded == disks)
3133                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3134                                 atomic_inc(&conf->pending_full_writes);
3135         } else {
3136                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3137                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3138                 BUG_ON(level == 6 &&
3139                         (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3140                            test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3141
3142                 for (i = disks; i--; ) {
3143                         struct r5dev *dev = &sh->dev[i];
3144                         if (i == pd_idx || i == qd_idx)
3145                                 continue;
3146
3147                         if (dev->towrite &&
3148                             (test_bit(R5_UPTODATE, &dev->flags) ||
3149                              test_bit(R5_Wantcompute, &dev->flags))) {
3150                                 set_bit(R5_Wantdrain, &dev->flags);
3151                                 set_bit(R5_LOCKED, &dev->flags);
3152                                 clear_bit(R5_UPTODATE, &dev->flags);
3153                                 s->locked++;
3154                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3155                                 set_bit(R5_LOCKED, &dev->flags);
3156                                 s->locked++;
3157                         }
3158                 }
3159                 if (!s->locked)
3160                         /* False alarm - nothing to do */
3161                         return;
3162                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3163                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3164                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3165                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3166         }
3167
3168         /* keep the parity disk(s) locked while asynchronous operations
3169          * are in flight
3170          */
3171         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3172         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3173         s->locked++;
3174
3175         if (level == 6) {
3176                 int qd_idx = sh->qd_idx;
3177                 struct r5dev *dev = &sh->dev[qd_idx];
3178
3179                 set_bit(R5_LOCKED, &dev->flags);
3180                 clear_bit(R5_UPTODATE, &dev->flags);
3181                 s->locked++;
3182         }
3183
3184         if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3185             test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3186             !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3187             test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3188                 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3189
3190         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3191                 __func__, (unsigned long long)sh->sector,
3192                 s->locked, s->ops_request);
3193 }
3194
3195 /*
3196  * Each stripe/dev can have one or more bion attached.
3197  * toread/towrite point to the first in a chain.
3198  * The bi_next chain must be in order.
3199  */
3200 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3201                           int forwrite, int previous)
3202 {
3203         struct bio **bip;
3204         struct r5conf *conf = sh->raid_conf;
3205         int firstwrite=0;
3206
3207         pr_debug("adding bi b#%llu to stripe s#%llu\n",
3208                 (unsigned long long)bi->bi_iter.bi_sector,
3209                 (unsigned long long)sh->sector);
3210
3211         spin_lock_irq(&sh->stripe_lock);
3212         sh->dev[dd_idx].write_hint = bi->bi_write_hint;
3213         /* Don't allow new IO added to stripes in batch list */
3214         if (sh->batch_head)
3215                 goto overlap;
3216         if (forwrite) {
3217                 bip = &sh->dev[dd_idx].towrite;
3218                 if (*bip == NULL)
3219                         firstwrite = 1;
3220         } else
3221                 bip = &sh->dev[dd_idx].toread;
3222         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3223                 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3224                         goto overlap;
3225                 bip = & (*bip)->bi_next;
3226         }
3227         if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3228                 goto overlap;
3229
3230         if (forwrite && raid5_has_ppl(conf)) {
3231                 /*
3232                  * With PPL only writes to consecutive data chunks within a
3233                  * stripe are allowed because for a single stripe_head we can
3234                  * only have one PPL entry at a time, which describes one data
3235                  * range. Not really an overlap, but wait_for_overlap can be
3236                  * used to handle this.
3237                  */
3238                 sector_t sector;
3239                 sector_t first = 0;
3240                 sector_t last = 0;
3241                 int count = 0;
3242                 int i;
3243
3244                 for (i = 0; i < sh->disks; i++) {
3245                         if (i != sh->pd_idx &&
3246                             (i == dd_idx || sh->dev[i].towrite)) {
3247                                 sector = sh->dev[i].sector;
3248                                 if (count == 0 || sector < first)
3249                                         first = sector;
3250                                 if (sector > last)
3251                                         last = sector;
3252                                 count++;
3253                         }
3254                 }
3255
3256                 if (first + conf->chunk_sectors * (count - 1) != last)
3257                         goto overlap;
3258         }
3259
3260         if (!forwrite || previous)
3261                 clear_bit(STRIPE_BATCH_READY, &sh->state);
3262
3263         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3264         if (*bip)
3265                 bi->bi_next = *bip;
3266         *bip = bi;
3267         bio_inc_remaining(bi);
3268         md_write_inc(conf->mddev, bi);
3269
3270         if (forwrite) {
3271                 /* check if page is covered */
3272                 sector_t sector = sh->dev[dd_idx].sector;
3273                 for (bi=sh->dev[dd_idx].towrite;
3274                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3275                              bi && bi->bi_iter.bi_sector <= sector;
3276                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3277                         if (bio_end_sector(bi) >= sector)
3278                                 sector = bio_end_sector(bi);
3279                 }
3280                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3281                         if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3282                                 sh->overwrite_disks++;
3283         }
3284
3285         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3286                 (unsigned long long)(*bip)->bi_iter.bi_sector,
3287                 (unsigned long long)sh->sector, dd_idx);
3288
3289         if (conf->mddev->bitmap && firstwrite) {
3290                 /* Cannot hold spinlock over bitmap_startwrite,
3291                  * but must ensure this isn't added to a batch until
3292                  * we have added to the bitmap and set bm_seq.
3293                  * So set STRIPE_BITMAP_PENDING to prevent
3294                  * batching.
3295                  * If multiple add_stripe_bio() calls race here they
3296                  * much all set STRIPE_BITMAP_PENDING.  So only the first one
3297                  * to complete "bitmap_startwrite" gets to set
3298                  * STRIPE_BIT_DELAY.  This is important as once a stripe
3299                  * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3300                  * any more.
3301                  */
3302                 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3303                 spin_unlock_irq(&sh->stripe_lock);
3304                 md_bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3305                                      STRIPE_SECTORS, 0);
3306                 spin_lock_irq(&sh->stripe_lock);
3307                 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3308                 if (!sh->batch_head) {
3309                         sh->bm_seq = conf->seq_flush+1;
3310                         set_bit(STRIPE_BIT_DELAY, &sh->state);
3311                 }
3312         }
3313         spin_unlock_irq(&sh->stripe_lock);
3314
3315         if (stripe_can_batch(sh))
3316                 stripe_add_to_batch_list(conf, sh);
3317         return 1;
3318
3319  overlap:
3320         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3321         spin_unlock_irq(&sh->stripe_lock);
3322         return 0;
3323 }
3324
3325 static void end_reshape(struct r5conf *conf);
3326
3327 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3328                             struct stripe_head *sh)
3329 {
3330         int sectors_per_chunk =
3331                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3332         int dd_idx;
3333         int chunk_offset = sector_div(stripe, sectors_per_chunk);
3334         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3335
3336         raid5_compute_sector(conf,
3337                              stripe * (disks - conf->max_degraded)
3338                              *sectors_per_chunk + chunk_offset,
3339                              previous,
3340                              &dd_idx, sh);
3341 }
3342
3343 static void
3344 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3345                      struct stripe_head_state *s, int disks)
3346 {
3347         int i;
3348         BUG_ON(sh->batch_head);
3349         for (i = disks; i--; ) {
3350                 struct bio *bi;
3351                 int bitmap_end = 0;
3352
3353                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3354                         struct md_rdev *rdev;
3355                         rcu_read_lock();
3356                         rdev = rcu_dereference(conf->disks[i].rdev);
3357                         if (rdev && test_bit(In_sync, &rdev->flags) &&
3358                             !test_bit(Faulty, &rdev->flags))
3359                                 atomic_inc(&rdev->nr_pending);
3360                         else
3361                                 rdev = NULL;
3362                         rcu_read_unlock();
3363                         if (rdev) {
3364                                 if (!rdev_set_badblocks(
3365                                             rdev,
3366                                             sh->sector,
3367                                             STRIPE_SECTORS, 0))
3368                                         md_error(conf->mddev, rdev);
3369                                 rdev_dec_pending(rdev, conf->mddev);
3370                         }
3371                 }
3372                 spin_lock_irq(&sh->stripe_lock);
3373                 /* fail all writes first */
3374                 bi = sh->dev[i].towrite;
3375                 sh->dev[i].towrite = NULL;
3376                 sh->overwrite_disks = 0;
3377                 spin_unlock_irq(&sh->stripe_lock);
3378                 if (bi)
3379                         bitmap_end = 1;
3380
3381                 log_stripe_write_finished(sh);
3382
3383                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3384                         wake_up(&conf->wait_for_overlap);
3385
3386                 while (bi && bi->bi_iter.bi_sector <
3387                         sh->dev[i].sector + STRIPE_SECTORS) {
3388                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3389
3390                         md_write_end(conf->mddev);
3391                         bio_io_error(bi);
3392                         bi = nextbi;
3393                 }
3394                 if (bitmap_end)
3395                         md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3396                                            STRIPE_SECTORS, 0, 0);
3397                 bitmap_end = 0;
3398                 /* and fail all 'written' */
3399                 bi = sh->dev[i].written;
3400                 sh->dev[i].written = NULL;
3401                 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3402                         WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3403                         sh->dev[i].page = sh->dev[i].orig_page;
3404                 }
3405
3406                 if (bi) bitmap_end = 1;
3407                 while (bi && bi->bi_iter.bi_sector <
3408                        sh->dev[i].sector + STRIPE_SECTORS) {
3409                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3410
3411                         md_write_end(conf->mddev);
3412                         bio_io_error(bi);
3413                         bi = bi2;
3414                 }
3415
3416                 /* fail any reads if this device is non-operational and
3417                  * the data has not reached the cache yet.
3418                  */
3419                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3420                     s->failed > conf->max_degraded &&
3421                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3422                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
3423                         spin_lock_irq(&sh->stripe_lock);
3424                         bi = sh->dev[i].toread;
3425                         sh->dev[i].toread = NULL;
3426                         spin_unlock_irq(&sh->stripe_lock);
3427                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3428                                 wake_up(&conf->wait_for_overlap);
3429                         if (bi)
3430                                 s->to_read--;
3431                         while (bi && bi->bi_iter.bi_sector <
3432                                sh->dev[i].sector + STRIPE_SECTORS) {
3433                                 struct bio *nextbi =
3434                                         r5_next_bio(bi, sh->dev[i].sector);
3435
3436                                 bio_io_error(bi);
3437                                 bi = nextbi;
3438                         }
3439                 }
3440                 if (bitmap_end)
3441                         md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3442                                            STRIPE_SECTORS, 0, 0);
3443                 /* If we were in the middle of a write the parity block might
3444                  * still be locked - so just clear all R5_LOCKED flags
3445                  */
3446                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3447         }
3448         s->to_write = 0;
3449         s->written = 0;
3450
3451         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3452                 if (atomic_dec_and_test(&conf->pending_full_writes))
3453                         md_wakeup_thread(conf->mddev->thread);
3454 }
3455
3456 static void
3457 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3458                    struct stripe_head_state *s)
3459 {
3460         int abort = 0;
3461         int i;
3462
3463         BUG_ON(sh->batch_head);
3464         clear_bit(STRIPE_SYNCING, &sh->state);
3465         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3466                 wake_up(&conf->wait_for_overlap);
3467         s->syncing = 0;
3468         s->replacing = 0;
3469         /* There is nothing more to do for sync/check/repair.
3470          * Don't even need to abort as that is handled elsewhere
3471          * if needed, and not always wanted e.g. if there is a known
3472          * bad block here.
3473          * For recover/replace we need to record a bad block on all
3474          * non-sync devices, or abort the recovery
3475          */
3476         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3477                 /* During recovery devices cannot be removed, so
3478                  * locking and refcounting of rdevs is not needed
3479                  */
3480                 rcu_read_lock();
3481                 for (i = 0; i < conf->raid_disks; i++) {
3482                         struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3483                         if (rdev
3484                             && !test_bit(Faulty, &rdev->flags)
3485                             && !test_bit(In_sync, &rdev->flags)
3486                             && !rdev_set_badblocks(rdev, sh->sector,
3487                                                    STRIPE_SECTORS, 0))
3488                                 abort = 1;
3489                         rdev = rcu_dereference(conf->disks[i].replacement);
3490                         if (rdev
3491                             && !test_bit(Faulty, &rdev->flags)
3492                             && !test_bit(In_sync, &rdev->flags)
3493                             && !rdev_set_badblocks(rdev, sh->sector,
3494                                                    STRIPE_SECTORS, 0))
3495                                 abort = 1;
3496                 }
3497                 rcu_read_unlock();
3498                 if (abort)
3499                         conf->recovery_disabled =
3500                                 conf->mddev->recovery_disabled;
3501         }
3502         md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3503 }
3504
3505 static int want_replace(struct stripe_head *sh, int disk_idx)
3506 {
3507         struct md_rdev *rdev;
3508         int rv = 0;
3509
3510         rcu_read_lock();
3511         rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3512         if (rdev
3513             && !test_bit(Faulty, &rdev->flags)
3514             && !test_bit(In_sync, &rdev->flags)
3515             && (rdev->recovery_offset <= sh->sector
3516                 || rdev->mddev->recovery_cp <= sh->sector))
3517                 rv = 1;
3518         rcu_read_unlock();
3519         return rv;
3520 }
3521
3522 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3523                            int disk_idx, int disks)
3524 {
3525         struct r5dev *dev = &sh->dev[disk_idx];
3526         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3527                                   &sh->dev[s->failed_num[1]] };
3528         int i;
3529
3530
3531         if (test_bit(R5_LOCKED, &dev->flags) ||
3532             test_bit(R5_UPTODATE, &dev->flags))
3533                 /* No point reading this as we already have it or have
3534                  * decided to get it.
3535                  */
3536                 return 0;
3537
3538         if (dev->toread ||
3539             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3540                 /* We need this block to directly satisfy a request */
3541                 return 1;
3542
3543         if (s->syncing || s->expanding ||
3544             (s->replacing && want_replace(sh, disk_idx)))
3545                 /* When syncing, or expanding we read everything.
3546                  * When replacing, we need the replaced block.
3547                  */
3548                 return 1;
3549
3550         if ((s->failed >= 1 && fdev[0]->toread) ||
3551             (s->failed >= 2 && fdev[1]->toread))
3552                 /* If we want to read from a failed device, then
3553                  * we need to actually read every other device.
3554                  */
3555                 return 1;
3556
3557         /* Sometimes neither read-modify-write nor reconstruct-write
3558          * cycles can work.  In those cases we read every block we
3559          * can.  Then the parity-update is certain to have enough to
3560          * work with.
3561          * This can only be a problem when we need to write something,
3562          * and some device has failed.  If either of those tests
3563          * fail we need look no further.
3564          */
3565         if (!s->failed || !s->to_write)
3566                 return 0;
3567
3568         if (test_bit(R5_Insync, &dev->flags) &&
3569             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3570                 /* Pre-reads at not permitted until after short delay
3571                  * to gather multiple requests.  However if this
3572                  * device is no Insync, the block could only be computed
3573                  * and there is no need to delay that.
3574                  */
3575                 return 0;
3576
3577         for (i = 0; i < s->failed && i < 2; i++) {
3578                 if (fdev[i]->towrite &&
3579                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3580                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3581                         /* If we have a partial write to a failed
3582                          * device, then we will need to reconstruct
3583                          * the content of that device, so all other
3584                          * devices must be read.
3585                          */
3586                         return 1;
3587         }
3588
3589         /* If we are forced to do a reconstruct-write, either because
3590          * the current RAID6 implementation only supports that, or
3591          * because parity cannot be trusted and we are currently
3592          * recovering it, there is extra need to be careful.
3593          * If one of the devices that we would need to read, because
3594          * it is not being overwritten (and maybe not written at all)
3595          * is missing/faulty, then we need to read everything we can.
3596          */
3597         if (sh->raid_conf->level != 6 &&
3598             sh->sector < sh->raid_conf->mddev->recovery_cp)
3599                 /* reconstruct-write isn't being forced */
3600                 return 0;
3601         for (i = 0; i < s->failed && i < 2; i++) {
3602                 if (s->failed_num[i] != sh->pd_idx &&
3603                     s->failed_num[i] != sh->qd_idx &&
3604                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3605                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3606                         return 1;
3607         }
3608
3609         return 0;
3610 }
3611
3612 /* fetch_block - checks the given member device to see if its data needs
3613  * to be read or computed to satisfy a request.
3614  *
3615  * Returns 1 when no more member devices need to be checked, otherwise returns
3616  * 0 to tell the loop in handle_stripe_fill to continue
3617  */
3618 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3619                        int disk_idx, int disks)
3620 {
3621         struct r5dev *dev = &sh->dev[disk_idx];
3622
3623         /* is the data in this block needed, and can we get it? */
3624         if (need_this_block(sh, s, disk_idx, disks)) {
3625                 /* we would like to get this block, possibly by computing it,
3626                  * otherwise read it if the backing disk is insync
3627                  */
3628                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3629                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3630                 BUG_ON(sh->batch_head);
3631
3632                 /*
3633                  * In the raid6 case if the only non-uptodate disk is P
3634                  * then we already trusted P to compute the other failed
3635                  * drives. It is safe to compute rather than re-read P.
3636                  * In other cases we only compute blocks from failed
3637                  * devices, otherwise check/repair might fail to detect
3638                  * a real inconsistency.
3639                  */
3640
3641                 if ((s->uptodate == disks - 1) &&
3642                     ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3643                     (s->failed && (disk_idx == s->failed_num[0] ||
3644                                    disk_idx == s->failed_num[1])))) {
3645                         /* have disk failed, and we're requested to fetch it;
3646                          * do compute it
3647                          */
3648                         pr_debug("Computing stripe %llu block %d\n",
3649                                (unsigned long long)sh->sector, disk_idx);
3650                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3651                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3652                         set_bit(R5_Wantcompute, &dev->flags);
3653                         sh->ops.target = disk_idx;
3654                         sh->ops.target2 = -1; /* no 2nd target */
3655                         s->req_compute = 1;
3656                         /* Careful: from this point on 'uptodate' is in the eye
3657                          * of raid_run_ops which services 'compute' operations
3658                          * before writes. R5_Wantcompute flags a block that will
3659                          * be R5_UPTODATE by the time it is needed for a
3660                          * subsequent operation.
3661                          */
3662                         s->uptodate++;
3663                         return 1;
3664                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3665                         /* Computing 2-failure is *very* expensive; only
3666                          * do it if failed >= 2
3667                          */
3668                         int other;
3669                         for (other = disks; other--; ) {
3670                                 if (other == disk_idx)
3671                                         continue;
3672                                 if (!test_bit(R5_UPTODATE,
3673                                       &sh->dev[other].flags))
3674                                         break;
3675                         }
3676                         BUG_ON(other < 0);
3677                         pr_debug("Computing stripe %llu blocks %d,%d\n",
3678                                (unsigned long long)sh->sector,
3679                                disk_idx, other);
3680                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3681                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3682                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3683                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
3684                         sh->ops.target = disk_idx;
3685                         sh->ops.target2 = other;
3686                         s->uptodate += 2;
3687                         s->req_compute = 1;
3688                         return 1;
3689                 } else if (test_bit(R5_Insync, &dev->flags)) {
3690                         set_bit(R5_LOCKED, &dev->flags);
3691                         set_bit(R5_Wantread, &dev->flags);
3692                         s->locked++;
3693                         pr_debug("Reading block %d (sync=%d)\n",
3694                                 disk_idx, s->syncing);
3695                 }
3696         }
3697
3698         return 0;
3699 }
3700
3701 /**
3702  * handle_stripe_fill - read or compute data to satisfy pending requests.
3703  */
3704 static void handle_stripe_fill(struct stripe_head *sh,
3705                                struct stripe_head_state *s,
3706                                int disks)
3707 {
3708         int i;
3709
3710         /* look for blocks to read/compute, skip this if a compute
3711          * is already in flight, or if the stripe contents are in the
3712          * midst of changing due to a write
3713          */
3714         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3715             !sh->reconstruct_state) {
3716
3717                 /*
3718                  * For degraded stripe with data in journal, do not handle
3719                  * read requests yet, instead, flush the stripe to raid
3720                  * disks first, this avoids handling complex rmw of write
3721                  * back cache (prexor with orig_page, and then xor with
3722                  * page) in the read path
3723                  */
3724                 if (s->injournal && s->failed) {
3725                         if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3726                                 r5c_make_stripe_write_out(sh);
3727                         goto out;
3728                 }
3729
3730                 for (i = disks; i--; )
3731                         if (fetch_block(sh, s, i, disks))
3732                                 break;
3733         }
3734 out:
3735         set_bit(STRIPE_HANDLE, &sh->state);
3736 }
3737
3738 static void break_stripe_batch_list(struct stripe_head *head_sh,
3739                                     unsigned long handle_flags);
3740 /* handle_stripe_clean_event
3741  * any written block on an uptodate or failed drive can be returned.
3742  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3743  * never LOCKED, so we don't need to test 'failed' directly.
3744  */
3745 static void handle_stripe_clean_event(struct r5conf *conf,
3746         struct stripe_head *sh, int disks)
3747 {
3748         int i;
3749         struct r5dev *dev;
3750         int discard_pending = 0;
3751         struct stripe_head *head_sh = sh;
3752         bool do_endio = false;
3753
3754         for (i = disks; i--; )
3755                 if (sh->dev[i].written) {
3756                         dev = &sh->dev[i];
3757                         if (!test_bit(R5_LOCKED, &dev->flags) &&
3758                             (test_bit(R5_UPTODATE, &dev->flags) ||
3759                              test_bit(R5_Discard, &dev->flags) ||
3760                              test_bit(R5_SkipCopy, &dev->flags))) {
3761                                 /* We can return any write requests */
3762                                 struct bio *wbi, *wbi2;
3763                                 pr_debug("Return write for disc %d\n", i);
3764                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
3765                                         clear_bit(R5_UPTODATE, &dev->flags);
3766                                 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3767                                         WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3768                                 }
3769                                 do_endio = true;
3770
3771 returnbi:
3772                                 dev->page = dev->orig_page;
3773                                 wbi = dev->written;
3774                                 dev->written = NULL;
3775                                 while (wbi && wbi->bi_iter.bi_sector <
3776                                         dev->sector + STRIPE_SECTORS) {
3777                                         wbi2 = r5_next_bio(wbi, dev->sector);
3778                                         md_write_end(conf->mddev);
3779                                         bio_endio(wbi);
3780                                         wbi = wbi2;
3781                                 }
3782                                 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3783                                                    STRIPE_SECTORS,
3784                                                    !test_bit(STRIPE_DEGRADED, &sh->state),
3785                                                    0);
3786                                 if (head_sh->batch_head) {
3787                                         sh = list_first_entry(&sh->batch_list,
3788                                                               struct stripe_head,
3789                                                               batch_list);
3790                                         if (sh != head_sh) {
3791                                                 dev = &sh->dev[i];
3792                                                 goto returnbi;
3793                                         }
3794                                 }
3795                                 sh = head_sh;
3796                                 dev = &sh->dev[i];
3797                         } else if (test_bit(R5_Discard, &dev->flags))
3798                                 discard_pending = 1;
3799                 }
3800
3801         log_stripe_write_finished(sh);
3802
3803         if (!discard_pending &&
3804             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3805                 int hash;
3806                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3807                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3808                 if (sh->qd_idx >= 0) {
3809                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3810                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3811                 }
3812                 /* now that discard is done we can proceed with any sync */
3813                 clear_bit(STRIPE_DISCARD, &sh->state);
3814                 /*
3815                  * SCSI discard will change some bio fields and the stripe has
3816                  * no updated data, so remove it from hash list and the stripe
3817                  * will be reinitialized
3818                  */
3819 unhash:
3820                 hash = sh->hash_lock_index;
3821                 spin_lock_irq(conf->hash_locks + hash);
3822                 remove_hash(sh);
3823                 spin_unlock_irq(conf->hash_locks + hash);
3824                 if (head_sh->batch_head) {
3825                         sh = list_first_entry(&sh->batch_list,
3826                                               struct stripe_head, batch_list);
3827                         if (sh != head_sh)
3828                                         goto unhash;
3829                 }
3830                 sh = head_sh;
3831
3832                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3833                         set_bit(STRIPE_HANDLE, &sh->state);
3834
3835         }
3836
3837         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3838                 if (atomic_dec_and_test(&conf->pending_full_writes))
3839                         md_wakeup_thread(conf->mddev->thread);
3840
3841         if (head_sh->batch_head && do_endio)
3842                 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3843 }
3844
3845 /*
3846  * For RMW in write back cache, we need extra page in prexor to store the
3847  * old data. This page is stored in dev->orig_page.
3848  *
3849  * This function checks whether we have data for prexor. The exact logic
3850  * is:
3851  *       R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3852  */
3853 static inline bool uptodate_for_rmw(struct r5dev *dev)
3854 {
3855         return (test_bit(R5_UPTODATE, &dev->flags)) &&
3856                 (!test_bit(R5_InJournal, &dev->flags) ||
3857                  test_bit(R5_OrigPageUPTDODATE, &dev->flags));
3858 }
3859
3860 static int handle_stripe_dirtying(struct r5conf *conf,
3861                                   struct stripe_head *sh,
3862                                   struct stripe_head_state *s,
3863                                   int disks)
3864 {
3865         int rmw = 0, rcw = 0, i;
3866         sector_t recovery_cp = conf->mddev->recovery_cp;
3867
3868         /* Check whether resync is now happening or should start.
3869          * If yes, then the array is dirty (after unclean shutdown or
3870          * initial creation), so parity in some stripes might be inconsistent.
3871          * In this case, we need to always do reconstruct-write, to ensure
3872          * that in case of drive failure or read-error correction, we
3873          * generate correct data from the parity.
3874          */
3875         if (conf->rmw_level == PARITY_DISABLE_RMW ||
3876             (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3877              s->failed == 0)) {
3878                 /* Calculate the real rcw later - for now make it
3879                  * look like rcw is cheaper
3880                  */
3881                 rcw = 1; rmw = 2;
3882                 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3883                          conf->rmw_level, (unsigned long long)recovery_cp,
3884                          (unsigned long long)sh->sector);
3885         } else for (i = disks; i--; ) {
3886                 /* would I have to read this buffer for read_modify_write */
3887                 struct r5dev *dev = &sh->dev[i];
3888                 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3889                      i == sh->pd_idx || i == sh->qd_idx ||
3890                      test_bit(R5_InJournal, &dev->flags)) &&
3891                     !test_bit(R5_LOCKED, &dev->flags) &&
3892                     !(uptodate_for_rmw(dev) ||
3893                       test_bit(R5_Wantcompute, &dev->flags))) {
3894                         if (test_bit(R5_Insync, &dev->flags))
3895                                 rmw++;
3896                         else
3897                                 rmw += 2*disks;  /* cannot read it */
3898                 }
3899                 /* Would I have to read this buffer for reconstruct_write */
3900                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3901                     i != sh->pd_idx && i != sh->qd_idx &&
3902                     !test_bit(R5_LOCKED, &dev->flags) &&
3903                     !(test_bit(R5_UPTODATE, &dev->flags) ||
3904                       test_bit(R5_Wantcompute, &dev->flags))) {
3905                         if (test_bit(R5_Insync, &dev->flags))
3906                                 rcw++;
3907                         else
3908                                 rcw += 2*disks;
3909                 }
3910         }
3911
3912         pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3913                  (unsigned long long)sh->sector, sh->state, rmw, rcw);
3914         set_bit(STRIPE_HANDLE, &sh->state);
3915         if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3916                 /* prefer read-modify-write, but need to get some data */
3917                 if (conf->mddev->queue)
3918                         blk_add_trace_msg(conf->mddev->queue,
3919                                           "raid5 rmw %llu %d",
3920                                           (unsigned long long)sh->sector, rmw);
3921                 for (i = disks; i--; ) {
3922                         struct r5dev *dev = &sh->dev[i];
3923                         if (test_bit(R5_InJournal, &dev->flags) &&
3924                             dev->page == dev->orig_page &&
3925                             !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
3926                                 /* alloc page for prexor */
3927                                 struct page *p = alloc_page(GFP_NOIO);
3928
3929                                 if (p) {
3930                                         dev->orig_page = p;
3931                                         continue;
3932                                 }
3933
3934                                 /*
3935                                  * alloc_page() failed, try use
3936                                  * disk_info->extra_page
3937                                  */
3938                                 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
3939                                                       &conf->cache_state)) {
3940                                         r5c_use_extra_page(sh);
3941                                         break;
3942                                 }
3943
3944                                 /* extra_page in use, add to delayed_list */
3945                                 set_bit(STRIPE_DELAYED, &sh->state);
3946                                 s->waiting_extra_page = 1;
3947                                 return -EAGAIN;
3948                         }
3949                 }
3950
3951                 for (i = disks; i--; ) {
3952                         struct r5dev *dev = &sh->dev[i];
3953                         if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3954                              i == sh->pd_idx || i == sh->qd_idx ||
3955                              test_bit(R5_InJournal, &dev->flags)) &&
3956                             !test_bit(R5_LOCKED, &dev->flags) &&
3957                             !(uptodate_for_rmw(dev) ||
3958                               test_bit(R5_Wantcompute, &dev->flags)) &&
3959                             test_bit(R5_Insync, &dev->flags)) {
3960                                 if (test_bit(STRIPE_PREREAD_ACTIVE,
3961                                              &sh->state)) {
3962                                         pr_debug("Read_old block %d for r-m-w\n",
3963                                                  i);
3964                                         set_bit(R5_LOCKED, &dev->flags);
3965                                         set_bit(R5_Wantread, &dev->flags);
3966                                         s->locked++;
3967                                 } else {
3968                                         set_bit(STRIPE_DELAYED, &sh->state);
3969                                         set_bit(STRIPE_HANDLE, &sh->state);
3970                                 }
3971                         }
3972                 }
3973         }
3974         if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3975                 /* want reconstruct write, but need to get some data */
3976                 int qread =0;
3977                 rcw = 0;
3978                 for (i = disks; i--; ) {
3979                         struct r5dev *dev = &sh->dev[i];
3980                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3981                             i != sh->pd_idx && i != sh->qd_idx &&
3982                             !test_bit(R5_LOCKED, &dev->flags) &&
3983                             !(test_bit(R5_UPTODATE, &dev->flags) ||
3984                               test_bit(R5_Wantcompute, &dev->flags))) {
3985                                 rcw++;
3986                                 if (test_bit(R5_Insync, &dev->flags) &&
3987                                     test_bit(STRIPE_PREREAD_ACTIVE,
3988                                              &sh->state)) {
3989                                         pr_debug("Read_old block "
3990                                                 "%d for Reconstruct\n", i);
3991                                         set_bit(R5_LOCKED, &dev->flags);
3992                                         set_bit(R5_Wantread, &dev->flags);
3993                                         s->locked++;
3994                                         qread++;
3995                                 } else {
3996                                         set_bit(STRIPE_DELAYED, &sh->state);
3997                                         set_bit(STRIPE_HANDLE, &sh->state);
3998                                 }
3999                         }
4000                 }
4001                 if (rcw && conf->mddev->queue)
4002                         blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
4003                                           (unsigned long long)sh->sector,
4004                                           rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
4005         }
4006
4007         if (rcw > disks && rmw > disks &&
4008             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4009                 set_bit(STRIPE_DELAYED, &sh->state);
4010
4011         /* now if nothing is locked, and if we have enough data,
4012          * we can start a write request
4013          */
4014         /* since handle_stripe can be called at any time we need to handle the
4015          * case where a compute block operation has been submitted and then a
4016          * subsequent call wants to start a write request.  raid_run_ops only
4017          * handles the case where compute block and reconstruct are requested
4018          * simultaneously.  If this is not the case then new writes need to be
4019          * held off until the compute completes.
4020          */
4021         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4022             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4023              !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4024                 schedule_reconstruction(sh, s, rcw == 0, 0);
4025         return 0;
4026 }
4027
4028 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4029                                 struct stripe_head_state *s, int disks)
4030 {
4031         struct r5dev *dev = NULL;
4032
4033         BUG_ON(sh->batch_head);
4034         set_bit(STRIPE_HANDLE, &sh->state);
4035
4036         switch (sh->check_state) {
4037         case check_state_idle:
4038                 /* start a new check operation if there are no failures */
4039                 if (s->failed == 0) {
4040                         BUG_ON(s->uptodate != disks);
4041                         sh->check_state = check_state_run;
4042                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4043                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4044                         s->uptodate--;
4045                         break;
4046                 }
4047                 dev = &sh->dev[s->failed_num[0]];
4048                 /* fall through */
4049         case check_state_compute_result:
4050                 sh->check_state = check_state_idle;
4051                 if (!dev)
4052                         dev = &sh->dev[sh->pd_idx];
4053
4054                 /* check that a write has not made the stripe insync */
4055                 if (test_bit(STRIPE_INSYNC, &sh->state))
4056                         break;
4057
4058                 /* either failed parity check, or recovery is happening */
4059                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4060                 BUG_ON(s->uptodate != disks);
4061
4062                 set_bit(R5_LOCKED, &dev->flags);
4063                 s->locked++;
4064                 set_bit(R5_Wantwrite, &dev->flags);
4065
4066                 clear_bit(STRIPE_DEGRADED, &sh->state);
4067                 set_bit(STRIPE_INSYNC, &sh->state);
4068                 break;
4069         case check_state_run:
4070                 break; /* we will be called again upon completion */
4071         case check_state_check_result:
4072                 sh->check_state = check_state_idle;
4073
4074                 /* if a failure occurred during the check operation, leave
4075                  * STRIPE_INSYNC not set and let the stripe be handled again
4076                  */
4077                 if (s->failed)
4078                         break;
4079
4080                 /* handle a successful check operation, if parity is correct
4081                  * we are done.  Otherwise update the mismatch count and repair
4082                  * parity if !MD_RECOVERY_CHECK
4083                  */
4084                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4085                         /* parity is correct (on disc,
4086                          * not in buffer any more)
4087                          */
4088                         set_bit(STRIPE_INSYNC, &sh->state);
4089                 else {
4090                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4091                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4092                                 /* don't try to repair!! */
4093                                 set_bit(STRIPE_INSYNC, &sh->state);
4094                                 pr_warn_ratelimited("%s: mismatch sector in range "
4095                                                     "%llu-%llu\n", mdname(conf->mddev),
4096                                                     (unsigned long long) sh->sector,
4097                                                     (unsigned long long) sh->sector +
4098                                                     STRIPE_SECTORS);
4099                         } else {
4100                                 sh->check_state = check_state_compute_run;
4101                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4102                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4103                                 set_bit(R5_Wantcompute,
4104                                         &sh->dev[sh->pd_idx].flags);
4105                                 sh->ops.target = sh->pd_idx;
4106                                 sh->ops.target2 = -1;
4107                                 s->uptodate++;
4108                         }
4109                 }
4110                 break;
4111         case check_state_compute_run:
4112                 break;
4113         default:
4114                 pr_err("%s: unknown check_state: %d sector: %llu\n",
4115                        __func__, sh->check_state,
4116                        (unsigned long long) sh->sector);
4117                 BUG();
4118         }
4119 }
4120
4121 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4122                                   struct stripe_head_state *s,
4123                                   int disks)
4124 {
4125         int pd_idx = sh->pd_idx;
4126         int qd_idx = sh->qd_idx;
4127         struct r5dev *dev;
4128
4129         BUG_ON(sh->batch_head);
4130         set_bit(STRIPE_HANDLE, &sh->state);
4131
4132         BUG_ON(s->failed > 2);
4133
4134         /* Want to check and possibly repair P and Q.
4135          * However there could be one 'failed' device, in which
4136          * case we can only check one of them, possibly using the
4137          * other to generate missing data
4138          */
4139
4140         switch (sh->check_state) {
4141         case check_state_idle:
4142                 /* start a new check operation if there are < 2 failures */
4143                 if (s->failed == s->q_failed) {
4144                         /* The only possible failed device holds Q, so it
4145                          * makes sense to check P (If anything else were failed,
4146                          * we would have used P to recreate it).
4147                          */
4148                         sh->check_state = check_state_run;
4149                 }
4150                 if (!s->q_failed && s->failed < 2) {
4151                         /* Q is not failed, and we didn't use it to generate
4152                          * anything, so it makes sense to check it
4153                          */
4154                         if (sh->check_state == check_state_run)
4155                                 sh->check_state = check_state_run_pq;
4156                         else
4157                                 sh->check_state = check_state_run_q;
4158                 }
4159
4160                 /* discard potentially stale zero_sum_result */
4161                 sh->ops.zero_sum_result = 0;
4162
4163                 if (sh->check_state == check_state_run) {
4164                         /* async_xor_zero_sum destroys the contents of P */
4165                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4166                         s->uptodate--;
4167                 }
4168                 if (sh->check_state >= check_state_run &&
4169                     sh->check_state <= check_state_run_pq) {
4170                         /* async_syndrome_zero_sum preserves P and Q, so
4171                          * no need to mark them !uptodate here
4172                          */
4173                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4174                         break;
4175                 }
4176
4177                 /* we have 2-disk failure */
4178                 BUG_ON(s->failed != 2);
4179                 /* fall through */
4180         case check_state_compute_result:
4181                 sh->check_state = check_state_idle;
4182
4183                 /* check that a write has not made the stripe insync */
4184                 if (test_bit(STRIPE_INSYNC, &sh->state))
4185                         break;
4186
4187                 /* now write out any block on a failed drive,
4188                  * or P or Q if they were recomputed
4189                  */
4190                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
4191                 if (s->failed == 2) {
4192                         dev = &sh->dev[s->failed_num[1]];
4193                         s->locked++;
4194                         set_bit(R5_LOCKED, &dev->flags);
4195                         set_bit(R5_Wantwrite, &dev->flags);
4196                 }
4197                 if (s->failed >= 1) {
4198                         dev = &sh->dev[s->failed_num[0]];
4199                         s->locked++;
4200                         set_bit(R5_LOCKED, &dev->flags);
4201                         set_bit(R5_Wantwrite, &dev->flags);
4202                 }
4203                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4204                         dev = &sh->dev[pd_idx];
4205                         s->locked++;
4206                         set_bit(R5_LOCKED, &dev->flags);
4207                         set_bit(R5_Wantwrite, &dev->flags);
4208                 }
4209                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4210                         dev = &sh->dev[qd_idx];
4211                         s->locked++;
4212                         set_bit(R5_LOCKED, &dev->flags);
4213                         set_bit(R5_Wantwrite, &dev->flags);
4214                 }
4215                 clear_bit(STRIPE_DEGRADED, &sh->state);
4216
4217                 set_bit(STRIPE_INSYNC, &sh->state);
4218                 break;
4219         case check_state_run:
4220         case check_state_run_q:
4221         case check_state_run_pq:
4222                 break; /* we will be called again upon completion */
4223         case check_state_check_result:
4224                 sh->check_state = check_state_idle;
4225
4226                 /* handle a successful check operation, if parity is correct
4227                  * we are done.  Otherwise update the mismatch count and repair
4228                  * parity if !MD_RECOVERY_CHECK
4229                  */
4230                 if (sh->ops.zero_sum_result == 0) {
4231                         /* both parities are correct */
4232                         if (!s->failed)
4233                                 set_bit(STRIPE_INSYNC, &sh->state);
4234                         else {
4235                                 /* in contrast to the raid5 case we can validate
4236                                  * parity, but still have a failure to write
4237                                  * back
4238                                  */
4239                                 sh->check_state = check_state_compute_result;
4240                                 /* Returning at this point means that we may go
4241                                  * off and bring p and/or q uptodate again so
4242                                  * we make sure to check zero_sum_result again
4243                                  * to verify if p or q need writeback
4244                                  */
4245                         }
4246                 } else {
4247                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4248                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4249                                 /* don't try to repair!! */
4250                                 set_bit(STRIPE_INSYNC, &sh->state);
4251                                 pr_warn_ratelimited("%s: mismatch sector in range "
4252                                                     "%llu-%llu\n", mdname(conf->mddev),
4253                                                     (unsigned long long) sh->sector,
4254                                                     (unsigned long long) sh->sector +
4255                                                     STRIPE_SECTORS);
4256                         } else {
4257                                 int *target = &sh->ops.target;
4258
4259                                 sh->ops.target = -1;
4260                                 sh->ops.target2 = -1;
4261                                 sh->check_state = check_state_compute_run;
4262                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4263                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4264                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4265                                         set_bit(R5_Wantcompute,
4266                                                 &sh->dev[pd_idx].flags);
4267                                         *target = pd_idx;
4268                                         target = &sh->ops.target2;
4269                                         s->uptodate++;
4270                                 }
4271                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4272                                         set_bit(R5_Wantcompute,
4273                                                 &sh->dev[qd_idx].flags);
4274                                         *target = qd_idx;
4275                                         s->uptodate++;
4276                                 }
4277                         }
4278                 }
4279                 break;
4280         case check_state_compute_run:
4281                 break;
4282         default:
4283                 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4284                         __func__, sh->check_state,
4285                         (unsigned long long) sh->sector);
4286                 BUG();
4287         }
4288 }
4289
4290 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4291 {
4292         int i;
4293
4294         /* We have read all the blocks in this stripe and now we need to
4295          * copy some of them into a target stripe for expand.
4296          */
4297         struct dma_async_tx_descriptor *tx = NULL;
4298         BUG_ON(sh->batch_head);
4299         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4300         for (i = 0; i < sh->disks; i++)
4301                 if (i != sh->pd_idx && i != sh->qd_idx) {
4302                         int dd_idx, j;
4303                         struct stripe_head *sh2;
4304                         struct async_submit_ctl submit;
4305
4306                         sector_t bn = raid5_compute_blocknr(sh, i, 1);
4307                         sector_t s = raid5_compute_sector(conf, bn, 0,
4308                                                           &dd_idx, NULL);
4309                         sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4310                         if (sh2 == NULL)
4311                                 /* so far only the early blocks of this stripe
4312                                  * have been requested.  When later blocks
4313                                  * get requested, we will try again
4314                                  */
4315                                 continue;
4316                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4317                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4318                                 /* must have already done this block */
4319                                 raid5_release_stripe(sh2);
4320                                 continue;
4321                         }
4322
4323                         /* place all the copies on one channel */
4324                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4325                         tx = async_memcpy(sh2->dev[dd_idx].page,
4326                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
4327                                           &submit);
4328
4329                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4330                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4331                         for (j = 0; j < conf->raid_disks; j++)
4332                                 if (j != sh2->pd_idx &&
4333                                     j != sh2->qd_idx &&
4334                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
4335                                         break;
4336                         if (j == conf->raid_disks) {
4337                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4338                                 set_bit(STRIPE_HANDLE, &sh2->state);
4339                         }
4340                         raid5_release_stripe(sh2);
4341
4342                 }
4343         /* done submitting copies, wait for them to complete */
4344         async_tx_quiesce(&tx);
4345 }
4346
4347 /*
4348  * handle_stripe - do things to a stripe.
4349  *
4350  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4351  * state of various bits to see what needs to be done.
4352  * Possible results:
4353  *    return some read requests which now have data
4354  *    return some write requests which are safely on storage
4355  *    schedule a read on some buffers
4356  *    schedule a write of some buffers
4357  *    return confirmation of parity correctness
4358  *
4359  */
4360
4361 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4362 {
4363         struct r5conf *conf = sh->raid_conf;
4364         int disks = sh->disks;
4365         struct r5dev *dev;
4366         int i;
4367         int do_recovery = 0;
4368
4369         memset(s, 0, sizeof(*s));
4370
4371         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4372         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4373         s->failed_num[0] = -1;
4374         s->failed_num[1] = -1;
4375         s->log_failed = r5l_log_disk_error(conf);
4376
4377         /* Now to look around and see what can be done */
4378         rcu_read_lock();
4379         for (i=disks; i--; ) {
4380                 struct md_rdev *rdev;
4381                 sector_t first_bad;
4382                 int bad_sectors;
4383                 int is_bad = 0;
4384
4385                 dev = &sh->dev[i];
4386
4387                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4388                          i, dev->flags,
4389                          dev->toread, dev->towrite, dev->written);
4390                 /* maybe we can reply to a read
4391                  *
4392                  * new wantfill requests are only permitted while
4393                  * ops_complete_biofill is guaranteed to be inactive
4394                  */
4395                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4396                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4397                         set_bit(R5_Wantfill, &dev->flags);
4398
4399                 /* now count some things */
4400                 if (test_bit(R5_LOCKED, &dev->flags))
4401                         s->locked++;
4402                 if (test_bit(R5_UPTODATE, &dev->flags))
4403                         s->uptodate++;
4404                 if (test_bit(R5_Wantcompute, &dev->flags)) {
4405                         s->compute++;
4406                         BUG_ON(s->compute > 2);
4407                 }
4408
4409                 if (test_bit(R5_Wantfill, &dev->flags))
4410                         s->to_fill++;
4411                 else if (dev->toread)
4412                         s->to_read++;
4413                 if (dev->towrite) {
4414                         s->to_write++;
4415                         if (!test_bit(R5_OVERWRITE, &dev->flags))
4416                                 s->non_overwrite++;
4417                 }
4418                 if (dev->written)
4419                         s->written++;
4420                 /* Prefer to use the replacement for reads, but only
4421                  * if it is recovered enough and has no bad blocks.
4422                  */
4423                 rdev = rcu_dereference(conf->disks[i].replacement);
4424                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4425                     rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4426                     !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4427                                  &first_bad, &bad_sectors))
4428                         set_bit(R5_ReadRepl, &dev->flags);
4429                 else {
4430                         if (rdev && !test_bit(Faulty, &rdev->flags))
4431                                 set_bit(R5_NeedReplace, &dev->flags);
4432                         else
4433                                 clear_bit(R5_NeedReplace, &dev->flags);
4434                         rdev = rcu_dereference(conf->disks[i].rdev);
4435                         clear_bit(R5_ReadRepl, &dev->flags);
4436                 }
4437                 if (rdev && test_bit(Faulty, &rdev->flags))
4438                         rdev = NULL;
4439                 if (rdev) {
4440                         is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4441                                              &first_bad, &bad_sectors);
4442                         if (s->blocked_rdev == NULL
4443                             && (test_bit(Blocked, &rdev->flags)
4444                                 || is_bad < 0)) {
4445                                 if (is_bad < 0)
4446                                         set_bit(BlockedBadBlocks,
4447                                                 &rdev->flags);
4448                                 s->blocked_rdev = rdev;
4449                                 atomic_inc(&rdev->nr_pending);
4450                         }
4451                 }
4452                 clear_bit(R5_Insync, &dev->flags);
4453                 if (!rdev)
4454                         /* Not in-sync */;
4455                 else if (is_bad) {
4456                         /* also not in-sync */
4457                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4458                             test_bit(R5_UPTODATE, &dev->flags)) {
4459                                 /* treat as in-sync, but with a read error
4460                                  * which we can now try to correct
4461                                  */
4462                                 set_bit(R5_Insync, &dev->flags);
4463                                 set_bit(R5_ReadError, &dev->flags);
4464                         }
4465                 } else if (test_bit(In_sync, &rdev->flags))
4466                         set_bit(R5_Insync, &dev->flags);
4467                 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4468                         /* in sync if before recovery_offset */
4469                         set_bit(R5_Insync, &dev->flags);
4470                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4471                          test_bit(R5_Expanded, &dev->flags))
4472                         /* If we've reshaped into here, we assume it is Insync.
4473                          * We will shortly update recovery_offset to make
4474                          * it official.
4475                          */
4476                         set_bit(R5_Insync, &dev->flags);
4477
4478                 if (test_bit(R5_WriteError, &dev->flags)) {
4479                         /* This flag does not apply to '.replacement'
4480                          * only to .rdev, so make sure to check that*/
4481                         struct md_rdev *rdev2 = rcu_dereference(
4482                                 conf->disks[i].rdev);
4483                         if (rdev2 == rdev)
4484                                 clear_bit(R5_Insync, &dev->flags);
4485                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4486                                 s->handle_bad_blocks = 1;
4487                                 atomic_inc(&rdev2->nr_pending);
4488                         } else
4489                                 clear_bit(R5_WriteError, &dev->flags);
4490                 }
4491                 if (test_bit(R5_MadeGood, &dev->flags)) {
4492                         /* This flag does not apply to '.replacement'
4493                          * only to .rdev, so make sure to check that*/
4494                         struct md_rdev *rdev2 = rcu_dereference(
4495                                 conf->disks[i].rdev);
4496                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4497                                 s->handle_bad_blocks = 1;
4498                                 atomic_inc(&rdev2->nr_pending);
4499                         } else
4500                                 clear_bit(R5_MadeGood, &dev->flags);
4501                 }
4502                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4503                         struct md_rdev *rdev2 = rcu_dereference(
4504                                 conf->disks[i].replacement);
4505                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4506                                 s->handle_bad_blocks = 1;
4507                                 atomic_inc(&rdev2->nr_pending);
4508                         } else
4509                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
4510                 }
4511                 if (!test_bit(R5_Insync, &dev->flags)) {
4512                         /* The ReadError flag will just be confusing now */
4513                         clear_bit(R5_ReadError, &dev->flags);
4514                         clear_bit(R5_ReWrite, &dev->flags);
4515                 }
4516                 if (test_bit(R5_ReadError, &dev->flags))
4517                         clear_bit(R5_Insync, &dev->flags);
4518                 if (!test_bit(R5_Insync, &dev->flags)) {
4519                         if (s->failed < 2)
4520                                 s->failed_num[s->failed] = i;
4521                         s->failed++;
4522                         if (rdev && !test_bit(Faulty, &rdev->flags))
4523                                 do_recovery = 1;
4524                         else if (!rdev) {
4525                                 rdev = rcu_dereference(
4526                                     conf->disks[i].replacement);
4527                                 if (rdev && !test_bit(Faulty, &rdev->flags))
4528                                         do_recovery = 1;
4529                         }
4530                 }
4531
4532                 if (test_bit(R5_InJournal, &dev->flags))
4533                         s->injournal++;
4534                 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4535                         s->just_cached++;
4536         }
4537         if (test_bit(STRIPE_SYNCING, &sh->state)) {
4538                 /* If there is a failed device being replaced,
4539                  *     we must be recovering.
4540                  * else if we are after recovery_cp, we must be syncing
4541                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4542                  * else we can only be replacing
4543                  * sync and recovery both need to read all devices, and so
4544                  * use the same flag.
4545                  */
4546                 if (do_recovery ||
4547                     sh->sector >= conf->mddev->recovery_cp ||
4548                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4549                         s->syncing = 1;
4550                 else
4551                         s->replacing = 1;
4552         }
4553         rcu_read_unlock();
4554 }
4555
4556 static int clear_batch_ready(struct stripe_head *sh)
4557 {
4558         /* Return '1' if this is a member of batch, or
4559          * '0' if it is a lone stripe or a head which can now be
4560          * handled.
4561          */
4562         struct stripe_head *tmp;
4563         if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4564                 return (sh->batch_head && sh->batch_head != sh);
4565         spin_lock(&sh->stripe_lock);
4566         if (!sh->batch_head) {
4567                 spin_unlock(&sh->stripe_lock);
4568                 return 0;
4569         }
4570
4571         /*
4572          * this stripe could be added to a batch list before we check
4573          * BATCH_READY, skips it
4574          */
4575         if (sh->batch_head != sh) {
4576                 spin_unlock(&sh->stripe_lock);
4577                 return 1;
4578         }
4579         spin_lock(&sh->batch_lock);
4580         list_for_each_entry(tmp, &sh->batch_list, batch_list)
4581                 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4582         spin_unlock(&sh->batch_lock);
4583         spin_unlock(&sh->stripe_lock);
4584
4585         /*
4586          * BATCH_READY is cleared, no new stripes can be added.
4587          * batch_list can be accessed without lock
4588          */
4589         return 0;
4590 }
4591
4592 static void break_stripe_batch_list(struct stripe_head *head_sh,
4593                                     unsigned long handle_flags)
4594 {
4595         struct stripe_head *sh, *next;
4596         int i;
4597         int do_wakeup = 0;
4598
4599         list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4600
4601                 list_del_init(&sh->batch_list);
4602
4603                 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4604                                           (1 << STRIPE_SYNCING) |
4605                                           (1 << STRIPE_REPLACED) |
4606                                           (1 << STRIPE_DELAYED) |
4607                                           (1 << STRIPE_BIT_DELAY) |
4608                                           (1 << STRIPE_FULL_WRITE) |
4609                                           (1 << STRIPE_BIOFILL_RUN) |
4610                                           (1 << STRIPE_COMPUTE_RUN)  |
4611                                           (1 << STRIPE_OPS_REQ_PENDING) |
4612                                           (1 << STRIPE_DISCARD) |
4613                                           (1 << STRIPE_BATCH_READY) |
4614                                           (1 << STRIPE_BATCH_ERR) |
4615                                           (1 << STRIPE_BITMAP_PENDING)),
4616                         "stripe state: %lx\n", sh->state);
4617                 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4618                                               (1 << STRIPE_REPLACED)),
4619                         "head stripe state: %lx\n", head_sh->state);
4620
4621                 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4622                                             (1 << STRIPE_PREREAD_ACTIVE) |
4623                                             (1 << STRIPE_DEGRADED) |
4624                                             (1 << STRIPE_ON_UNPLUG_LIST)),
4625                               head_sh->state & (1 << STRIPE_INSYNC));
4626
4627                 sh->check_state = head_sh->check_state;
4628                 sh->reconstruct_state = head_sh->reconstruct_state;
4629                 spin_lock_irq(&sh->stripe_lock);
4630                 sh->batch_head = NULL;
4631                 spin_unlock_irq(&sh->stripe_lock);
4632                 for (i = 0; i < sh->disks; i++) {
4633                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4634                                 do_wakeup = 1;
4635                         sh->dev[i].flags = head_sh->dev[i].flags &
4636                                 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4637                 }
4638                 if (handle_flags == 0 ||
4639                     sh->state & handle_flags)
4640                         set_bit(STRIPE_HANDLE, &sh->state);
4641                 raid5_release_stripe(sh);
4642         }
4643         spin_lock_irq(&head_sh->stripe_lock);
4644         head_sh->batch_head = NULL;
4645         spin_unlock_irq(&head_sh->stripe_lock);
4646         for (i = 0; i < head_sh->disks; i++)
4647                 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4648                         do_wakeup = 1;
4649         if (head_sh->state & handle_flags)
4650                 set_bit(STRIPE_HANDLE, &head_sh->state);
4651
4652         if (do_wakeup)
4653                 wake_up(&head_sh->raid_conf->wait_for_overlap);
4654 }
4655
4656 static void handle_stripe(struct stripe_head *sh)
4657 {
4658         struct stripe_head_state s;
4659         struct r5conf *conf = sh->raid_conf;
4660         int i;
4661         int prexor;
4662         int disks = sh->disks;
4663         struct r5dev *pdev, *qdev;
4664
4665         clear_bit(STRIPE_HANDLE, &sh->state);
4666         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4667                 /* already being handled, ensure it gets handled
4668                  * again when current action finishes */
4669                 set_bit(STRIPE_HANDLE, &sh->state);
4670                 return;
4671         }
4672
4673         if (clear_batch_ready(sh) ) {
4674                 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4675                 return;
4676         }
4677
4678         if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4679                 break_stripe_batch_list(sh, 0);
4680
4681         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4682                 spin_lock(&sh->stripe_lock);
4683                 /*
4684                  * Cannot process 'sync' concurrently with 'discard'.
4685                  * Flush data in r5cache before 'sync'.
4686                  */
4687                 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4688                     !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4689                     !test_bit(STRIPE_DISCARD, &sh->state) &&
4690                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4691                         set_bit(STRIPE_SYNCING, &sh->state);
4692                         clear_bit(STRIPE_INSYNC, &sh->state);
4693                         clear_bit(STRIPE_REPLACED, &sh->state);
4694                 }
4695                 spin_unlock(&sh->stripe_lock);
4696         }
4697         clear_bit(STRIPE_DELAYED, &sh->state);
4698
4699         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4700                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4701                (unsigned long long)sh->sector, sh->state,
4702                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4703                sh->check_state, sh->reconstruct_state);
4704
4705         analyse_stripe(sh, &s);
4706
4707         if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4708                 goto finish;
4709
4710         if (s.handle_bad_blocks ||
4711             test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4712                 set_bit(STRIPE_HANDLE, &sh->state);
4713                 goto finish;
4714         }
4715
4716         if (unlikely(s.blocked_rdev)) {
4717                 if (s.syncing || s.expanding || s.expanded ||
4718                     s.replacing || s.to_write || s.written) {
4719                         set_bit(STRIPE_HANDLE, &sh->state);
4720                         goto finish;
4721                 }
4722                 /* There is nothing for the blocked_rdev to block */
4723                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4724                 s.blocked_rdev = NULL;
4725         }
4726
4727         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4728                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4729                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4730         }
4731
4732         pr_debug("locked=%d uptodate=%d to_read=%d"
4733                " to_write=%d failed=%d failed_num=%d,%d\n",
4734                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4735                s.failed_num[0], s.failed_num[1]);
4736         /*
4737          * check if the array has lost more than max_degraded devices and,
4738          * if so, some requests might need to be failed.
4739          *
4740          * When journal device failed (log_failed), we will only process
4741          * the stripe if there is data need write to raid disks
4742          */
4743         if (s.failed > conf->max_degraded ||
4744             (s.log_failed && s.injournal == 0)) {
4745                 sh->check_state = 0;
4746                 sh->reconstruct_state = 0;
4747                 break_stripe_batch_list(sh, 0);
4748                 if (s.to_read+s.to_write+s.written)
4749                         handle_failed_stripe(conf, sh, &s, disks);
4750                 if (s.syncing + s.replacing)
4751                         handle_failed_sync(conf, sh, &s);
4752         }
4753
4754         /* Now we check to see if any write operations have recently
4755          * completed
4756          */
4757         prexor = 0;
4758         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4759                 prexor = 1;
4760         if (sh->reconstruct_state == reconstruct_state_drain_result ||
4761             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4762                 sh->reconstruct_state = reconstruct_state_idle;
4763
4764                 /* All the 'written' buffers and the parity block are ready to
4765                  * be written back to disk
4766                  */
4767                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4768                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4769                 BUG_ON(sh->qd_idx >= 0 &&
4770                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4771                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4772                 for (i = disks; i--; ) {
4773                         struct r5dev *dev = &sh->dev[i];
4774                         if (test_bit(R5_LOCKED, &dev->flags) &&
4775                                 (i == sh->pd_idx || i == sh->qd_idx ||
4776                                  dev->written || test_bit(R5_InJournal,
4777                                                           &dev->flags))) {
4778                                 pr_debug("Writing block %d\n", i);
4779                                 set_bit(R5_Wantwrite, &dev->flags);
4780                                 if (prexor)
4781                                         continue;
4782                                 if (s.failed > 1)
4783                                         continue;
4784                                 if (!test_bit(R5_Insync, &dev->flags) ||
4785                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
4786                                      s.failed == 0))
4787                                         set_bit(STRIPE_INSYNC, &sh->state);
4788                         }
4789                 }
4790                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4791                         s.dec_preread_active = 1;
4792         }
4793
4794         /*
4795          * might be able to return some write requests if the parity blocks
4796          * are safe, or on a failed drive
4797          */
4798         pdev = &sh->dev[sh->pd_idx];
4799         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4800                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4801         qdev = &sh->dev[sh->qd_idx];
4802         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4803                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4804                 || conf->level < 6;
4805
4806         if (s.written &&
4807             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4808                              && !test_bit(R5_LOCKED, &pdev->flags)
4809                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
4810                                  test_bit(R5_Discard, &pdev->flags))))) &&
4811             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4812                              && !test_bit(R5_LOCKED, &qdev->flags)
4813                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
4814                                  test_bit(R5_Discard, &qdev->flags))))))
4815                 handle_stripe_clean_event(conf, sh, disks);
4816
4817         if (s.just_cached)
4818                 r5c_handle_cached_data_endio(conf, sh, disks);
4819         log_stripe_write_finished(sh);
4820
4821         /* Now we might consider reading some blocks, either to check/generate
4822          * parity, or to satisfy requests
4823          * or to load a block that is being partially written.
4824          */
4825         if (s.to_read || s.non_overwrite
4826             || (conf->level == 6 && s.to_write && s.failed)
4827             || (s.syncing && (s.uptodate + s.compute < disks))
4828             || s.replacing
4829             || s.expanding)
4830                 handle_stripe_fill(sh, &s, disks);
4831
4832         /*
4833          * When the stripe finishes full journal write cycle (write to journal
4834          * and raid disk), this is the clean up procedure so it is ready for
4835          * next operation.
4836          */
4837         r5c_finish_stripe_write_out(conf, sh, &s);
4838
4839         /*
4840          * Now to consider new write requests, cache write back and what else,
4841          * if anything should be read.  We do not handle new writes when:
4842          * 1/ A 'write' operation (copy+xor) is already in flight.
4843          * 2/ A 'check' operation is in flight, as it may clobber the parity
4844          *    block.
4845          * 3/ A r5c cache log write is in flight.
4846          */
4847
4848         if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
4849                 if (!r5c_is_writeback(conf->log)) {
4850                         if (s.to_write)
4851                                 handle_stripe_dirtying(conf, sh, &s, disks);
4852                 } else { /* write back cache */
4853                         int ret = 0;
4854
4855                         /* First, try handle writes in caching phase */
4856                         if (s.to_write)
4857                                 ret = r5c_try_caching_write(conf, sh, &s,
4858                                                             disks);
4859                         /*
4860                          * If caching phase failed: ret == -EAGAIN
4861                          *    OR
4862                          * stripe under reclaim: !caching && injournal
4863                          *
4864                          * fall back to handle_stripe_dirtying()
4865                          */
4866                         if (ret == -EAGAIN ||
4867                             /* stripe under reclaim: !caching && injournal */
4868                             (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
4869                              s.injournal > 0)) {
4870                                 ret = handle_stripe_dirtying(conf, sh, &s,
4871                                                              disks);
4872                                 if (ret == -EAGAIN)
4873                                         goto finish;
4874                         }
4875                 }
4876         }
4877
4878         /* maybe we need to check and possibly fix the parity for this stripe
4879          * Any reads will already have been scheduled, so we just see if enough
4880          * data is available.  The parity check is held off while parity
4881          * dependent operations are in flight.
4882          */
4883         if (sh->check_state ||
4884             (s.syncing && s.locked == 0 &&
4885              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4886              !test_bit(STRIPE_INSYNC, &sh->state))) {
4887                 if (conf->level == 6)
4888                         handle_parity_checks6(conf, sh, &s, disks);
4889                 else
4890                         handle_parity_checks5(conf, sh, &s, disks);
4891         }
4892
4893         if ((s.replacing || s.syncing) && s.locked == 0
4894             && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4895             && !test_bit(STRIPE_REPLACED, &sh->state)) {
4896                 /* Write out to replacement devices where possible */
4897                 for (i = 0; i < conf->raid_disks; i++)
4898                         if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4899                                 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4900                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
4901                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
4902                                 s.locked++;
4903                         }
4904                 if (s.replacing)
4905                         set_bit(STRIPE_INSYNC, &sh->state);
4906                 set_bit(STRIPE_REPLACED, &sh->state);
4907         }
4908         if ((s.syncing || s.replacing) && s.locked == 0 &&
4909             !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4910             test_bit(STRIPE_INSYNC, &sh->state)) {
4911                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4912                 clear_bit(STRIPE_SYNCING, &sh->state);
4913                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4914                         wake_up(&conf->wait_for_overlap);
4915         }
4916
4917         /* If the failed drives are just a ReadError, then we might need
4918          * to progress the repair/check process
4919          */
4920         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4921                 for (i = 0; i < s.failed; i++) {
4922                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
4923                         if (test_bit(R5_ReadError, &dev->flags)
4924                             && !test_bit(R5_LOCKED, &dev->flags)
4925                             && test_bit(R5_UPTODATE, &dev->flags)
4926                                 ) {
4927                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
4928                                         set_bit(R5_Wantwrite, &dev->flags);
4929                                         set_bit(R5_ReWrite, &dev->flags);
4930                                         set_bit(R5_LOCKED, &dev->flags);
4931                                         s.locked++;
4932                                 } else {
4933                                         /* let's read it back */
4934                                         set_bit(R5_Wantread, &dev->flags);
4935                                         set_bit(R5_LOCKED, &dev->flags);
4936                                         s.locked++;
4937                                 }
4938                         }
4939                 }
4940
4941         /* Finish reconstruct operations initiated by the expansion process */
4942         if (sh->reconstruct_state == reconstruct_state_result) {
4943                 struct stripe_head *sh_src
4944                         = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4945                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4946                         /* sh cannot be written until sh_src has been read.
4947                          * so arrange for sh to be delayed a little
4948                          */
4949                         set_bit(STRIPE_DELAYED, &sh->state);
4950                         set_bit(STRIPE_HANDLE, &sh->state);
4951                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4952                                               &sh_src->state))
4953                                 atomic_inc(&conf->preread_active_stripes);
4954                         raid5_release_stripe(sh_src);
4955                         goto finish;
4956                 }
4957                 if (sh_src)
4958                         raid5_release_stripe(sh_src);
4959
4960                 sh->reconstruct_state = reconstruct_state_idle;
4961                 clear_bit(STRIPE_EXPANDING, &sh->state);
4962                 for (i = conf->raid_disks; i--; ) {
4963                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
4964                         set_bit(R5_LOCKED, &sh->dev[i].flags);
4965                         s.locked++;
4966                 }
4967         }
4968
4969         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4970             !sh->reconstruct_state) {
4971                 /* Need to write out all blocks after computing parity */
4972                 sh->disks = conf->raid_disks;
4973                 stripe_set_idx(sh->sector, conf, 0, sh);
4974                 schedule_reconstruction(sh, &s, 1, 1);
4975         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4976                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4977                 atomic_dec(&conf->reshape_stripes);
4978                 wake_up(&conf->wait_for_overlap);
4979                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4980         }
4981
4982         if (s.expanding && s.locked == 0 &&
4983             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4984                 handle_stripe_expansion(conf, sh);
4985
4986 finish:
4987         /* wait for this device to become unblocked */
4988         if (unlikely(s.blocked_rdev)) {
4989                 if (conf->mddev->external)
4990                         md_wait_for_blocked_rdev(s.blocked_rdev,
4991                                                  conf->mddev);
4992                 else
4993                         /* Internal metadata will immediately
4994                          * be written by raid5d, so we don't
4995                          * need to wait here.
4996                          */
4997                         rdev_dec_pending(s.blocked_rdev,
4998                                          conf->mddev);
4999         }
5000
5001         if (s.handle_bad_blocks)
5002                 for (i = disks; i--; ) {
5003                         struct md_rdev *rdev;
5004                         struct r5dev *dev = &sh->dev[i];
5005                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
5006                                 /* We own a safe reference to the rdev */
5007                                 rdev = conf->disks[i].rdev;
5008                                 if (!rdev_set_badblocks(rdev, sh->sector,
5009                                                         STRIPE_SECTORS, 0))
5010                                         md_error(conf->mddev, rdev);
5011                                 rdev_dec_pending(rdev, conf->mddev);
5012                         }
5013                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5014                                 rdev = conf->disks[i].rdev;
5015                                 rdev_clear_badblocks(rdev, sh->sector,
5016                                                      STRIPE_SECTORS, 0);
5017                                 rdev_dec_pending(rdev, conf->mddev);
5018                         }
5019                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5020                                 rdev = conf->disks[i].replacement;
5021                                 if (!rdev)
5022                                         /* rdev have been moved down */
5023                                         rdev = conf->disks[i].rdev;
5024                                 rdev_clear_badblocks(rdev, sh->sector,
5025                                                      STRIPE_SECTORS, 0);
5026                                 rdev_dec_pending(rdev, conf->mddev);
5027                         }
5028                 }
5029
5030         if (s.ops_request)
5031                 raid_run_ops(sh, s.ops_request);
5032
5033         ops_run_io(sh, &s);
5034
5035         if (s.dec_preread_active) {
5036                 /* We delay this until after ops_run_io so that if make_request
5037                  * is waiting on a flush, it won't continue until the writes
5038                  * have actually been submitted.
5039                  */
5040                 atomic_dec(&conf->preread_active_stripes);
5041                 if (atomic_read(&conf->preread_active_stripes) <
5042                     IO_THRESHOLD)
5043                         md_wakeup_thread(conf->mddev->thread);
5044         }
5045
5046         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5047 }
5048
5049 static void raid5_activate_delayed(struct r5conf *conf)
5050 {
5051         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5052                 while (!list_empty(&conf->delayed_list)) {
5053                         struct list_head *l = conf->delayed_list.next;
5054                         struct stripe_head *sh;
5055                         sh = list_entry(l, struct stripe_head, lru);
5056                         list_del_init(l);
5057                         clear_bit(STRIPE_DELAYED, &sh->state);
5058                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5059                                 atomic_inc(&conf->preread_active_stripes);
5060                         list_add_tail(&sh->lru, &conf->hold_list);
5061                         raid5_wakeup_stripe_thread(sh);
5062                 }
5063         }
5064 }
5065
5066 static void activate_bit_delay(struct r5conf *conf,
5067         struct list_head *temp_inactive_list)
5068 {
5069         /* device_lock is held */
5070         struct list_head head;
5071         list_add(&head, &conf->bitmap_list);
5072         list_del_init(&conf->bitmap_list);
5073         while (!list_empty(&head)) {
5074                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5075                 int hash;
5076                 list_del_init(&sh->lru);
5077                 atomic_inc(&sh->count);
5078                 hash = sh->hash_lock_index;
5079                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5080         }
5081 }
5082
5083 static int raid5_congested(struct mddev *mddev, int bits)
5084 {
5085         struct r5conf *conf = mddev->private;
5086
5087         /* No difference between reads and writes.  Just check
5088          * how busy the stripe_cache is
5089          */
5090
5091         if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
5092                 return 1;
5093
5094         /* Also checks whether there is pressure on r5cache log space */
5095         if (test_bit(R5C_LOG_TIGHT, &conf->cache_state))
5096                 return 1;
5097         if (conf->quiesce)
5098                 return 1;
5099         if (atomic_read(&conf->empty_inactive_list_nr))
5100                 return 1;
5101
5102         return 0;
5103 }
5104
5105 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5106 {
5107         struct r5conf *conf = mddev->private;
5108         sector_t sector = bio->bi_iter.bi_sector;
5109         unsigned int chunk_sectors;
5110         unsigned int bio_sectors = bio_sectors(bio);
5111
5112         WARN_ON_ONCE(bio->bi_partno);
5113
5114         chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5115         return  chunk_sectors >=
5116                 ((sector & (chunk_sectors - 1)) + bio_sectors);
5117 }
5118
5119 /*
5120  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
5121  *  later sampled by raid5d.
5122  */
5123 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5124 {
5125         unsigned long flags;
5126
5127         spin_lock_irqsave(&conf->device_lock, flags);
5128
5129         bi->bi_next = conf->retry_read_aligned_list;
5130         conf->retry_read_aligned_list = bi;
5131
5132         spin_unlock_irqrestore(&conf->device_lock, flags);
5133         md_wakeup_thread(conf->mddev->thread);
5134 }
5135
5136 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5137                                          unsigned int *offset)
5138 {
5139         struct bio *bi;
5140
5141         bi = conf->retry_read_aligned;
5142         if (bi) {
5143                 *offset = conf->retry_read_offset;
5144                 conf->retry_read_aligned = NULL;
5145                 return bi;
5146         }
5147         bi = conf->retry_read_aligned_list;
5148         if(bi) {
5149                 conf->retry_read_aligned_list = bi->bi_next;
5150                 bi->bi_next = NULL;
5151                 *offset = 0;
5152         }
5153
5154         return bi;
5155 }
5156
5157 /*
5158  *  The "raid5_align_endio" should check if the read succeeded and if it
5159  *  did, call bio_endio on the original bio (having bio_put the new bio
5160  *  first).
5161  *  If the read failed..
5162  */
5163 static void raid5_align_endio(struct bio *bi)
5164 {
5165         struct bio* raid_bi  = bi->bi_private;
5166         struct mddev *mddev;
5167         struct r5conf *conf;
5168         struct md_rdev *rdev;
5169         blk_status_t error = bi->bi_status;
5170
5171         bio_put(bi);
5172
5173         rdev = (void*)raid_bi->bi_next;
5174         raid_bi->bi_next = NULL;
5175         mddev = rdev->mddev;
5176         conf = mddev->private;
5177
5178         rdev_dec_pending(rdev, conf->mddev);
5179
5180         if (!error) {
5181                 bio_endio(raid_bi);
5182                 if (atomic_dec_and_test(&conf->active_aligned_reads))
5183                         wake_up(&conf->wait_for_quiescent);
5184                 return;
5185         }
5186
5187         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5188
5189         add_bio_to_retry(raid_bi, conf);
5190 }
5191
5192 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5193 {
5194         struct r5conf *conf = mddev->private;
5195         int dd_idx;
5196         struct bio* align_bi;
5197         struct md_rdev *rdev;
5198         sector_t end_sector;
5199
5200         if (!in_chunk_boundary(mddev, raid_bio)) {
5201                 pr_debug("%s: non aligned\n", __func__);
5202                 return 0;
5203         }
5204         /*
5205          * use bio_clone_fast to make a copy of the bio
5206          */
5207         align_bi = bio_clone_fast(raid_bio, GFP_NOIO, &mddev->bio_set);
5208         if (!align_bi)
5209                 return 0;
5210         /*
5211          *   set bi_end_io to a new function, and set bi_private to the
5212          *     original bio.
5213          */
5214         align_bi->bi_end_io  = raid5_align_endio;
5215         align_bi->bi_private = raid_bio;
5216         /*
5217          *      compute position
5218          */
5219         align_bi->bi_iter.bi_sector =
5220                 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
5221                                      0, &dd_idx, NULL);
5222
5223         end_sector = bio_end_sector(align_bi);
5224         rcu_read_lock();
5225         rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5226         if (!rdev || test_bit(Faulty, &rdev->flags) ||
5227             rdev->recovery_offset < end_sector) {
5228                 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5229                 if (rdev &&
5230                     (test_bit(Faulty, &rdev->flags) ||
5231                     !(test_bit(In_sync, &rdev->flags) ||
5232                       rdev->recovery_offset >= end_sector)))
5233                         rdev = NULL;
5234         }
5235
5236         if (r5c_big_stripe_cached(conf, align_bi->bi_iter.bi_sector)) {
5237                 rcu_read_unlock();
5238                 bio_put(align_bi);
5239                 return 0;
5240         }
5241
5242         if (rdev) {
5243                 sector_t first_bad;
5244                 int bad_sectors;
5245
5246                 atomic_inc(&rdev->nr_pending);
5247                 rcu_read_unlock();
5248                 raid_bio->bi_next = (void*)rdev;
5249                 bio_set_dev(align_bi, rdev->bdev);
5250                 bio_clear_flag(align_bi, BIO_SEG_VALID);
5251
5252                 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
5253                                 bio_sectors(align_bi),
5254                                 &first_bad, &bad_sectors)) {
5255                         bio_put(align_bi);
5256                         rdev_dec_pending(rdev, mddev);
5257                         return 0;
5258                 }
5259
5260                 /* No reshape active, so we can trust rdev->data_offset */
5261                 align_bi->bi_iter.bi_sector += rdev->data_offset;
5262
5263                 spin_lock_irq(&conf->device_lock);
5264                 wait_event_lock_irq(conf->wait_for_quiescent,
5265                                     conf->quiesce == 0,
5266                                     conf->device_lock);
5267                 atomic_inc(&conf->active_aligned_reads);
5268                 spin_unlock_irq(&conf->device_lock);
5269
5270                 if (mddev->gendisk)
5271                         trace_block_bio_remap(align_bi->bi_disk->queue,
5272                                               align_bi, disk_devt(mddev->gendisk),
5273                                               raid_bio->bi_iter.bi_sector);
5274                 generic_make_request(align_bi);
5275                 return 1;
5276         } else {
5277                 rcu_read_unlock();
5278                 bio_put(align_bi);
5279                 return 0;
5280         }
5281 }
5282
5283 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5284 {
5285         struct bio *split;
5286         sector_t sector = raid_bio->bi_iter.bi_sector;
5287         unsigned chunk_sects = mddev->chunk_sectors;
5288         unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5289
5290         if (sectors < bio_sectors(raid_bio)) {
5291                 struct r5conf *conf = mddev->private;
5292                 split = bio_split(raid_bio, sectors, GFP_NOIO, &conf->bio_split);
5293                 bio_chain(split, raid_bio);
5294                 generic_make_request(raid_bio);
5295                 raid_bio = split;
5296         }
5297
5298         if (!raid5_read_one_chunk(mddev, raid_bio))
5299                 return raid_bio;
5300
5301         return NULL;
5302 }
5303
5304 /* __get_priority_stripe - get the next stripe to process
5305  *
5306  * Full stripe writes are allowed to pass preread active stripes up until
5307  * the bypass_threshold is exceeded.  In general the bypass_count
5308  * increments when the handle_list is handled before the hold_list; however, it
5309  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5310  * stripe with in flight i/o.  The bypass_count will be reset when the
5311  * head of the hold_list has changed, i.e. the head was promoted to the
5312  * handle_list.
5313  */
5314 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5315 {
5316         struct stripe_head *sh, *tmp;
5317         struct list_head *handle_list = NULL;
5318         struct r5worker_group *wg;
5319         bool second_try = !r5c_is_writeback(conf->log) &&
5320                 !r5l_log_disk_error(conf);
5321         bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5322                 r5l_log_disk_error(conf);
5323
5324 again:
5325         wg = NULL;
5326         sh = NULL;
5327         if (conf->worker_cnt_per_group == 0) {
5328                 handle_list = try_loprio ? &conf->loprio_list :
5329                                         &conf->handle_list;
5330         } else if (group != ANY_GROUP) {
5331                 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5332                                 &conf->worker_groups[group].handle_list;
5333                 wg = &conf->worker_groups[group];
5334         } else {
5335                 int i;
5336                 for (i = 0; i < conf->group_cnt; i++) {
5337                         handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5338                                 &conf->worker_groups[i].handle_list;
5339                         wg = &conf->worker_groups[i];
5340                         if (!list_empty(handle_list))
5341                                 break;
5342                 }
5343         }
5344
5345         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5346                   __func__,
5347                   list_empty(handle_list) ? "empty" : "busy",
5348                   list_empty(&conf->hold_list) ? "empty" : "busy",
5349                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
5350
5351         if (!list_empty(handle_list)) {
5352                 sh = list_entry(handle_list->next, typeof(*sh), lru);
5353
5354                 if (list_empty(&conf->hold_list))
5355                         conf->bypass_count = 0;
5356                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5357                         if (conf->hold_list.next == conf->last_hold)
5358                                 conf->bypass_count++;
5359                         else {
5360                                 conf->last_hold = conf->hold_list.next;
5361                                 conf->bypass_count -= conf->bypass_threshold;
5362                                 if (conf->bypass_count < 0)
5363                                         conf->bypass_count = 0;
5364                         }
5365                 }
5366         } else if (!list_empty(&conf->hold_list) &&
5367                    ((conf->bypass_threshold &&
5368                      conf->bypass_count > conf->bypass_threshold) ||
5369                     atomic_read(&conf->pending_full_writes) == 0)) {
5370
5371                 list_for_each_entry(tmp, &conf->hold_list,  lru) {
5372                         if (conf->worker_cnt_per_group == 0 ||
5373                             group == ANY_GROUP ||
5374                             !cpu_online(tmp->cpu) ||
5375                             cpu_to_group(tmp->cpu) == group) {
5376                                 sh = tmp;
5377                                 break;
5378                         }
5379                 }
5380
5381                 if (sh) {
5382                         conf->bypass_count -= conf->bypass_threshold;
5383                         if (conf->bypass_count < 0)
5384                                 conf->bypass_count = 0;
5385                 }
5386                 wg = NULL;
5387         }
5388
5389         if (!sh) {
5390                 if (second_try)
5391                         return NULL;
5392                 second_try = true;
5393                 try_loprio = !try_loprio;
5394                 goto again;
5395         }
5396
5397         if (wg) {
5398                 wg->stripes_cnt--;
5399                 sh->group = NULL;
5400         }
5401         list_del_init(&sh->lru);
5402         BUG_ON(atomic_inc_return(&sh->count) != 1);
5403         return sh;
5404 }
5405
5406 struct raid5_plug_cb {
5407         struct blk_plug_cb      cb;
5408         struct list_head        list;
5409         struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5410 };
5411
5412 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5413 {
5414         struct raid5_plug_cb *cb = container_of(
5415                 blk_cb, struct raid5_plug_cb, cb);
5416         struct stripe_head *sh;
5417         struct mddev *mddev = cb->cb.data;
5418         struct r5conf *conf = mddev->private;
5419         int cnt = 0;
5420         int hash;
5421
5422         if (cb->list.next && !list_empty(&cb->list)) {
5423                 spin_lock_irq(&conf->device_lock);
5424                 while (!list_empty(&cb->list)) {
5425                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
5426                         list_del_init(&sh->lru);
5427                         /*
5428                          * avoid race release_stripe_plug() sees
5429                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
5430                          * is still in our list
5431                          */
5432                         smp_mb__before_atomic();
5433                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5434                         /*
5435                          * STRIPE_ON_RELEASE_LIST could be set here. In that
5436                          * case, the count is always > 1 here
5437                          */
5438                         hash = sh->hash_lock_index;
5439                         __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5440                         cnt++;
5441                 }
5442                 spin_unlock_irq(&conf->device_lock);
5443         }
5444         release_inactive_stripe_list(conf, cb->temp_inactive_list,
5445                                      NR_STRIPE_HASH_LOCKS);
5446         if (mddev->queue)
5447                 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5448         kfree(cb);
5449 }
5450
5451 static void release_stripe_plug(struct mddev *mddev,
5452                                 struct stripe_head *sh)
5453 {
5454         struct blk_plug_cb *blk_cb = blk_check_plugged(
5455                 raid5_unplug, mddev,
5456                 sizeof(struct raid5_plug_cb));
5457         struct raid5_plug_cb *cb;
5458
5459         if (!blk_cb) {
5460                 raid5_release_stripe(sh);
5461                 return;
5462         }
5463
5464         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5465
5466         if (cb->list.next == NULL) {
5467                 int i;
5468                 INIT_LIST_HEAD(&cb->list);
5469                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5470                         INIT_LIST_HEAD(cb->temp_inactive_list + i);
5471         }
5472
5473         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5474                 list_add_tail(&sh->lru, &cb->list);
5475         else
5476                 raid5_release_stripe(sh);
5477 }
5478
5479 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5480 {
5481         struct r5conf *conf = mddev->private;
5482         sector_t logical_sector, last_sector;
5483         struct stripe_head *sh;
5484         int stripe_sectors;
5485
5486         if (mddev->reshape_position != MaxSector)
5487                 /* Skip discard while reshape is happening */
5488                 return;
5489
5490         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5491         last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5492
5493         bi->bi_next = NULL;
5494
5495         stripe_sectors = conf->chunk_sectors *
5496                 (conf->raid_disks - conf->max_degraded);
5497         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5498                                                stripe_sectors);
5499         sector_div(last_sector, stripe_sectors);
5500
5501         logical_sector *= conf->chunk_sectors;
5502         last_sector *= conf->chunk_sectors;
5503
5504         for (; logical_sector < last_sector;
5505              logical_sector += STRIPE_SECTORS) {
5506                 DEFINE_WAIT(w);
5507                 int d;
5508         again:
5509                 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5510                 prepare_to_wait(&conf->wait_for_overlap, &w,
5511                                 TASK_UNINTERRUPTIBLE);
5512                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5513                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5514                         raid5_release_stripe(sh);
5515                         schedule();
5516                         goto again;
5517                 }
5518                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5519                 spin_lock_irq(&sh->stripe_lock);
5520                 for (d = 0; d < conf->raid_disks; d++) {
5521                         if (d == sh->pd_idx || d == sh->qd_idx)
5522                                 continue;
5523                         if (sh->dev[d].towrite || sh->dev[d].toread) {
5524                                 set_bit(R5_Overlap, &sh->dev[d].flags);
5525                                 spin_unlock_irq(&sh->stripe_lock);
5526                                 raid5_release_stripe(sh);
5527                                 schedule();
5528                                 goto again;
5529                         }
5530                 }
5531                 set_bit(STRIPE_DISCARD, &sh->state);
5532                 finish_wait(&conf->wait_for_overlap, &w);
5533                 sh->overwrite_disks = 0;
5534                 for (d = 0; d < conf->raid_disks; d++) {
5535                         if (d == sh->pd_idx || d == sh->qd_idx)
5536                                 continue;
5537                         sh->dev[d].towrite = bi;
5538                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5539                         bio_inc_remaining(bi);
5540                         md_write_inc(mddev, bi);
5541                         sh->overwrite_disks++;
5542                 }
5543                 spin_unlock_irq(&sh->stripe_lock);
5544                 if (conf->mddev->bitmap) {
5545                         for (d = 0;
5546                              d < conf->raid_disks - conf->max_degraded;
5547                              d++)
5548                                 md_bitmap_startwrite(mddev->bitmap,
5549                                                      sh->sector,
5550                                                      STRIPE_SECTORS,
5551                                                      0);
5552                         sh->bm_seq = conf->seq_flush + 1;
5553                         set_bit(STRIPE_BIT_DELAY, &sh->state);
5554                 }
5555
5556                 set_bit(STRIPE_HANDLE, &sh->state);
5557                 clear_bit(STRIPE_DELAYED, &sh->state);
5558                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5559                         atomic_inc(&conf->preread_active_stripes);
5560                 release_stripe_plug(mddev, sh);
5561         }
5562
5563         bio_endio(bi);
5564 }
5565
5566 static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
5567 {
5568         struct r5conf *conf = mddev->private;
5569         int dd_idx;
5570         sector_t new_sector;
5571         sector_t logical_sector, last_sector;
5572         struct stripe_head *sh;
5573         const int rw = bio_data_dir(bi);
5574         DEFINE_WAIT(w);
5575         bool do_prepare;
5576         bool do_flush = false;
5577
5578         if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5579                 int ret = log_handle_flush_request(conf, bi);
5580
5581                 if (ret == 0)
5582                         return true;
5583                 if (ret == -ENODEV) {
5584                         md_flush_request(mddev, bi);
5585                         return true;
5586                 }
5587                 /* ret == -EAGAIN, fallback */
5588                 /*
5589                  * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5590                  * we need to flush journal device
5591                  */
5592                 do_flush = bi->bi_opf & REQ_PREFLUSH;
5593         }
5594
5595         if (!md_write_start(mddev, bi))
5596                 return false;
5597         /*
5598          * If array is degraded, better not do chunk aligned read because
5599          * later we might have to read it again in order to reconstruct
5600          * data on failed drives.
5601          */
5602         if (rw == READ && mddev->degraded == 0 &&
5603             mddev->reshape_position == MaxSector) {
5604                 bi = chunk_aligned_read(mddev, bi);
5605                 if (!bi)
5606                         return true;
5607         }
5608
5609         if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5610                 make_discard_request(mddev, bi);
5611                 md_write_end(mddev);
5612                 return true;
5613         }
5614
5615         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5616         last_sector = bio_end_sector(bi);
5617         bi->bi_next = NULL;
5618
5619         prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5620         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5621                 int previous;
5622                 int seq;
5623
5624                 do_prepare = false;
5625         retry:
5626                 seq = read_seqcount_begin(&conf->gen_lock);
5627                 previous = 0;
5628                 if (do_prepare)
5629                         prepare_to_wait(&conf->wait_for_overlap, &w,
5630                                 TASK_UNINTERRUPTIBLE);
5631                 if (unlikely(conf->reshape_progress != MaxSector)) {
5632                         /* spinlock is needed as reshape_progress may be
5633                          * 64bit on a 32bit platform, and so it might be
5634                          * possible to see a half-updated value
5635                          * Of course reshape_progress could change after
5636                          * the lock is dropped, so once we get a reference
5637                          * to the stripe that we think it is, we will have
5638                          * to check again.
5639                          */
5640                         spin_lock_irq(&conf->device_lock);
5641                         if (mddev->reshape_backwards
5642                             ? logical_sector < conf->reshape_progress
5643                             : logical_sector >= conf->reshape_progress) {
5644                                 previous = 1;
5645                         } else {
5646                                 if (mddev->reshape_backwards
5647                                     ? logical_sector < conf->reshape_safe
5648                                     : logical_sector >= conf->reshape_safe) {
5649                                         spin_unlock_irq(&conf->device_lock);
5650                                         schedule();
5651                                         do_prepare = true;
5652                                         goto retry;
5653                                 }
5654                         }
5655                         spin_unlock_irq(&conf->device_lock);
5656                 }
5657
5658                 new_sector = raid5_compute_sector(conf, logical_sector,
5659                                                   previous,
5660                                                   &dd_idx, NULL);
5661                 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5662                         (unsigned long long)new_sector,
5663                         (unsigned long long)logical_sector);
5664
5665                 sh = raid5_get_active_stripe(conf, new_sector, previous,
5666                                        (bi->bi_opf & REQ_RAHEAD), 0);
5667                 if (sh) {
5668                         if (unlikely(previous)) {
5669                                 /* expansion might have moved on while waiting for a
5670                                  * stripe, so we must do the range check again.
5671                                  * Expansion could still move past after this
5672                                  * test, but as we are holding a reference to
5673                                  * 'sh', we know that if that happens,
5674                                  *  STRIPE_EXPANDING will get set and the expansion
5675                                  * won't proceed until we finish with the stripe.
5676                                  */
5677                                 int must_retry = 0;
5678                                 spin_lock_irq(&conf->device_lock);
5679                                 if (mddev->reshape_backwards
5680                                     ? logical_sector >= conf->reshape_progress
5681                                     : logical_sector < conf->reshape_progress)
5682                                         /* mismatch, need to try again */
5683                                         must_retry = 1;
5684                                 spin_unlock_irq(&conf->device_lock);
5685                                 if (must_retry) {
5686                                         raid5_release_stripe(sh);
5687                                         schedule();
5688                                         do_prepare = true;
5689                                         goto retry;
5690                                 }
5691                         }
5692                         if (read_seqcount_retry(&conf->gen_lock, seq)) {
5693                                 /* Might have got the wrong stripe_head
5694                                  * by accident
5695                                  */
5696                                 raid5_release_stripe(sh);
5697                                 goto retry;
5698                         }
5699
5700                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5701                             !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5702                                 /* Stripe is busy expanding or
5703                                  * add failed due to overlap.  Flush everything
5704                                  * and wait a while
5705                                  */
5706                                 md_wakeup_thread(mddev->thread);
5707                                 raid5_release_stripe(sh);
5708                                 schedule();
5709                                 do_prepare = true;
5710                                 goto retry;
5711                         }
5712                         if (do_flush) {
5713                                 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5714                                 /* we only need flush for one stripe */
5715                                 do_flush = false;
5716                         }
5717
5718                         set_bit(STRIPE_HANDLE, &sh->state);
5719                         clear_bit(STRIPE_DELAYED, &sh->state);
5720                         if ((!sh->batch_head || sh == sh->batch_head) &&
5721                             (bi->bi_opf & REQ_SYNC) &&
5722                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5723                                 atomic_inc(&conf->preread_active_stripes);
5724                         release_stripe_plug(mddev, sh);
5725                 } else {
5726                         /* cannot get stripe for read-ahead, just give-up */
5727                         bi->bi_status = BLK_STS_IOERR;
5728                         break;
5729                 }
5730         }
5731         finish_wait(&conf->wait_for_overlap, &w);
5732
5733         if (rw == WRITE)
5734                 md_write_end(mddev);
5735         bio_endio(bi);
5736         return true;
5737 }
5738
5739 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5740
5741 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5742 {
5743         /* reshaping is quite different to recovery/resync so it is
5744          * handled quite separately ... here.
5745          *
5746          * On each call to sync_request, we gather one chunk worth of
5747          * destination stripes and flag them as expanding.
5748          * Then we find all the source stripes and request reads.
5749          * As the reads complete, handle_stripe will copy the data
5750          * into the destination stripe and release that stripe.
5751          */
5752         struct r5conf *conf = mddev->private;
5753         struct stripe_head *sh;
5754         struct md_rdev *rdev;
5755         sector_t first_sector, last_sector;
5756         int raid_disks = conf->previous_raid_disks;
5757         int data_disks = raid_disks - conf->max_degraded;
5758         int new_data_disks = conf->raid_disks - conf->max_degraded;
5759         int i;
5760         int dd_idx;
5761         sector_t writepos, readpos, safepos;
5762         sector_t stripe_addr;
5763         int reshape_sectors;
5764         struct list_head stripes;
5765         sector_t retn;
5766
5767         if (sector_nr == 0) {
5768                 /* If restarting in the middle, skip the initial sectors */
5769                 if (mddev->reshape_backwards &&
5770                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5771                         sector_nr = raid5_size(mddev, 0, 0)
5772                                 - conf->reshape_progress;
5773                 } else if (mddev->reshape_backwards &&
5774                            conf->reshape_progress == MaxSector) {
5775                         /* shouldn't happen, but just in case, finish up.*/
5776                         sector_nr = MaxSector;
5777                 } else if (!mddev->reshape_backwards &&
5778                            conf->reshape_progress > 0)
5779                         sector_nr = conf->reshape_progress;
5780                 sector_div(sector_nr, new_data_disks);
5781                 if (sector_nr) {
5782                         mddev->curr_resync_completed = sector_nr;
5783                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5784                         *skipped = 1;
5785                         retn = sector_nr;
5786                         goto finish;
5787                 }
5788         }
5789
5790         /* We need to process a full chunk at a time.
5791          * If old and new chunk sizes differ, we need to process the
5792          * largest of these
5793          */
5794
5795         reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5796
5797         /* We update the metadata at least every 10 seconds, or when
5798          * the data about to be copied would over-write the source of
5799          * the data at the front of the range.  i.e. one new_stripe
5800          * along from reshape_progress new_maps to after where
5801          * reshape_safe old_maps to
5802          */
5803         writepos = conf->reshape_progress;
5804         sector_div(writepos, new_data_disks);
5805         readpos = conf->reshape_progress;
5806         sector_div(readpos, data_disks);
5807         safepos = conf->reshape_safe;
5808         sector_div(safepos, data_disks);
5809         if (mddev->reshape_backwards) {
5810                 BUG_ON(writepos < reshape_sectors);
5811                 writepos -= reshape_sectors;
5812                 readpos += reshape_sectors;
5813                 safepos += reshape_sectors;
5814         } else {
5815                 writepos += reshape_sectors;
5816                 /* readpos and safepos are worst-case calculations.
5817                  * A negative number is overly pessimistic, and causes
5818                  * obvious problems for unsigned storage.  So clip to 0.
5819                  */
5820                 readpos -= min_t(sector_t, reshape_sectors, readpos);
5821                 safepos -= min_t(sector_t, reshape_sectors, safepos);
5822         }
5823
5824         /* Having calculated the 'writepos' possibly use it
5825          * to set 'stripe_addr' which is where we will write to.
5826          */
5827         if (mddev->reshape_backwards) {
5828                 BUG_ON(conf->reshape_progress == 0);
5829                 stripe_addr = writepos;
5830                 BUG_ON((mddev->dev_sectors &
5831                         ~((sector_t)reshape_sectors - 1))
5832                        - reshape_sectors - stripe_addr
5833                        != sector_nr);
5834         } else {
5835                 BUG_ON(writepos != sector_nr + reshape_sectors);
5836                 stripe_addr = sector_nr;
5837         }
5838
5839         /* 'writepos' is the most advanced device address we might write.
5840          * 'readpos' is the least advanced device address we might read.
5841          * 'safepos' is the least address recorded in the metadata as having
5842          *     been reshaped.
5843          * If there is a min_offset_diff, these are adjusted either by
5844          * increasing the safepos/readpos if diff is negative, or
5845          * increasing writepos if diff is positive.
5846          * If 'readpos' is then behind 'writepos', there is no way that we can
5847          * ensure safety in the face of a crash - that must be done by userspace
5848          * making a backup of the data.  So in that case there is no particular
5849          * rush to update metadata.
5850          * Otherwise if 'safepos' is behind 'writepos', then we really need to
5851          * update the metadata to advance 'safepos' to match 'readpos' so that
5852          * we can be safe in the event of a crash.
5853          * So we insist on updating metadata if safepos is behind writepos and
5854          * readpos is beyond writepos.
5855          * In any case, update the metadata every 10 seconds.
5856          * Maybe that number should be configurable, but I'm not sure it is
5857          * worth it.... maybe it could be a multiple of safemode_delay???
5858          */
5859         if (conf->min_offset_diff < 0) {
5860                 safepos += -conf->min_offset_diff;
5861                 readpos += -conf->min_offset_diff;
5862         } else
5863                 writepos += conf->min_offset_diff;
5864
5865         if ((mddev->reshape_backwards
5866              ? (safepos > writepos && readpos < writepos)
5867              : (safepos < writepos && readpos > writepos)) ||
5868             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5869                 /* Cannot proceed until we've updated the superblock... */
5870                 wait_event(conf->wait_for_overlap,
5871                            atomic_read(&conf->reshape_stripes)==0
5872                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5873                 if (atomic_read(&conf->reshape_stripes) != 0)
5874                         return 0;
5875                 mddev->reshape_position = conf->reshape_progress;
5876                 mddev->curr_resync_completed = sector_nr;
5877                 if (!mddev->reshape_backwards)
5878                         /* Can update recovery_offset */
5879                         rdev_for_each(rdev, mddev)
5880                                 if (rdev->raid_disk >= 0 &&
5881                                     !test_bit(Journal, &rdev->flags) &&
5882                                     !test_bit(In_sync, &rdev->flags) &&
5883                                     rdev->recovery_offset < sector_nr)
5884                                         rdev->recovery_offset = sector_nr;
5885
5886                 conf->reshape_checkpoint = jiffies;
5887                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5888                 md_wakeup_thread(mddev->thread);
5889                 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
5890                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5891                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5892                         return 0;
5893                 spin_lock_irq(&conf->device_lock);
5894                 conf->reshape_safe = mddev->reshape_position;
5895                 spin_unlock_irq(&conf->device_lock);
5896                 wake_up(&conf->wait_for_overlap);
5897                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5898         }
5899
5900         INIT_LIST_HEAD(&stripes);
5901         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5902                 int j;
5903                 int skipped_disk = 0;
5904                 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5905                 set_bit(STRIPE_EXPANDING, &sh->state);
5906                 atomic_inc(&conf->reshape_stripes);
5907                 /* If any of this stripe is beyond the end of the old
5908                  * array, then we need to zero those blocks
5909                  */
5910                 for (j=sh->disks; j--;) {
5911                         sector_t s;
5912                         if (j == sh->pd_idx)
5913                                 continue;
5914                         if (conf->level == 6 &&
5915                             j == sh->qd_idx)
5916                                 continue;
5917                         s = raid5_compute_blocknr(sh, j, 0);
5918                         if (s < raid5_size(mddev, 0, 0)) {
5919                                 skipped_disk = 1;
5920                                 continue;
5921                         }
5922                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5923                         set_bit(R5_Expanded, &sh->dev[j].flags);
5924                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
5925                 }
5926                 if (!skipped_disk) {
5927                         set_bit(STRIPE_EXPAND_READY, &sh->state);
5928                         set_bit(STRIPE_HANDLE, &sh->state);
5929                 }
5930                 list_add(&sh->lru, &stripes);
5931         }
5932         spin_lock_irq(&conf->device_lock);
5933         if (mddev->reshape_backwards)
5934                 conf->reshape_progress -= reshape_sectors * new_data_disks;
5935         else
5936                 conf->reshape_progress += reshape_sectors * new_data_disks;
5937         spin_unlock_irq(&conf->device_lock);
5938         /* Ok, those stripe are ready. We can start scheduling
5939          * reads on the source stripes.
5940          * The source stripes are determined by mapping the first and last
5941          * block on the destination stripes.
5942          */
5943         first_sector =
5944                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5945                                      1, &dd_idx, NULL);
5946         last_sector =
5947                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5948                                             * new_data_disks - 1),
5949                                      1, &dd_idx, NULL);
5950         if (last_sector >= mddev->dev_sectors)
5951                 last_sector = mddev->dev_sectors - 1;
5952         while (first_sector <= last_sector) {
5953                 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5954                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5955                 set_bit(STRIPE_HANDLE, &sh->state);
5956                 raid5_release_stripe(sh);
5957                 first_sector += STRIPE_SECTORS;
5958         }
5959         /* Now that the sources are clearly marked, we can release
5960          * the destination stripes
5961          */
5962         while (!list_empty(&stripes)) {
5963                 sh = list_entry(stripes.next, struct stripe_head, lru);
5964                 list_del_init(&sh->lru);
5965                 raid5_release_stripe(sh);
5966         }
5967         /* If this takes us to the resync_max point where we have to pause,
5968          * then we need to write out the superblock.
5969          */
5970         sector_nr += reshape_sectors;
5971         retn = reshape_sectors;
5972 finish:
5973         if (mddev->curr_resync_completed > mddev->resync_max ||
5974             (sector_nr - mddev->curr_resync_completed) * 2
5975             >= mddev->resync_max - mddev->curr_resync_completed) {
5976                 /* Cannot proceed until we've updated the superblock... */
5977                 wait_event(conf->wait_for_overlap,
5978                            atomic_read(&conf->reshape_stripes) == 0
5979                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5980                 if (atomic_read(&conf->reshape_stripes) != 0)
5981                         goto ret;
5982                 mddev->reshape_position = conf->reshape_progress;
5983                 mddev->curr_resync_completed = sector_nr;
5984                 if (!mddev->reshape_backwards)
5985                         /* Can update recovery_offset */
5986                         rdev_for_each(rdev, mddev)
5987                                 if (rdev->raid_disk >= 0 &&
5988                                     !test_bit(Journal, &rdev->flags) &&
5989                                     !test_bit(In_sync, &rdev->flags) &&
5990                                     rdev->recovery_offset < sector_nr)
5991                                         rdev->recovery_offset = sector_nr;
5992                 conf->reshape_checkpoint = jiffies;
5993                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5994                 md_wakeup_thread(mddev->thread);
5995                 wait_event(mddev->sb_wait,
5996                            !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
5997                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5998                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5999                         goto ret;
6000                 spin_lock_irq(&conf->device_lock);
6001                 conf->reshape_safe = mddev->reshape_position;
6002                 spin_unlock_irq(&conf->device_lock);
6003                 wake_up(&conf->wait_for_overlap);
6004                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
6005         }
6006 ret:
6007         return retn;
6008 }
6009
6010 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6011                                           int *skipped)
6012 {
6013         struct r5conf *conf = mddev->private;
6014         struct stripe_head *sh;
6015         sector_t max_sector = mddev->dev_sectors;
6016         sector_t sync_blocks;
6017         int still_degraded = 0;
6018         int i;
6019
6020         if (sector_nr >= max_sector) {
6021                 /* just being told to finish up .. nothing much to do */
6022
6023                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6024                         end_reshape(conf);
6025                         return 0;
6026                 }
6027
6028                 if (mddev->curr_resync < max_sector) /* aborted */
6029                         md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6030                                            &sync_blocks, 1);
6031                 else /* completed sync */
6032                         conf->fullsync = 0;
6033                 md_bitmap_close_sync(mddev->bitmap);
6034
6035                 return 0;
6036         }
6037
6038         /* Allow raid5_quiesce to complete */
6039         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6040
6041         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6042                 return reshape_request(mddev, sector_nr, skipped);
6043
6044         /* No need to check resync_max as we never do more than one
6045          * stripe, and as resync_max will always be on a chunk boundary,
6046          * if the check in md_do_sync didn't fire, there is no chance
6047          * of overstepping resync_max here
6048          */
6049
6050         /* if there is too many failed drives and we are trying
6051          * to resync, then assert that we are finished, because there is
6052          * nothing we can do.
6053          */
6054         if (mddev->degraded >= conf->max_degraded &&
6055             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6056                 sector_t rv = mddev->dev_sectors - sector_nr;
6057                 *skipped = 1;
6058                 return rv;
6059         }
6060         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6061             !conf->fullsync &&
6062             !md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6063             sync_blocks >= STRIPE_SECTORS) {
6064                 /* we can skip this block, and probably more */
6065                 sync_blocks /= STRIPE_SECTORS;
6066                 *skipped = 1;
6067                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
6068         }
6069
6070         md_bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6071
6072         sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
6073         if (sh == NULL) {
6074                 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
6075                 /* make sure we don't swamp the stripe cache if someone else
6076                  * is trying to get access
6077                  */
6078                 schedule_timeout_uninterruptible(1);
6079         }
6080         /* Need to check if array will still be degraded after recovery/resync
6081          * Note in case of > 1 drive failures it's possible we're rebuilding
6082          * one drive while leaving another faulty drive in array.
6083          */
6084         rcu_read_lock();
6085         for (i = 0; i < conf->raid_disks; i++) {
6086                 struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
6087
6088                 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6089                         still_degraded = 1;
6090         }
6091         rcu_read_unlock();
6092
6093         md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6094
6095         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6096         set_bit(STRIPE_HANDLE, &sh->state);
6097
6098         raid5_release_stripe(sh);
6099
6100         return STRIPE_SECTORS;
6101 }
6102
6103 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6104                                unsigned int offset)
6105 {
6106         /* We may not be able to submit a whole bio at once as there
6107          * may not be enough stripe_heads available.
6108          * We cannot pre-allocate enough stripe_heads as we may need
6109          * more than exist in the cache (if we allow ever large chunks).
6110          * So we do one stripe head at a time and record in
6111          * ->bi_hw_segments how many have been done.
6112          *
6113          * We *know* that this entire raid_bio is in one chunk, so
6114          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6115          */
6116         struct stripe_head *sh;
6117         int dd_idx;
6118         sector_t sector, logical_sector, last_sector;
6119         int scnt = 0;
6120         int handled = 0;
6121
6122         logical_sector = raid_bio->bi_iter.bi_sector &
6123                 ~((sector_t)STRIPE_SECTORS-1);
6124         sector = raid5_compute_sector(conf, logical_sector,
6125                                       0, &dd_idx, NULL);
6126         last_sector = bio_end_sector(raid_bio);
6127
6128         for (; logical_sector < last_sector;
6129              logical_sector += STRIPE_SECTORS,
6130                      sector += STRIPE_SECTORS,
6131                      scnt++) {
6132
6133                 if (scnt < offset)
6134                         /* already done this stripe */
6135                         continue;
6136
6137                 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
6138
6139                 if (!sh) {
6140                         /* failed to get a stripe - must wait */
6141                         conf->retry_read_aligned = raid_bio;
6142                         conf->retry_read_offset = scnt;
6143                         return handled;
6144                 }
6145
6146                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6147                         raid5_release_stripe(sh);
6148                         conf->retry_read_aligned = raid_bio;
6149                         conf->retry_read_offset = scnt;
6150                         return handled;
6151                 }
6152
6153                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6154                 handle_stripe(sh);
6155                 raid5_release_stripe(sh);
6156                 handled++;
6157         }
6158
6159         bio_endio(raid_bio);
6160
6161         if (atomic_dec_and_test(&conf->active_aligned_reads))
6162                 wake_up(&conf->wait_for_quiescent);
6163         return handled;
6164 }
6165
6166 static int handle_active_stripes(struct r5conf *conf, int group,
6167                                  struct r5worker *worker,
6168                                  struct list_head *temp_inactive_list)
6169 {
6170         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6171         int i, batch_size = 0, hash;
6172         bool release_inactive = false;
6173
6174         while (batch_size < MAX_STRIPE_BATCH &&
6175                         (sh = __get_priority_stripe(conf, group)) != NULL)
6176                 batch[batch_size++] = sh;
6177
6178         if (batch_size == 0) {
6179                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6180                         if (!list_empty(temp_inactive_list + i))
6181                                 break;
6182                 if (i == NR_STRIPE_HASH_LOCKS) {
6183                         spin_unlock_irq(&conf->device_lock);
6184                         log_flush_stripe_to_raid(conf);
6185                         spin_lock_irq(&conf->device_lock);
6186                         return batch_size;
6187                 }
6188                 release_inactive = true;
6189         }
6190         spin_unlock_irq(&conf->device_lock);
6191
6192         release_inactive_stripe_list(conf, temp_inactive_list,
6193                                      NR_STRIPE_HASH_LOCKS);
6194
6195         r5l_flush_stripe_to_raid(conf->log);
6196         if (release_inactive) {
6197                 spin_lock_irq(&conf->device_lock);
6198                 return 0;
6199         }
6200
6201         for (i = 0; i < batch_size; i++)
6202                 handle_stripe(batch[i]);
6203         log_write_stripe_run(conf);
6204
6205         cond_resched();
6206
6207         spin_lock_irq(&conf->device_lock);
6208         for (i = 0; i < batch_size; i++) {
6209                 hash = batch[i]->hash_lock_index;
6210                 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6211         }
6212         return batch_size;
6213 }
6214
6215 static void raid5_do_work(struct work_struct *work)
6216 {
6217         struct r5worker *worker = container_of(work, struct r5worker, work);
6218         struct r5worker_group *group = worker->group;
6219         struct r5conf *conf = group->conf;
6220         struct mddev *mddev = conf->mddev;
6221         int group_id = group - conf->worker_groups;
6222         int handled;
6223         struct blk_plug plug;
6224
6225         pr_debug("+++ raid5worker active\n");
6226
6227         blk_start_plug(&plug);
6228         handled = 0;
6229         spin_lock_irq(&conf->device_lock);
6230         while (1) {
6231                 int batch_size, released;
6232
6233                 released = release_stripe_list(conf, worker->temp_inactive_list);
6234
6235                 batch_size = handle_active_stripes(conf, group_id, worker,
6236                                                    worker->temp_inactive_list);
6237                 worker->working = false;
6238                 if (!batch_size && !released)
6239                         break;
6240                 handled += batch_size;
6241                 wait_event_lock_irq(mddev->sb_wait,
6242                         !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6243                         conf->device_lock);
6244         }
6245         pr_debug("%d stripes handled\n", handled);
6246
6247         spin_unlock_irq(&conf->device_lock);
6248
6249         flush_deferred_bios(conf);
6250
6251         r5l_flush_stripe_to_raid(conf->log);
6252
6253         async_tx_issue_pending_all();
6254         blk_finish_plug(&plug);
6255
6256         pr_debug("--- raid5worker inactive\n");
6257 }
6258
6259 /*
6260  * This is our raid5 kernel thread.
6261  *
6262  * We scan the hash table for stripes which can be handled now.
6263  * During the scan, completed stripes are saved for us by the interrupt
6264  * handler, so that they will not have to wait for our next wakeup.
6265  */
6266 static void raid5d(struct md_thread *thread)
6267 {
6268         struct mddev *mddev = thread->mddev;
6269         struct r5conf *conf = mddev->private;
6270         int handled;
6271         struct blk_plug plug;
6272
6273         pr_debug("+++ raid5d active\n");
6274
6275         md_check_recovery(mddev);
6276
6277         blk_start_plug(&plug);
6278         handled = 0;
6279         spin_lock_irq(&conf->device_lock);
6280         while (1) {
6281                 struct bio *bio;
6282                 int batch_size, released;
6283                 unsigned int offset;
6284
6285                 released = release_stripe_list(conf, conf->temp_inactive_list);
6286                 if (released)
6287                         clear_bit(R5_DID_ALLOC, &conf->cache_state);
6288
6289                 if (
6290                     !list_empty(&conf->bitmap_list)) {
6291                         /* Now is a good time to flush some bitmap updates */
6292                         conf->seq_flush++;
6293                         spin_unlock_irq(&conf->device_lock);
6294                         md_bitmap_unplug(mddev->bitmap);
6295                         spin_lock_irq(&conf->device_lock);
6296                         conf->seq_write = conf->seq_flush;
6297                         activate_bit_delay(conf, conf->temp_inactive_list);
6298                 }
6299                 raid5_activate_delayed(conf);
6300
6301                 while ((bio = remove_bio_from_retry(conf, &offset))) {
6302                         int ok;
6303                         spin_unlock_irq(&conf->device_lock);
6304                         ok = retry_aligned_read(conf, bio, offset);
6305                         spin_lock_irq(&conf->device_lock);
6306                         if (!ok)
6307                                 break;
6308                         handled++;
6309                 }
6310
6311                 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6312                                                    conf->temp_inactive_list);
6313                 if (!batch_size && !released)
6314                         break;
6315                 handled += batch_size;
6316
6317                 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6318                         spin_unlock_irq(&conf->device_lock);
6319                         md_check_recovery(mddev);
6320                         spin_lock_irq(&conf->device_lock);
6321                 }
6322         }
6323         pr_debug("%d stripes handled\n", handled);
6324
6325         spin_unlock_irq(&conf->device_lock);
6326         if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6327             mutex_trylock(&conf->cache_size_mutex)) {
6328                 grow_one_stripe(conf, __GFP_NOWARN);
6329                 /* Set flag even if allocation failed.  This helps
6330                  * slow down allocation requests when mem is short
6331                  */
6332                 set_bit(R5_DID_ALLOC, &conf->cache_state);
6333                 mutex_unlock(&conf->cache_size_mutex);
6334         }
6335
6336         flush_deferred_bios(conf);
6337
6338         r5l_flush_stripe_to_raid(conf->log);
6339
6340         async_tx_issue_pending_all();
6341         blk_finish_plug(&plug);
6342
6343         pr_debug("--- raid5d inactive\n");
6344 }
6345
6346 static ssize_t
6347 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6348 {
6349         struct r5conf *conf;
6350         int ret = 0;
6351         spin_lock(&mddev->lock);
6352         conf = mddev->private;
6353         if (conf)
6354                 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6355         spin_unlock(&mddev->lock);
6356         return ret;
6357 }
6358
6359 int
6360 raid5_set_cache_size(struct mddev *mddev, int size)
6361 {
6362         int result = 0;
6363         struct r5conf *conf = mddev->private;
6364
6365         if (size <= 16 || size > 32768)
6366                 return -EINVAL;
6367
6368         conf->min_nr_stripes = size;
6369         mutex_lock(&conf->cache_size_mutex);
6370         while (size < conf->max_nr_stripes &&
6371                drop_one_stripe(conf))
6372                 ;
6373         mutex_unlock(&conf->cache_size_mutex);
6374
6375         md_allow_write(mddev);
6376
6377         mutex_lock(&conf->cache_size_mutex);
6378         while (size > conf->max_nr_stripes)
6379                 if (!grow_one_stripe(conf, GFP_KERNEL)) {
6380                         conf->min_nr_stripes = conf->max_nr_stripes;
6381                         result = -ENOMEM;
6382                         break;
6383                 }
6384         mutex_unlock(&conf->cache_size_mutex);
6385
6386         return result;
6387 }
6388 EXPORT_SYMBOL(raid5_set_cache_size);
6389
6390 static ssize_t
6391 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6392 {
6393         struct r5conf *conf;
6394         unsigned long new;
6395         int err;
6396
6397         if (len >= PAGE_SIZE)
6398                 return -EINVAL;
6399         if (kstrtoul(page, 10, &new))
6400                 return -EINVAL;
6401         err = mddev_lock(mddev);
6402         if (err)
6403                 return err;
6404         conf = mddev->private;
6405         if (!conf)
6406                 err = -ENODEV;
6407         else
6408                 err = raid5_set_cache_size(mddev, new);
6409         mddev_unlock(mddev);
6410
6411         return err ?: len;
6412 }
6413
6414 static struct md_sysfs_entry
6415 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6416                                 raid5_show_stripe_cache_size,
6417                                 raid5_store_stripe_cache_size);
6418
6419 static ssize_t
6420 raid5_show_rmw_level(struct mddev  *mddev, char *page)
6421 {
6422         struct r5conf *conf = mddev->private;
6423         if (conf)
6424                 return sprintf(page, "%d\n", conf->rmw_level);
6425         else
6426                 return 0;
6427 }
6428
6429 static ssize_t
6430 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6431 {
6432         struct r5conf *conf = mddev->private;
6433         unsigned long new;
6434
6435         if (!conf)
6436                 return -ENODEV;
6437
6438         if (len >= PAGE_SIZE)
6439                 return -EINVAL;
6440
6441         if (kstrtoul(page, 10, &new))
6442                 return -EINVAL;
6443
6444         if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6445                 return -EINVAL;
6446
6447         if (new != PARITY_DISABLE_RMW &&
6448             new != PARITY_ENABLE_RMW &&
6449             new != PARITY_PREFER_RMW)
6450                 return -EINVAL;
6451
6452         conf->rmw_level = new;
6453         return len;
6454 }
6455
6456 static struct md_sysfs_entry
6457 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6458                          raid5_show_rmw_level,
6459                          raid5_store_rmw_level);
6460
6461
6462 static ssize_t
6463 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6464 {
6465         struct r5conf *conf;
6466         int ret = 0;
6467         spin_lock(&mddev->lock);
6468         conf = mddev->private;
6469         if (conf)
6470                 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6471         spin_unlock(&mddev->lock);
6472         return ret;
6473 }
6474
6475 static ssize_t
6476 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6477 {
6478         struct r5conf *conf;
6479         unsigned long new;
6480         int err;
6481
6482         if (len >= PAGE_SIZE)
6483                 return -EINVAL;
6484         if (kstrtoul(page, 10, &new))
6485                 return -EINVAL;
6486
6487         err = mddev_lock(mddev);
6488         if (err)
6489                 return err;
6490         conf = mddev->private;
6491         if (!conf)
6492                 err = -ENODEV;
6493         else if (new > conf->min_nr_stripes)
6494                 err = -EINVAL;
6495         else
6496                 conf->bypass_threshold = new;
6497         mddev_unlock(mddev);
6498         return err ?: len;
6499 }
6500
6501 static struct md_sysfs_entry
6502 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6503                                         S_IRUGO | S_IWUSR,
6504                                         raid5_show_preread_threshold,
6505                                         raid5_store_preread_threshold);
6506
6507 static ssize_t
6508 raid5_show_skip_copy(struct mddev *mddev, char *page)
6509 {
6510         struct r5conf *conf;
6511         int ret = 0;
6512         spin_lock(&mddev->lock);
6513         conf = mddev->private;
6514         if (conf)
6515                 ret = sprintf(page, "%d\n", conf->skip_copy);
6516         spin_unlock(&mddev->lock);
6517         return ret;
6518 }
6519
6520 static ssize_t
6521 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6522 {
6523         struct r5conf *conf;
6524         unsigned long new;
6525         int err;
6526
6527         if (len >= PAGE_SIZE)
6528                 return -EINVAL;
6529         if (kstrtoul(page, 10, &new))
6530                 return -EINVAL;
6531         new = !!new;
6532
6533         err = mddev_lock(mddev);
6534         if (err)
6535                 return err;
6536         conf = mddev->private;
6537         if (!conf)
6538                 err = -ENODEV;
6539         else if (new != conf->skip_copy) {
6540                 mddev_suspend(mddev);
6541                 conf->skip_copy = new;
6542                 if (new)
6543                         mddev->queue->backing_dev_info->capabilities |=
6544                                 BDI_CAP_STABLE_WRITES;
6545                 else
6546                         mddev->queue->backing_dev_info->capabilities &=
6547                                 ~BDI_CAP_STABLE_WRITES;
6548                 mddev_resume(mddev);
6549         }
6550         mddev_unlock(mddev);
6551         return err ?: len;
6552 }
6553
6554 static struct md_sysfs_entry
6555 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6556                                         raid5_show_skip_copy,
6557                                         raid5_store_skip_copy);
6558
6559 static ssize_t
6560 stripe_cache_active_show(struct mddev *mddev, char *page)
6561 {
6562         struct r5conf *conf = mddev->private;
6563         if (conf)
6564                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6565         else
6566                 return 0;
6567 }
6568
6569 static struct md_sysfs_entry
6570 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6571
6572 static ssize_t
6573 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6574 {
6575         struct r5conf *conf;
6576         int ret = 0;
6577         spin_lock(&mddev->lock);
6578         conf = mddev->private;
6579         if (conf)
6580                 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6581         spin_unlock(&mddev->lock);
6582         return ret;
6583 }
6584
6585 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6586                                int *group_cnt,
6587                                int *worker_cnt_per_group,
6588                                struct r5worker_group **worker_groups);
6589 static ssize_t
6590 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6591 {
6592         struct r5conf *conf;
6593         unsigned int new;
6594         int err;
6595         struct r5worker_group *new_groups, *old_groups;
6596         int group_cnt, worker_cnt_per_group;
6597
6598         if (len >= PAGE_SIZE)
6599                 return -EINVAL;
6600         if (kstrtouint(page, 10, &new))
6601                 return -EINVAL;
6602         /* 8192 should be big enough */
6603         if (new > 8192)
6604                 return -EINVAL;
6605
6606         err = mddev_lock(mddev);
6607         if (err)
6608                 return err;
6609         conf = mddev->private;
6610         if (!conf)
6611                 err = -ENODEV;
6612         else if (new != conf->worker_cnt_per_group) {
6613                 mddev_suspend(mddev);
6614
6615                 old_groups = conf->worker_groups;
6616                 if (old_groups)
6617                         flush_workqueue(raid5_wq);
6618
6619                 err = alloc_thread_groups(conf, new,
6620                                           &group_cnt, &worker_cnt_per_group,
6621                                           &new_groups);
6622                 if (!err) {
6623                         spin_lock_irq(&conf->device_lock);
6624                         conf->group_cnt = group_cnt;
6625                         conf->worker_cnt_per_group = worker_cnt_per_group;
6626                         conf->worker_groups = new_groups;
6627                         spin_unlock_irq(&conf->device_lock);
6628
6629                         if (old_groups)
6630                                 kfree(old_groups[0].workers);
6631                         kfree(old_groups);
6632                 }
6633                 mddev_resume(mddev);
6634         }
6635         mddev_unlock(mddev);
6636
6637         return err ?: len;
6638 }
6639
6640 static struct md_sysfs_entry
6641 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6642                                 raid5_show_group_thread_cnt,
6643                                 raid5_store_group_thread_cnt);
6644
6645 static struct attribute *raid5_attrs[] =  {
6646         &raid5_stripecache_size.attr,
6647         &raid5_stripecache_active.attr,
6648         &raid5_preread_bypass_threshold.attr,
6649         &raid5_group_thread_cnt.attr,
6650         &raid5_skip_copy.attr,
6651         &raid5_rmw_level.attr,
6652         &r5c_journal_mode.attr,
6653         NULL,
6654 };
6655 static struct attribute_group raid5_attrs_group = {
6656         .name = NULL,
6657         .attrs = raid5_attrs,
6658 };
6659
6660 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6661                                int *group_cnt,
6662                                int *worker_cnt_per_group,
6663                                struct r5worker_group **worker_groups)
6664 {
6665         int i, j, k;
6666         ssize_t size;
6667         struct r5worker *workers;
6668
6669         *worker_cnt_per_group = cnt;
6670         if (cnt == 0) {
6671                 *group_cnt = 0;
6672                 *worker_groups = NULL;
6673                 return 0;
6674         }
6675         *group_cnt = num_possible_nodes();
6676         size = sizeof(struct r5worker) * cnt;
6677         workers = kcalloc(size, *group_cnt, GFP_NOIO);
6678         *worker_groups = kcalloc(*group_cnt, sizeof(struct r5worker_group),
6679                                  GFP_NOIO);
6680         if (!*worker_groups || !workers) {
6681                 kfree(workers);
6682                 kfree(*worker_groups);
6683                 return -ENOMEM;
6684         }
6685
6686         for (i = 0; i < *group_cnt; i++) {
6687                 struct r5worker_group *group;
6688
6689                 group = &(*worker_groups)[i];
6690                 INIT_LIST_HEAD(&group->handle_list);
6691                 INIT_LIST_HEAD(&group->loprio_list);
6692                 group->conf = conf;
6693                 group->workers = workers + i * cnt;
6694
6695                 for (j = 0; j < cnt; j++) {
6696                         struct r5worker *worker = group->workers + j;
6697                         worker->group = group;
6698                         INIT_WORK(&worker->work, raid5_do_work);
6699
6700                         for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6701                                 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6702                 }
6703         }
6704
6705         return 0;
6706 }
6707
6708 static void free_thread_groups(struct r5conf *conf)
6709 {
6710         if (conf->worker_groups)
6711                 kfree(conf->worker_groups[0].workers);
6712         kfree(conf->worker_groups);
6713         conf->worker_groups = NULL;
6714 }
6715
6716 static sector_t
6717 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6718 {
6719         struct r5conf *conf = mddev->private;
6720
6721         if (!sectors)
6722                 sectors = mddev->dev_sectors;
6723         if (!raid_disks)
6724                 /* size is defined by the smallest of previous and new size */
6725                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6726
6727         sectors &= ~((sector_t)conf->chunk_sectors - 1);
6728         sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6729         return sectors * (raid_disks - conf->max_degraded);
6730 }
6731
6732 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6733 {
6734         safe_put_page(percpu->spare_page);
6735         percpu->spare_page = NULL;
6736         kvfree(percpu->scribble);
6737         percpu->scribble = NULL;
6738 }
6739
6740 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6741 {
6742         if (conf->level == 6 && !percpu->spare_page) {
6743                 percpu->spare_page = alloc_page(GFP_KERNEL);
6744                 if (!percpu->spare_page)
6745                         return -ENOMEM;
6746         }
6747
6748         if (scribble_alloc(percpu,
6749                            max(conf->raid_disks,
6750                                conf->previous_raid_disks),
6751                            max(conf->chunk_sectors,
6752                                conf->prev_chunk_sectors)
6753                            / STRIPE_SECTORS,
6754                            GFP_KERNEL)) {
6755                 free_scratch_buffer(conf, percpu);
6756                 return -ENOMEM;
6757         }
6758
6759         return 0;
6760 }
6761
6762 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
6763 {
6764         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6765
6766         free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6767         return 0;
6768 }
6769
6770 static void raid5_free_percpu(struct r5conf *conf)
6771 {
6772         if (!conf->percpu)
6773                 return;
6774
6775         cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6776         free_percpu(conf->percpu);
6777 }
6778
6779 static void free_conf(struct r5conf *conf)
6780 {
6781         int i;
6782
6783         log_exit(conf);
6784
6785         unregister_shrinker(&conf->shrinker);
6786         free_thread_groups(conf);
6787         shrink_stripes(conf);
6788         raid5_free_percpu(conf);
6789         for (i = 0; i < conf->pool_size; i++)
6790                 if (conf->disks[i].extra_page)
6791                         put_page(conf->disks[i].extra_page);
6792         kfree(conf->disks);
6793         bioset_exit(&conf->bio_split);
6794         kfree(conf->stripe_hashtbl);
6795         kfree(conf->pending_data);
6796         kfree(conf);
6797 }
6798
6799 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
6800 {
6801         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6802         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6803
6804         if (alloc_scratch_buffer(conf, percpu)) {
6805                 pr_warn("%s: failed memory allocation for cpu%u\n",
6806                         __func__, cpu);
6807                 return -ENOMEM;
6808         }
6809         return 0;
6810 }
6811
6812 static int raid5_alloc_percpu(struct r5conf *conf)
6813 {
6814         int err = 0;
6815
6816         conf->percpu = alloc_percpu(struct raid5_percpu);
6817         if (!conf->percpu)
6818                 return -ENOMEM;
6819
6820         err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6821         if (!err) {
6822                 conf->scribble_disks = max(conf->raid_disks,
6823                         conf->previous_raid_disks);
6824                 conf->scribble_sectors = max(conf->chunk_sectors,
6825                         conf->prev_chunk_sectors);
6826         }
6827         return err;
6828 }
6829
6830 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6831                                       struct shrink_control *sc)
6832 {
6833         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6834         unsigned long ret = SHRINK_STOP;
6835
6836         if (mutex_trylock(&conf->cache_size_mutex)) {
6837                 ret= 0;
6838                 while (ret < sc->nr_to_scan &&
6839                        conf->max_nr_stripes > conf->min_nr_stripes) {
6840                         if (drop_one_stripe(conf) == 0) {
6841                                 ret = SHRINK_STOP;
6842                                 break;
6843                         }
6844                         ret++;
6845                 }
6846                 mutex_unlock(&conf->cache_size_mutex);
6847         }
6848         return ret;
6849 }
6850
6851 static unsigned long raid5_cache_count(struct shrinker *shrink,
6852                                        struct shrink_control *sc)
6853 {
6854         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6855
6856         if (conf->max_nr_stripes < conf->min_nr_stripes)
6857                 /* unlikely, but not impossible */
6858                 return 0;
6859         return conf->max_nr_stripes - conf->min_nr_stripes;
6860 }
6861
6862 static struct r5conf *setup_conf(struct mddev *mddev)
6863 {
6864         struct r5conf *conf;
6865         int raid_disk, memory, max_disks;
6866         struct md_rdev *rdev;
6867         struct disk_info *disk;
6868         char pers_name[6];
6869         int i;
6870         int group_cnt, worker_cnt_per_group;
6871         struct r5worker_group *new_group;
6872         int ret;
6873
6874         if (mddev->new_level != 5
6875             && mddev->new_level != 4
6876             && mddev->new_level != 6) {
6877                 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6878                         mdname(mddev), mddev->new_level);
6879                 return ERR_PTR(-EIO);
6880         }
6881         if ((mddev->new_level == 5
6882              && !algorithm_valid_raid5(mddev->new_layout)) ||
6883             (mddev->new_level == 6
6884              && !algorithm_valid_raid6(mddev->new_layout))) {
6885                 pr_warn("md/raid:%s: layout %d not supported\n",
6886                         mdname(mddev), mddev->new_layout);
6887                 return ERR_PTR(-EIO);
6888         }
6889         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6890                 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6891                         mdname(mddev), mddev->raid_disks);
6892                 return ERR_PTR(-EINVAL);
6893         }
6894
6895         if (!mddev->new_chunk_sectors ||
6896             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6897             !is_power_of_2(mddev->new_chunk_sectors)) {
6898                 pr_warn("md/raid:%s: invalid chunk size %d\n",
6899                         mdname(mddev), mddev->new_chunk_sectors << 9);
6900                 return ERR_PTR(-EINVAL);
6901         }
6902
6903         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6904         if (conf == NULL)
6905                 goto abort;
6906         INIT_LIST_HEAD(&conf->free_list);
6907         INIT_LIST_HEAD(&conf->pending_list);
6908         conf->pending_data = kcalloc(PENDING_IO_MAX,
6909                                      sizeof(struct r5pending_data),
6910                                      GFP_KERNEL);
6911         if (!conf->pending_data)
6912                 goto abort;
6913         for (i = 0; i < PENDING_IO_MAX; i++)
6914                 list_add(&conf->pending_data[i].sibling, &conf->free_list);
6915         /* Don't enable multi-threading by default*/
6916         if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6917                                  &new_group)) {
6918                 conf->group_cnt = group_cnt;
6919                 conf->worker_cnt_per_group = worker_cnt_per_group;
6920                 conf->worker_groups = new_group;
6921         } else
6922                 goto abort;
6923         spin_lock_init(&conf->device_lock);
6924         seqcount_init(&conf->gen_lock);
6925         mutex_init(&conf->cache_size_mutex);
6926         init_waitqueue_head(&conf->wait_for_quiescent);
6927         init_waitqueue_head(&conf->wait_for_stripe);
6928         init_waitqueue_head(&conf->wait_for_overlap);
6929         INIT_LIST_HEAD(&conf->handle_list);
6930         INIT_LIST_HEAD(&conf->loprio_list);
6931         INIT_LIST_HEAD(&conf->hold_list);
6932         INIT_LIST_HEAD(&conf->delayed_list);
6933         INIT_LIST_HEAD(&conf->bitmap_list);
6934         init_llist_head(&conf->released_stripes);
6935         atomic_set(&conf->active_stripes, 0);
6936         atomic_set(&conf->preread_active_stripes, 0);
6937         atomic_set(&conf->active_aligned_reads, 0);
6938         spin_lock_init(&conf->pending_bios_lock);
6939         conf->batch_bio_dispatch = true;
6940         rdev_for_each(rdev, mddev) {
6941                 if (test_bit(Journal, &rdev->flags))
6942                         continue;
6943                 if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
6944                         conf->batch_bio_dispatch = false;
6945                         break;
6946                 }
6947         }
6948
6949         conf->bypass_threshold = BYPASS_THRESHOLD;
6950         conf->recovery_disabled = mddev->recovery_disabled - 1;
6951
6952         conf->raid_disks = mddev->raid_disks;
6953         if (mddev->reshape_position == MaxSector)
6954                 conf->previous_raid_disks = mddev->raid_disks;
6955         else
6956                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6957         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6958
6959         conf->disks = kcalloc(max_disks, sizeof(struct disk_info),
6960                               GFP_KERNEL);
6961
6962         if (!conf->disks)
6963                 goto abort;
6964
6965         for (i = 0; i < max_disks; i++) {
6966                 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
6967                 if (!conf->disks[i].extra_page)
6968                         goto abort;
6969         }
6970
6971         ret = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
6972         if (ret)
6973                 goto abort;
6974         conf->mddev = mddev;
6975
6976         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6977                 goto abort;
6978
6979         /* We init hash_locks[0] separately to that it can be used
6980          * as the reference lock in the spin_lock_nest_lock() call
6981          * in lock_all_device_hash_locks_irq in order to convince
6982          * lockdep that we know what we are doing.
6983          */
6984         spin_lock_init(conf->hash_locks);
6985         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6986                 spin_lock_init(conf->hash_locks + i);
6987
6988         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6989                 INIT_LIST_HEAD(conf->inactive_list + i);
6990
6991         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6992                 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6993
6994         atomic_set(&conf->r5c_cached_full_stripes, 0);
6995         INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
6996         atomic_set(&conf->r5c_cached_partial_stripes, 0);
6997         INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
6998         atomic_set(&conf->r5c_flushing_full_stripes, 0);
6999         atomic_set(&conf->r5c_flushing_partial_stripes, 0);
7000
7001         conf->level = mddev->new_level;
7002         conf->chunk_sectors = mddev->new_chunk_sectors;
7003         if (raid5_alloc_percpu(conf) != 0)
7004                 goto abort;
7005
7006         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
7007
7008         rdev_for_each(rdev, mddev) {
7009                 raid_disk = rdev->raid_disk;
7010                 if (raid_disk >= max_disks
7011                     || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7012                         continue;
7013                 disk = conf->disks + raid_disk;
7014
7015                 if (test_bit(Replacement, &rdev->flags)) {
7016                         if (disk->replacement)
7017                                 goto abort;
7018                         disk->replacement = rdev;
7019                 } else {
7020                         if (disk->rdev)
7021                                 goto abort;
7022                         disk->rdev = rdev;
7023                 }
7024
7025                 if (test_bit(In_sync, &rdev->flags)) {
7026                         char b[BDEVNAME_SIZE];
7027                         pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7028                                 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
7029                 } else if (rdev->saved_raid_disk != raid_disk)
7030                         /* Cannot rely on bitmap to complete recovery */
7031                         conf->fullsync = 1;
7032         }
7033
7034         conf->level = mddev->new_level;
7035         if (conf->level == 6) {
7036                 conf->max_degraded = 2;
7037                 if (raid6_call.xor_syndrome)
7038                         conf->rmw_level = PARITY_ENABLE_RMW;
7039                 else
7040                         conf->rmw_level = PARITY_DISABLE_RMW;
7041         } else {
7042                 conf->max_degraded = 1;
7043                 conf->rmw_level = PARITY_ENABLE_RMW;
7044         }
7045         conf->algorithm = mddev->new_layout;
7046         conf->reshape_progress = mddev->reshape_position;
7047         if (conf->reshape_progress != MaxSector) {
7048                 conf->prev_chunk_sectors = mddev->chunk_sectors;
7049                 conf->prev_algo = mddev->layout;
7050         } else {
7051                 conf->prev_chunk_sectors = conf->chunk_sectors;
7052                 conf->prev_algo = conf->algorithm;
7053         }
7054
7055         conf->min_nr_stripes = NR_STRIPES;
7056         if (mddev->reshape_position != MaxSector) {
7057                 int stripes = max_t(int,
7058                         ((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4,
7059                         ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4);
7060                 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7061                 if (conf->min_nr_stripes != NR_STRIPES)
7062                         pr_info("md/raid:%s: force stripe size %d for reshape\n",
7063                                 mdname(mddev), conf->min_nr_stripes);
7064         }
7065         memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7066                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7067         atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7068         if (grow_stripes(conf, conf->min_nr_stripes)) {
7069                 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7070                         mdname(mddev), memory);
7071                 goto abort;
7072         } else
7073                 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7074         /*
7075          * Losing a stripe head costs more than the time to refill it,
7076          * it reduces the queue depth and so can hurt throughput.
7077          * So set it rather large, scaled by number of devices.
7078          */
7079         conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7080         conf->shrinker.scan_objects = raid5_cache_scan;
7081         conf->shrinker.count_objects = raid5_cache_count;
7082         conf->shrinker.batch = 128;
7083         conf->shrinker.flags = 0;
7084         if (register_shrinker(&conf->shrinker)) {
7085                 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7086                         mdname(mddev));
7087                 goto abort;
7088         }
7089
7090         sprintf(pers_name, "raid%d", mddev->new_level);
7091         conf->thread = md_register_thread(raid5d, mddev, pers_name);
7092         if (!conf->thread) {
7093                 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7094                         mdname(mddev));
7095                 goto abort;
7096         }
7097
7098         return conf;
7099
7100  abort:
7101         if (conf) {
7102                 free_conf(conf);
7103                 return ERR_PTR(-EIO);
7104         } else
7105                 return ERR_PTR(-ENOMEM);
7106 }
7107
7108 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7109 {
7110         switch (algo) {
7111         case ALGORITHM_PARITY_0:
7112                 if (raid_disk < max_degraded)
7113                         return 1;
7114                 break;
7115         case ALGORITHM_PARITY_N:
7116                 if (raid_disk >= raid_disks - max_degraded)
7117                         return 1;
7118                 break;
7119         case ALGORITHM_PARITY_0_6:
7120                 if (raid_disk == 0 ||
7121                     raid_disk == raid_disks - 1)
7122                         return 1;
7123                 break;
7124         case ALGORITHM_LEFT_ASYMMETRIC_6:
7125         case ALGORITHM_RIGHT_ASYMMETRIC_6:
7126         case ALGORITHM_LEFT_SYMMETRIC_6:
7127         case ALGORITHM_RIGHT_SYMMETRIC_6:
7128                 if (raid_disk == raid_disks - 1)
7129                         return 1;
7130         }
7131         return 0;
7132 }
7133
7134 static int raid5_run(struct mddev *mddev)
7135 {
7136         struct r5conf *conf;
7137         int working_disks = 0;
7138         int dirty_parity_disks = 0;
7139         struct md_rdev *rdev;
7140         struct md_rdev *journal_dev = NULL;
7141         sector_t reshape_offset = 0;
7142         int i;
7143         long long min_offset_diff = 0;
7144         int first = 1;
7145
7146         if (mddev_init_writes_pending(mddev) < 0)
7147                 return -ENOMEM;
7148
7149         if (mddev->recovery_cp != MaxSector)
7150                 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7151                           mdname(mddev));
7152
7153         rdev_for_each(rdev, mddev) {
7154                 long long diff;
7155
7156                 if (test_bit(Journal, &rdev->flags)) {
7157                         journal_dev = rdev;
7158                         continue;
7159                 }
7160                 if (rdev->raid_disk < 0)
7161                         continue;
7162                 diff = (rdev->new_data_offset - rdev->data_offset);
7163                 if (first) {
7164                         min_offset_diff = diff;
7165                         first = 0;
7166                 } else if (mddev->reshape_backwards &&
7167                          diff < min_offset_diff)
7168                         min_offset_diff = diff;
7169                 else if (!mddev->reshape_backwards &&
7170                          diff > min_offset_diff)
7171                         min_offset_diff = diff;
7172         }
7173
7174         if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7175             (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7176                 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7177                           mdname(mddev));
7178                 return -EINVAL;
7179         }
7180
7181         if (mddev->reshape_position != MaxSector) {
7182                 /* Check that we can continue the reshape.
7183                  * Difficulties arise if the stripe we would write to
7184                  * next is at or after the stripe we would read from next.
7185                  * For a reshape that changes the number of devices, this
7186                  * is only possible for a very short time, and mdadm makes
7187                  * sure that time appears to have past before assembling
7188                  * the array.  So we fail if that time hasn't passed.
7189                  * For a reshape that keeps the number of devices the same
7190                  * mdadm must be monitoring the reshape can keeping the
7191                  * critical areas read-only and backed up.  It will start
7192                  * the array in read-only mode, so we check for that.
7193                  */
7194                 sector_t here_new, here_old;
7195                 int old_disks;
7196                 int max_degraded = (mddev->level == 6 ? 2 : 1);
7197                 int chunk_sectors;
7198                 int new_data_disks;
7199
7200                 if (journal_dev) {
7201                         pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7202                                 mdname(mddev));
7203                         return -EINVAL;
7204                 }
7205
7206                 if (mddev->new_level != mddev->level) {
7207                         pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7208                                 mdname(mddev));
7209                         return -EINVAL;
7210                 }
7211                 old_disks = mddev->raid_disks - mddev->delta_disks;
7212                 /* reshape_position must be on a new-stripe boundary, and one
7213                  * further up in new geometry must map after here in old
7214                  * geometry.
7215                  * If the chunk sizes are different, then as we perform reshape
7216                  * in units of the largest of the two, reshape_position needs
7217                  * be a multiple of the largest chunk size times new data disks.
7218                  */
7219                 here_new = mddev->reshape_position;
7220                 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7221                 new_data_disks = mddev->raid_disks - max_degraded;
7222                 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7223                         pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7224                                 mdname(mddev));
7225                         return -EINVAL;
7226                 }
7227                 reshape_offset = here_new * chunk_sectors;
7228                 /* here_new is the stripe we will write to */
7229                 here_old = mddev->reshape_position;
7230                 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7231                 /* here_old is the first stripe that we might need to read
7232                  * from */
7233                 if (mddev->delta_disks == 0) {
7234                         /* We cannot be sure it is safe to start an in-place
7235                          * reshape.  It is only safe if user-space is monitoring
7236                          * and taking constant backups.
7237                          * mdadm always starts a situation like this in
7238                          * readonly mode so it can take control before
7239                          * allowing any writes.  So just check for that.
7240                          */
7241                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7242                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
7243                                 /* not really in-place - so OK */;
7244                         else if (mddev->ro == 0) {
7245                                 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7246                                         mdname(mddev));
7247                                 return -EINVAL;
7248                         }
7249                 } else if (mddev->reshape_backwards
7250                     ? (here_new * chunk_sectors + min_offset_diff <=
7251                        here_old * chunk_sectors)
7252                     : (here_new * chunk_sectors >=
7253                        here_old * chunk_sectors + (-min_offset_diff))) {
7254                         /* Reading from the same stripe as writing to - bad */
7255                         pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7256                                 mdname(mddev));
7257                         return -EINVAL;
7258                 }
7259                 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7260                 /* OK, we should be able to continue; */
7261         } else {
7262                 BUG_ON(mddev->level != mddev->new_level);
7263                 BUG_ON(mddev->layout != mddev->new_layout);
7264                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7265                 BUG_ON(mddev->delta_disks != 0);
7266         }
7267
7268         if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7269             test_bit(MD_HAS_PPL, &mddev->flags)) {
7270                 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7271                         mdname(mddev));
7272                 clear_bit(MD_HAS_PPL, &mddev->flags);
7273                 clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7274         }
7275
7276         if (mddev->private == NULL)
7277                 conf = setup_conf(mddev);
7278         else
7279                 conf = mddev->private;
7280
7281         if (IS_ERR(conf))
7282                 return PTR_ERR(conf);
7283
7284         if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7285                 if (!journal_dev) {
7286                         pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7287                                 mdname(mddev));
7288                         mddev->ro = 1;
7289                         set_disk_ro(mddev->gendisk, 1);
7290                 } else if (mddev->recovery_cp == MaxSector)
7291                         set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7292         }
7293
7294         conf->min_offset_diff = min_offset_diff;
7295         mddev->thread = conf->thread;
7296         conf->thread = NULL;
7297         mddev->private = conf;
7298
7299         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7300              i++) {
7301                 rdev = conf->disks[i].rdev;
7302                 if (!rdev && conf->disks[i].replacement) {
7303                         /* The replacement is all we have yet */
7304                         rdev = conf->disks[i].replacement;
7305                         conf->disks[i].replacement = NULL;
7306                         clear_bit(Replacement, &rdev->flags);
7307                         conf->disks[i].rdev = rdev;
7308                 }
7309                 if (!rdev)
7310                         continue;
7311                 if (conf->disks[i].replacement &&
7312                     conf->reshape_progress != MaxSector) {
7313                         /* replacements and reshape simply do not mix. */
7314                         pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7315                         goto abort;
7316                 }
7317                 if (test_bit(In_sync, &rdev->flags)) {
7318                         working_disks++;
7319                         continue;
7320                 }
7321                 /* This disc is not fully in-sync.  However if it
7322                  * just stored parity (beyond the recovery_offset),
7323                  * when we don't need to be concerned about the
7324                  * array being dirty.
7325                  * When reshape goes 'backwards', we never have
7326                  * partially completed devices, so we only need
7327                  * to worry about reshape going forwards.
7328                  */
7329                 /* Hack because v0.91 doesn't store recovery_offset properly. */
7330                 if (mddev->major_version == 0 &&
7331                     mddev->minor_version > 90)
7332                         rdev->recovery_offset = reshape_offset;
7333
7334                 if (rdev->recovery_offset < reshape_offset) {
7335                         /* We need to check old and new layout */
7336                         if (!only_parity(rdev->raid_disk,
7337                                          conf->algorithm,
7338                                          conf->raid_disks,
7339                                          conf->max_degraded))
7340                                 continue;
7341                 }
7342                 if (!only_parity(rdev->raid_disk,
7343                                  conf->prev_algo,
7344                                  conf->previous_raid_disks,
7345                                  conf->max_degraded))
7346                         continue;
7347                 dirty_parity_disks++;
7348         }
7349
7350         /*
7351          * 0 for a fully functional array, 1 or 2 for a degraded array.
7352          */
7353         mddev->degraded = raid5_calc_degraded(conf);
7354
7355         if (has_failed(conf)) {
7356                 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7357                         mdname(mddev), mddev->degraded, conf->raid_disks);
7358                 goto abort;
7359         }
7360
7361         /* device size must be a multiple of chunk size */
7362         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
7363         mddev->resync_max_sectors = mddev->dev_sectors;
7364
7365         if (mddev->degraded > dirty_parity_disks &&
7366             mddev->recovery_cp != MaxSector) {
7367                 if (test_bit(MD_HAS_PPL, &mddev->flags))
7368                         pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7369                                 mdname(mddev));
7370                 else if (mddev->ok_start_degraded)
7371                         pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7372                                 mdname(mddev));
7373                 else {
7374                         pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7375                                 mdname(mddev));
7376                         goto abort;
7377                 }
7378         }
7379
7380         pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7381                 mdname(mddev), conf->level,
7382                 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7383                 mddev->new_layout);
7384
7385         print_raid5_conf(conf);
7386
7387         if (conf->reshape_progress != MaxSector) {
7388                 conf->reshape_safe = conf->reshape_progress;
7389                 atomic_set(&conf->reshape_stripes, 0);
7390                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7391                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7392                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7393                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7394                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7395                                                         "reshape");
7396         }
7397
7398         /* Ok, everything is just fine now */
7399         if (mddev->to_remove == &raid5_attrs_group)
7400                 mddev->to_remove = NULL;
7401         else if (mddev->kobj.sd &&
7402             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7403                 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7404                         mdname(mddev));
7405         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7406
7407         if (mddev->queue) {
7408                 int chunk_size;
7409                 /* read-ahead size must cover two whole stripes, which
7410                  * is 2 * (datadisks) * chunksize where 'n' is the
7411                  * number of raid devices
7412                  */
7413                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7414                 int stripe = data_disks *
7415                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7416                 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7417                         mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7418
7419                 chunk_size = mddev->chunk_sectors << 9;
7420                 blk_queue_io_min(mddev->queue, chunk_size);
7421                 blk_queue_io_opt(mddev->queue, chunk_size *
7422                                  (conf->raid_disks - conf->max_degraded));
7423                 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7424                 /*
7425                  * We can only discard a whole stripe. It doesn't make sense to
7426                  * discard data disk but write parity disk
7427                  */
7428                 stripe = stripe * PAGE_SIZE;
7429                 /* Round up to power of 2, as discard handling
7430                  * currently assumes that */
7431                 while ((stripe-1) & stripe)
7432                         stripe = (stripe | (stripe-1)) + 1;
7433                 mddev->queue->limits.discard_alignment = stripe;
7434                 mddev->queue->limits.discard_granularity = stripe;
7435
7436                 blk_queue_max_write_same_sectors(mddev->queue, 0);
7437                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
7438
7439                 rdev_for_each(rdev, mddev) {
7440                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7441                                           rdev->data_offset << 9);
7442                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7443                                           rdev->new_data_offset << 9);
7444                 }
7445
7446                 /*
7447                  * zeroing is required, otherwise data
7448                  * could be lost. Consider a scenario: discard a stripe
7449                  * (the stripe could be inconsistent if
7450                  * discard_zeroes_data is 0); write one disk of the
7451                  * stripe (the stripe could be inconsistent again
7452                  * depending on which disks are used to calculate
7453                  * parity); the disk is broken; The stripe data of this
7454                  * disk is lost.
7455                  *
7456                  * We only allow DISCARD if the sysadmin has confirmed that
7457                  * only safe devices are in use by setting a module parameter.
7458                  * A better idea might be to turn DISCARD into WRITE_ZEROES
7459                  * requests, as that is required to be safe.
7460                  */
7461                 if (devices_handle_discard_safely &&
7462                     mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7463                     mddev->queue->limits.discard_granularity >= stripe)
7464                         blk_queue_flag_set(QUEUE_FLAG_DISCARD,
7465                                                 mddev->queue);
7466                 else
7467                         blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
7468                                                 mddev->queue);
7469
7470                 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7471         }
7472
7473         if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
7474                 goto abort;
7475
7476         return 0;
7477 abort:
7478         md_unregister_thread(&mddev->thread);
7479         print_raid5_conf(conf);
7480         free_conf(conf);
7481         mddev->private = NULL;
7482         pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7483         return -EIO;
7484 }
7485
7486 static void raid5_free(struct mddev *mddev, void *priv)
7487 {
7488         struct r5conf *conf = priv;
7489
7490         free_conf(conf);
7491         mddev->to_remove = &raid5_attrs_group;
7492 }
7493
7494 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7495 {
7496         struct r5conf *conf = mddev->private;
7497         int i;
7498
7499         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7500                 conf->chunk_sectors / 2, mddev->layout);
7501         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7502         rcu_read_lock();
7503         for (i = 0; i < conf->raid_disks; i++) {
7504                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7505                 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7506         }
7507         rcu_read_unlock();
7508         seq_printf (seq, "]");
7509 }
7510
7511 static void print_raid5_conf (struct r5conf *conf)
7512 {
7513         int i;
7514         struct disk_info *tmp;
7515
7516         pr_debug("RAID conf printout:\n");
7517         if (!conf) {
7518                 pr_debug("(conf==NULL)\n");
7519                 return;
7520         }
7521         pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7522                conf->raid_disks,
7523                conf->raid_disks - conf->mddev->degraded);
7524
7525         for (i = 0; i < conf->raid_disks; i++) {
7526                 char b[BDEVNAME_SIZE];
7527                 tmp = conf->disks + i;
7528                 if (tmp->rdev)
7529                         pr_debug(" disk %d, o:%d, dev:%s\n",
7530                                i, !test_bit(Faulty, &tmp->rdev->flags),
7531                                bdevname(tmp->rdev->bdev, b));
7532         }
7533 }
7534
7535 static int raid5_spare_active(struct mddev *mddev)
7536 {
7537         int i;
7538         struct r5conf *conf = mddev->private;
7539         struct disk_info *tmp;
7540         int count = 0;
7541         unsigned long flags;
7542
7543         for (i = 0; i < conf->raid_disks; i++) {
7544                 tmp = conf->disks + i;
7545                 if (tmp->replacement
7546                     && tmp->replacement->recovery_offset == MaxSector
7547                     && !test_bit(Faulty, &tmp->replacement->flags)
7548                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7549                         /* Replacement has just become active. */
7550                         if (!tmp->rdev
7551                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7552                                 count++;
7553                         if (tmp->rdev) {
7554                                 /* Replaced device not technically faulty,
7555                                  * but we need to be sure it gets removed
7556                                  * and never re-added.
7557                                  */
7558                                 set_bit(Faulty, &tmp->rdev->flags);
7559                                 sysfs_notify_dirent_safe(
7560                                         tmp->rdev->sysfs_state);
7561                         }
7562                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7563                 } else if (tmp->rdev
7564                     && tmp->rdev->recovery_offset == MaxSector
7565                     && !test_bit(Faulty, &tmp->rdev->flags)
7566                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7567                         count++;
7568                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7569                 }
7570         }
7571         spin_lock_irqsave(&conf->device_lock, flags);
7572         mddev->degraded = raid5_calc_degraded(conf);
7573         spin_unlock_irqrestore(&conf->device_lock, flags);
7574         print_raid5_conf(conf);
7575         return count;
7576 }
7577
7578 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7579 {
7580         struct r5conf *conf = mddev->private;
7581         int err = 0;
7582         int number = rdev->raid_disk;
7583         struct md_rdev **rdevp;
7584         struct disk_info *p = conf->disks + number;
7585
7586         print_raid5_conf(conf);
7587         if (test_bit(Journal, &rdev->flags) && conf->log) {
7588                 /*
7589                  * we can't wait pending write here, as this is called in
7590                  * raid5d, wait will deadlock.
7591                  * neilb: there is no locking about new writes here,
7592                  * so this cannot be safe.
7593                  */
7594                 if (atomic_read(&conf->active_stripes) ||
7595                     atomic_read(&conf->r5c_cached_full_stripes) ||
7596                     atomic_read(&conf->r5c_cached_partial_stripes)) {
7597                         return -EBUSY;
7598                 }
7599                 log_exit(conf);
7600                 return 0;
7601         }
7602         if (rdev == p->rdev)
7603                 rdevp = &p->rdev;
7604         else if (rdev == p->replacement)
7605                 rdevp = &p->replacement;
7606         else
7607                 return 0;
7608
7609         if (number >= conf->raid_disks &&
7610             conf->reshape_progress == MaxSector)
7611                 clear_bit(In_sync, &rdev->flags);
7612
7613         if (test_bit(In_sync, &rdev->flags) ||
7614             atomic_read(&rdev->nr_pending)) {
7615                 err = -EBUSY;
7616                 goto abort;
7617         }
7618         /* Only remove non-faulty devices if recovery
7619          * isn't possible.
7620          */
7621         if (!test_bit(Faulty, &rdev->flags) &&
7622             mddev->recovery_disabled != conf->recovery_disabled &&
7623             !has_failed(conf) &&
7624             (!p->replacement || p->replacement == rdev) &&
7625             number < conf->raid_disks) {
7626                 err = -EBUSY;
7627                 goto abort;
7628         }
7629         *rdevp = NULL;
7630         if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7631                 synchronize_rcu();
7632                 if (atomic_read(&rdev->nr_pending)) {
7633                         /* lost the race, try later */
7634                         err = -EBUSY;
7635                         *rdevp = rdev;
7636                 }
7637         }
7638         if (!err) {
7639                 err = log_modify(conf, rdev, false);
7640                 if (err)
7641                         goto abort;
7642         }
7643         if (p->replacement) {
7644                 /* We must have just cleared 'rdev' */
7645                 p->rdev = p->replacement;
7646                 clear_bit(Replacement, &p->replacement->flags);
7647                 smp_mb(); /* Make sure other CPUs may see both as identical
7648                            * but will never see neither - if they are careful
7649                            */
7650                 p->replacement = NULL;
7651
7652                 if (!err)
7653                         err = log_modify(conf, p->rdev, true);
7654         }
7655
7656         clear_bit(WantReplacement, &rdev->flags);
7657 abort:
7658
7659         print_raid5_conf(conf);
7660         return err;
7661 }
7662
7663 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7664 {
7665         struct r5conf *conf = mddev->private;
7666         int err = -EEXIST;
7667         int disk;
7668         struct disk_info *p;
7669         int first = 0;
7670         int last = conf->raid_disks - 1;
7671
7672         if (test_bit(Journal, &rdev->flags)) {
7673                 if (conf->log)
7674                         return -EBUSY;
7675
7676                 rdev->raid_disk = 0;
7677                 /*
7678                  * The array is in readonly mode if journal is missing, so no
7679                  * write requests running. We should be safe
7680                  */
7681                 log_init(conf, rdev, false);
7682                 return 0;
7683         }
7684         if (mddev->recovery_disabled == conf->recovery_disabled)
7685                 return -EBUSY;
7686
7687         if (rdev->saved_raid_disk < 0 && has_failed(conf))
7688                 /* no point adding a device */
7689                 return -EINVAL;
7690
7691         if (rdev->raid_disk >= 0)
7692                 first = last = rdev->raid_disk;
7693
7694         /*
7695          * find the disk ... but prefer rdev->saved_raid_disk
7696          * if possible.
7697          */
7698         if (rdev->saved_raid_disk >= 0 &&
7699             rdev->saved_raid_disk >= first &&
7700             conf->disks[rdev->saved_raid_disk].rdev == NULL)
7701                 first = rdev->saved_raid_disk;
7702
7703         for (disk = first; disk <= last; disk++) {
7704                 p = conf->disks + disk;
7705                 if (p->rdev == NULL) {
7706                         clear_bit(In_sync, &rdev->flags);
7707                         rdev->raid_disk = disk;
7708                         if (rdev->saved_raid_disk != disk)
7709                                 conf->fullsync = 1;
7710                         rcu_assign_pointer(p->rdev, rdev);
7711
7712                         err = log_modify(conf, rdev, true);
7713
7714                         goto out;
7715                 }
7716         }
7717         for (disk = first; disk <= last; disk++) {
7718                 p = conf->disks + disk;
7719                 if (test_bit(WantReplacement, &p->rdev->flags) &&
7720                     p->replacement == NULL) {
7721                         clear_bit(In_sync, &rdev->flags);
7722                         set_bit(Replacement, &rdev->flags);
7723                         rdev->raid_disk = disk;
7724                         err = 0;
7725                         conf->fullsync = 1;
7726                         rcu_assign_pointer(p->replacement, rdev);
7727                         break;
7728                 }
7729         }
7730 out:
7731         print_raid5_conf(conf);
7732         return err;
7733 }
7734
7735 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7736 {
7737         /* no resync is happening, and there is enough space
7738          * on all devices, so we can resize.
7739          * We need to make sure resync covers any new space.
7740          * If the array is shrinking we should possibly wait until
7741          * any io in the removed space completes, but it hardly seems
7742          * worth it.
7743          */
7744         sector_t newsize;
7745         struct r5conf *conf = mddev->private;
7746
7747         if (raid5_has_log(conf) || raid5_has_ppl(conf))
7748                 return -EINVAL;
7749         sectors &= ~((sector_t)conf->chunk_sectors - 1);
7750         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7751         if (mddev->external_size &&
7752             mddev->array_sectors > newsize)
7753                 return -EINVAL;
7754         if (mddev->bitmap) {
7755                 int ret = md_bitmap_resize(mddev->bitmap, sectors, 0, 0);
7756                 if (ret)
7757                         return ret;
7758         }
7759         md_set_array_sectors(mddev, newsize);
7760         if (sectors > mddev->dev_sectors &&
7761             mddev->recovery_cp > mddev->dev_sectors) {
7762                 mddev->recovery_cp = mddev->dev_sectors;
7763                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7764         }
7765         mddev->dev_sectors = sectors;
7766         mddev->resync_max_sectors = sectors;
7767         return 0;
7768 }
7769
7770 static int check_stripe_cache(struct mddev *mddev)
7771 {
7772         /* Can only proceed if there are plenty of stripe_heads.
7773          * We need a minimum of one full stripe,, and for sensible progress
7774          * it is best to have about 4 times that.
7775          * If we require 4 times, then the default 256 4K stripe_heads will
7776          * allow for chunk sizes up to 256K, which is probably OK.
7777          * If the chunk size is greater, user-space should request more
7778          * stripe_heads first.
7779          */
7780         struct r5conf *conf = mddev->private;
7781         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7782             > conf->min_nr_stripes ||
7783             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7784             > conf->min_nr_stripes) {
7785                 pr_warn("md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
7786                         mdname(mddev),
7787                         ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7788                          / STRIPE_SIZE)*4);
7789                 return 0;
7790         }
7791         return 1;
7792 }
7793
7794 static int check_reshape(struct mddev *mddev)
7795 {
7796         struct r5conf *conf = mddev->private;
7797
7798         if (raid5_has_log(conf) || raid5_has_ppl(conf))
7799                 return -EINVAL;
7800         if (mddev->delta_disks == 0 &&
7801             mddev->new_layout == mddev->layout &&
7802             mddev->new_chunk_sectors == mddev->chunk_sectors)
7803                 return 0; /* nothing to do */
7804         if (has_failed(conf))
7805                 return -EINVAL;
7806         if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7807                 /* We might be able to shrink, but the devices must
7808                  * be made bigger first.
7809                  * For raid6, 4 is the minimum size.
7810                  * Otherwise 2 is the minimum
7811                  */
7812                 int min = 2;
7813                 if (mddev->level == 6)
7814                         min = 4;
7815                 if (mddev->raid_disks + mddev->delta_disks < min)
7816                         return -EINVAL;
7817         }
7818
7819         if (!check_stripe_cache(mddev))
7820                 return -ENOSPC;
7821
7822         if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7823             mddev->delta_disks > 0)
7824                 if (resize_chunks(conf,
7825                                   conf->previous_raid_disks
7826                                   + max(0, mddev->delta_disks),
7827                                   max(mddev->new_chunk_sectors,
7828                                       mddev->chunk_sectors)
7829                             ) < 0)
7830                         return -ENOMEM;
7831
7832         if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
7833                 return 0; /* never bother to shrink */
7834         return resize_stripes(conf, (conf->previous_raid_disks
7835                                      + mddev->delta_disks));
7836 }
7837
7838 static int raid5_start_reshape(struct mddev *mddev)
7839 {
7840         struct r5conf *conf = mddev->private;
7841         struct md_rdev *rdev;
7842         int spares = 0;
7843         unsigned long flags;
7844
7845         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7846                 return -EBUSY;
7847
7848         if (!check_stripe_cache(mddev))
7849                 return -ENOSPC;
7850
7851         if (has_failed(conf))
7852                 return -EINVAL;
7853
7854         rdev_for_each(rdev, mddev) {
7855                 if (!test_bit(In_sync, &rdev->flags)
7856                     && !test_bit(Faulty, &rdev->flags))
7857                         spares++;
7858         }
7859
7860         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7861                 /* Not enough devices even to make a degraded array
7862                  * of that size
7863                  */
7864                 return -EINVAL;
7865
7866         /* Refuse to reduce size of the array.  Any reductions in
7867          * array size must be through explicit setting of array_size
7868          * attribute.
7869          */
7870         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7871             < mddev->array_sectors) {
7872                 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7873                         mdname(mddev));
7874                 return -EINVAL;
7875         }
7876
7877         atomic_set(&conf->reshape_stripes, 0);
7878         spin_lock_irq(&conf->device_lock);
7879         write_seqcount_begin(&conf->gen_lock);
7880         conf->previous_raid_disks = conf->raid_disks;
7881         conf->raid_disks += mddev->delta_disks;
7882         conf->prev_chunk_sectors = conf->chunk_sectors;
7883         conf->chunk_sectors = mddev->new_chunk_sectors;
7884         conf->prev_algo = conf->algorithm;
7885         conf->algorithm = mddev->new_layout;
7886         conf->generation++;
7887         /* Code that selects data_offset needs to see the generation update
7888          * if reshape_progress has been set - so a memory barrier needed.
7889          */
7890         smp_mb();
7891         if (mddev->reshape_backwards)
7892                 conf->reshape_progress = raid5_size(mddev, 0, 0);
7893         else
7894                 conf->reshape_progress = 0;
7895         conf->reshape_safe = conf->reshape_progress;
7896         write_seqcount_end(&conf->gen_lock);
7897         spin_unlock_irq(&conf->device_lock);
7898
7899         /* Now make sure any requests that proceeded on the assumption
7900          * the reshape wasn't running - like Discard or Read - have
7901          * completed.
7902          */
7903         mddev_suspend(mddev);
7904         mddev_resume(mddev);
7905
7906         /* Add some new drives, as many as will fit.
7907          * We know there are enough to make the newly sized array work.
7908          * Don't add devices if we are reducing the number of
7909          * devices in the array.  This is because it is not possible
7910          * to correctly record the "partially reconstructed" state of
7911          * such devices during the reshape and confusion could result.
7912          */
7913         if (mddev->delta_disks >= 0) {
7914                 rdev_for_each(rdev, mddev)
7915                         if (rdev->raid_disk < 0 &&
7916                             !test_bit(Faulty, &rdev->flags)) {
7917                                 if (raid5_add_disk(mddev, rdev) == 0) {
7918                                         if (rdev->raid_disk
7919                                             >= conf->previous_raid_disks)
7920                                                 set_bit(In_sync, &rdev->flags);
7921                                         else
7922                                                 rdev->recovery_offset = 0;
7923
7924                                         if (sysfs_link_rdev(mddev, rdev))
7925                                                 /* Failure here is OK */;
7926                                 }
7927                         } else if (rdev->raid_disk >= conf->previous_raid_disks
7928                                    && !test_bit(Faulty, &rdev->flags)) {
7929                                 /* This is a spare that was manually added */
7930                                 set_bit(In_sync, &rdev->flags);
7931                         }
7932
7933                 /* When a reshape changes the number of devices,
7934                  * ->degraded is measured against the larger of the
7935                  * pre and post number of devices.
7936                  */
7937                 spin_lock_irqsave(&conf->device_lock, flags);
7938                 mddev->degraded = raid5_calc_degraded(conf);
7939                 spin_unlock_irqrestore(&conf->device_lock, flags);
7940         }
7941         mddev->raid_disks = conf->raid_disks;
7942         mddev->reshape_position = conf->reshape_progress;
7943         set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
7944
7945         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7946         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7947         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7948         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7949         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7950         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7951                                                 "reshape");
7952         if (!mddev->sync_thread) {
7953                 mddev->recovery = 0;
7954                 spin_lock_irq(&conf->device_lock);
7955                 write_seqcount_begin(&conf->gen_lock);
7956                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7957                 mddev->new_chunk_sectors =
7958                         conf->chunk_sectors = conf->prev_chunk_sectors;
7959                 mddev->new_layout = conf->algorithm = conf->prev_algo;
7960                 rdev_for_each(rdev, mddev)
7961                         rdev->new_data_offset = rdev->data_offset;
7962                 smp_wmb();
7963                 conf->generation --;
7964                 conf->reshape_progress = MaxSector;
7965                 mddev->reshape_position = MaxSector;
7966                 write_seqcount_end(&conf->gen_lock);
7967                 spin_unlock_irq(&conf->device_lock);
7968                 return -EAGAIN;
7969         }
7970         conf->reshape_checkpoint = jiffies;
7971         md_wakeup_thread(mddev->sync_thread);
7972         md_new_event(mddev);
7973         return 0;
7974 }
7975
7976 /* This is called from the reshape thread and should make any
7977  * changes needed in 'conf'
7978  */
7979 static void end_reshape(struct r5conf *conf)
7980 {
7981
7982         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7983                 struct md_rdev *rdev;
7984
7985                 spin_lock_irq(&conf->device_lock);
7986                 conf->previous_raid_disks = conf->raid_disks;
7987                 md_finish_reshape(conf->mddev);
7988                 smp_wmb();
7989                 conf->reshape_progress = MaxSector;
7990                 conf->mddev->reshape_position = MaxSector;
7991                 rdev_for_each(rdev, conf->mddev)
7992                         if (rdev->raid_disk >= 0 &&
7993                             !test_bit(Journal, &rdev->flags) &&
7994                             !test_bit(In_sync, &rdev->flags))
7995                                 rdev->recovery_offset = MaxSector;
7996                 spin_unlock_irq(&conf->device_lock);
7997                 wake_up(&conf->wait_for_overlap);
7998
7999                 /* read-ahead size must cover two whole stripes, which is
8000                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
8001                  */
8002                 if (conf->mddev->queue) {
8003                         int data_disks = conf->raid_disks - conf->max_degraded;
8004                         int stripe = data_disks * ((conf->chunk_sectors << 9)
8005                                                    / PAGE_SIZE);
8006                         if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
8007                                 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
8008                 }
8009         }
8010 }
8011
8012 /* This is called from the raid5d thread with mddev_lock held.
8013  * It makes config changes to the device.
8014  */
8015 static void raid5_finish_reshape(struct mddev *mddev)
8016 {
8017         struct r5conf *conf = mddev->private;
8018
8019         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8020
8021                 if (mddev->delta_disks <= 0) {
8022                         int d;
8023                         spin_lock_irq(&conf->device_lock);
8024                         mddev->degraded = raid5_calc_degraded(conf);
8025                         spin_unlock_irq(&conf->device_lock);
8026                         for (d = conf->raid_disks ;
8027                              d < conf->raid_disks - mddev->delta_disks;
8028                              d++) {
8029                                 struct md_rdev *rdev = conf->disks[d].rdev;
8030                                 if (rdev)
8031                                         clear_bit(In_sync, &rdev->flags);
8032                                 rdev = conf->disks[d].replacement;
8033                                 if (rdev)
8034                                         clear_bit(In_sync, &rdev->flags);
8035                         }
8036                 }
8037                 mddev->layout = conf->algorithm;
8038                 mddev->chunk_sectors = conf->chunk_sectors;
8039                 mddev->reshape_position = MaxSector;
8040                 mddev->delta_disks = 0;
8041                 mddev->reshape_backwards = 0;
8042         }
8043 }
8044
8045 static void raid5_quiesce(struct mddev *mddev, int quiesce)
8046 {
8047         struct r5conf *conf = mddev->private;
8048
8049         if (quiesce) {
8050                 /* stop all writes */
8051                 lock_all_device_hash_locks_irq(conf);
8052                 /* '2' tells resync/reshape to pause so that all
8053                  * active stripes can drain
8054                  */
8055                 r5c_flush_cache(conf, INT_MAX);
8056                 conf->quiesce = 2;
8057                 wait_event_cmd(conf->wait_for_quiescent,
8058                                     atomic_read(&conf->active_stripes) == 0 &&
8059                                     atomic_read(&conf->active_aligned_reads) == 0,
8060                                     unlock_all_device_hash_locks_irq(conf),
8061                                     lock_all_device_hash_locks_irq(conf));
8062                 conf->quiesce = 1;
8063                 unlock_all_device_hash_locks_irq(conf);
8064                 /* allow reshape to continue */
8065                 wake_up(&conf->wait_for_overlap);
8066         } else {
8067                 /* re-enable writes */
8068                 lock_all_device_hash_locks_irq(conf);
8069                 conf->quiesce = 0;
8070                 wake_up(&conf->wait_for_quiescent);
8071                 wake_up(&conf->wait_for_overlap);
8072                 unlock_all_device_hash_locks_irq(conf);
8073         }
8074         log_quiesce(conf, quiesce);
8075 }
8076
8077 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8078 {
8079         struct r0conf *raid0_conf = mddev->private;
8080         sector_t sectors;
8081
8082         /* for raid0 takeover only one zone is supported */
8083         if (raid0_conf->nr_strip_zones > 1) {
8084                 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8085                         mdname(mddev));
8086                 return ERR_PTR(-EINVAL);
8087         }
8088
8089         sectors = raid0_conf->strip_zone[0].zone_end;
8090         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8091         mddev->dev_sectors = sectors;
8092         mddev->new_level = level;
8093         mddev->new_layout = ALGORITHM_PARITY_N;
8094         mddev->new_chunk_sectors = mddev->chunk_sectors;
8095         mddev->raid_disks += 1;
8096         mddev->delta_disks = 1;
8097         /* make sure it will be not marked as dirty */
8098         mddev->recovery_cp = MaxSector;
8099
8100         return setup_conf(mddev);
8101 }
8102
8103 static void *raid5_takeover_raid1(struct mddev *mddev)
8104 {
8105         int chunksect;
8106         void *ret;
8107
8108         if (mddev->raid_disks != 2 ||
8109             mddev->degraded > 1)
8110                 return ERR_PTR(-EINVAL);
8111
8112         /* Should check if there are write-behind devices? */
8113
8114         chunksect = 64*2; /* 64K by default */
8115
8116         /* The array must be an exact multiple of chunksize */
8117         while (chunksect && (mddev->array_sectors & (chunksect-1)))
8118                 chunksect >>= 1;
8119
8120         if ((chunksect<<9) < STRIPE_SIZE)
8121                 /* array size does not allow a suitable chunk size */
8122                 return ERR_PTR(-EINVAL);
8123
8124         mddev->new_level = 5;
8125         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8126         mddev->new_chunk_sectors = chunksect;
8127
8128         ret = setup_conf(mddev);
8129         if (!IS_ERR(ret))
8130                 mddev_clear_unsupported_flags(mddev,
8131                         UNSUPPORTED_MDDEV_FLAGS);
8132         return ret;
8133 }
8134
8135 static void *raid5_takeover_raid6(struct mddev *mddev)
8136 {
8137         int new_layout;
8138
8139         switch (mddev->layout) {
8140         case ALGORITHM_LEFT_ASYMMETRIC_6:
8141                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8142                 break;
8143         case ALGORITHM_RIGHT_ASYMMETRIC_6:
8144                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8145                 break;
8146         case ALGORITHM_LEFT_SYMMETRIC_6:
8147                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8148                 break;
8149         case ALGORITHM_RIGHT_SYMMETRIC_6:
8150                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8151                 break;
8152         case ALGORITHM_PARITY_0_6:
8153                 new_layout = ALGORITHM_PARITY_0;
8154                 break;
8155         case ALGORITHM_PARITY_N:
8156                 new_layout = ALGORITHM_PARITY_N;
8157                 break;
8158         default:
8159                 return ERR_PTR(-EINVAL);
8160         }
8161         mddev->new_level = 5;
8162         mddev->new_layout = new_layout;
8163         mddev->delta_disks = -1;
8164         mddev->raid_disks -= 1;
8165         return setup_conf(mddev);
8166 }
8167
8168 static int raid5_check_reshape(struct mddev *mddev)
8169 {
8170         /* For a 2-drive array, the layout and chunk size can be changed
8171          * immediately as not restriping is needed.
8172          * For larger arrays we record the new value - after validation
8173          * to be used by a reshape pass.
8174          */
8175         struct r5conf *conf = mddev->private;
8176         int new_chunk = mddev->new_chunk_sectors;
8177
8178         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8179                 return -EINVAL;
8180         if (new_chunk > 0) {
8181                 if (!is_power_of_2(new_chunk))
8182                         return -EINVAL;
8183                 if (new_chunk < (PAGE_SIZE>>9))
8184                         return -EINVAL;
8185                 if (mddev->array_sectors & (new_chunk-1))
8186                         /* not factor of array size */
8187                         return -EINVAL;
8188         }
8189
8190         /* They look valid */
8191
8192         if (mddev->raid_disks == 2) {
8193                 /* can make the change immediately */
8194                 if (mddev->new_layout >= 0) {
8195                         conf->algorithm = mddev->new_layout;
8196                         mddev->layout = mddev->new_layout;
8197                 }
8198                 if (new_chunk > 0) {
8199                         conf->chunk_sectors = new_chunk ;
8200                         mddev->chunk_sectors = new_chunk;
8201                 }
8202                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8203                 md_wakeup_thread(mddev->thread);
8204         }
8205         return check_reshape(mddev);
8206 }
8207
8208 static int raid6_check_reshape(struct mddev *mddev)
8209 {
8210         int new_chunk = mddev->new_chunk_sectors;
8211
8212         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8213                 return -EINVAL;
8214         if (new_chunk > 0) {
8215                 if (!is_power_of_2(new_chunk))
8216                         return -EINVAL;
8217                 if (new_chunk < (PAGE_SIZE >> 9))
8218                         return -EINVAL;
8219                 if (mddev->array_sectors & (new_chunk-1))
8220                         /* not factor of array size */
8221                         return -EINVAL;
8222         }
8223
8224         /* They look valid */
8225         return check_reshape(mddev);
8226 }
8227
8228 static void *raid5_takeover(struct mddev *mddev)
8229 {
8230         /* raid5 can take over:
8231          *  raid0 - if there is only one strip zone - make it a raid4 layout
8232          *  raid1 - if there are two drives.  We need to know the chunk size
8233          *  raid4 - trivial - just use a raid4 layout.
8234          *  raid6 - Providing it is a *_6 layout
8235          */
8236         if (mddev->level == 0)
8237                 return raid45_takeover_raid0(mddev, 5);
8238         if (mddev->level == 1)
8239                 return raid5_takeover_raid1(mddev);
8240         if (mddev->level == 4) {
8241                 mddev->new_layout = ALGORITHM_PARITY_N;
8242                 mddev->new_level = 5;
8243                 return setup_conf(mddev);
8244         }
8245         if (mddev->level == 6)
8246                 return raid5_takeover_raid6(mddev);
8247
8248         return ERR_PTR(-EINVAL);
8249 }
8250
8251 static void *raid4_takeover(struct mddev *mddev)
8252 {
8253         /* raid4 can take over:
8254          *  raid0 - if there is only one strip zone
8255          *  raid5 - if layout is right
8256          */
8257         if (mddev->level == 0)
8258                 return raid45_takeover_raid0(mddev, 4);
8259         if (mddev->level == 5 &&
8260             mddev->layout == ALGORITHM_PARITY_N) {
8261                 mddev->new_layout = 0;
8262                 mddev->new_level = 4;
8263                 return setup_conf(mddev);
8264         }
8265         return ERR_PTR(-EINVAL);
8266 }
8267
8268 static struct md_personality raid5_personality;
8269
8270 static void *raid6_takeover(struct mddev *mddev)
8271 {
8272         /* Currently can only take over a raid5.  We map the
8273          * personality to an equivalent raid6 personality
8274          * with the Q block at the end.
8275          */
8276         int new_layout;
8277
8278         if (mddev->pers != &raid5_personality)
8279                 return ERR_PTR(-EINVAL);
8280         if (mddev->degraded > 1)
8281                 return ERR_PTR(-EINVAL);
8282         if (mddev->raid_disks > 253)
8283                 return ERR_PTR(-EINVAL);
8284         if (mddev->raid_disks < 3)
8285                 return ERR_PTR(-EINVAL);
8286
8287         switch (mddev->layout) {
8288         case ALGORITHM_LEFT_ASYMMETRIC:
8289                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8290                 break;
8291         case ALGORITHM_RIGHT_ASYMMETRIC:
8292                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8293                 break;
8294         case ALGORITHM_LEFT_SYMMETRIC:
8295                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8296                 break;
8297         case ALGORITHM_RIGHT_SYMMETRIC:
8298                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8299                 break;
8300         case ALGORITHM_PARITY_0:
8301                 new_layout = ALGORITHM_PARITY_0_6;
8302                 break;
8303         case ALGORITHM_PARITY_N:
8304                 new_layout = ALGORITHM_PARITY_N;
8305                 break;
8306         default:
8307                 return ERR_PTR(-EINVAL);
8308         }
8309         mddev->new_level = 6;
8310         mddev->new_layout = new_layout;
8311         mddev->delta_disks = 1;
8312         mddev->raid_disks += 1;
8313         return setup_conf(mddev);
8314 }
8315
8316 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8317 {
8318         struct r5conf *conf;
8319         int err;
8320
8321         err = mddev_lock(mddev);
8322         if (err)
8323                 return err;
8324         conf = mddev->private;
8325         if (!conf) {
8326                 mddev_unlock(mddev);
8327                 return -ENODEV;
8328         }
8329
8330         if (strncmp(buf, "ppl", 3) == 0) {
8331                 /* ppl only works with RAID 5 */
8332                 if (!raid5_has_ppl(conf) && conf->level == 5) {
8333                         err = log_init(conf, NULL, true);
8334                         if (!err) {
8335                                 err = resize_stripes(conf, conf->pool_size);
8336                                 if (err)
8337                                         log_exit(conf);
8338                         }
8339                 } else
8340                         err = -EINVAL;
8341         } else if (strncmp(buf, "resync", 6) == 0) {
8342                 if (raid5_has_ppl(conf)) {
8343                         mddev_suspend(mddev);
8344                         log_exit(conf);
8345                         mddev_resume(mddev);
8346                         err = resize_stripes(conf, conf->pool_size);
8347                 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8348                            r5l_log_disk_error(conf)) {
8349                         bool journal_dev_exists = false;
8350                         struct md_rdev *rdev;
8351
8352                         rdev_for_each(rdev, mddev)
8353                                 if (test_bit(Journal, &rdev->flags)) {
8354                                         journal_dev_exists = true;
8355                                         break;
8356                                 }
8357
8358                         if (!journal_dev_exists) {
8359                                 mddev_suspend(mddev);
8360                                 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8361                                 mddev_resume(mddev);
8362                         } else  /* need remove journal device first */
8363                                 err = -EBUSY;
8364                 } else
8365                         err = -EINVAL;
8366         } else {
8367                 err = -EINVAL;
8368         }
8369
8370         if (!err)
8371                 md_update_sb(mddev, 1);
8372
8373         mddev_unlock(mddev);
8374
8375         return err;
8376 }
8377
8378 static int raid5_start(struct mddev *mddev)
8379 {
8380         struct r5conf *conf = mddev->private;
8381
8382         return r5l_start(conf->log);
8383 }
8384
8385 static struct md_personality raid6_personality =
8386 {
8387         .name           = "raid6",
8388         .level          = 6,
8389         .owner          = THIS_MODULE,
8390         .make_request   = raid5_make_request,
8391         .run            = raid5_run,
8392         .start          = raid5_start,
8393         .free           = raid5_free,
8394         .status         = raid5_status,
8395         .error_handler  = raid5_error,
8396         .hot_add_disk   = raid5_add_disk,
8397         .hot_remove_disk= raid5_remove_disk,
8398         .spare_active   = raid5_spare_active,
8399         .sync_request   = raid5_sync_request,
8400         .resize         = raid5_resize,
8401         .size           = raid5_size,
8402         .check_reshape  = raid6_check_reshape,
8403         .start_reshape  = raid5_start_reshape,
8404         .finish_reshape = raid5_finish_reshape,
8405         .quiesce        = raid5_quiesce,
8406         .takeover       = raid6_takeover,
8407         .congested      = raid5_congested,
8408         .change_consistency_policy = raid5_change_consistency_policy,
8409 };
8410 static struct md_personality raid5_personality =
8411 {
8412         .name           = "raid5",
8413         .level          = 5,
8414         .owner          = THIS_MODULE,
8415         .make_request   = raid5_make_request,
8416         .run            = raid5_run,
8417         .start          = raid5_start,
8418         .free           = raid5_free,
8419         .status         = raid5_status,
8420         .error_handler  = raid5_error,
8421         .hot_add_disk   = raid5_add_disk,
8422         .hot_remove_disk= raid5_remove_disk,
8423         .spare_active   = raid5_spare_active,
8424         .sync_request   = raid5_sync_request,
8425         .resize         = raid5_resize,
8426         .size           = raid5_size,
8427         .check_reshape  = raid5_check_reshape,
8428         .start_reshape  = raid5_start_reshape,
8429         .finish_reshape = raid5_finish_reshape,
8430         .quiesce        = raid5_quiesce,
8431         .takeover       = raid5_takeover,
8432         .congested      = raid5_congested,
8433         .change_consistency_policy = raid5_change_consistency_policy,
8434 };
8435
8436 static struct md_personality raid4_personality =
8437 {
8438         .name           = "raid4",
8439         .level          = 4,
8440         .owner          = THIS_MODULE,
8441         .make_request   = raid5_make_request,
8442         .run            = raid5_run,
8443         .start          = raid5_start,
8444         .free           = raid5_free,
8445         .status         = raid5_status,
8446         .error_handler  = raid5_error,
8447         .hot_add_disk   = raid5_add_disk,
8448         .hot_remove_disk= raid5_remove_disk,
8449         .spare_active   = raid5_spare_active,
8450         .sync_request   = raid5_sync_request,
8451         .resize         = raid5_resize,
8452         .size           = raid5_size,
8453         .check_reshape  = raid5_check_reshape,
8454         .start_reshape  = raid5_start_reshape,
8455         .finish_reshape = raid5_finish_reshape,
8456         .quiesce        = raid5_quiesce,
8457         .takeover       = raid4_takeover,
8458         .congested      = raid5_congested,
8459         .change_consistency_policy = raid5_change_consistency_policy,
8460 };
8461
8462 static int __init raid5_init(void)
8463 {
8464         int ret;
8465
8466         raid5_wq = alloc_workqueue("raid5wq",
8467                 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8468         if (!raid5_wq)
8469                 return -ENOMEM;
8470
8471         ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8472                                       "md/raid5:prepare",
8473                                       raid456_cpu_up_prepare,
8474                                       raid456_cpu_dead);
8475         if (ret) {
8476                 destroy_workqueue(raid5_wq);
8477                 return ret;
8478         }
8479         register_md_personality(&raid6_personality);
8480         register_md_personality(&raid5_personality);
8481         register_md_personality(&raid4_personality);
8482         return 0;
8483 }
8484
8485 static void raid5_exit(void)
8486 {
8487         unregister_md_personality(&raid6_personality);
8488         unregister_md_personality(&raid5_personality);
8489         unregister_md_personality(&raid4_personality);
8490         cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8491         destroy_workqueue(raid5_wq);
8492 }
8493
8494 module_init(raid5_init);
8495 module_exit(raid5_exit);
8496 MODULE_LICENSE("GPL");
8497 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8498 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8499 MODULE_ALIAS("md-raid5");
8500 MODULE_ALIAS("md-raid4");
8501 MODULE_ALIAS("md-level-5");
8502 MODULE_ALIAS("md-level-4");
8503 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8504 MODULE_ALIAS("md-raid6");
8505 MODULE_ALIAS("md-level-6");
8506
8507 /* This used to be two separate modules, they were: */
8508 MODULE_ALIAS("raid5");
8509 MODULE_ALIAS("raid6");