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 #include <linux/flex_array.h>
  58 #include <trace/events/block.h>
  59
  60 #include "md.h"
  61 #include "raid5.h"
  62 #include "raid0.h"
  63 #include "bitmap.h"
  64
  65 #define cpu_to_group(cpu) cpu_to_node(cpu)
  66 #define ANY_GROUP NUMA_NO_NODE
  67
  68 static bool devices_handle_discard_safely = false;
  69 module_param(devices_handle_discard_safely, bool, 0644);
  70 MODULE_PARM_DESC(devices_handle_discard_safely,
  71                  "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
  72 static struct workqueue_struct *raid5_wq;
  73 /*
  74  * Stripe cache
  75  */
  76
  77 #define NR_STRIPES              256
  78 #define STRIPE_SIZE             PAGE_SIZE
  79 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
  80 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
  81 #define IO_THRESHOLD            1
  82 #define BYPASS_THRESHOLD        1
  83 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
  84 #define HASH_MASK               (NR_HASH - 1)
  85 #define MAX_STRIPE_BATCH        8
  86
  87 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
  88 {
  89         int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
  90         return &conf->stripe_hashtbl[hash];
  91 }
  92
  93 static inline int stripe_hash_locks_hash(sector_t sect)
  94 {
  95         return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
  96 }
  97
  98 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
  99 {
 100         spin_lock_irq(conf->hash_locks + hash);
 101         spin_lock(&conf->device_lock);
 102 }
 103
 104 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
 105 {
 106         spin_unlock(&conf->device_lock);
 107         spin_unlock_irq(conf->hash_locks + hash);
 108 }
 109
 110 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
 111 {
 112         int i;
 113         local_irq_disable();
 114         spin_lock(conf->hash_locks);
 115         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
 116                 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
 117         spin_lock(&conf->device_lock);
 118 }
 119
 120 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
 121 {
 122         int i;
 123         spin_unlock(&conf->device_lock);
 124         for (i = NR_STRIPE_HASH_LOCKS; i; i--)
 125                 spin_unlock(conf->hash_locks + i - 1);
 126         local_irq_enable();
 127 }
 128
 129 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
 130  * order without overlap.  There may be several bio's per stripe+device, and
 131  * a bio could span several devices.
 132  * When walking this list for a particular stripe+device, we must never proceed
 133  * beyond a bio that extends past this device, as the next bio might no longer
 134  * be valid.
 135  * This function is used to determine the 'next' bio in the list, given the sector
 136  * of the current stripe+device
 137  */
 138 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
 139 {
 140         int sectors = bio_sectors(bio);
 141         if (bio->bi_iter.bi_sector + sectors < sector + STRIPE_SECTORS)
 142                 return bio->bi_next;
 143         else
 144                 return NULL;
 145 }
 146
 147 /*
 148  * We maintain a biased count of active stripes in the bottom 16 bits of
 149  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
 150  */
 151 static inline int raid5_bi_processed_stripes(struct bio *bio)
 152 {
 153         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 154         return (atomic_read(segments) >> 16) & 0xffff;
 155 }
 156
 157 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
 158 {
 159         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 160         return atomic_sub_return(1, segments) & 0xffff;
 161 }
 162
 163 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
 164 {
 165         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 166         atomic_inc(segments);
 167 }
 168
 169 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
 170         unsigned int cnt)
 171 {
 172         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 173         int old, new;
 174
 175         do {
 176                 old = atomic_read(segments);
 177                 new = (old & 0xffff) | (cnt << 16);
 178         } while (atomic_cmpxchg(segments, old, new) != old);
 179 }
 180
 181 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
 182 {
 183         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 184         atomic_set(segments, cnt);
 185 }
 186
 187 /* Find first data disk in a raid6 stripe */
 188 static inline int raid6_d0(struct stripe_head *sh)
 189 {
 190         if (sh->ddf_layout)
 191                 /* ddf always start from first device */
 192                 return 0;
 193         /* md starts just after Q block */
 194         if (sh->qd_idx == sh->disks - 1)
 195                 return 0;
 196         else
 197                 return sh->qd_idx + 1;
 198 }
 199 static inline int raid6_next_disk(int disk, int raid_disks)
 200 {
 201         disk++;
 202         return (disk < raid_disks) ? disk : 0;
 203 }
 204
 205 /* When walking through the disks in a raid5, starting at raid6_d0,
 206  * We need to map each disk to a 'slot', where the data disks are slot
 207  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 208  * is raid_disks-1.  This help does that mapping.
 209  */
 210 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 211                              int *count, int syndrome_disks)
 212 {
 213         int slot = *count;
 214
 215         if (sh->ddf_layout)
 216                 (*count)++;
 217         if (idx == sh->pd_idx)
 218                 return syndrome_disks;
 219         if (idx == sh->qd_idx)
 220                 return syndrome_disks + 1;
 221         if (!sh->ddf_layout)
 222                 (*count)++;
 223         return slot;
 224 }
 225
 226 static void return_io(struct bio_list *return_bi)
 227 {
 228         struct bio *bi;
 229         while ((bi = bio_list_pop(return_bi)) != NULL) {
 230                 bi->bi_iter.bi_size = 0;
 231                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
 232                                          bi, 0);
 233                 bio_endio(bi);
 234         }
 235 }
 236
 237 static void print_raid5_conf (struct r5conf *conf);
 238
 239 static int stripe_operations_active(struct stripe_head *sh)
 240 {
 241         return sh->check_state || sh->reconstruct_state ||
 242                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 243                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 244 }
 245
 246 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
 247 {
 248         struct r5conf *conf = sh->raid_conf;
 249         struct r5worker_group *group;
 250         int thread_cnt;
 251         int i, cpu = sh->cpu;
 252
 253         if (!cpu_online(cpu)) {
 254                 cpu = cpumask_any(cpu_online_mask);
 255                 sh->cpu = cpu;
 256         }
 257
 258         if (list_empty(&sh->lru)) {
 259                 struct r5worker_group *group;
 260                 group = conf->worker_groups + cpu_to_group(cpu);
 261                 list_add_tail(&sh->lru, &group->handle_list);
 262                 group->stripes_cnt++;
 263                 sh->group = group;
 264         }
 265
 266         if (conf->worker_cnt_per_group == 0) {
 267                 md_wakeup_thread(conf->mddev->thread);
 268                 return;
 269         }
 270
 271         group = conf->worker_groups + cpu_to_group(sh->cpu);
 272
 273         group->workers[0].working = true;
 274         /* at least one worker should run to avoid race */
 275         queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
 276
 277         thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
 278         /* wakeup more workers */
 279         for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
 280                 if (group->workers[i].working == false) {
 281                         group->workers[i].working = true;
 282                         queue_work_on(sh->cpu, raid5_wq,
 283                                       &group->workers[i].work);
 284                         thread_cnt--;
 285                 }
 286         }
 287 }
 288
 289 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
 290                               struct list_head *temp_inactive_list)
 291 {
 292         BUG_ON(!list_empty(&sh->lru));
 293         BUG_ON(atomic_read(&conf->active_stripes)==0);
 294         if (test_bit(STRIPE_HANDLE, &sh->state)) {
 295                 if (test_bit(STRIPE_DELAYED, &sh->state) &&
 296                     !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 297                         list_add_tail(&sh->lru, &conf->delayed_list);
 298                 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 299                            sh->bm_seq - conf->seq_write > 0)
 300                         list_add_tail(&sh->lru, &conf->bitmap_list);
 301                 else {
 302                         clear_bit(STRIPE_DELAYED, &sh->state);
 303                         clear_bit(STRIPE_BIT_DELAY, &sh->state);
 304                         if (conf->worker_cnt_per_group == 0) {
 305                                 list_add_tail(&sh->lru, &conf->handle_list);
 306                         } else {
 307                                 raid5_wakeup_stripe_thread(sh);
 308                                 return;
 309                         }
 310                 }
 311                 md_wakeup_thread(conf->mddev->thread);
 312         } else {
 313                 BUG_ON(stripe_operations_active(sh));
 314                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 315                         if (atomic_dec_return(&conf->preread_active_stripes)
 316                             < IO_THRESHOLD)
 317                                 md_wakeup_thread(conf->mddev->thread);
 318                 atomic_dec(&conf->active_stripes);
 319                 if (!test_bit(STRIPE_EXPANDING, &sh->state))
 320                         list_add_tail(&sh->lru, temp_inactive_list);
 321         }
 322 }
 323
 324 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
 325                              struct list_head *temp_inactive_list)
 326 {
 327         if (atomic_dec_and_test(&sh->count))
 328                 do_release_stripe(conf, sh, temp_inactive_list);
 329 }
 330
 331 /*
 332  * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
 333  *
 334  * Be careful: Only one task can add/delete stripes from temp_inactive_list at
 335  * given time. Adding stripes only takes device lock, while deleting stripes
 336  * only takes hash lock.
 337  */
 338 static void release_inactive_stripe_list(struct r5conf *conf,
 339                                          struct list_head *temp_inactive_list,
 340                                          int hash)
 341 {
 342         int size;
 343         bool do_wakeup = false;
 344         unsigned long flags;
 345
 346         if (hash == NR_STRIPE_HASH_LOCKS) {
 347                 size = NR_STRIPE_HASH_LOCKS;
 348                 hash = NR_STRIPE_HASH_LOCKS - 1;
 349         } else
 350                 size = 1;
 351         while (size) {
 352                 struct list_head *list = &temp_inactive_list[size - 1];
 353
 354                 /*
 355                  * We don't hold any lock here yet, raid5_get_active_stripe() might
 356                  * remove stripes from the list
 357                  */
 358                 if (!list_empty_careful(list)) {
 359                         spin_lock_irqsave(conf->hash_locks + hash, flags);
 360                         if (list_empty(conf->inactive_list + hash) &&
 361                             !list_empty(list))
 362                                 atomic_dec(&conf->empty_inactive_list_nr);
 363                         list_splice_tail_init(list, conf->inactive_list + hash);
 364                         do_wakeup = true;
 365                         spin_unlock_irqrestore(conf->hash_locks + hash, flags);
 366                 }
 367                 size--;
 368                 hash--;
 369         }
 370
 371         if (do_wakeup) {
 372                 wake_up(&conf->wait_for_stripe);
 373                 if (atomic_read(&conf->active_stripes) == 0)
 374                         wake_up(&conf->wait_for_quiescent);
 375                 if (conf->retry_read_aligned)
 376                         md_wakeup_thread(conf->mddev->thread);
 377         }
 378 }
 379
 380 /* should hold conf->device_lock already */
 381 static int release_stripe_list(struct r5conf *conf,
 382                                struct list_head *temp_inactive_list)
 383 {
 384         struct stripe_head *sh;
 385         int count = 0;
 386         struct llist_node *head;
 387
 388         head = llist_del_all(&conf->released_stripes);
 389         head = llist_reverse_order(head);
 390         while (head) {
 391                 int hash;
 392
 393                 sh = llist_entry(head, struct stripe_head, release_list);
 394                 head = llist_next(head);
 395                 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
 396                 smp_mb();
 397                 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
 398                 /*
 399                  * Don't worry the bit is set here, because if the bit is set
 400                  * again, the count is always > 1. This is true for
 401                  * STRIPE_ON_UNPLUG_LIST bit too.
 402                  */
 403                 hash = sh->hash_lock_index;
 404                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
 405                 count++;
 406         }
 407
 408         return count;
 409 }
 410
 411 void raid5_release_stripe(struct stripe_head *sh)
 412 {
 413         struct r5conf *conf = sh->raid_conf;
 414         unsigned long flags;
 415         struct list_head list;
 416         int hash;
 417         bool wakeup;
 418
 419         /* Avoid release_list until the last reference.
 420          */
 421         if (atomic_add_unless(&sh->count, -1, 1))
 422                 return;
 423
 424         if (unlikely(!conf->mddev->thread) ||
 425                 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
 426                 goto slow_path;
 427         wakeup = llist_add(&sh->release_list, &conf->released_stripes);
 428         if (wakeup)
 429                 md_wakeup_thread(conf->mddev->thread);
 430         return;
 431 slow_path:
 432         local_irq_save(flags);
 433         /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
 434         if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
 435                 INIT_LIST_HEAD(&list);
 436                 hash = sh->hash_lock_index;
 437                 do_release_stripe(conf, sh, &list);
 438                 spin_unlock(&conf->device_lock);
 439                 release_inactive_stripe_list(conf, &list, hash);
 440         }
 441         local_irq_restore(flags);
 442 }
 443
 444 static inline void remove_hash(struct stripe_head *sh)
 445 {
 446         pr_debug("remove_hash(), stripe %llu\n",
 447                 (unsigned long long)sh->sector);
 448
 449         hlist_del_init(&sh->hash);
 450 }
 451
 452 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 453 {
 454         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 455
 456         pr_debug("insert_hash(), stripe %llu\n",
 457                 (unsigned long long)sh->sector);
 458
 459         hlist_add_head(&sh->hash, hp);
 460 }
 461
 462 /* find an idle stripe, make sure it is unhashed, and return it. */
 463 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
 464 {
 465         struct stripe_head *sh = NULL;
 466         struct list_head *first;
 467
 468         if (list_empty(conf->inactive_list + hash))
 469                 goto out;
 470         first = (conf->inactive_list + hash)->next;
 471         sh = list_entry(first, struct stripe_head, lru);
 472         list_del_init(first);
 473         remove_hash(sh);
 474         atomic_inc(&conf->active_stripes);
 475         BUG_ON(hash != sh->hash_lock_index);
 476         if (list_empty(conf->inactive_list + hash))
 477                 atomic_inc(&conf->empty_inactive_list_nr);
 478 out:
 479         return sh;
 480 }
 481
 482 static void shrink_buffers(struct stripe_head *sh)
 483 {
 484         struct page *p;
 485         int i;
 486         int num = sh->raid_conf->pool_size;
 487
 488         for (i = 0; i < num ; i++) {
 489                 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
 490                 p = sh->dev[i].page;
 491                 if (!p)
 492                         continue;
 493                 sh->dev[i].page = NULL;
 494                 put_page(p);
 495         }
 496 }
 497
 498 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
 499 {
 500         int i;
 501         int num = sh->raid_conf->pool_size;
 502
 503         for (i = 0; i < num; i++) {
 504                 struct page *page;
 505
 506                 if (!(page = alloc_page(gfp))) {
 507                         return 1;
 508                 }
 509                 sh->dev[i].page = page;
 510                 sh->dev[i].orig_page = page;
 511         }
 512         return 0;
 513 }
 514
 515 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
 516 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 517                             struct stripe_head *sh);
 518
 519 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 520 {
 521         struct r5conf *conf = sh->raid_conf;
 522         int i, seq;
 523
 524         BUG_ON(atomic_read(&sh->count) != 0);
 525         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 526         BUG_ON(stripe_operations_active(sh));
 527         BUG_ON(sh->batch_head);
 528
 529         pr_debug("init_stripe called, stripe %llu\n",
 530                 (unsigned long long)sector);
 531 retry:
 532         seq = read_seqcount_begin(&conf->gen_lock);
 533         sh->generation = conf->generation - previous;
 534         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 535         sh->sector = sector;
 536         stripe_set_idx(sector, conf, previous, sh);
 537         sh->state = 0;
 538
 539         for (i = sh->disks; i--; ) {
 540                 struct r5dev *dev = &sh->dev[i];
 541
 542                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 543                     test_bit(R5_LOCKED, &dev->flags)) {
 544                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
 545                                (unsigned long long)sh->sector, i, dev->toread,
 546                                dev->read, dev->towrite, dev->written,
 547                                test_bit(R5_LOCKED, &dev->flags));
 548                         WARN_ON(1);
 549                 }
 550                 dev->flags = 0;
 551                 raid5_build_block(sh, i, previous);
 552         }
 553         if (read_seqcount_retry(&conf->gen_lock, seq))
 554                 goto retry;
 555         sh->overwrite_disks = 0;
 556         insert_hash(conf, sh);
 557         sh->cpu = smp_processor_id();
 558         set_bit(STRIPE_BATCH_READY, &sh->state);
 559 }
 560
 561 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 562                                          short generation)
 563 {
 564         struct stripe_head *sh;
 565
 566         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 567         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 568                 if (sh->sector == sector && sh->generation == generation)
 569                         return sh;
 570         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 571         return NULL;
 572 }
 573
 574 /*
 575  * Need to check if array has failed when deciding whether to:
 576  *  - start an array
 577  *  - remove non-faulty devices
 578  *  - add a spare
 579  *  - allow a reshape
 580  * This determination is simple when no reshape is happening.
 581  * However if there is a reshape, we need to carefully check
 582  * both the before and after sections.
 583  * This is because some failed devices may only affect one
 584  * of the two sections, and some non-in_sync devices may
 585  * be insync in the section most affected by failed devices.
 586  */
 587 static int calc_degraded(struct r5conf *conf)
 588 {
 589         int degraded, degraded2;
 590         int i;
 591
 592         rcu_read_lock();
 593         degraded = 0;
 594         for (i = 0; i < conf->previous_raid_disks; i++) {
 595                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 596                 if (rdev && test_bit(Faulty, &rdev->flags))
 597                         rdev = rcu_dereference(conf->disks[i].replacement);
 598                 if (!rdev || test_bit(Faulty, &rdev->flags))
 599                         degraded++;
 600                 else if (test_bit(In_sync, &rdev->flags))
 601                         ;
 602                 else
 603                         /* not in-sync or faulty.
 604                          * If the reshape increases the number of devices,
 605                          * this is being recovered by the reshape, so
 606                          * this 'previous' section is not in_sync.
 607                          * If the number of devices is being reduced however,
 608                          * the device can only be part of the array if
 609                          * we are reverting a reshape, so this section will
 610                          * be in-sync.
 611                          */
 612                         if (conf->raid_disks >= conf->previous_raid_disks)
 613                                 degraded++;
 614         }
 615         rcu_read_unlock();
 616         if (conf->raid_disks == conf->previous_raid_disks)
 617                 return degraded;
 618         rcu_read_lock();
 619         degraded2 = 0;
 620         for (i = 0; i < conf->raid_disks; i++) {
 621                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 622                 if (rdev && test_bit(Faulty, &rdev->flags))
 623                         rdev = rcu_dereference(conf->disks[i].replacement);
 624                 if (!rdev || test_bit(Faulty, &rdev->flags))
 625                         degraded2++;
 626                 else if (test_bit(In_sync, &rdev->flags))
 627                         ;
 628                 else
 629                         /* not in-sync or faulty.
 630                          * If reshape increases the number of devices, this
 631                          * section has already been recovered, else it
 632                          * almost certainly hasn't.
 633                          */
 634                         if (conf->raid_disks <= conf->previous_raid_disks)
 635                                 degraded2++;
 636         }
 637         rcu_read_unlock();
 638         if (degraded2 > degraded)
 639                 return degraded2;
 640         return degraded;
 641 }
 642
 643 static int has_failed(struct r5conf *conf)
 644 {
 645         int degraded;
 646
 647         if (conf->mddev->reshape_position == MaxSector)
 648                 return conf->mddev->degraded > conf->max_degraded;
 649
 650         degraded = calc_degraded(conf);
 651         if (degraded > conf->max_degraded)
 652                 return 1;
 653         return 0;
 654 }
 655
 656 struct stripe_head *
 657 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
 658                         int previous, int noblock, int noquiesce)
 659 {
 660         struct stripe_head *sh;
 661         int hash = stripe_hash_locks_hash(sector);
 662
 663         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 664
 665         spin_lock_irq(conf->hash_locks + hash);
 666
 667         do {
 668                 wait_event_lock_irq(conf->wait_for_quiescent,
 669                                     conf->quiesce == 0 || noquiesce,
 670                                     *(conf->hash_locks + hash));
 671                 sh = __find_stripe(conf, sector, conf->generation - previous);
 672                 if (!sh) {
 673                         if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 674                                 sh = get_free_stripe(conf, hash);
 675                                 if (!sh && !test_bit(R5_DID_ALLOC,
 676                                                      &conf->cache_state))
 677                                         set_bit(R5_ALLOC_MORE,
 678                                                 &conf->cache_state);
 679                         }
 680                         if (noblock && sh == NULL)
 681                                 break;
 682                         if (!sh) {
 683                                 set_bit(R5_INACTIVE_BLOCKED,
 684                                         &conf->cache_state);
 685                                 wait_event_lock_irq(
 686                                         conf->wait_for_stripe,
 687                                         !list_empty(conf->inactive_list + hash) &&
 688                                         (atomic_read(&conf->active_stripes)
 689                                          < (conf->max_nr_stripes * 3 / 4)
 690                                          || !test_bit(R5_INACTIVE_BLOCKED,
 691                                                       &conf->cache_state)),
 692                                         *(conf->hash_locks + hash));
 693                                 clear_bit(R5_INACTIVE_BLOCKED,
 694                                           &conf->cache_state);
 695                         } else {
 696                                 init_stripe(sh, sector, previous);
 697                                 atomic_inc(&sh->count);
 698                         }
 699                 } else if (!atomic_inc_not_zero(&sh->count)) {
 700                         spin_lock(&conf->device_lock);
 701                         if (!atomic_read(&sh->count)) {
 702                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 703                                         atomic_inc(&conf->active_stripes);
 704                                 BUG_ON(list_empty(&sh->lru) &&
 705                                        !test_bit(STRIPE_EXPANDING, &sh->state));
 706                                 list_del_init(&sh->lru);
 707                                 if (sh->group) {
 708                                         sh->group->stripes_cnt--;
 709                                         sh->group = NULL;
 710                                 }
 711                         }
 712                         atomic_inc(&sh->count);
 713                         spin_unlock(&conf->device_lock);
 714                 }
 715         } while (sh == NULL);
 716
 717         spin_unlock_irq(conf->hash_locks + hash);
 718         return sh;
 719 }
 720
 721 static bool is_full_stripe_write(struct stripe_head *sh)
 722 {
 723         BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 724         return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 725 }
 726
 727 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 728 {
 729         local_irq_disable();
 730         if (sh1 > sh2) {
 731                 spin_lock(&sh2->stripe_lock);
 732                 spin_lock_nested(&sh1->stripe_lock, 1);
 733         } else {
 734                 spin_lock(&sh1->stripe_lock);
 735                 spin_lock_nested(&sh2->stripe_lock, 1);
 736         }
 737 }
 738
 739 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 740 {
 741         spin_unlock(&sh1->stripe_lock);
 742         spin_unlock(&sh2->stripe_lock);
 743         local_irq_enable();
 744 }
 745
 746 /* Only freshly new full stripe normal write stripe can be added to a batch list */
 747 static bool stripe_can_batch(struct stripe_head *sh)
 748 {
 749         struct r5conf *conf = sh->raid_conf;
 750
 751         if (conf->log)
 752                 return false;
 753         return test_bit(STRIPE_BATCH_READY, &sh->state) &&
 754                 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
 755                 is_full_stripe_write(sh);
 756 }
 757
 758 /* we only do back search */
 759 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
 760 {
 761         struct stripe_head *head;
 762         sector_t head_sector, tmp_sec;
 763         int hash;
 764         int dd_idx;
 765
 766         /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
 767         tmp_sec = sh->sector;
 768         if (!sector_div(tmp_sec, conf->chunk_sectors))
 769                 return;
 770         head_sector = sh->sector - STRIPE_SECTORS;
 771
 772         hash = stripe_hash_locks_hash(head_sector);
 773         spin_lock_irq(conf->hash_locks + hash);
 774         head = __find_stripe(conf, head_sector, conf->generation);
 775         if (head && !atomic_inc_not_zero(&head->count)) {
 776                 spin_lock(&conf->device_lock);
 777                 if (!atomic_read(&head->count)) {
 778                         if (!test_bit(STRIPE_HANDLE, &head->state))
 779                                 atomic_inc(&conf->active_stripes);
 780                         BUG_ON(list_empty(&head->lru) &&
 781                                !test_bit(STRIPE_EXPANDING, &head->state));
 782                         list_del_init(&head->lru);
 783                         if (head->group) {
 784                                 head->group->stripes_cnt--;
 785                                 head->group = NULL;
 786                         }
 787                 }
 788                 atomic_inc(&head->count);
 789                 spin_unlock(&conf->device_lock);
 790         }
 791         spin_unlock_irq(conf->hash_locks + hash);
 792
 793         if (!head)
 794                 return;
 795         if (!stripe_can_batch(head))
 796                 goto out;
 797
 798         lock_two_stripes(head, sh);
 799         /* clear_batch_ready clear the flag */
 800         if (!stripe_can_batch(head) || !stripe_can_batch(sh))
 801                 goto unlock_out;
 802
 803         if (sh->batch_head)
 804                 goto unlock_out;
 805
 806         dd_idx = 0;
 807         while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
 808                 dd_idx++;
 809         if (head->dev[dd_idx].towrite->bi_rw != sh->dev[dd_idx].towrite->bi_rw ||
 810             bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
 811                 goto unlock_out;
 812
 813         if (head->batch_head) {
 814                 spin_lock(&head->batch_head->batch_lock);
 815                 /* This batch list is already running */
 816                 if (!stripe_can_batch(head)) {
 817                         spin_unlock(&head->batch_head->batch_lock);
 818                         goto unlock_out;
 819                 }
 820
 821                 /*
 822                  * at this point, head's BATCH_READY could be cleared, but we
 823                  * can still add the stripe to batch list
 824                  */
 825                 list_add(&sh->batch_list, &head->batch_list);
 826                 spin_unlock(&head->batch_head->batch_lock);
 827
 828                 sh->batch_head = head->batch_head;
 829         } else {
 830                 head->batch_head = head;
 831                 sh->batch_head = head->batch_head;
 832                 spin_lock(&head->batch_lock);
 833                 list_add_tail(&sh->batch_list, &head->batch_list);
 834                 spin_unlock(&head->batch_lock);
 835         }
 836
 837         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 838                 if (atomic_dec_return(&conf->preread_active_stripes)
 839                     < IO_THRESHOLD)
 840                         md_wakeup_thread(conf->mddev->thread);
 841
 842         if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
 843                 int seq = sh->bm_seq;
 844                 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
 845                     sh->batch_head->bm_seq > seq)
 846                         seq = sh->batch_head->bm_seq;
 847                 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
 848                 sh->batch_head->bm_seq = seq;
 849         }
 850
 851         atomic_inc(&sh->count);
 852 unlock_out:
 853         unlock_two_stripes(head, sh);
 854 out:
 855         raid5_release_stripe(head);
 856 }
 857
 858 /* Determine if 'data_offset' or 'new_data_offset' should be used
 859  * in this stripe_head.
 860  */
 861 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
 862 {
 863         sector_t progress = conf->reshape_progress;
 864         /* Need a memory barrier to make sure we see the value
 865          * of conf->generation, or ->data_offset that was set before
 866          * reshape_progress was updated.
 867          */
 868         smp_rmb();
 869         if (progress == MaxSector)
 870                 return 0;
 871         if (sh->generation == conf->generation - 1)
 872                 return 0;
 873         /* We are in a reshape, and this is a new-generation stripe,
 874          * so use new_data_offset.
 875          */
 876         return 1;
 877 }
 878
 879 static void
 880 raid5_end_read_request(struct bio *bi);
 881 static void
 882 raid5_end_write_request(struct bio *bi);
 883
 884 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
 885 {
 886         struct r5conf *conf = sh->raid_conf;
 887         int i, disks = sh->disks;
 888         struct stripe_head *head_sh = sh;
 889
 890         might_sleep();
 891
 892         if (r5l_write_stripe(conf->log, sh) == 0)
 893                 return;
 894         for (i = disks; i--; ) {
 895                 int op, op_flags = 0;
 896                 int replace_only = 0;
 897                 struct bio *bi, *rbi;
 898                 struct md_rdev *rdev, *rrdev = NULL;
 899
 900                 sh = head_sh;
 901                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
 902                         op = REQ_OP_WRITE;
 903                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
 904                                 op_flags = WRITE_FUA;
 905                         if (test_bit(R5_Discard, &sh->dev[i].flags))
 906                                 op = REQ_OP_DISCARD;
 907                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
 908                         op = REQ_OP_READ;
 909                 else if (test_and_clear_bit(R5_WantReplace,
 910                                             &sh->dev[i].flags)) {
 911                         op = REQ_OP_WRITE;
 912                         replace_only = 1;
 913                 } else
 914                         continue;
 915                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
 916                         op_flags |= REQ_SYNC;
 917
 918 again:
 919                 bi = &sh->dev[i].req;
 920                 rbi = &sh->dev[i].rreq; /* For writing to replacement */
 921
 922                 rcu_read_lock();
 923                 rrdev = rcu_dereference(conf->disks[i].replacement);
 924                 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
 925                 rdev = rcu_dereference(conf->disks[i].rdev);
 926                 if (!rdev) {
 927                         rdev = rrdev;
 928                         rrdev = NULL;
 929                 }
 930                 if (op_is_write(op)) {
 931                         if (replace_only)
 932                                 rdev = NULL;
 933                         if (rdev == rrdev)
 934                                 /* We raced and saw duplicates */
 935                                 rrdev = NULL;
 936                 } else {
 937                         if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
 938                                 rdev = rrdev;
 939                         rrdev = NULL;
 940                 }
 941
 942                 if (rdev && test_bit(Faulty, &rdev->flags))
 943                         rdev = NULL;
 944                 if (rdev)
 945                         atomic_inc(&rdev->nr_pending);
 946                 if (rrdev && test_bit(Faulty, &rrdev->flags))
 947                         rrdev = NULL;
 948                 if (rrdev)
 949                         atomic_inc(&rrdev->nr_pending);
 950                 rcu_read_unlock();
 951
 952                 /* We have already checked bad blocks for reads.  Now
 953                  * need to check for writes.  We never accept write errors
 954                  * on the replacement, so we don't to check rrdev.
 955                  */
 956                 while (op_is_write(op) && rdev &&
 957                        test_bit(WriteErrorSeen, &rdev->flags)) {
 958                         sector_t first_bad;
 959                         int bad_sectors;
 960                         int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
 961                                               &first_bad, &bad_sectors);
 962                         if (!bad)
 963                                 break;
 964
 965                         if (bad < 0) {
 966                                 set_bit(BlockedBadBlocks, &rdev->flags);
 967                                 if (!conf->mddev->external &&
 968                                     conf->mddev->flags) {
 969                                         /* It is very unlikely, but we might
 970                                          * still need to write out the
 971                                          * bad block log - better give it
 972                                          * a chance*/
 973                                         md_check_recovery(conf->mddev);
 974                                 }
 975                                 /*
 976                                  * Because md_wait_for_blocked_rdev
 977                                  * will dec nr_pending, we must
 978                                  * increment it first.
 979                                  */
 980                                 atomic_inc(&rdev->nr_pending);
 981                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
 982                         } else {
 983                                 /* Acknowledged bad block - skip the write */
 984                                 rdev_dec_pending(rdev, conf->mddev);
 985                                 rdev = NULL;
 986                         }
 987                 }
 988
 989                 if (rdev) {
 990                         if (s->syncing || s->expanding || s->expanded
 991                             || s->replacing)
 992                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
 993
 994                         set_bit(STRIPE_IO_STARTED, &sh->state);
 995
 996                         bio_reset(bi);
 997                         bi->bi_bdev = rdev->bdev;
 998                         bio_set_op_attrs(bi, op, op_flags);
 999                         bi->bi_end_io = op_is_write(op)
1000                                 ? raid5_end_write_request
1001                                 : raid5_end_read_request;
1002                         bi->bi_private = sh;
1003
1004                         pr_debug("%s: for %llu schedule op %d on disc %d\n",
1005                                 __func__, (unsigned long long)sh->sector,
1006                                 bi->bi_rw, i);
1007                         atomic_inc(&sh->count);
1008                         if (sh != head_sh)
1009                                 atomic_inc(&head_sh->count);
1010                         if (use_new_offset(conf, sh))
1011                                 bi->bi_iter.bi_sector = (sh->sector
1012                                                  + rdev->new_data_offset);
1013                         else
1014                                 bi->bi_iter.bi_sector = (sh->sector
1015                                                  + rdev->data_offset);
1016                         if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1017                                 bi->bi_rw |= REQ_NOMERGE;
1018
1019                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1020                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1021                         sh->dev[i].vec.bv_page = sh->dev[i].page;
1022                         bi->bi_vcnt = 1;
1023                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1024                         bi->bi_io_vec[0].bv_offset = 0;
1025                         bi->bi_iter.bi_size = STRIPE_SIZE;
1026                         /*
1027                          * If this is discard request, set bi_vcnt 0. We don't
1028                          * want to confuse SCSI because SCSI will replace payload
1029                          */
1030                         if (op == REQ_OP_DISCARD)
1031                                 bi->bi_vcnt = 0;
1032                         if (rrdev)
1033                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1034
1035                         if (conf->mddev->gendisk)
1036                                 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1037                                                       bi, disk_devt(conf->mddev->gendisk),
1038                                                       sh->dev[i].sector);
1039                         generic_make_request(bi);
1040                 }
1041                 if (rrdev) {
1042                         if (s->syncing || s->expanding || s->expanded
1043                             || s->replacing)
1044                                 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1045
1046                         set_bit(STRIPE_IO_STARTED, &sh->state);
1047
1048                         bio_reset(rbi);
1049                         rbi->bi_bdev = rrdev->bdev;
1050                         bio_set_op_attrs(rbi, op, op_flags);
1051                         BUG_ON(!op_is_write(op));
1052                         rbi->bi_end_io = raid5_end_write_request;
1053                         rbi->bi_private = sh;
1054
1055                         pr_debug("%s: for %llu schedule op %d on "
1056                                  "replacement disc %d\n",
1057                                 __func__, (unsigned long long)sh->sector,
1058                                 rbi->bi_rw, i);
1059                         atomic_inc(&sh->count);
1060                         if (sh != head_sh)
1061                                 atomic_inc(&head_sh->count);
1062                         if (use_new_offset(conf, sh))
1063                                 rbi->bi_iter.bi_sector = (sh->sector
1064                                                   + rrdev->new_data_offset);
1065                         else
1066                                 rbi->bi_iter.bi_sector = (sh->sector
1067                                                   + rrdev->data_offset);
1068                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1069                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1070                         sh->dev[i].rvec.bv_page = sh->dev[i].page;
1071                         rbi->bi_vcnt = 1;
1072                         rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1073                         rbi->bi_io_vec[0].bv_offset = 0;
1074                         rbi->bi_iter.bi_size = STRIPE_SIZE;
1075                         /*
1076                          * If this is discard request, set bi_vcnt 0. We don't
1077                          * want to confuse SCSI because SCSI will replace payload
1078                          */
1079                         if (op == REQ_OP_DISCARD)
1080                                 rbi->bi_vcnt = 0;
1081                         if (conf->mddev->gendisk)
1082                                 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1083                                                       rbi, disk_devt(conf->mddev->gendisk),
1084                                                       sh->dev[i].sector);
1085                         generic_make_request(rbi);
1086                 }
1087                 if (!rdev && !rrdev) {
1088                         if (op_is_write(op))
1089                                 set_bit(STRIPE_DEGRADED, &sh->state);
1090                         pr_debug("skip op %d on disc %d for sector %llu\n",
1091                                 bi->bi_rw, i, (unsigned long long)sh->sector);
1092                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1093                         set_bit(STRIPE_HANDLE, &sh->state);
1094                 }
1095
1096                 if (!head_sh->batch_head)
1097                         continue;
1098                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1099                                       batch_list);
1100                 if (sh != head_sh)
1101                         goto again;
1102         }
1103 }
1104
1105 static struct dma_async_tx_descriptor *
1106 async_copy_data(int frombio, struct bio *bio, struct page **page,
1107         sector_t sector, struct dma_async_tx_descriptor *tx,
1108         struct stripe_head *sh)
1109 {
1110         struct bio_vec bvl;
1111         struct bvec_iter iter;
1112         struct page *bio_page;
1113         int page_offset;
1114         struct async_submit_ctl submit;
1115         enum async_tx_flags flags = 0;
1116
1117         if (bio->bi_iter.bi_sector >= sector)
1118                 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1119         else
1120                 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1121
1122         if (frombio)
1123                 flags |= ASYNC_TX_FENCE;
1124         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1125
1126         bio_for_each_segment(bvl, bio, iter) {
1127                 int len = bvl.bv_len;
1128                 int clen;
1129                 int b_offset = 0;
1130
1131                 if (page_offset < 0) {
1132                         b_offset = -page_offset;
1133                         page_offset += b_offset;
1134                         len -= b_offset;
1135                 }
1136
1137                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1138                         clen = STRIPE_SIZE - page_offset;
1139                 else
1140                         clen = len;
1141
1142                 if (clen > 0) {
1143                         b_offset += bvl.bv_offset;
1144                         bio_page = bvl.bv_page;
1145                         if (frombio) {
1146                                 if (sh->raid_conf->skip_copy &&
1147                                     b_offset == 0 && page_offset == 0 &&
1148                                     clen == STRIPE_SIZE)
1149                                         *page = bio_page;
1150                                 else
1151                                         tx = async_memcpy(*page, bio_page, page_offset,
1152                                                   b_offset, clen, &submit);
1153                         } else
1154                                 tx = async_memcpy(bio_page, *page, b_offset,
1155                                                   page_offset, clen, &submit);
1156                 }
1157                 /* chain the operations */
1158                 submit.depend_tx = tx;
1159
1160                 if (clen < len) /* hit end of page */
1161                         break;
1162                 page_offset +=  len;
1163         }
1164
1165         return tx;
1166 }
1167
1168 static void ops_complete_biofill(void *stripe_head_ref)
1169 {
1170         struct stripe_head *sh = stripe_head_ref;
1171         struct bio_list return_bi = BIO_EMPTY_LIST;
1172         int i;
1173
1174         pr_debug("%s: stripe %llu\n", __func__,
1175                 (unsigned long long)sh->sector);
1176
1177         /* clear completed biofills */
1178         for (i = sh->disks; i--; ) {
1179                 struct r5dev *dev = &sh->dev[i];
1180
1181                 /* acknowledge completion of a biofill operation */
1182                 /* and check if we need to reply to a read request,
1183                  * new R5_Wantfill requests are held off until
1184                  * !STRIPE_BIOFILL_RUN
1185                  */
1186                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1187                         struct bio *rbi, *rbi2;
1188
1189                         BUG_ON(!dev->read);
1190                         rbi = dev->read;
1191                         dev->read = NULL;
1192                         while (rbi && rbi->bi_iter.bi_sector <
1193                                 dev->sector + STRIPE_SECTORS) {
1194                                 rbi2 = r5_next_bio(rbi, dev->sector);
1195                                 if (!raid5_dec_bi_active_stripes(rbi))
1196                                         bio_list_add(&return_bi, rbi);
1197                                 rbi = rbi2;
1198                         }
1199                 }
1200         }
1201         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1202
1203         return_io(&return_bi);
1204
1205         set_bit(STRIPE_HANDLE, &sh->state);
1206         raid5_release_stripe(sh);
1207 }
1208
1209 static void ops_run_biofill(struct stripe_head *sh)
1210 {
1211         struct dma_async_tx_descriptor *tx = NULL;
1212         struct async_submit_ctl submit;
1213         int i;
1214
1215         BUG_ON(sh->batch_head);
1216         pr_debug("%s: stripe %llu\n", __func__,
1217                 (unsigned long long)sh->sector);
1218
1219         for (i = sh->disks; i--; ) {
1220                 struct r5dev *dev = &sh->dev[i];
1221                 if (test_bit(R5_Wantfill, &dev->flags)) {
1222                         struct bio *rbi;
1223                         spin_lock_irq(&sh->stripe_lock);
1224                         dev->read = rbi = dev->toread;
1225                         dev->toread = NULL;
1226                         spin_unlock_irq(&sh->stripe_lock);
1227                         while (rbi && rbi->bi_iter.bi_sector <
1228                                 dev->sector + STRIPE_SECTORS) {
1229                                 tx = async_copy_data(0, rbi, &dev->page,
1230                                         dev->sector, tx, sh);
1231                                 rbi = r5_next_bio(rbi, dev->sector);
1232                         }
1233                 }
1234         }
1235
1236         atomic_inc(&sh->count);
1237         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1238         async_trigger_callback(&submit);
1239 }
1240
1241 static void mark_target_uptodate(struct stripe_head *sh, int target)
1242 {
1243         struct r5dev *tgt;
1244
1245         if (target < 0)
1246                 return;
1247
1248         tgt = &sh->dev[target];
1249         set_bit(R5_UPTODATE, &tgt->flags);
1250         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1251         clear_bit(R5_Wantcompute, &tgt->flags);
1252 }
1253
1254 static void ops_complete_compute(void *stripe_head_ref)
1255 {
1256         struct stripe_head *sh = stripe_head_ref;
1257
1258         pr_debug("%s: stripe %llu\n", __func__,
1259                 (unsigned long long)sh->sector);
1260
1261         /* mark the computed target(s) as uptodate */
1262         mark_target_uptodate(sh, sh->ops.target);
1263         mark_target_uptodate(sh, sh->ops.target2);
1264
1265         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1266         if (sh->check_state == check_state_compute_run)
1267                 sh->check_state = check_state_compute_result;
1268         set_bit(STRIPE_HANDLE, &sh->state);
1269         raid5_release_stripe(sh);
1270 }
1271
1272 /* return a pointer to the address conversion region of the scribble buffer */
1273 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1274                                  struct raid5_percpu *percpu, int i)
1275 {
1276         void *addr;
1277
1278         addr = flex_array_get(percpu->scribble, i);
1279         return addr + sizeof(struct page *) * (sh->disks + 2);
1280 }
1281
1282 /* return a pointer to the address conversion region of the scribble buffer */
1283 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1284 {
1285         void *addr;
1286
1287         addr = flex_array_get(percpu->scribble, i);
1288         return addr;
1289 }
1290
1291 static struct dma_async_tx_descriptor *
1292 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1293 {
1294         int disks = sh->disks;
1295         struct page **xor_srcs = to_addr_page(percpu, 0);
1296         int target = sh->ops.target;
1297         struct r5dev *tgt = &sh->dev[target];
1298         struct page *xor_dest = tgt->page;
1299         int count = 0;
1300         struct dma_async_tx_descriptor *tx;
1301         struct async_submit_ctl submit;
1302         int i;
1303
1304         BUG_ON(sh->batch_head);
1305
1306         pr_debug("%s: stripe %llu block: %d\n",
1307                 __func__, (unsigned long long)sh->sector, target);
1308         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1309
1310         for (i = disks; i--; )
1311                 if (i != target)
1312                         xor_srcs[count++] = sh->dev[i].page;
1313
1314         atomic_inc(&sh->count);
1315
1316         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1317                           ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1318         if (unlikely(count == 1))
1319                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1320         else
1321                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1322
1323         return tx;
1324 }
1325
1326 /* set_syndrome_sources - populate source buffers for gen_syndrome
1327  * @srcs - (struct page *) array of size sh->disks
1328  * @sh - stripe_head to parse
1329  *
1330  * Populates srcs in proper layout order for the stripe and returns the
1331  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1332  * destination buffer is recorded in srcs[count] and the Q destination
1333  * is recorded in srcs[count+1]].
1334  */
1335 static int set_syndrome_sources(struct page **srcs,
1336                                 struct stripe_head *sh,
1337                                 int srctype)
1338 {
1339         int disks = sh->disks;
1340         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1341         int d0_idx = raid6_d0(sh);
1342         int count;
1343         int i;
1344
1345         for (i = 0; i < disks; i++)
1346                 srcs[i] = NULL;
1347
1348         count = 0;
1349         i = d0_idx;
1350         do {
1351                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1352                 struct r5dev *dev = &sh->dev[i];
1353
1354                 if (i == sh->qd_idx || i == sh->pd_idx ||
1355                     (srctype == SYNDROME_SRC_ALL) ||
1356                     (srctype == SYNDROME_SRC_WANT_DRAIN &&
1357                      test_bit(R5_Wantdrain, &dev->flags)) ||
1358                     (srctype == SYNDROME_SRC_WRITTEN &&
1359                      dev->written))
1360                         srcs[slot] = sh->dev[i].page;
1361                 i = raid6_next_disk(i, disks);
1362         } while (i != d0_idx);
1363
1364         return syndrome_disks;
1365 }
1366
1367 static struct dma_async_tx_descriptor *
1368 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1369 {
1370         int disks = sh->disks;
1371         struct page **blocks = to_addr_page(percpu, 0);
1372         int target;
1373         int qd_idx = sh->qd_idx;
1374         struct dma_async_tx_descriptor *tx;
1375         struct async_submit_ctl submit;
1376         struct r5dev *tgt;
1377         struct page *dest;
1378         int i;
1379         int count;
1380
1381         BUG_ON(sh->batch_head);
1382         if (sh->ops.target < 0)
1383                 target = sh->ops.target2;
1384         else if (sh->ops.target2 < 0)
1385                 target = sh->ops.target;
1386         else
1387                 /* we should only have one valid target */
1388                 BUG();
1389         BUG_ON(target < 0);
1390         pr_debug("%s: stripe %llu block: %d\n",
1391                 __func__, (unsigned long long)sh->sector, target);
1392
1393         tgt = &sh->dev[target];
1394         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1395         dest = tgt->page;
1396
1397         atomic_inc(&sh->count);
1398
1399         if (target == qd_idx) {
1400                 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1401                 blocks[count] = NULL; /* regenerating p is not necessary */
1402                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1403                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1404                                   ops_complete_compute, sh,
1405                                   to_addr_conv(sh, percpu, 0));
1406                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1407         } else {
1408                 /* Compute any data- or p-drive using XOR */
1409                 count = 0;
1410                 for (i = disks; i-- ; ) {
1411                         if (i == target || i == qd_idx)
1412                                 continue;
1413                         blocks[count++] = sh->dev[i].page;
1414                 }
1415
1416                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1417                                   NULL, ops_complete_compute, sh,
1418                                   to_addr_conv(sh, percpu, 0));
1419                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1420         }
1421
1422         return tx;
1423 }
1424
1425 static struct dma_async_tx_descriptor *
1426 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1427 {
1428         int i, count, disks = sh->disks;
1429         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1430         int d0_idx = raid6_d0(sh);
1431         int faila = -1, failb = -1;
1432         int target = sh->ops.target;
1433         int target2 = sh->ops.target2;
1434         struct r5dev *tgt = &sh->dev[target];
1435         struct r5dev *tgt2 = &sh->dev[target2];
1436         struct dma_async_tx_descriptor *tx;
1437         struct page **blocks = to_addr_page(percpu, 0);
1438         struct async_submit_ctl submit;
1439
1440         BUG_ON(sh->batch_head);
1441         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1442                  __func__, (unsigned long long)sh->sector, target, target2);
1443         BUG_ON(target < 0 || target2 < 0);
1444         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1445         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1446
1447         /* we need to open-code set_syndrome_sources to handle the
1448          * slot number conversion for 'faila' and 'failb'
1449          */
1450         for (i = 0; i < disks ; i++)
1451                 blocks[i] = NULL;
1452         count = 0;
1453         i = d0_idx;
1454         do {
1455                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1456
1457                 blocks[slot] = sh->dev[i].page;
1458
1459                 if (i == target)
1460                         faila = slot;
1461                 if (i == target2)
1462                         failb = slot;
1463                 i = raid6_next_disk(i, disks);
1464         } while (i != d0_idx);
1465
1466         BUG_ON(faila == failb);
1467         if (failb < faila)
1468                 swap(faila, failb);
1469         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1470                  __func__, (unsigned long long)sh->sector, faila, failb);
1471
1472         atomic_inc(&sh->count);
1473
1474         if (failb == syndrome_disks+1) {
1475                 /* Q disk is one of the missing disks */
1476                 if (faila == syndrome_disks) {
1477                         /* Missing P+Q, just recompute */
1478                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1479                                           ops_complete_compute, sh,
1480                                           to_addr_conv(sh, percpu, 0));
1481                         return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1482                                                   STRIPE_SIZE, &submit);
1483                 } else {
1484                         struct page *dest;
1485                         int data_target;
1486                         int qd_idx = sh->qd_idx;
1487
1488                         /* Missing D+Q: recompute D from P, then recompute Q */
1489                         if (target == qd_idx)
1490                                 data_target = target2;
1491                         else
1492                                 data_target = target;
1493
1494                         count = 0;
1495                         for (i = disks; i-- ; ) {
1496                                 if (i == data_target || i == qd_idx)
1497                                         continue;
1498                                 blocks[count++] = sh->dev[i].page;
1499                         }
1500                         dest = sh->dev[data_target].page;
1501                         init_async_submit(&submit,
1502                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1503                                           NULL, NULL, NULL,
1504                                           to_addr_conv(sh, percpu, 0));
1505                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1506                                        &submit);
1507
1508                         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1509                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1510                                           ops_complete_compute, sh,
1511                                           to_addr_conv(sh, percpu, 0));
1512                         return async_gen_syndrome(blocks, 0, count+2,
1513                                                   STRIPE_SIZE, &submit);
1514                 }
1515         } else {
1516                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1517                                   ops_complete_compute, sh,
1518                                   to_addr_conv(sh, percpu, 0));
1519                 if (failb == syndrome_disks) {
1520                         /* We're missing D+P. */
1521                         return async_raid6_datap_recov(syndrome_disks+2,
1522                                                        STRIPE_SIZE, faila,
1523                                                        blocks, &submit);
1524                 } else {
1525                         /* We're missing D+D. */
1526                         return async_raid6_2data_recov(syndrome_disks+2,
1527                                                        STRIPE_SIZE, faila, failb,
1528                                                        blocks, &submit);
1529                 }
1530         }
1531 }
1532
1533 static void ops_complete_prexor(void *stripe_head_ref)
1534 {
1535         struct stripe_head *sh = stripe_head_ref;
1536
1537         pr_debug("%s: stripe %llu\n", __func__,
1538                 (unsigned long long)sh->sector);
1539 }
1540
1541 static struct dma_async_tx_descriptor *
1542 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1543                 struct dma_async_tx_descriptor *tx)
1544 {
1545         int disks = sh->disks;
1546         struct page **xor_srcs = to_addr_page(percpu, 0);
1547         int count = 0, pd_idx = sh->pd_idx, i;
1548         struct async_submit_ctl submit;
1549
1550         /* existing parity data subtracted */
1551         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1552
1553         BUG_ON(sh->batch_head);
1554         pr_debug("%s: stripe %llu\n", __func__,
1555                 (unsigned long long)sh->sector);
1556
1557         for (i = disks; i--; ) {
1558                 struct r5dev *dev = &sh->dev[i];
1559                 /* Only process blocks that are known to be uptodate */
1560                 if (test_bit(R5_Wantdrain, &dev->flags))
1561                         xor_srcs[count++] = dev->page;
1562         }
1563
1564         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1565                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1566         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1567
1568         return tx;
1569 }
1570
1571 static struct dma_async_tx_descriptor *
1572 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1573                 struct dma_async_tx_descriptor *tx)
1574 {
1575         struct page **blocks = to_addr_page(percpu, 0);
1576         int count;
1577         struct async_submit_ctl submit;
1578
1579         pr_debug("%s: stripe %llu\n", __func__,
1580                 (unsigned long long)sh->sector);
1581
1582         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1583
1584         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1585                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1586         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1587
1588         return tx;
1589 }
1590
1591 static struct dma_async_tx_descriptor *
1592 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1593 {
1594         int disks = sh->disks;
1595         int i;
1596         struct stripe_head *head_sh = sh;
1597
1598         pr_debug("%s: stripe %llu\n", __func__,
1599                 (unsigned long long)sh->sector);
1600
1601         for (i = disks; i--; ) {
1602                 struct r5dev *dev;
1603                 struct bio *chosen;
1604
1605                 sh = head_sh;
1606                 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1607                         struct bio *wbi;
1608
1609 again:
1610                         dev = &sh->dev[i];
1611                         spin_lock_irq(&sh->stripe_lock);
1612                         chosen = dev->towrite;
1613                         dev->towrite = NULL;
1614                         sh->overwrite_disks = 0;
1615                         BUG_ON(dev->written);
1616                         wbi = dev->written = chosen;
1617                         spin_unlock_irq(&sh->stripe_lock);
1618                         WARN_ON(dev->page != dev->orig_page);
1619
1620                         while (wbi && wbi->bi_iter.bi_sector <
1621                                 dev->sector + STRIPE_SECTORS) {
1622                                 if (wbi->bi_rw & REQ_FUA)
1623                                         set_bit(R5_WantFUA, &dev->flags);
1624                                 if (wbi->bi_rw & REQ_SYNC)
1625                                         set_bit(R5_SyncIO, &dev->flags);
1626                                 if (bio_op(wbi) == REQ_OP_DISCARD)
1627                                         set_bit(R5_Discard, &dev->flags);
1628                                 else {
1629                                         tx = async_copy_data(1, wbi, &dev->page,
1630                                                 dev->sector, tx, sh);
1631                                         if (dev->page != dev->orig_page) {
1632                                                 set_bit(R5_SkipCopy, &dev->flags);
1633                                                 clear_bit(R5_UPTODATE, &dev->flags);
1634                                                 clear_bit(R5_OVERWRITE, &dev->flags);
1635                                         }
1636                                 }
1637                                 wbi = r5_next_bio(wbi, dev->sector);
1638                         }
1639
1640                         if (head_sh->batch_head) {
1641                                 sh = list_first_entry(&sh->batch_list,
1642                                                       struct stripe_head,
1643                                                       batch_list);
1644                                 if (sh == head_sh)
1645                                         continue;
1646                                 goto again;
1647                         }
1648                 }
1649         }
1650
1651         return tx;
1652 }
1653
1654 static void ops_complete_reconstruct(void *stripe_head_ref)
1655 {
1656         struct stripe_head *sh = stripe_head_ref;
1657         int disks = sh->disks;
1658         int pd_idx = sh->pd_idx;
1659         int qd_idx = sh->qd_idx;
1660         int i;
1661         bool fua = false, sync = false, discard = false;
1662
1663         pr_debug("%s: stripe %llu\n", __func__,
1664                 (unsigned long long)sh->sector);
1665
1666         for (i = disks; i--; ) {
1667                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1668                 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1669                 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1670         }
1671
1672         for (i = disks; i--; ) {
1673                 struct r5dev *dev = &sh->dev[i];
1674
1675                 if (dev->written || i == pd_idx || i == qd_idx) {
1676                         if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1677                                 set_bit(R5_UPTODATE, &dev->flags);
1678                         if (fua)
1679                                 set_bit(R5_WantFUA, &dev->flags);
1680                         if (sync)
1681                                 set_bit(R5_SyncIO, &dev->flags);
1682                 }
1683         }
1684
1685         if (sh->reconstruct_state == reconstruct_state_drain_run)
1686                 sh->reconstruct_state = reconstruct_state_drain_result;
1687         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1688                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1689         else {
1690                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1691                 sh->reconstruct_state = reconstruct_state_result;
1692         }
1693
1694         set_bit(STRIPE_HANDLE, &sh->state);
1695         raid5_release_stripe(sh);
1696 }
1697
1698 static void
1699 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1700                      struct dma_async_tx_descriptor *tx)
1701 {
1702         int disks = sh->disks;
1703         struct page **xor_srcs;
1704         struct async_submit_ctl submit;
1705         int count, pd_idx = sh->pd_idx, i;
1706         struct page *xor_dest;
1707         int prexor = 0;
1708         unsigned long flags;
1709         int j = 0;
1710         struct stripe_head *head_sh = sh;
1711         int last_stripe;
1712
1713         pr_debug("%s: stripe %llu\n", __func__,
1714                 (unsigned long long)sh->sector);
1715
1716         for (i = 0; i < sh->disks; i++) {
1717                 if (pd_idx == i)
1718                         continue;
1719                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1720                         break;
1721         }
1722         if (i >= sh->disks) {
1723                 atomic_inc(&sh->count);
1724                 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1725                 ops_complete_reconstruct(sh);
1726                 return;
1727         }
1728 again:
1729         count = 0;
1730         xor_srcs = to_addr_page(percpu, j);
1731         /* check if prexor is active which means only process blocks
1732          * that are part of a read-modify-write (written)
1733          */
1734         if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1735                 prexor = 1;
1736                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1737                 for (i = disks; i--; ) {
1738                         struct r5dev *dev = &sh->dev[i];
1739                         if (head_sh->dev[i].written)
1740                                 xor_srcs[count++] = dev->page;
1741                 }
1742         } else {
1743                 xor_dest = sh->dev[pd_idx].page;
1744                 for (i = disks; i--; ) {
1745                         struct r5dev *dev = &sh->dev[i];
1746                         if (i != pd_idx)
1747                                 xor_srcs[count++] = dev->page;
1748                 }
1749         }
1750
1751         /* 1/ if we prexor'd then the dest is reused as a source
1752          * 2/ if we did not prexor then we are redoing the parity
1753          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1754          * for the synchronous xor case
1755          */
1756         last_stripe = !head_sh->batch_head ||
1757                 list_first_entry(&sh->batch_list,
1758                                  struct stripe_head, batch_list) == head_sh;
1759         if (last_stripe) {
1760                 flags = ASYNC_TX_ACK |
1761                         (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1762
1763                 atomic_inc(&head_sh->count);
1764                 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1765                                   to_addr_conv(sh, percpu, j));
1766         } else {
1767                 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1768                 init_async_submit(&submit, flags, tx, NULL, NULL,
1769                                   to_addr_conv(sh, percpu, j));
1770         }
1771
1772         if (unlikely(count == 1))
1773                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1774         else
1775                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1776         if (!last_stripe) {
1777                 j++;
1778                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1779                                       batch_list);
1780                 goto again;
1781         }
1782 }
1783
1784 static void
1785 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1786                      struct dma_async_tx_descriptor *tx)
1787 {
1788         struct async_submit_ctl submit;
1789         struct page **blocks;
1790         int count, i, j = 0;
1791         struct stripe_head *head_sh = sh;
1792         int last_stripe;
1793         int synflags;
1794         unsigned long txflags;
1795
1796         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1797
1798         for (i = 0; i < sh->disks; i++) {
1799                 if (sh->pd_idx == i || sh->qd_idx == i)
1800                         continue;
1801                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1802                         break;
1803         }
1804         if (i >= sh->disks) {
1805                 atomic_inc(&sh->count);
1806                 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1807                 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1808                 ops_complete_reconstruct(sh);
1809                 return;
1810         }
1811
1812 again:
1813         blocks = to_addr_page(percpu, j);
1814
1815         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1816                 synflags = SYNDROME_SRC_WRITTEN;
1817                 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1818         } else {
1819                 synflags = SYNDROME_SRC_ALL;
1820                 txflags = ASYNC_TX_ACK;
1821         }
1822
1823         count = set_syndrome_sources(blocks, sh, synflags);
1824         last_stripe = !head_sh->batch_head ||
1825                 list_first_entry(&sh->batch_list,
1826                                  struct stripe_head, batch_list) == head_sh;
1827
1828         if (last_stripe) {
1829                 atomic_inc(&head_sh->count);
1830                 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1831                                   head_sh, to_addr_conv(sh, percpu, j));
1832         } else
1833                 init_async_submit(&submit, 0, tx, NULL, NULL,
1834                                   to_addr_conv(sh, percpu, j));
1835         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1836         if (!last_stripe) {
1837                 j++;
1838                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1839                                       batch_list);
1840                 goto again;
1841         }
1842 }
1843
1844 static void ops_complete_check(void *stripe_head_ref)
1845 {
1846         struct stripe_head *sh = stripe_head_ref;
1847
1848         pr_debug("%s: stripe %llu\n", __func__,
1849                 (unsigned long long)sh->sector);
1850
1851         sh->check_state = check_state_check_result;
1852         set_bit(STRIPE_HANDLE, &sh->state);
1853         raid5_release_stripe(sh);
1854 }
1855
1856 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1857 {
1858         int disks = sh->disks;
1859         int pd_idx = sh->pd_idx;
1860         int qd_idx = sh->qd_idx;
1861         struct page *xor_dest;
1862         struct page **xor_srcs = to_addr_page(percpu, 0);
1863         struct dma_async_tx_descriptor *tx;
1864         struct async_submit_ctl submit;
1865         int count;
1866         int i;
1867
1868         pr_debug("%s: stripe %llu\n", __func__,
1869                 (unsigned long long)sh->sector);
1870
1871         BUG_ON(sh->batch_head);
1872         count = 0;
1873         xor_dest = sh->dev[pd_idx].page;
1874         xor_srcs[count++] = xor_dest;
1875         for (i = disks; i--; ) {
1876                 if (i == pd_idx || i == qd_idx)
1877                         continue;
1878                 xor_srcs[count++] = sh->dev[i].page;
1879         }
1880
1881         init_async_submit(&submit, 0, NULL, NULL, NULL,
1882                           to_addr_conv(sh, percpu, 0));
1883         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1884                            &sh->ops.zero_sum_result, &submit);
1885
1886         atomic_inc(&sh->count);
1887         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1888         tx = async_trigger_callback(&submit);
1889 }
1890
1891 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1892 {
1893         struct page **srcs = to_addr_page(percpu, 0);
1894         struct async_submit_ctl submit;
1895         int count;
1896
1897         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1898                 (unsigned long long)sh->sector, checkp);
1899
1900         BUG_ON(sh->batch_head);
1901         count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
1902         if (!checkp)
1903                 srcs[count] = NULL;
1904
1905         atomic_inc(&sh->count);
1906         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1907                           sh, to_addr_conv(sh, percpu, 0));
1908         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1909                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1910 }
1911
1912 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1913 {
1914         int overlap_clear = 0, i, disks = sh->disks;
1915         struct dma_async_tx_descriptor *tx = NULL;
1916         struct r5conf *conf = sh->raid_conf;
1917         int level = conf->level;
1918         struct raid5_percpu *percpu;
1919         unsigned long cpu;
1920
1921         cpu = get_cpu();
1922         percpu = per_cpu_ptr(conf->percpu, cpu);
1923         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1924                 ops_run_biofill(sh);
1925                 overlap_clear++;
1926         }
1927
1928         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1929                 if (level < 6)
1930                         tx = ops_run_compute5(sh, percpu);
1931                 else {
1932                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
1933                                 tx = ops_run_compute6_1(sh, percpu);
1934                         else
1935                                 tx = ops_run_compute6_2(sh, percpu);
1936                 }
1937                 /* terminate the chain if reconstruct is not set to be run */
1938                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1939                         async_tx_ack(tx);
1940         }
1941
1942         if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
1943                 if (level < 6)
1944                         tx = ops_run_prexor5(sh, percpu, tx);
1945                 else
1946                         tx = ops_run_prexor6(sh, percpu, tx);
1947         }
1948
1949         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1950                 tx = ops_run_biodrain(sh, tx);
1951                 overlap_clear++;
1952         }
1953
1954         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1955                 if (level < 6)
1956                         ops_run_reconstruct5(sh, percpu, tx);
1957                 else
1958                         ops_run_reconstruct6(sh, percpu, tx);
1959         }
1960
1961         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1962                 if (sh->check_state == check_state_run)
1963                         ops_run_check_p(sh, percpu);
1964                 else if (sh->check_state == check_state_run_q)
1965                         ops_run_check_pq(sh, percpu, 0);
1966                 else if (sh->check_state == check_state_run_pq)
1967                         ops_run_check_pq(sh, percpu, 1);
1968                 else
1969                         BUG();
1970         }
1971
1972         if (overlap_clear && !sh->batch_head)
1973                 for (i = disks; i--; ) {
1974                         struct r5dev *dev = &sh->dev[i];
1975                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1976                                 wake_up(&sh->raid_conf->wait_for_overlap);
1977                 }
1978         put_cpu();
1979 }
1980
1981 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp)
1982 {
1983         struct stripe_head *sh;
1984
1985         sh = kmem_cache_zalloc(sc, gfp);
1986         if (sh) {
1987                 spin_lock_init(&sh->stripe_lock);
1988                 spin_lock_init(&sh->batch_lock);
1989                 INIT_LIST_HEAD(&sh->batch_list);
1990                 INIT_LIST_HEAD(&sh->lru);
1991                 atomic_set(&sh->count, 1);
1992         }
1993         return sh;
1994 }
1995 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
1996 {
1997         struct stripe_head *sh;
1998
1999         sh = alloc_stripe(conf->slab_cache, gfp);
2000         if (!sh)
2001                 return 0;
2002
2003         sh->raid_conf = conf;
2004
2005         if (grow_buffers(sh, gfp)) {
2006                 shrink_buffers(sh);
2007                 kmem_cache_free(conf->slab_cache, sh);
2008                 return 0;
2009         }
2010         sh->hash_lock_index =
2011                 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2012         /* we just created an active stripe so... */
2013         atomic_inc(&conf->active_stripes);
2014
2015         raid5_release_stripe(sh);
2016         conf->max_nr_stripes++;
2017         return 1;
2018 }
2019
2020 static int grow_stripes(struct r5conf *conf, int num)
2021 {
2022         struct kmem_cache *sc;
2023         int devs = max(conf->raid_disks, conf->previous_raid_disks);
2024
2025         if (conf->mddev->gendisk)
2026                 sprintf(conf->cache_name[0],
2027                         "raid%d-%s", conf->level, mdname(conf->mddev));
2028         else
2029                 sprintf(conf->cache_name[0],
2030                         "raid%d-%p", conf->level, conf->mddev);
2031         sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2032
2033         conf->active_name = 0;
2034         sc = kmem_cache_create(conf->cache_name[conf->active_name],
2035                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2036                                0, 0, NULL);
2037         if (!sc)
2038                 return 1;
2039         conf->slab_cache = sc;
2040         conf->pool_size = devs;
2041         while (num--)
2042                 if (!grow_one_stripe(conf, GFP_KERNEL))
2043                         return 1;
2044
2045         return 0;
2046 }
2047
2048 /**
2049  * scribble_len - return the required size of the scribble region
2050  * @num - total number of disks in the array
2051  *
2052  * The size must be enough to contain:
2053  * 1/ a struct page pointer for each device in the array +2
2054  * 2/ room to convert each entry in (1) to its corresponding dma
2055  *    (dma_map_page()) or page (page_address()) address.
2056  *
2057  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2058  * calculate over all devices (not just the data blocks), using zeros in place
2059  * of the P and Q blocks.
2060  */
2061 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2062 {
2063         struct flex_array *ret;
2064         size_t len;
2065
2066         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2067         ret = flex_array_alloc(len, cnt, flags);
2068         if (!ret)
2069                 return NULL;
2070         /* always prealloc all elements, so no locking is required */
2071         if (flex_array_prealloc(ret, 0, cnt, flags)) {
2072                 flex_array_free(ret);
2073                 return NULL;
2074         }
2075         return ret;
2076 }
2077
2078 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2079 {
2080         unsigned long cpu;
2081         int err = 0;
2082
2083         /*
2084          * Never shrink. And mddev_suspend() could deadlock if this is called
2085          * from raid5d. In that case, scribble_disks and scribble_sectors
2086          * should equal to new_disks and new_sectors
2087          */
2088         if (conf->scribble_disks >= new_disks &&
2089             conf->scribble_sectors >= new_sectors)
2090                 return 0;
2091         mddev_suspend(conf->mddev);
2092         get_online_cpus();
2093         for_each_present_cpu(cpu) {
2094                 struct raid5_percpu *percpu;
2095                 struct flex_array *scribble;
2096
2097                 percpu = per_cpu_ptr(conf->percpu, cpu);
2098                 scribble = scribble_alloc(new_disks,
2099                                           new_sectors / STRIPE_SECTORS,
2100                                           GFP_NOIO);
2101
2102                 if (scribble) {
2103                         flex_array_free(percpu->scribble);
2104                         percpu->scribble = scribble;
2105                 } else {
2106                         err = -ENOMEM;
2107                         break;
2108                 }
2109         }
2110         put_online_cpus();
2111         mddev_resume(conf->mddev);
2112         if (!err) {
2113                 conf->scribble_disks = new_disks;
2114                 conf->scribble_sectors = new_sectors;
2115         }
2116         return err;
2117 }
2118
2119 static int resize_stripes(struct r5conf *conf, int newsize)
2120 {
2121         /* Make all the stripes able to hold 'newsize' devices.
2122          * New slots in each stripe get 'page' set to a new page.
2123          *
2124          * This happens in stages:
2125          * 1/ create a new kmem_cache and allocate the required number of
2126          *    stripe_heads.
2127          * 2/ gather all the old stripe_heads and transfer the pages across
2128          *    to the new stripe_heads.  This will have the side effect of
2129          *    freezing the array as once all stripe_heads have been collected,
2130          *    no IO will be possible.  Old stripe heads are freed once their
2131          *    pages have been transferred over, and the old kmem_cache is
2132          *    freed when all stripes are done.
2133          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2134          *    we simple return a failre status - no need to clean anything up.
2135          * 4/ allocate new pages for the new slots in the new stripe_heads.
2136          *    If this fails, we don't bother trying the shrink the
2137          *    stripe_heads down again, we just leave them as they are.
2138          *    As each stripe_head is processed the new one is released into
2139          *    active service.
2140          *
2141          * Once step2 is started, we cannot afford to wait for a write,
2142          * so we use GFP_NOIO allocations.
2143          */
2144         struct stripe_head *osh, *nsh;
2145         LIST_HEAD(newstripes);
2146         struct disk_info *ndisks;
2147         int err;
2148         struct kmem_cache *sc;
2149         int i;
2150         int hash, cnt;
2151
2152         if (newsize <= conf->pool_size)
2153                 return 0; /* never bother to shrink */
2154
2155         err = md_allow_write(conf->mddev);
2156         if (err)
2157                 return err;
2158
2159         /* Step 1 */
2160         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2161                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2162                                0, 0, NULL);
2163         if (!sc)
2164                 return -ENOMEM;
2165
2166         /* Need to ensure auto-resizing doesn't interfere */
2167         mutex_lock(&conf->cache_size_mutex);
2168
2169         for (i = conf->max_nr_stripes; i; i--) {
2170                 nsh = alloc_stripe(sc, GFP_KERNEL);
2171                 if (!nsh)
2172                         break;
2173
2174                 nsh->raid_conf = conf;
2175                 list_add(&nsh->lru, &newstripes);
2176         }
2177         if (i) {
2178                 /* didn't get enough, give up */
2179                 while (!list_empty(&newstripes)) {
2180                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
2181                         list_del(&nsh->lru);
2182                         kmem_cache_free(sc, nsh);
2183                 }
2184                 kmem_cache_destroy(sc);
2185                 mutex_unlock(&conf->cache_size_mutex);
2186                 return -ENOMEM;
2187         }
2188         /* Step 2 - Must use GFP_NOIO now.
2189          * OK, we have enough stripes, start collecting inactive
2190          * stripes and copying them over
2191          */
2192         hash = 0;
2193         cnt = 0;
2194         list_for_each_entry(nsh, &newstripes, lru) {
2195                 lock_device_hash_lock(conf, hash);
2196                 wait_event_cmd(conf->wait_for_stripe,
2197                                     !list_empty(conf->inactive_list + hash),
2198                                     unlock_device_hash_lock(conf, hash),
2199                                     lock_device_hash_lock(conf, hash));
2200                 osh = get_free_stripe(conf, hash);
2201                 unlock_device_hash_lock(conf, hash);
2202
2203                 for(i=0; i<conf->pool_size; i++) {
2204                         nsh->dev[i].page = osh->dev[i].page;
2205                         nsh->dev[i].orig_page = osh->dev[i].page;
2206                 }
2207                 nsh->hash_lock_index = hash;
2208                 kmem_cache_free(conf->slab_cache, osh);
2209                 cnt++;
2210                 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2211                     !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2212                         hash++;
2213                         cnt = 0;
2214                 }
2215         }
2216         kmem_cache_destroy(conf->slab_cache);
2217
2218         /* Step 3.
2219          * At this point, we are holding all the stripes so the array
2220          * is completely stalled, so now is a good time to resize
2221          * conf->disks and the scribble region
2222          */
2223         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2224         if (ndisks) {
2225                 for (i=0; i<conf->raid_disks; i++)
2226                         ndisks[i] = conf->disks[i];
2227                 kfree(conf->disks);
2228                 conf->disks = ndisks;
2229         } else
2230                 err = -ENOMEM;
2231
2232         mutex_unlock(&conf->cache_size_mutex);
2233         /* Step 4, return new stripes to service */
2234         while(!list_empty(&newstripes)) {
2235                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2236                 list_del_init(&nsh->lru);
2237
2238                 for (i=conf->raid_disks; i < newsize; i++)
2239                         if (nsh->dev[i].page == NULL) {
2240                                 struct page *p = alloc_page(GFP_NOIO);
2241                                 nsh->dev[i].page = p;
2242                                 nsh->dev[i].orig_page = p;
2243                                 if (!p)
2244                                         err = -ENOMEM;
2245                         }
2246                 raid5_release_stripe(nsh);
2247         }
2248         /* critical section pass, GFP_NOIO no longer needed */
2249
2250         conf->slab_cache = sc;
2251         conf->active_name = 1-conf->active_name;
2252         if (!err)
2253                 conf->pool_size = newsize;
2254         return err;
2255 }
2256
2257 static int drop_one_stripe(struct r5conf *conf)
2258 {
2259         struct stripe_head *sh;
2260         int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2261
2262         spin_lock_irq(conf->hash_locks + hash);
2263         sh = get_free_stripe(conf, hash);
2264         spin_unlock_irq(conf->hash_locks + hash);
2265         if (!sh)
2266                 return 0;
2267         BUG_ON(atomic_read(&sh->count));
2268         shrink_buffers(sh);
2269         kmem_cache_free(conf->slab_cache, sh);
2270         atomic_dec(&conf->active_stripes);
2271         conf->max_nr_stripes--;
2272         return 1;
2273 }
2274
2275 static void shrink_stripes(struct r5conf *conf)
2276 {
2277         while (conf->max_nr_stripes &&
2278                drop_one_stripe(conf))
2279                 ;
2280
2281         kmem_cache_destroy(conf->slab_cache);
2282         conf->slab_cache = NULL;
2283 }
2284
2285 static void raid5_end_read_request(struct bio * bi)
2286 {
2287         struct stripe_head *sh = bi->bi_private;
2288         struct r5conf *conf = sh->raid_conf;
2289         int disks = sh->disks, i;
2290         char b[BDEVNAME_SIZE];
2291         struct md_rdev *rdev = NULL;
2292         sector_t s;
2293
2294         for (i=0 ; i<disks; i++)
2295                 if (bi == &sh->dev[i].req)
2296                         break;
2297
2298         pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2299                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2300                 bi->bi_error);
2301         if (i == disks) {
2302                 BUG();
2303                 return;
2304         }
2305         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2306                 /* If replacement finished while this request was outstanding,
2307                  * 'replacement' might be NULL already.
2308                  * In that case it moved down to 'rdev'.
2309                  * rdev is not removed until all requests are finished.
2310                  */
2311                 rdev = conf->disks[i].replacement;
2312         if (!rdev)
2313                 rdev = conf->disks[i].rdev;
2314
2315         if (use_new_offset(conf, sh))
2316                 s = sh->sector + rdev->new_data_offset;
2317         else
2318                 s = sh->sector + rdev->data_offset;
2319         if (!bi->bi_error) {
2320                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2321                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2322                         /* Note that this cannot happen on a
2323                          * replacement device.  We just fail those on
2324                          * any error
2325                          */
2326                         printk_ratelimited(
2327                                 KERN_INFO
2328                                 "md/raid:%s: read error corrected"
2329                                 " (%lu sectors at %llu on %s)\n",
2330                                 mdname(conf->mddev), STRIPE_SECTORS,
2331                                 (unsigned long long)s,
2332                                 bdevname(rdev->bdev, b));
2333                         atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2334                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2335                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2336                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2337                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2338
2339                 if (atomic_read(&rdev->read_errors))
2340                         atomic_set(&rdev->read_errors, 0);
2341         } else {
2342                 const char *bdn = bdevname(rdev->bdev, b);
2343                 int retry = 0;
2344                 int set_bad = 0;
2345
2346                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2347                 atomic_inc(&rdev->read_errors);
2348                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2349                         printk_ratelimited(
2350                                 KERN_WARNING
2351                                 "md/raid:%s: read error on replacement device "
2352                                 "(sector %llu on %s).\n",
2353                                 mdname(conf->mddev),
2354                                 (unsigned long long)s,
2355                                 bdn);
2356                 else if (conf->mddev->degraded >= conf->max_degraded) {
2357                         set_bad = 1;
2358                         printk_ratelimited(
2359                                 KERN_WARNING
2360                                 "md/raid:%s: read error not correctable "
2361                                 "(sector %llu on %s).\n",
2362                                 mdname(conf->mddev),
2363                                 (unsigned long long)s,
2364                                 bdn);
2365                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2366                         /* Oh, no!!! */
2367                         set_bad = 1;
2368                         printk_ratelimited(
2369                                 KERN_WARNING
2370                                 "md/raid:%s: read error NOT corrected!! "
2371                                 "(sector %llu on %s).\n",
2372                                 mdname(conf->mddev),
2373                                 (unsigned long long)s,
2374                                 bdn);
2375                 } else if (atomic_read(&rdev->read_errors)
2376                          > conf->max_nr_stripes)
2377                         printk(KERN_WARNING
2378                                "md/raid:%s: Too many read errors, failing device %s.\n",
2379                                mdname(conf->mddev), bdn);
2380                 else
2381                         retry = 1;
2382                 if (set_bad && test_bit(In_sync, &rdev->flags)
2383                     && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2384                         retry = 1;
2385                 if (retry)
2386                         if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2387                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2388                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2389                         } else
2390                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2391                 else {
2392                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2393                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2394                         if (!(set_bad
2395                               && test_bit(In_sync, &rdev->flags)
2396                               && rdev_set_badblocks(
2397                                       rdev, sh->sector, STRIPE_SECTORS, 0)))
2398                                 md_error(conf->mddev, rdev);
2399                 }
2400         }
2401         rdev_dec_pending(rdev, conf->mddev);
2402         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2403         set_bit(STRIPE_HANDLE, &sh->state);
2404         raid5_release_stripe(sh);
2405 }
2406
2407 static void raid5_end_write_request(struct bio *bi)
2408 {
2409         struct stripe_head *sh = bi->bi_private;
2410         struct r5conf *conf = sh->raid_conf;
2411         int disks = sh->disks, i;
2412         struct md_rdev *uninitialized_var(rdev);
2413         sector_t first_bad;
2414         int bad_sectors;
2415         int replacement = 0;
2416
2417         for (i = 0 ; i < disks; i++) {
2418                 if (bi == &sh->dev[i].req) {
2419                         rdev = conf->disks[i].rdev;
2420                         break;
2421                 }
2422                 if (bi == &sh->dev[i].rreq) {
2423                         rdev = conf->disks[i].replacement;
2424                         if (rdev)
2425                                 replacement = 1;
2426                         else
2427                                 /* rdev was removed and 'replacement'
2428                                  * replaced it.  rdev is not removed
2429                                  * until all requests are finished.
2430                                  */
2431                                 rdev = conf->disks[i].rdev;
2432                         break;
2433                 }
2434         }
2435         pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2436                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2437                 bi->bi_error);
2438         if (i == disks) {
2439                 BUG();
2440                 return;
2441         }
2442
2443         if (replacement) {
2444                 if (bi->bi_error)
2445                         md_error(conf->mddev, rdev);
2446                 else if (is_badblock(rdev, sh->sector,
2447                                      STRIPE_SECTORS,
2448                                      &first_bad, &bad_sectors))
2449                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2450         } else {
2451                 if (bi->bi_error) {
2452                         set_bit(STRIPE_DEGRADED, &sh->state);
2453                         set_bit(WriteErrorSeen, &rdev->flags);
2454                         set_bit(R5_WriteError, &sh->dev[i].flags);
2455                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2456                                 set_bit(MD_RECOVERY_NEEDED,
2457                                         &rdev->mddev->recovery);
2458                 } else if (is_badblock(rdev, sh->sector,
2459                                        STRIPE_SECTORS,
2460                                        &first_bad, &bad_sectors)) {
2461                         set_bit(R5_MadeGood, &sh->dev[i].flags);
2462                         if (test_bit(R5_ReadError, &sh->dev[i].flags))
2463                                 /* That was a successful write so make
2464                                  * sure it looks like we already did
2465                                  * a re-write.
2466                                  */
2467                                 set_bit(R5_ReWrite, &sh->dev[i].flags);
2468                 }
2469         }
2470         rdev_dec_pending(rdev, conf->mddev);
2471
2472         if (sh->batch_head && bi->bi_error && !replacement)
2473                 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2474
2475         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2476                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2477         set_bit(STRIPE_HANDLE, &sh->state);
2478         raid5_release_stripe(sh);
2479
2480         if (sh->batch_head && sh != sh->batch_head)
2481                 raid5_release_stripe(sh->batch_head);
2482 }
2483
2484 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2485 {
2486         struct r5dev *dev = &sh->dev[i];
2487
2488         bio_init(&dev->req);
2489         dev->req.bi_io_vec = &dev->vec;
2490         dev->req.bi_max_vecs = 1;
2491         dev->req.bi_private = sh;
2492
2493         bio_init(&dev->rreq);
2494         dev->rreq.bi_io_vec = &dev->rvec;
2495         dev->rreq.bi_max_vecs = 1;
2496         dev->rreq.bi_private = sh;
2497
2498         dev->flags = 0;
2499         dev->sector = raid5_compute_blocknr(sh, i, previous);
2500 }
2501
2502 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2503 {
2504         char b[BDEVNAME_SIZE];
2505         struct r5conf *conf = mddev->private;
2506         unsigned long flags;
2507         pr_debug("raid456: error called\n");
2508
2509         spin_lock_irqsave(&conf->device_lock, flags);
2510         clear_bit(In_sync, &rdev->flags);
2511         mddev->degraded = calc_degraded(conf);
2512         spin_unlock_irqrestore(&conf->device_lock, flags);
2513         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2514
2515         set_bit(Blocked, &rdev->flags);
2516         set_bit(Faulty, &rdev->flags);
2517         set_mask_bits(&mddev->flags, 0,
2518                       BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING));
2519         printk(KERN_ALERT
2520                "md/raid:%s: Disk failure on %s, disabling device.\n"
2521                "md/raid:%s: Operation continuing on %d devices.\n",
2522                mdname(mddev),
2523                bdevname(rdev->bdev, b),
2524                mdname(mddev),
2525                conf->raid_disks - mddev->degraded);
2526 }
2527
2528 /*
2529  * Input: a 'big' sector number,
2530  * Output: index of the data and parity disk, and the sector # in them.
2531  */
2532 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2533                               int previous, int *dd_idx,
2534                               struct stripe_head *sh)
2535 {
2536         sector_t stripe, stripe2;
2537         sector_t chunk_number;
2538         unsigned int chunk_offset;
2539         int pd_idx, qd_idx;
2540         int ddf_layout = 0;
2541         sector_t new_sector;
2542         int algorithm = previous ? conf->prev_algo
2543                                  : conf->algorithm;
2544         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2545                                          : conf->chunk_sectors;
2546         int raid_disks = previous ? conf->previous_raid_disks
2547                                   : conf->raid_disks;
2548         int data_disks = raid_disks - conf->max_degraded;
2549
2550         /* First compute the information on this sector */
2551
2552         /*
2553          * Compute the chunk number and the sector offset inside the chunk
2554          */
2555         chunk_offset = sector_div(r_sector, sectors_per_chunk);
2556         chunk_number = r_sector;
2557
2558         /*
2559          * Compute the stripe number
2560          */
2561         stripe = chunk_number;
2562         *dd_idx = sector_div(stripe, data_disks);
2563         stripe2 = stripe;
2564         /*
2565          * Select the parity disk based on the user selected algorithm.
2566          */
2567         pd_idx = qd_idx = -1;
2568         switch(conf->level) {
2569         case 4:
2570                 pd_idx = data_disks;
2571                 break;
2572         case 5:
2573                 switch (algorithm) {
2574                 case ALGORITHM_LEFT_ASYMMETRIC:
2575                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2576                         if (*dd_idx >= pd_idx)
2577                                 (*dd_idx)++;
2578                         break;
2579                 case ALGORITHM_RIGHT_ASYMMETRIC:
2580                         pd_idx = sector_div(stripe2, raid_disks);
2581                         if (*dd_idx >= pd_idx)
2582                                 (*dd_idx)++;
2583                         break;
2584                 case ALGORITHM_LEFT_SYMMETRIC:
2585                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2586                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2587                         break;
2588                 case ALGORITHM_RIGHT_SYMMETRIC:
2589                         pd_idx = sector_div(stripe2, raid_disks);
2590                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2591                         break;
2592                 case ALGORITHM_PARITY_0:
2593                         pd_idx = 0;
2594                         (*dd_idx)++;
2595                         break;
2596                 case ALGORITHM_PARITY_N:
2597                         pd_idx = data_disks;
2598                         break;
2599                 default:
2600                         BUG();
2601                 }
2602                 break;
2603         case 6:
2604
2605                 switch (algorithm) {
2606                 case ALGORITHM_LEFT_ASYMMETRIC:
2607                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2608                         qd_idx = pd_idx + 1;
2609                         if (pd_idx == raid_disks-1) {
2610                                 (*dd_idx)++;    /* Q D D D P */
2611                                 qd_idx = 0;
2612                         } else if (*dd_idx >= pd_idx)
2613                                 (*dd_idx) += 2; /* D D P Q D */
2614                         break;
2615                 case ALGORITHM_RIGHT_ASYMMETRIC:
2616                         pd_idx = sector_div(stripe2, raid_disks);
2617                         qd_idx = pd_idx + 1;
2618                         if (pd_idx == raid_disks-1) {
2619                                 (*dd_idx)++;    /* Q D D D P */
2620                                 qd_idx = 0;
2621                         } else if (*dd_idx >= pd_idx)
2622                                 (*dd_idx) += 2; /* D D P Q D */
2623                         break;
2624                 case ALGORITHM_LEFT_SYMMETRIC:
2625                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2626                         qd_idx = (pd_idx + 1) % raid_disks;
2627                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2628                         break;
2629                 case ALGORITHM_RIGHT_SYMMETRIC:
2630                         pd_idx = sector_div(stripe2, raid_disks);
2631                         qd_idx = (pd_idx + 1) % raid_disks;
2632                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2633                         break;
2634
2635                 case ALGORITHM_PARITY_0:
2636                         pd_idx = 0;
2637                         qd_idx = 1;
2638                         (*dd_idx) += 2;
2639                         break;
2640                 case ALGORITHM_PARITY_N:
2641                         pd_idx = data_disks;
2642                         qd_idx = data_disks + 1;
2643                         break;
2644
2645                 case ALGORITHM_ROTATING_ZERO_RESTART:
2646                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
2647                          * of blocks for computing Q is different.
2648                          */
2649                         pd_idx = sector_div(stripe2, raid_disks);
2650                         qd_idx = pd_idx + 1;
2651                         if (pd_idx == raid_disks-1) {
2652                                 (*dd_idx)++;    /* Q D D D P */
2653                                 qd_idx = 0;
2654                         } else if (*dd_idx >= pd_idx)
2655                                 (*dd_idx) += 2; /* D D P Q D */
2656                         ddf_layout = 1;
2657                         break;
2658
2659                 case ALGORITHM_ROTATING_N_RESTART:
2660                         /* Same a left_asymmetric, by first stripe is
2661                          * D D D P Q  rather than
2662                          * Q D D D P
2663                          */
2664                         stripe2 += 1;
2665                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2666                         qd_idx = pd_idx + 1;
2667                         if (pd_idx == raid_disks-1) {
2668                                 (*dd_idx)++;    /* Q D D D P */
2669                                 qd_idx = 0;
2670                         } else if (*dd_idx >= pd_idx)
2671                                 (*dd_idx) += 2; /* D D P Q D */
2672                         ddf_layout = 1;
2673                         break;
2674
2675                 case ALGORITHM_ROTATING_N_CONTINUE:
2676                         /* Same as left_symmetric but Q is before P */
2677                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2678                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2679                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2680                         ddf_layout = 1;
2681                         break;
2682
2683                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2684                         /* RAID5 left_asymmetric, with Q on last device */
2685                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2686                         if (*dd_idx >= pd_idx)
2687                                 (*dd_idx)++;
2688                         qd_idx = raid_disks - 1;
2689                         break;
2690
2691                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2692                         pd_idx = sector_div(stripe2, raid_disks-1);
2693                         if (*dd_idx >= pd_idx)
2694                                 (*dd_idx)++;
2695                         qd_idx = raid_disks - 1;
2696                         break;
2697
2698                 case ALGORITHM_LEFT_SYMMETRIC_6:
2699                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2700                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2701                         qd_idx = raid_disks - 1;
2702                         break;
2703
2704                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2705                         pd_idx = sector_div(stripe2, raid_disks-1);
2706                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2707                         qd_idx = raid_disks - 1;
2708                         break;
2709
2710                 case ALGORITHM_PARITY_0_6:
2711                         pd_idx = 0;
2712                         (*dd_idx)++;
2713                         qd_idx = raid_disks - 1;
2714                         break;
2715
2716                 default:
2717                         BUG();
2718                 }
2719                 break;
2720         }
2721
2722         if (sh) {
2723                 sh->pd_idx = pd_idx;
2724                 sh->qd_idx = qd_idx;
2725                 sh->ddf_layout = ddf_layout;
2726         }
2727         /*
2728          * Finally, compute the new sector number
2729          */
2730         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2731         return new_sector;
2732 }
2733
2734 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2735 {
2736         struct r5conf *conf = sh->raid_conf;
2737         int raid_disks = sh->disks;
2738         int data_disks = raid_disks - conf->max_degraded;
2739         sector_t new_sector = sh->sector, check;
2740         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2741                                          : conf->chunk_sectors;
2742         int algorithm = previous ? conf->prev_algo
2743                                  : conf->algorithm;
2744         sector_t stripe;
2745         int chunk_offset;
2746         sector_t chunk_number;
2747         int dummy1, dd_idx = i;
2748         sector_t r_sector;
2749         struct stripe_head sh2;
2750
2751         chunk_offset = sector_div(new_sector, sectors_per_chunk);
2752         stripe = new_sector;
2753
2754         if (i == sh->pd_idx)
2755                 return 0;
2756         switch(conf->level) {
2757         case 4: break;
2758         case 5:
2759                 switch (algorithm) {
2760                 case ALGORITHM_LEFT_ASYMMETRIC:
2761                 case ALGORITHM_RIGHT_ASYMMETRIC:
2762                         if (i > sh->pd_idx)
2763                                 i--;
2764                         break;
2765                 case ALGORITHM_LEFT_SYMMETRIC:
2766                 case ALGORITHM_RIGHT_SYMMETRIC:
2767                         if (i < sh->pd_idx)
2768                                 i += raid_disks;
2769                         i -= (sh->pd_idx + 1);
2770                         break;
2771                 case ALGORITHM_PARITY_0:
2772                         i -= 1;
2773                         break;
2774                 case ALGORITHM_PARITY_N:
2775                         break;
2776                 default:
2777                         BUG();
2778                 }
2779                 break;
2780         case 6:
2781                 if (i == sh->qd_idx)
2782                         return 0; /* It is the Q disk */
2783                 switch (algorithm) {
2784                 case ALGORITHM_LEFT_ASYMMETRIC:
2785                 case ALGORITHM_RIGHT_ASYMMETRIC:
2786                 case ALGORITHM_ROTATING_ZERO_RESTART:
2787                 case ALGORITHM_ROTATING_N_RESTART:
2788                         if (sh->pd_idx == raid_disks-1)
2789                                 i--;    /* Q D D D P */
2790                         else if (i > sh->pd_idx)
2791                                 i -= 2; /* D D P Q D */
2792                         break;
2793                 case ALGORITHM_LEFT_SYMMETRIC:
2794                 case ALGORITHM_RIGHT_SYMMETRIC:
2795                         if (sh->pd_idx == raid_disks-1)
2796                                 i--; /* Q D D D P */
2797                         else {
2798                                 /* D D P Q D */
2799                                 if (i < sh->pd_idx)
2800                                         i += raid_disks;
2801                                 i -= (sh->pd_idx + 2);
2802                         }
2803                         break;
2804                 case ALGORITHM_PARITY_0:
2805                         i -= 2;
2806                         break;
2807                 case ALGORITHM_PARITY_N:
2808                         break;
2809                 case ALGORITHM_ROTATING_N_CONTINUE:
2810                         /* Like left_symmetric, but P is before Q */
2811                         if (sh->pd_idx == 0)
2812                                 i--;    /* P D D D Q */
2813                         else {
2814                                 /* D D Q P D */
2815                                 if (i < sh->pd_idx)
2816                                         i += raid_disks;
2817                                 i -= (sh->pd_idx + 1);
2818                         }
2819                         break;
2820                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2821                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2822                         if (i > sh->pd_idx)
2823                                 i--;
2824                         break;
2825                 case ALGORITHM_LEFT_SYMMETRIC_6:
2826                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2827                         if (i < sh->pd_idx)
2828                                 i += data_disks + 1;
2829                         i -= (sh->pd_idx + 1);
2830                         break;
2831                 case ALGORITHM_PARITY_0_6:
2832                         i -= 1;
2833                         break;
2834                 default:
2835                         BUG();
2836                 }
2837                 break;
2838         }
2839
2840         chunk_number = stripe * data_disks + i;
2841         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2842
2843         check = raid5_compute_sector(conf, r_sector,
2844                                      previous, &dummy1, &sh2);
2845         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2846                 || sh2.qd_idx != sh->qd_idx) {
2847                 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2848                        mdname(conf->mddev));
2849                 return 0;
2850         }
2851         return r_sector;
2852 }
2853
2854 static void
2855 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2856                          int rcw, int expand)
2857 {
2858         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2859         struct r5conf *conf = sh->raid_conf;
2860         int level = conf->level;
2861
2862         if (rcw) {
2863
2864                 for (i = disks; i--; ) {
2865                         struct r5dev *dev = &sh->dev[i];
2866
2867                         if (dev->towrite) {
2868                                 set_bit(R5_LOCKED, &dev->flags);
2869                                 set_bit(R5_Wantdrain, &dev->flags);
2870                                 if (!expand)
2871                                         clear_bit(R5_UPTODATE, &dev->flags);
2872                                 s->locked++;
2873                         }
2874                 }
2875                 /* if we are not expanding this is a proper write request, and
2876                  * there will be bios with new data to be drained into the
2877                  * stripe cache
2878                  */
2879                 if (!expand) {
2880                         if (!s->locked)
2881                                 /* False alarm, nothing to do */
2882                                 return;
2883                         sh->reconstruct_state = reconstruct_state_drain_run;
2884                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2885                 } else
2886                         sh->reconstruct_state = reconstruct_state_run;
2887
2888                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2889
2890                 if (s->locked + conf->max_degraded == disks)
2891                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2892                                 atomic_inc(&conf->pending_full_writes);
2893         } else {
2894                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2895                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2896                 BUG_ON(level == 6 &&
2897                         (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2898                            test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2899
2900                 for (i = disks; i--; ) {
2901                         struct r5dev *dev = &sh->dev[i];
2902                         if (i == pd_idx || i == qd_idx)
2903                                 continue;
2904
2905                         if (dev->towrite &&
2906                             (test_bit(R5_UPTODATE, &dev->flags) ||
2907                              test_bit(R5_Wantcompute, &dev->flags))) {
2908                                 set_bit(R5_Wantdrain, &dev->flags);
2909                                 set_bit(R5_LOCKED, &dev->flags);
2910                                 clear_bit(R5_UPTODATE, &dev->flags);
2911                                 s->locked++;
2912                         }
2913                 }
2914                 if (!s->locked)
2915                         /* False alarm - nothing to do */
2916                         return;
2917                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2918                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2919                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2920                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2921         }
2922
2923         /* keep the parity disk(s) locked while asynchronous operations
2924          * are in flight
2925          */
2926         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2927         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2928         s->locked++;
2929
2930         if (level == 6) {
2931                 int qd_idx = sh->qd_idx;
2932                 struct r5dev *dev = &sh->dev[qd_idx];
2933
2934                 set_bit(R5_LOCKED, &dev->flags);
2935                 clear_bit(R5_UPTODATE, &dev->flags);
2936                 s->locked++;
2937         }
2938
2939         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2940                 __func__, (unsigned long long)sh->sector,
2941                 s->locked, s->ops_request);
2942 }
2943
2944 /*
2945  * Each stripe/dev can have one or more bion attached.
2946  * toread/towrite point to the first in a chain.
2947  * The bi_next chain must be in order.
2948  */
2949 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2950                           int forwrite, int previous)
2951 {
2952         struct bio **bip;
2953         struct r5conf *conf = sh->raid_conf;
2954         int firstwrite=0;
2955
2956         pr_debug("adding bi b#%llu to stripe s#%llu\n",
2957                 (unsigned long long)bi->bi_iter.bi_sector,
2958                 (unsigned long long)sh->sector);
2959
2960         /*
2961          * If several bio share a stripe. The bio bi_phys_segments acts as a
2962          * reference count to avoid race. The reference count should already be
2963          * increased before this function is called (for example, in
2964          * raid5_make_request()), so other bio sharing this stripe will not free the
2965          * stripe. If a stripe is owned by one stripe, the stripe lock will
2966          * protect it.
2967          */
2968         spin_lock_irq(&sh->stripe_lock);
2969         /* Don't allow new IO added to stripes in batch list */
2970         if (sh->batch_head)
2971                 goto overlap;
2972         if (forwrite) {
2973                 bip = &sh->dev[dd_idx].towrite;
2974                 if (*bip == NULL)
2975                         firstwrite = 1;
2976         } else
2977                 bip = &sh->dev[dd_idx].toread;
2978         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2979                 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2980                         goto overlap;
2981                 bip = & (*bip)->bi_next;
2982         }
2983         if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2984                 goto overlap;
2985
2986         if (!forwrite || previous)
2987                 clear_bit(STRIPE_BATCH_READY, &sh->state);
2988
2989         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2990         if (*bip)
2991                 bi->bi_next = *bip;
2992         *bip = bi;
2993         raid5_inc_bi_active_stripes(bi);
2994
2995         if (forwrite) {
2996                 /* check if page is covered */
2997                 sector_t sector = sh->dev[dd_idx].sector;
2998                 for (bi=sh->dev[dd_idx].towrite;
2999                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3000                              bi && bi->bi_iter.bi_sector <= sector;
3001                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3002                         if (bio_end_sector(bi) >= sector)
3003                                 sector = bio_end_sector(bi);
3004                 }
3005                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3006                         if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3007                                 sh->overwrite_disks++;
3008         }
3009
3010         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3011                 (unsigned long long)(*bip)->bi_iter.bi_sector,
3012                 (unsigned long long)sh->sector, dd_idx);
3013
3014         if (conf->mddev->bitmap && firstwrite) {
3015                 /* Cannot hold spinlock over bitmap_startwrite,
3016                  * but must ensure this isn't added to a batch until
3017                  * we have added to the bitmap and set bm_seq.
3018                  * So set STRIPE_BITMAP_PENDING to prevent
3019                  * batching.
3020                  * If multiple add_stripe_bio() calls race here they
3021                  * much all set STRIPE_BITMAP_PENDING.  So only the first one
3022                  * to complete "bitmap_startwrite" gets to set
3023                  * STRIPE_BIT_DELAY.  This is important as once a stripe
3024                  * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3025                  * any more.
3026                  */
3027                 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3028                 spin_unlock_irq(&sh->stripe_lock);
3029                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3030                                   STRIPE_SECTORS, 0);
3031                 spin_lock_irq(&sh->stripe_lock);
3032                 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3033                 if (!sh->batch_head) {
3034                         sh->bm_seq = conf->seq_flush+1;
3035                         set_bit(STRIPE_BIT_DELAY, &sh->state);
3036                 }
3037         }
3038         spin_unlock_irq(&sh->stripe_lock);
3039
3040         if (stripe_can_batch(sh))
3041                 stripe_add_to_batch_list(conf, sh);
3042         return 1;
3043
3044  overlap:
3045         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3046         spin_unlock_irq(&sh->stripe_lock);
3047         return 0;
3048 }
3049
3050 static void end_reshape(struct r5conf *conf);
3051
3052 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3053                             struct stripe_head *sh)
3054 {
3055         int sectors_per_chunk =
3056                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3057         int dd_idx;
3058         int chunk_offset = sector_div(stripe, sectors_per_chunk);
3059         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3060
3061         raid5_compute_sector(conf,
3062                              stripe * (disks - conf->max_degraded)
3063                              *sectors_per_chunk + chunk_offset,
3064                              previous,
3065                              &dd_idx, sh);
3066 }
3067
3068 static void
3069 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3070                                 struct stripe_head_state *s, int disks,
3071                                 struct bio_list *return_bi)
3072 {
3073         int i;
3074         BUG_ON(sh->batch_head);
3075         for (i = disks; i--; ) {
3076                 struct bio *bi;
3077                 int bitmap_end = 0;
3078
3079                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3080                         struct md_rdev *rdev;
3081                         rcu_read_lock();
3082                         rdev = rcu_dereference(conf->disks[i].rdev);
3083                         if (rdev && test_bit(In_sync, &rdev->flags))
3084                                 atomic_inc(&rdev->nr_pending);
3085                         else
3086                                 rdev = NULL;
3087                         rcu_read_unlock();
3088                         if (rdev) {
3089                                 if (!rdev_set_badblocks(
3090                                             rdev,
3091                                             sh->sector,
3092                                             STRIPE_SECTORS, 0))
3093                                         md_error(conf->mddev, rdev);
3094                                 rdev_dec_pending(rdev, conf->mddev);
3095                         }
3096                 }
3097                 spin_lock_irq(&sh->stripe_lock);
3098                 /* fail all writes first */
3099                 bi = sh->dev[i].towrite;
3100                 sh->dev[i].towrite = NULL;
3101                 sh->overwrite_disks = 0;
3102                 spin_unlock_irq(&sh->stripe_lock);
3103                 if (bi)
3104                         bitmap_end = 1;
3105
3106                 r5l_stripe_write_finished(sh);
3107
3108                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3109                         wake_up(&conf->wait_for_overlap);
3110
3111                 while (bi && bi->bi_iter.bi_sector <
3112                         sh->dev[i].sector + STRIPE_SECTORS) {
3113                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3114
3115                         bi->bi_error = -EIO;
3116                         if (!raid5_dec_bi_active_stripes(bi)) {
3117                                 md_write_end(conf->mddev);
3118                                 bio_list_add(return_bi, bi);
3119                         }
3120                         bi = nextbi;
3121                 }
3122                 if (bitmap_end)
3123                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3124                                 STRIPE_SECTORS, 0, 0);
3125                 bitmap_end = 0;
3126                 /* and fail all 'written' */
3127                 bi = sh->dev[i].written;
3128                 sh->dev[i].written = NULL;
3129                 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3130                         WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3131                         sh->dev[i].page = sh->dev[i].orig_page;
3132                 }
3133
3134                 if (bi) bitmap_end = 1;
3135                 while (bi && bi->bi_iter.bi_sector <
3136                        sh->dev[i].sector + STRIPE_SECTORS) {
3137                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3138
3139                         bi->bi_error = -EIO;
3140                         if (!raid5_dec_bi_active_stripes(bi)) {
3141                                 md_write_end(conf->mddev);
3142                                 bio_list_add(return_bi, bi);
3143                         }
3144                         bi = bi2;
3145                 }
3146
3147                 /* fail any reads if this device is non-operational and
3148                  * the data has not reached the cache yet.
3149                  */
3150                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3151                     s->failed > conf->max_degraded &&
3152                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3153                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
3154                         spin_lock_irq(&sh->stripe_lock);
3155                         bi = sh->dev[i].toread;
3156                         sh->dev[i].toread = NULL;
3157                         spin_unlock_irq(&sh->stripe_lock);
3158                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3159                                 wake_up(&conf->wait_for_overlap);
3160                         if (bi)
3161                                 s->to_read--;
3162                         while (bi && bi->bi_iter.bi_sector <
3163                                sh->dev[i].sector + STRIPE_SECTORS) {
3164                                 struct bio *nextbi =
3165                                         r5_next_bio(bi, sh->dev[i].sector);
3166
3167                                 bi->bi_error = -EIO;
3168                                 if (!raid5_dec_bi_active_stripes(bi))
3169                                         bio_list_add(return_bi, bi);
3170                                 bi = nextbi;
3171                         }
3172                 }
3173                 if (bitmap_end)
3174                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3175                                         STRIPE_SECTORS, 0, 0);
3176                 /* If we were in the middle of a write the parity block might
3177                  * still be locked - so just clear all R5_LOCKED flags
3178                  */
3179                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3180         }
3181         s->to_write = 0;
3182         s->written = 0;
3183
3184         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3185                 if (atomic_dec_and_test(&conf->pending_full_writes))
3186                         md_wakeup_thread(conf->mddev->thread);
3187 }
3188
3189 static void
3190 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3191                    struct stripe_head_state *s)
3192 {
3193         int abort = 0;
3194         int i;
3195
3196         BUG_ON(sh->batch_head);
3197         clear_bit(STRIPE_SYNCING, &sh->state);
3198         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3199                 wake_up(&conf->wait_for_overlap);
3200         s->syncing = 0;
3201         s->replacing = 0;
3202         /* There is nothing more to do for sync/check/repair.
3203          * Don't even need to abort as that is handled elsewhere
3204          * if needed, and not always wanted e.g. if there is a known
3205          * bad block here.
3206          * For recover/replace we need to record a bad block on all
3207          * non-sync devices, or abort the recovery
3208          */
3209         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3210                 /* During recovery devices cannot be removed, so
3211                  * locking and refcounting of rdevs is not needed
3212                  */
3213                 for (i = 0; i < conf->raid_disks; i++) {
3214                         struct md_rdev *rdev = conf->disks[i].rdev;
3215                         if (rdev
3216                             && !test_bit(Faulty, &rdev->flags)
3217                             && !test_bit(In_sync, &rdev->flags)
3218                             && !rdev_set_badblocks(rdev, sh->sector,
3219                                                    STRIPE_SECTORS, 0))
3220                                 abort = 1;
3221                         rdev = conf->disks[i].replacement;
3222                         if (rdev
3223                             && !test_bit(Faulty, &rdev->flags)
3224                             && !test_bit(In_sync, &rdev->flags)
3225                             && !rdev_set_badblocks(rdev, sh->sector,
3226                                                    STRIPE_SECTORS, 0))
3227                                 abort = 1;
3228                 }
3229                 if (abort)
3230                         conf->recovery_disabled =
3231                                 conf->mddev->recovery_disabled;
3232         }
3233         md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3234 }
3235
3236 static int want_replace(struct stripe_head *sh, int disk_idx)
3237 {
3238         struct md_rdev *rdev;
3239         int rv = 0;
3240         /* Doing recovery so rcu locking not required */
3241         rdev = sh->raid_conf->disks[disk_idx].replacement;
3242         if (rdev
3243             && !test_bit(Faulty, &rdev->flags)
3244             && !test_bit(In_sync, &rdev->flags)
3245             && (rdev->recovery_offset <= sh->sector
3246                 || rdev->mddev->recovery_cp <= sh->sector))
3247                 rv = 1;
3248
3249         return rv;
3250 }
3251
3252 /* fetch_block - checks the given member device to see if its data needs
3253  * to be read or computed to satisfy a request.
3254  *
3255  * Returns 1 when no more member devices need to be checked, otherwise returns
3256  * 0 to tell the loop in handle_stripe_fill to continue
3257  */
3258
3259 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3260                            int disk_idx, int disks)
3261 {
3262         struct r5dev *dev = &sh->dev[disk_idx];
3263         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3264                                   &sh->dev[s->failed_num[1]] };
3265         int i;
3266
3267
3268         if (test_bit(R5_LOCKED, &dev->flags) ||
3269             test_bit(R5_UPTODATE, &dev->flags))
3270                 /* No point reading this as we already have it or have
3271                  * decided to get it.
3272                  */
3273                 return 0;
3274
3275         if (dev->toread ||
3276             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3277                 /* We need this block to directly satisfy a request */
3278                 return 1;
3279
3280         if (s->syncing || s->expanding ||
3281             (s->replacing && want_replace(sh, disk_idx)))
3282                 /* When syncing, or expanding we read everything.
3283                  * When replacing, we need the replaced block.
3284                  */
3285                 return 1;
3286
3287         if ((s->failed >= 1 && fdev[0]->toread) ||
3288             (s->failed >= 2 && fdev[1]->toread))
3289                 /* If we want to read from a failed device, then
3290                  * we need to actually read every other device.
3291                  */
3292                 return 1;
3293
3294         /* Sometimes neither read-modify-write nor reconstruct-write
3295          * cycles can work.  In those cases we read every block we
3296          * can.  Then the parity-update is certain to have enough to
3297          * work with.
3298          * This can only be a problem when we need to write something,
3299          * and some device has failed.  If either of those tests
3300          * fail we need look no further.
3301          */
3302         if (!s->failed || !s->to_write)
3303                 return 0;
3304
3305         if (test_bit(R5_Insync, &dev->flags) &&
3306             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3307                 /* Pre-reads at not permitted until after short delay
3308                  * to gather multiple requests.  However if this
3309                  * device is no Insync, the block could only be be computed
3310                  * and there is no need to delay that.
3311                  */
3312                 return 0;
3313
3314         for (i = 0; i < s->failed && i < 2; i++) {
3315                 if (fdev[i]->towrite &&
3316                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3317                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3318                         /* If we have a partial write to a failed
3319                          * device, then we will need to reconstruct
3320                          * the content of that device, so all other
3321                          * devices must be read.
3322                          */
3323                         return 1;
3324         }
3325
3326         /* If we are forced to do a reconstruct-write, either because
3327          * the current RAID6 implementation only supports that, or
3328          * or because parity cannot be trusted and we are currently
3329          * recovering it, there is extra need to be careful.
3330          * If one of the devices that we would need to read, because
3331          * it is not being overwritten (and maybe not written at all)
3332          * is missing/faulty, then we need to read everything we can.
3333          */
3334         if (sh->raid_conf->level != 6 &&
3335             sh->sector < sh->raid_conf->mddev->recovery_cp)
3336                 /* reconstruct-write isn't being forced */
3337                 return 0;
3338         for (i = 0; i < s->failed && i < 2; i++) {
3339                 if (s->failed_num[i] != sh->pd_idx &&
3340                     s->failed_num[i] != sh->qd_idx &&
3341                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3342                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3343                         return 1;
3344         }
3345
3346         return 0;
3347 }
3348
3349 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3350                        int disk_idx, int disks)
3351 {
3352         struct r5dev *dev = &sh->dev[disk_idx];
3353
3354         /* is the data in this block needed, and can we get it? */
3355         if (need_this_block(sh, s, disk_idx, disks)) {
3356                 /* we would like to get this block, possibly by computing it,
3357                  * otherwise read it if the backing disk is insync
3358                  */
3359                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3360                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3361                 BUG_ON(sh->batch_head);
3362                 if ((s->uptodate == disks - 1) &&
3363                     (s->failed && (disk_idx == s->failed_num[0] ||
3364                                    disk_idx == s->failed_num[1]))) {
3365                         /* have disk failed, and we're requested to fetch it;
3366                          * do compute it
3367                          */
3368                         pr_debug("Computing stripe %llu block %d\n",
3369                                (unsigned long long)sh->sector, disk_idx);
3370                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3371                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3372                         set_bit(R5_Wantcompute, &dev->flags);
3373                         sh->ops.target = disk_idx;
3374                         sh->ops.target2 = -1; /* no 2nd target */
3375                         s->req_compute = 1;
3376                         /* Careful: from this point on 'uptodate' is in the eye
3377                          * of raid_run_ops which services 'compute' operations
3378                          * before writes. R5_Wantcompute flags a block that will
3379                          * be R5_UPTODATE by the time it is needed for a
3380                          * subsequent operation.
3381                          */
3382                         s->uptodate++;
3383                         return 1;
3384                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3385                         /* Computing 2-failure is *very* expensive; only
3386                          * do it if failed >= 2
3387                          */
3388                         int other;
3389                         for (other = disks; other--; ) {
3390                                 if (other == disk_idx)
3391                                         continue;
3392                                 if (!test_bit(R5_UPTODATE,
3393                                       &sh->dev[other].flags))
3394                                         break;
3395                         }
3396                         BUG_ON(other < 0);
3397                         pr_debug("Computing stripe %llu blocks %d,%d\n",
3398                                (unsigned long long)sh->sector,
3399                                disk_idx, other);
3400                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3401                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3402                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3403                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
3404                         sh->ops.target = disk_idx;
3405                         sh->ops.target2 = other;
3406                         s->uptodate += 2;
3407                         s->req_compute = 1;
3408                         return 1;
3409                 } else if (test_bit(R5_Insync, &dev->flags)) {
3410                         set_bit(R5_LOCKED, &dev->flags);
3411                         set_bit(R5_Wantread, &dev->flags);
3412                         s->locked++;
3413                         pr_debug("Reading block %d (sync=%d)\n",
3414                                 disk_idx, s->syncing);
3415                 }
3416         }
3417
3418         return 0;
3419 }
3420
3421 /**
3422  * handle_stripe_fill - read or compute data to satisfy pending requests.
3423  */
3424 static void handle_stripe_fill(struct stripe_head *sh,
3425                                struct stripe_head_state *s,
3426                                int disks)
3427 {
3428         int i;
3429
3430         /* look for blocks to read/compute, skip this if a compute
3431          * is already in flight, or if the stripe contents are in the
3432          * midst of changing due to a write
3433          */
3434         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3435             !sh->reconstruct_state)
3436                 for (i = disks; i--; )
3437                         if (fetch_block(sh, s, i, disks))
3438                                 break;
3439         set_bit(STRIPE_HANDLE, &sh->state);
3440 }
3441
3442 static void break_stripe_batch_list(struct stripe_head *head_sh,
3443                                     unsigned long handle_flags);
3444 /* handle_stripe_clean_event
3445  * any written block on an uptodate or failed drive can be returned.
3446  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3447  * never LOCKED, so we don't need to test 'failed' directly.
3448  */
3449 static void handle_stripe_clean_event(struct r5conf *conf,
3450         struct stripe_head *sh, int disks, struct bio_list *return_bi)
3451 {
3452         int i;
3453         struct r5dev *dev;
3454         int discard_pending = 0;
3455         struct stripe_head *head_sh = sh;
3456         bool do_endio = false;
3457
3458         for (i = disks; i--; )
3459                 if (sh->dev[i].written) {
3460                         dev = &sh->dev[i];
3461                         if (!test_bit(R5_LOCKED, &dev->flags) &&
3462                             (test_bit(R5_UPTODATE, &dev->flags) ||
3463                              test_bit(R5_Discard, &dev->flags) ||
3464                              test_bit(R5_SkipCopy, &dev->flags))) {
3465                                 /* We can return any write requests */
3466                                 struct bio *wbi, *wbi2;
3467                                 pr_debug("Return write for disc %d\n", i);
3468                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
3469                                         clear_bit(R5_UPTODATE, &dev->flags);
3470                                 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3471                                         WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3472                                 }
3473                                 do_endio = true;
3474
3475 returnbi:
3476                                 dev->page = dev->orig_page;
3477                                 wbi = dev->written;
3478                                 dev->written = NULL;
3479                                 while (wbi && wbi->bi_iter.bi_sector <
3480                                         dev->sector + STRIPE_SECTORS) {
3481                                         wbi2 = r5_next_bio(wbi, dev->sector);
3482                                         if (!raid5_dec_bi_active_stripes(wbi)) {
3483                                                 md_write_end(conf->mddev);
3484                                                 bio_list_add(return_bi, wbi);
3485                                         }
3486                                         wbi = wbi2;
3487                                 }
3488                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3489                                                 STRIPE_SECTORS,
3490                                          !test_bit(STRIPE_DEGRADED, &sh->state),
3491                                                 0);
3492                                 if (head_sh->batch_head) {
3493                                         sh = list_first_entry(&sh->batch_list,
3494                                                               struct stripe_head,
3495                                                               batch_list);
3496                                         if (sh != head_sh) {
3497                                                 dev = &sh->dev[i];
3498                                                 goto returnbi;
3499                                         }
3500                                 }
3501                                 sh = head_sh;
3502                                 dev = &sh->dev[i];
3503                         } else if (test_bit(R5_Discard, &dev->flags))
3504                                 discard_pending = 1;
3505                 }
3506
3507         r5l_stripe_write_finished(sh);
3508
3509         if (!discard_pending &&
3510             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3511                 int hash;
3512                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3513                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3514                 if (sh->qd_idx >= 0) {
3515                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3516                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3517                 }
3518                 /* now that discard is done we can proceed with any sync */
3519                 clear_bit(STRIPE_DISCARD, &sh->state);
3520                 /*
3521                  * SCSI discard will change some bio fields and the stripe has
3522                  * no updated data, so remove it from hash list and the stripe
3523                  * will be reinitialized
3524                  */
3525 unhash:
3526                 hash = sh->hash_lock_index;
3527                 spin_lock_irq(conf->hash_locks + hash);
3528                 remove_hash(sh);
3529                 spin_unlock_irq(conf->hash_locks + hash);
3530                 if (head_sh->batch_head) {
3531                         sh = list_first_entry(&sh->batch_list,
3532                                               struct stripe_head, batch_list);
3533                         if (sh != head_sh)
3534                                         goto unhash;
3535                 }
3536                 sh = head_sh;
3537
3538                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3539                         set_bit(STRIPE_HANDLE, &sh->state);
3540
3541         }
3542
3543         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3544                 if (atomic_dec_and_test(&conf->pending_full_writes))
3545                         md_wakeup_thread(conf->mddev->thread);
3546
3547         if (head_sh->batch_head && do_endio)
3548                 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3549 }
3550
3551 static void handle_stripe_dirtying(struct r5conf *conf,
3552                                    struct stripe_head *sh,
3553                                    struct stripe_head_state *s,
3554                                    int disks)
3555 {
3556         int rmw = 0, rcw = 0, i;
3557         sector_t recovery_cp = conf->mddev->recovery_cp;
3558
3559         /* Check whether resync is now happening or should start.
3560          * If yes, then the array is dirty (after unclean shutdown or
3561          * initial creation), so parity in some stripes might be inconsistent.
3562          * In this case, we need to always do reconstruct-write, to ensure
3563          * that in case of drive failure or read-error correction, we
3564          * generate correct data from the parity.
3565          */
3566         if (conf->rmw_level == PARITY_DISABLE_RMW ||
3567             (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3568              s->failed == 0)) {
3569                 /* Calculate the real rcw later - for now make it
3570                  * look like rcw is cheaper
3571                  */
3572                 rcw = 1; rmw = 2;
3573                 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3574                          conf->rmw_level, (unsigned long long)recovery_cp,
3575                          (unsigned long long)sh->sector);
3576         } else for (i = disks; i--; ) {
3577                 /* would I have to read this buffer for read_modify_write */
3578                 struct r5dev *dev = &sh->dev[i];
3579                 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3580                     !test_bit(R5_LOCKED, &dev->flags) &&
3581                     !(test_bit(R5_UPTODATE, &dev->flags) ||
3582                       test_bit(R5_Wantcompute, &dev->flags))) {
3583                         if (test_bit(R5_Insync, &dev->flags))
3584                                 rmw++;
3585                         else
3586                                 rmw += 2*disks;  /* cannot read it */
3587                 }
3588                 /* Would I have to read this buffer for reconstruct_write */
3589                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3590                     i != sh->pd_idx && i != sh->qd_idx &&
3591                     !test_bit(R5_LOCKED, &dev->flags) &&
3592                     !(test_bit(R5_UPTODATE, &dev->flags) ||
3593                     test_bit(R5_Wantcompute, &dev->flags))) {
3594                         if (test_bit(R5_Insync, &dev->flags))
3595                                 rcw++;
3596                         else
3597                                 rcw += 2*disks;
3598                 }
3599         }
3600         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3601                 (unsigned long long)sh->sector, rmw, rcw);
3602         set_bit(STRIPE_HANDLE, &sh->state);
3603         if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_ENABLE_RMW)) && rmw > 0) {
3604                 /* prefer read-modify-write, but need to get some data */
3605                 if (conf->mddev->queue)
3606                         blk_add_trace_msg(conf->mddev->queue,
3607                                           "raid5 rmw %llu %d",
3608                                           (unsigned long long)sh->sector, rmw);
3609                 for (i = disks; i--; ) {
3610                         struct r5dev *dev = &sh->dev[i];
3611                         if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3612                             !test_bit(R5_LOCKED, &dev->flags) &&
3613                             !(test_bit(R5_UPTODATE, &dev->flags) ||
3614                             test_bit(R5_Wantcompute, &dev->flags)) &&
3615                             test_bit(R5_Insync, &dev->flags)) {
3616                                 if (test_bit(STRIPE_PREREAD_ACTIVE,
3617                                              &sh->state)) {
3618                                         pr_debug("Read_old block %d for r-m-w\n",
3619                                                  i);
3620                                         set_bit(R5_LOCKED, &dev->flags);
3621                                         set_bit(R5_Wantread, &dev->flags);
3622                                         s->locked++;
3623                                 } else {
3624                                         set_bit(STRIPE_DELAYED, &sh->state);
3625                                         set_bit(STRIPE_HANDLE, &sh->state);
3626                                 }
3627                         }
3628                 }
3629         }
3630         if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_ENABLE_RMW)) && rcw > 0) {
3631                 /* want reconstruct write, but need to get some data */
3632                 int qread =0;
3633                 rcw = 0;
3634                 for (i = disks; i--; ) {
3635                         struct r5dev *dev = &sh->dev[i];
3636                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3637                             i != sh->pd_idx && i != sh->qd_idx &&
3638                             !test_bit(R5_LOCKED, &dev->flags) &&
3639                             !(test_bit(R5_UPTODATE, &dev->flags) ||
3640                               test_bit(R5_Wantcompute, &dev->flags))) {
3641                                 rcw++;
3642                                 if (test_bit(R5_Insync, &dev->flags) &&
3643                                     test_bit(STRIPE_PREREAD_ACTIVE,
3644                                              &sh->state)) {
3645                                         pr_debug("Read_old block "
3646                                                 "%d for Reconstruct\n", i);
3647                                         set_bit(R5_LOCKED, &dev->flags);
3648                                         set_bit(R5_Wantread, &dev->flags);
3649                                         s->locked++;
3650                                         qread++;
3651                                 } else {
3652                                         set_bit(STRIPE_DELAYED, &sh->state);
3653                                         set_bit(STRIPE_HANDLE, &sh->state);
3654                                 }
3655                         }
3656                 }
3657                 if (rcw && conf->mddev->queue)
3658                         blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3659                                           (unsigned long long)sh->sector,
3660                                           rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3661         }
3662
3663         if (rcw > disks && rmw > disks &&
3664             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3665                 set_bit(STRIPE_DELAYED, &sh->state);
3666
3667         /* now if nothing is locked, and if we have enough data,
3668          * we can start a write request
3669          */
3670         /* since handle_stripe can be called at any time we need to handle the
3671          * case where a compute block operation has been submitted and then a
3672          * subsequent call wants to start a write request.  raid_run_ops only
3673          * handles the case where compute block and reconstruct are requested
3674          * simultaneously.  If this is not the case then new writes need to be
3675          * held off until the compute completes.
3676          */
3677         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3678             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3679             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3680                 schedule_reconstruction(sh, s, rcw == 0, 0);
3681 }
3682
3683 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3684                                 struct stripe_head_state *s, int disks)
3685 {
3686         struct r5dev *dev = NULL;
3687
3688         BUG_ON(sh->batch_head);
3689         set_bit(STRIPE_HANDLE, &sh->state);
3690
3691         switch (sh->check_state) {
3692         case check_state_idle:
3693                 /* start a new check operation if there are no failures */
3694                 if (s->failed == 0) {
3695                         BUG_ON(s->uptodate != disks);
3696                         sh->check_state = check_state_run;
3697                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
3698                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3699                         s->uptodate--;
3700                         break;
3701                 }
3702                 dev = &sh->dev[s->failed_num[0]];
3703                 /* fall through */
3704         case check_state_compute_result:
3705                 sh->check_state = check_state_idle;
3706                 if (!dev)
3707                         dev = &sh->dev[sh->pd_idx];
3708
3709                 /* check that a write has not made the stripe insync */
3710                 if (test_bit(STRIPE_INSYNC, &sh->state))
3711                         break;
3712
3713                 /* either failed parity check, or recovery is happening */
3714                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3715                 BUG_ON(s->uptodate != disks);
3716
3717                 set_bit(R5_LOCKED, &dev->flags);
3718                 s->locked++;
3719                 set_bit(R5_Wantwrite, &dev->flags);
3720
3721                 clear_bit(STRIPE_DEGRADED, &sh->state);
3722                 set_bit(STRIPE_INSYNC, &sh->state);
3723                 break;
3724         case check_state_run:
3725                 break; /* we will be called again upon completion */
3726         case check_state_check_result:
3727                 sh->check_state = check_state_idle;
3728
3729                 /* if a failure occurred during the check operation, leave
3730                  * STRIPE_INSYNC not set and let the stripe be handled again
3731                  */
3732                 if (s->failed)
3733                         break;
3734
3735                 /* handle a successful check operation, if parity is correct
3736                  * we are done.  Otherwise update the mismatch count and repair
3737                  * parity if !MD_RECOVERY_CHECK
3738                  */
3739                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3740                         /* parity is correct (on disc,
3741                          * not in buffer any more)
3742                          */
3743                         set_bit(STRIPE_INSYNC, &sh->state);
3744                 else {
3745                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3746                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3747                                 /* don't try to repair!! */
3748                                 set_bit(STRIPE_INSYNC, &sh->state);
3749                         else {
3750                                 sh->check_state = check_state_compute_run;
3751                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3752                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3753                                 set_bit(R5_Wantcompute,
3754                                         &sh->dev[sh->pd_idx].flags);
3755                                 sh->ops.target = sh->pd_idx;
3756                                 sh->ops.target2 = -1;
3757                                 s->uptodate++;
3758                         }
3759                 }
3760                 break;
3761         case check_state_compute_run:
3762                 break;
3763         default:
3764                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3765                        __func__, sh->check_state,
3766                        (unsigned long long) sh->sector);
3767                 BUG();
3768         }
3769 }
3770
3771 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3772                                   struct stripe_head_state *s,
3773                                   int disks)
3774 {
3775         int pd_idx = sh->pd_idx;
3776         int qd_idx = sh->qd_idx;
3777         struct r5dev *dev;
3778
3779         BUG_ON(sh->batch_head);
3780         set_bit(STRIPE_HANDLE, &sh->state);
3781
3782         BUG_ON(s->failed > 2);
3783
3784         /* Want to check and possibly repair P and Q.
3785          * However there could be one 'failed' device, in which
3786          * case we can only check one of them, possibly using the
3787          * other to generate missing data
3788          */
3789
3790         switch (sh->check_state) {
3791         case check_state_idle:
3792                 /* start a new check operation if there are < 2 failures */
3793                 if (s->failed == s->q_failed) {
3794                         /* The only possible failed device holds Q, so it
3795                          * makes sense to check P (If anything else were failed,
3796                          * we would have used P to recreate it).
3797                          */
3798                         sh->check_state = check_state_run;
3799                 }
3800                 if (!s->q_failed && s->failed < 2) {
3801                         /* Q is not failed, and we didn't use it to generate
3802                          * anything, so it makes sense to check it
3803                          */
3804                         if (sh->check_state == check_state_run)
3805                                 sh->check_state = check_state_run_pq;
3806                         else
3807                                 sh->check_state = check_state_run_q;
3808                 }
3809
3810                 /* discard potentially stale zero_sum_result */
3811                 sh->ops.zero_sum_result = 0;
3812
3813                 if (sh->check_state == check_state_run) {
3814                         /* async_xor_zero_sum destroys the contents of P */
3815                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3816                         s->uptodate--;
3817                 }
3818                 if (sh->check_state >= check_state_run &&
3819                     sh->check_state <= check_state_run_pq) {
3820                         /* async_syndrome_zero_sum preserves P and Q, so
3821                          * no need to mark them !uptodate here
3822                          */
3823                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
3824                         break;
3825                 }
3826
3827                 /* we have 2-disk failure */
3828                 BUG_ON(s->failed != 2);
3829                 /* fall through */
3830         case check_state_compute_result:
3831                 sh->check_state = check_state_idle;
3832
3833                 /* check that a write has not made the stripe insync */
3834                 if (test_bit(STRIPE_INSYNC, &sh->state))
3835                         break;
3836
3837                 /* now write out any block on a failed drive,
3838                  * or P or Q if they were recomputed
3839                  */
3840                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3841                 if (s->failed == 2) {
3842                         dev = &sh->dev[s->failed_num[1]];
3843                         s->locked++;
3844                         set_bit(R5_LOCKED, &dev->flags);
3845                         set_bit(R5_Wantwrite, &dev->flags);
3846                 }
3847                 if (s->failed >= 1) {
3848                         dev = &sh->dev[s->failed_num[0]];
3849                         s->locked++;
3850                         set_bit(R5_LOCKED, &dev->flags);
3851                         set_bit(R5_Wantwrite, &dev->flags);
3852                 }
3853                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3854                         dev = &sh->dev[pd_idx];
3855                         s->locked++;
3856                         set_bit(R5_LOCKED, &dev->flags);
3857                         set_bit(R5_Wantwrite, &dev->flags);
3858                 }
3859                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3860                         dev = &sh->dev[qd_idx];
3861                         s->locked++;
3862                         set_bit(R5_LOCKED, &dev->flags);
3863                         set_bit(R5_Wantwrite, &dev->flags);
3864                 }
3865                 clear_bit(STRIPE_DEGRADED, &sh->state);
3866
3867                 set_bit(STRIPE_INSYNC, &sh->state);
3868                 break;
3869         case check_state_run:
3870         case check_state_run_q:
3871         case check_state_run_pq:
3872                 break; /* we will be called again upon completion */
3873         case check_state_check_result:
3874                 sh->check_state = check_state_idle;
3875
3876                 /* handle a successful check operation, if parity is correct
3877                  * we are done.  Otherwise update the mismatch count and repair
3878                  * parity if !MD_RECOVERY_CHECK
3879                  */
3880                 if (sh->ops.zero_sum_result == 0) {
3881                         /* both parities are correct */
3882                         if (!s->failed)
3883                                 set_bit(STRIPE_INSYNC, &sh->state);
3884                         else {
3885                                 /* in contrast to the raid5 case we can validate
3886                                  * parity, but still have a failure to write
3887                                  * back
3888                                  */
3889                                 sh->check_state = check_state_compute_result;
3890                                 /* Returning at this point means that we may go
3891                                  * off and bring p and/or q uptodate again so
3892                                  * we make sure to check zero_sum_result again
3893                                  * to verify if p or q need writeback
3894                                  */
3895                         }
3896                 } else {
3897                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3898                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3899                                 /* don't try to repair!! */
3900                                 set_bit(STRIPE_INSYNC, &sh->state);
3901                         else {
3902                                 int *target = &sh->ops.target;
3903
3904                                 sh->ops.target = -1;
3905                                 sh->ops.target2 = -1;
3906                                 sh->check_state = check_state_compute_run;
3907                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3908                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3909                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3910                                         set_bit(R5_Wantcompute,
3911                                                 &sh->dev[pd_idx].flags);
3912                                         *target = pd_idx;
3913                                         target = &sh->ops.target2;
3914                                         s->uptodate++;
3915                                 }
3916                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3917                                         set_bit(R5_Wantcompute,
3918                                                 &sh->dev[qd_idx].flags);
3919                                         *target = qd_idx;
3920                                         s->uptodate++;
3921                                 }
3922                         }
3923                 }
3924                 break;
3925         case check_state_compute_run:
3926                 break;
3927         default:
3928                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3929                        __func__, sh->check_state,
3930                        (unsigned long long) sh->sector);
3931                 BUG();
3932         }
3933 }
3934
3935 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3936 {
3937         int i;
3938
3939         /* We have read all the blocks in this stripe and now we need to
3940          * copy some of them into a target stripe for expand.
3941          */
3942         struct dma_async_tx_descriptor *tx = NULL;
3943         BUG_ON(sh->batch_head);
3944         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3945         for (i = 0; i < sh->disks; i++)
3946                 if (i != sh->pd_idx && i != sh->qd_idx) {
3947                         int dd_idx, j;
3948                         struct stripe_head *sh2;
3949                         struct async_submit_ctl submit;
3950
3951                         sector_t bn = raid5_compute_blocknr(sh, i, 1);
3952                         sector_t s = raid5_compute_sector(conf, bn, 0,
3953                                                           &dd_idx, NULL);
3954                         sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
3955                         if (sh2 == NULL)
3956                                 /* so far only the early blocks of this stripe
3957                                  * have been requested.  When later blocks
3958                                  * get requested, we will try again
3959                                  */
3960                                 continue;
3961                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3962                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3963                                 /* must have already done this block */
3964                                 raid5_release_stripe(sh2);
3965                                 continue;
3966                         }
3967
3968                         /* place all the copies on one channel */
3969                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3970                         tx = async_memcpy(sh2->dev[dd_idx].page,
3971                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
3972                                           &submit);
3973
3974                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3975                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3976                         for (j = 0; j < conf->raid_disks; j++)
3977                                 if (j != sh2->pd_idx &&
3978                                     j != sh2->qd_idx &&
3979                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
3980                                         break;
3981                         if (j == conf->raid_disks) {
3982                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3983                                 set_bit(STRIPE_HANDLE, &sh2->state);
3984                         }
3985                         raid5_release_stripe(sh2);
3986
3987                 }
3988         /* done submitting copies, wait for them to complete */
3989         async_tx_quiesce(&tx);
3990 }
3991
3992 /*
3993  * handle_stripe - do things to a stripe.
3994  *
3995  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3996  * state of various bits to see what needs to be done.
3997  * Possible results:
3998  *    return some read requests which now have data
3999  *    return some write requests which are safely on storage
4000  *    schedule a read on some buffers
4001  *    schedule a write of some buffers
4002  *    return confirmation of parity correctness
4003  *
4004  */
4005
4006 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4007 {
4008         struct r5conf *conf = sh->raid_conf;
4009         int disks = sh->disks;
4010         struct r5dev *dev;
4011         int i;
4012         int do_recovery = 0;
4013
4014         memset(s, 0, sizeof(*s));
4015
4016         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4017         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4018         s->failed_num[0] = -1;
4019         s->failed_num[1] = -1;
4020         s->log_failed = r5l_log_disk_error(conf);
4021
4022         /* Now to look around and see what can be done */
4023         rcu_read_lock();
4024         for (i=disks; i--; ) {
4025                 struct md_rdev *rdev;
4026                 sector_t first_bad;
4027                 int bad_sectors;
4028                 int is_bad = 0;
4029
4030                 dev = &sh->dev[i];
4031
4032                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4033                          i, dev->flags,
4034                          dev->toread, dev->towrite, dev->written);
4035                 /* maybe we can reply to a read
4036                  *
4037                  * new wantfill requests are only permitted while
4038                  * ops_complete_biofill is guaranteed to be inactive
4039                  */
4040                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4041                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4042                         set_bit(R5_Wantfill, &dev->flags);
4043
4044                 /* now count some things */
4045                 if (test_bit(R5_LOCKED, &dev->flags))
4046                         s->locked++;
4047                 if (test_bit(R5_UPTODATE, &dev->flags))
4048                         s->uptodate++;
4049                 if (test_bit(R5_Wantcompute, &dev->flags)) {
4050                         s->compute++;
4051                         BUG_ON(s->compute > 2);
4052                 }
4053
4054                 if (test_bit(R5_Wantfill, &dev->flags))
4055                         s->to_fill++;
4056                 else if (dev->toread)
4057                         s->to_read++;
4058                 if (dev->towrite) {
4059                         s->to_write++;
4060                         if (!test_bit(R5_OVERWRITE, &dev->flags))
4061                                 s->non_overwrite++;
4062                 }
4063                 if (dev->written)
4064                         s->written++;
4065                 /* Prefer to use the replacement for reads, but only
4066                  * if it is recovered enough and has no bad blocks.
4067                  */
4068                 rdev = rcu_dereference(conf->disks[i].replacement);
4069                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4070                     rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4071                     !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4072                                  &first_bad, &bad_sectors))
4073                         set_bit(R5_ReadRepl, &dev->flags);
4074                 else {
4075                         if (rdev && !test_bit(Faulty, &rdev->flags))
4076                                 set_bit(R5_NeedReplace, &dev->flags);
4077                         else
4078                                 clear_bit(R5_NeedReplace, &dev->flags);
4079                         rdev = rcu_dereference(conf->disks[i].rdev);
4080                         clear_bit(R5_ReadRepl, &dev->flags);
4081                 }
4082                 if (rdev && test_bit(Faulty, &rdev->flags))
4083                         rdev = NULL;
4084                 if (rdev) {
4085                         is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4086                                              &first_bad, &bad_sectors);
4087                         if (s->blocked_rdev == NULL
4088                             && (test_bit(Blocked, &rdev->flags)
4089                                 || is_bad < 0)) {
4090                                 if (is_bad < 0)
4091                                         set_bit(BlockedBadBlocks,
4092                                                 &rdev->flags);
4093                                 s->blocked_rdev = rdev;
4094                                 atomic_inc(&rdev->nr_pending);
4095                         }
4096                 }
4097                 clear_bit(R5_Insync, &dev->flags);
4098                 if (!rdev)
4099                         /* Not in-sync */;
4100                 else if (is_bad) {
4101                         /* also not in-sync */
4102                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4103                             test_bit(R5_UPTODATE, &dev->flags)) {
4104                                 /* treat as in-sync, but with a read error
4105                                  * which we can now try to correct
4106                                  */
4107                                 set_bit(R5_Insync, &dev->flags);
4108                                 set_bit(R5_ReadError, &dev->flags);
4109                         }
4110                 } else if (test_bit(In_sync, &rdev->flags))
4111                         set_bit(R5_Insync, &dev->flags);
4112                 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4113                         /* in sync if before recovery_offset */
4114                         set_bit(R5_Insync, &dev->flags);
4115                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4116                          test_bit(R5_Expanded, &dev->flags))
4117                         /* If we've reshaped into here, we assume it is Insync.
4118                          * We will shortly update recovery_offset to make
4119                          * it official.
4120                          */
4121                         set_bit(R5_Insync, &dev->flags);
4122
4123                 if (test_bit(R5_WriteError, &dev->flags)) {
4124                         /* This flag does not apply to '.replacement'
4125                          * only to .rdev, so make sure to check that*/
4126                         struct md_rdev *rdev2 = rcu_dereference(
4127                                 conf->disks[i].rdev);
4128                         if (rdev2 == rdev)
4129                                 clear_bit(R5_Insync, &dev->flags);
4130                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4131                                 s->handle_bad_blocks = 1;
4132                                 atomic_inc(&rdev2->nr_pending);
4133                         } else
4134                                 clear_bit(R5_WriteError, &dev->flags);
4135                 }
4136                 if (test_bit(R5_MadeGood, &dev->flags)) {
4137                         /* This flag does not apply to '.replacement'
4138                          * only to .rdev, so make sure to check that*/
4139                         struct md_rdev *rdev2 = rcu_dereference(
4140                                 conf->disks[i].rdev);
4141                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4142                                 s->handle_bad_blocks = 1;
4143                                 atomic_inc(&rdev2->nr_pending);
4144                         } else
4145                                 clear_bit(R5_MadeGood, &dev->flags);
4146                 }
4147                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4148                         struct md_rdev *rdev2 = rcu_dereference(
4149                                 conf->disks[i].replacement);
4150                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4151                                 s->handle_bad_blocks = 1;
4152                                 atomic_inc(&rdev2->nr_pending);
4153                         } else
4154                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
4155                 }
4156                 if (!test_bit(R5_Insync, &dev->flags)) {
4157                         /* The ReadError flag will just be confusing now */
4158                         clear_bit(R5_ReadError, &dev->flags);
4159                         clear_bit(R5_ReWrite, &dev->flags);
4160                 }
4161                 if (test_bit(R5_ReadError, &dev->flags))
4162                         clear_bit(R5_Insync, &dev->flags);
4163                 if (!test_bit(R5_Insync, &dev->flags)) {
4164                         if (s->failed < 2)
4165                                 s->failed_num[s->failed] = i;
4166                         s->failed++;
4167                         if (rdev && !test_bit(Faulty, &rdev->flags))
4168                                 do_recovery = 1;
4169                 }
4170         }
4171         if (test_bit(STRIPE_SYNCING, &sh->state)) {
4172                 /* If there is a failed device being replaced,
4173                  *     we must be recovering.
4174                  * else if we are after recovery_cp, we must be syncing
4175                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4176                  * else we can only be replacing
4177                  * sync and recovery both need to read all devices, and so
4178                  * use the same flag.
4179                  */
4180                 if (do_recovery ||
4181                     sh->sector >= conf->mddev->recovery_cp ||
4182                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4183                         s->syncing = 1;
4184                 else
4185                         s->replacing = 1;
4186         }
4187         rcu_read_unlock();
4188 }
4189
4190 static int clear_batch_ready(struct stripe_head *sh)
4191 {
4192         /* Return '1' if this is a member of batch, or
4193          * '0' if it is a lone stripe or a head which can now be
4194          * handled.
4195          */
4196         struct stripe_head *tmp;
4197         if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4198                 return (sh->batch_head && sh->batch_head != sh);
4199         spin_lock(&sh->stripe_lock);
4200         if (!sh->batch_head) {
4201                 spin_unlock(&sh->stripe_lock);
4202                 return 0;
4203         }
4204
4205         /*
4206          * this stripe could be added to a batch list before we check
4207          * BATCH_READY, skips it
4208          */
4209         if (sh->batch_head != sh) {
4210                 spin_unlock(&sh->stripe_lock);
4211                 return 1;
4212         }
4213         spin_lock(&sh->batch_lock);
4214         list_for_each_entry(tmp, &sh->batch_list, batch_list)
4215                 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4216         spin_unlock(&sh->batch_lock);
4217         spin_unlock(&sh->stripe_lock);
4218
4219         /*
4220          * BATCH_READY is cleared, no new stripes can be added.
4221          * batch_list can be accessed without lock
4222          */
4223         return 0;
4224 }
4225
4226 static void break_stripe_batch_list(struct stripe_head *head_sh,
4227                                     unsigned long handle_flags)
4228 {
4229         struct stripe_head *sh, *next;
4230         int i;
4231         int do_wakeup = 0;
4232
4233         list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4234
4235                 list_del_init(&sh->batch_list);
4236
4237                 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4238                                           (1 << STRIPE_SYNCING) |
4239                                           (1 << STRIPE_REPLACED) |
4240                                           (1 << STRIPE_DELAYED) |
4241                                           (1 << STRIPE_BIT_DELAY) |
4242                                           (1 << STRIPE_FULL_WRITE) |
4243                                           (1 << STRIPE_BIOFILL_RUN) |
4244                                           (1 << STRIPE_COMPUTE_RUN)  |
4245                                           (1 << STRIPE_OPS_REQ_PENDING) |
4246                                           (1 << STRIPE_DISCARD) |
4247                                           (1 << STRIPE_BATCH_READY) |
4248                                           (1 << STRIPE_BATCH_ERR) |
4249                                           (1 << STRIPE_BITMAP_PENDING)),
4250                         "stripe state: %lx\n", sh->state);
4251                 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4252                                               (1 << STRIPE_REPLACED)),
4253                         "head stripe state: %lx\n", head_sh->state);
4254
4255                 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4256                                             (1 << STRIPE_PREREAD_ACTIVE) |
4257                                             (1 << STRIPE_DEGRADED)),
4258                               head_sh->state & (1 << STRIPE_INSYNC));
4259
4260                 sh->check_state = head_sh->check_state;
4261                 sh->reconstruct_state = head_sh->reconstruct_state;
4262                 for (i = 0; i < sh->disks; i++) {
4263                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4264                                 do_wakeup = 1;
4265                         sh->dev[i].flags = head_sh->dev[i].flags &
4266                                 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4267                 }
4268                 spin_lock_irq(&sh->stripe_lock);
4269                 sh->batch_head = NULL;
4270                 spin_unlock_irq(&sh->stripe_lock);
4271                 if (handle_flags == 0 ||
4272                     sh->state & handle_flags)
4273                         set_bit(STRIPE_HANDLE, &sh->state);
4274                 raid5_release_stripe(sh);
4275         }
4276         spin_lock_irq(&head_sh->stripe_lock);
4277         head_sh->batch_head = NULL;
4278         spin_unlock_irq(&head_sh->stripe_lock);
4279         for (i = 0; i < head_sh->disks; i++)
4280                 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4281                         do_wakeup = 1;
4282         if (head_sh->state & handle_flags)
4283                 set_bit(STRIPE_HANDLE, &head_sh->state);
4284
4285         if (do_wakeup)
4286                 wake_up(&head_sh->raid_conf->wait_for_overlap);
4287 }
4288
4289 static void handle_stripe(struct stripe_head *sh)
4290 {
4291         struct stripe_head_state s;
4292         struct r5conf *conf = sh->raid_conf;
4293         int i;
4294         int prexor;
4295         int disks = sh->disks;
4296         struct r5dev *pdev, *qdev;
4297
4298         clear_bit(STRIPE_HANDLE, &sh->state);
4299         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4300                 /* already being handled, ensure it gets handled
4301                  * again when current action finishes */
4302                 set_bit(STRIPE_HANDLE, &sh->state);
4303                 return;
4304         }
4305
4306         if (clear_batch_ready(sh) ) {
4307                 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4308                 return;
4309         }
4310
4311         if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4312                 break_stripe_batch_list(sh, 0);
4313
4314         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4315                 spin_lock(&sh->stripe_lock);
4316                 /* Cannot process 'sync' concurrently with 'discard' */
4317                 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4318                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4319                         set_bit(STRIPE_SYNCING, &sh->state);
4320                         clear_bit(STRIPE_INSYNC, &sh->state);
4321                         clear_bit(STRIPE_REPLACED, &sh->state);
4322                 }
4323                 spin_unlock(&sh->stripe_lock);
4324         }
4325         clear_bit(STRIPE_DELAYED, &sh->state);
4326
4327         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4328                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4329                (unsigned long long)sh->sector, sh->state,
4330                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4331                sh->check_state, sh->reconstruct_state);
4332
4333         analyse_stripe(sh, &s);
4334
4335         if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4336                 goto finish;
4337
4338         if (s.handle_bad_blocks) {
4339                 set_bit(STRIPE_HANDLE, &sh->state);
4340                 goto finish;
4341         }
4342
4343         if (unlikely(s.blocked_rdev)) {
4344                 if (s.syncing || s.expanding || s.expanded ||
4345                     s.replacing || s.to_write || s.written) {
4346                         set_bit(STRIPE_HANDLE, &sh->state);
4347                         goto finish;
4348                 }
4349                 /* There is nothing for the blocked_rdev to block */
4350                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4351                 s.blocked_rdev = NULL;
4352         }
4353
4354         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4355                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4356                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4357         }
4358
4359         pr_debug("locked=%d uptodate=%d to_read=%d"
4360                " to_write=%d failed=%d failed_num=%d,%d\n",
4361                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4362                s.failed_num[0], s.failed_num[1]);
4363         /* check if the array has lost more than max_degraded devices and,
4364          * if so, some requests might need to be failed.
4365          */
4366         if (s.failed > conf->max_degraded || s.log_failed) {
4367                 sh->check_state = 0;
4368                 sh->reconstruct_state = 0;
4369                 break_stripe_batch_list(sh, 0);
4370                 if (s.to_read+s.to_write+s.written)
4371                         handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4372                 if (s.syncing + s.replacing)
4373                         handle_failed_sync(conf, sh, &s);
4374         }
4375
4376         /* Now we check to see if any write operations have recently
4377          * completed
4378          */
4379         prexor = 0;
4380         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4381                 prexor = 1;
4382         if (sh->reconstruct_state == reconstruct_state_drain_result ||
4383             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4384                 sh->reconstruct_state = reconstruct_state_idle;
4385
4386                 /* All the 'written' buffers and the parity block are ready to
4387                  * be written back to disk
4388                  */
4389                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4390                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4391                 BUG_ON(sh->qd_idx >= 0 &&
4392                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4393                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4394                 for (i = disks; i--; ) {
4395                         struct r5dev *dev = &sh->dev[i];
4396                         if (test_bit(R5_LOCKED, &dev->flags) &&
4397                                 (i == sh->pd_idx || i == sh->qd_idx ||
4398                                  dev->written)) {
4399                                 pr_debug("Writing block %d\n", i);
4400                                 set_bit(R5_Wantwrite, &dev->flags);
4401                                 if (prexor)
4402                                         continue;
4403                                 if (s.failed > 1)
4404                                         continue;
4405                                 if (!test_bit(R5_Insync, &dev->flags) ||
4406                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
4407                                      s.failed == 0))
4408                                         set_bit(STRIPE_INSYNC, &sh->state);
4409                         }
4410                 }
4411                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4412                         s.dec_preread_active = 1;
4413         }
4414
4415         /*
4416          * might be able to return some write requests if the parity blocks
4417          * are safe, or on a failed drive
4418          */
4419         pdev = &sh->dev[sh->pd_idx];
4420         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4421                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4422         qdev = &sh->dev[sh->qd_idx];
4423         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4424                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4425                 || conf->level < 6;
4426
4427         if (s.written &&
4428             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4429                              && !test_bit(R5_LOCKED, &pdev->flags)
4430                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
4431                                  test_bit(R5_Discard, &pdev->flags))))) &&
4432             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4433                              && !test_bit(R5_LOCKED, &qdev->flags)
4434                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
4435                                  test_bit(R5_Discard, &qdev->flags))))))
4436                 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4437
4438         /* Now we might consider reading some blocks, either to check/generate
4439          * parity, or to satisfy requests
4440          * or to load a block that is being partially written.
4441          */
4442         if (s.to_read || s.non_overwrite
4443             || (conf->level == 6 && s.to_write && s.failed)
4444             || (s.syncing && (s.uptodate + s.compute < disks))
4445             || s.replacing
4446             || s.expanding)
4447                 handle_stripe_fill(sh, &s, disks);
4448
4449         /* Now to consider new write requests and what else, if anything
4450          * should be read.  We do not handle new writes when:
4451          * 1/ A 'write' operation (copy+xor) is already in flight.
4452          * 2/ A 'check' operation is in flight, as it may clobber the parity
4453          *    block.
4454          */
4455         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4456                 handle_stripe_dirtying(conf, sh, &s, disks);
4457
4458         /* maybe we need to check and possibly fix the parity for this stripe
4459          * Any reads will already have been scheduled, so we just see if enough
4460          * data is available.  The parity check is held off while parity
4461          * dependent operations are in flight.
4462          */
4463         if (sh->check_state ||
4464             (s.syncing && s.locked == 0 &&
4465              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4466              !test_bit(STRIPE_INSYNC, &sh->state))) {
4467                 if (conf->level == 6)
4468                         handle_parity_checks6(conf, sh, &s, disks);
4469                 else
4470                         handle_parity_checks5(conf, sh, &s, disks);
4471         }
4472
4473         if ((s.replacing || s.syncing) && s.locked == 0
4474             && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4475             && !test_bit(STRIPE_REPLACED, &sh->state)) {
4476                 /* Write out to replacement devices where possible */
4477                 for (i = 0; i < conf->raid_disks; i++)
4478                         if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4479                                 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4480                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
4481                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
4482                                 s.locked++;
4483                         }
4484                 if (s.replacing)
4485                         set_bit(STRIPE_INSYNC, &sh->state);
4486                 set_bit(STRIPE_REPLACED, &sh->state);
4487         }
4488         if ((s.syncing || s.replacing) && s.locked == 0 &&
4489             !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4490             test_bit(STRIPE_INSYNC, &sh->state)) {
4491                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4492                 clear_bit(STRIPE_SYNCING, &sh->state);
4493                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4494                         wake_up(&conf->wait_for_overlap);
4495         }
4496
4497         /* If the failed drives are just a ReadError, then we might need
4498          * to progress the repair/check process
4499          */
4500         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4501                 for (i = 0; i < s.failed; i++) {
4502                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
4503                         if (test_bit(R5_ReadError, &dev->flags)
4504                             && !test_bit(R5_LOCKED, &dev->flags)
4505                             && test_bit(R5_UPTODATE, &dev->flags)
4506                                 ) {
4507                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
4508                                         set_bit(R5_Wantwrite, &dev->flags);
4509                                         set_bit(R5_ReWrite, &dev->flags);
4510                                         set_bit(R5_LOCKED, &dev->flags);
4511                                         s.locked++;
4512                                 } else {
4513                                         /* let's read it back */
4514                                         set_bit(R5_Wantread, &dev->flags);
4515                                         set_bit(R5_LOCKED, &dev->flags);
4516                                         s.locked++;
4517                                 }
4518                         }
4519                 }
4520
4521         /* Finish reconstruct operations initiated by the expansion process */
4522         if (sh->reconstruct_state == reconstruct_state_result) {
4523                 struct stripe_head *sh_src
4524                         = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4525                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4526                         /* sh cannot be written until sh_src has been read.
4527                          * so arrange for sh to be delayed a little
4528                          */
4529                         set_bit(STRIPE_DELAYED, &sh->state);
4530                         set_bit(STRIPE_HANDLE, &sh->state);
4531                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4532                                               &sh_src->state))
4533                                 atomic_inc(&conf->preread_active_stripes);
4534                         raid5_release_stripe(sh_src);
4535                         goto finish;
4536                 }
4537                 if (sh_src)
4538                         raid5_release_stripe(sh_src);
4539
4540                 sh->reconstruct_state = reconstruct_state_idle;
4541                 clear_bit(STRIPE_EXPANDING, &sh->state);
4542                 for (i = conf->raid_disks; i--; ) {
4543                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
4544                         set_bit(R5_LOCKED, &sh->dev[i].flags);
4545                         s.locked++;
4546                 }
4547         }
4548
4549         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4550             !sh->reconstruct_state) {
4551                 /* Need to write out all blocks after computing parity */
4552                 sh->disks = conf->raid_disks;
4553                 stripe_set_idx(sh->sector, conf, 0, sh);
4554                 schedule_reconstruction(sh, &s, 1, 1);
4555         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4556                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4557                 atomic_dec(&conf->reshape_stripes);
4558                 wake_up(&conf->wait_for_overlap);
4559                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4560         }
4561
4562         if (s.expanding && s.locked == 0 &&
4563             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4564                 handle_stripe_expansion(conf, sh);
4565
4566 finish:
4567         /* wait for this device to become unblocked */
4568         if (unlikely(s.blocked_rdev)) {
4569                 if (conf->mddev->external)
4570                         md_wait_for_blocked_rdev(s.blocked_rdev,
4571                                                  conf->mddev);
4572                 else
4573                         /* Internal metadata will immediately
4574                          * be written by raid5d, so we don't
4575                          * need to wait here.
4576                          */
4577                         rdev_dec_pending(s.blocked_rdev,
4578                                          conf->mddev);
4579         }
4580
4581         if (s.handle_bad_blocks)
4582                 for (i = disks; i--; ) {
4583                         struct md_rdev *rdev;
4584                         struct r5dev *dev = &sh->dev[i];
4585                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4586                                 /* We own a safe reference to the rdev */
4587                                 rdev = conf->disks[i].rdev;
4588                                 if (!rdev_set_badblocks(rdev, sh->sector,
4589                                                         STRIPE_SECTORS, 0))
4590                                         md_error(conf->mddev, rdev);
4591                                 rdev_dec_pending(rdev, conf->mddev);
4592                         }
4593                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4594                                 rdev = conf->disks[i].rdev;
4595                                 rdev_clear_badblocks(rdev, sh->sector,
4596                                                      STRIPE_SECTORS, 0);
4597                                 rdev_dec_pending(rdev, conf->mddev);
4598                         }
4599                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4600                                 rdev = conf->disks[i].replacement;
4601                                 if (!rdev)
4602                                         /* rdev have been moved down */
4603                                         rdev = conf->disks[i].rdev;
4604                                 rdev_clear_badblocks(rdev, sh->sector,
4605                                                      STRIPE_SECTORS, 0);
4606                                 rdev_dec_pending(rdev, conf->mddev);
4607                         }
4608                 }
4609
4610         if (s.ops_request)
4611                 raid_run_ops(sh, s.ops_request);
4612
4613         ops_run_io(sh, &s);
4614
4615         if (s.dec_preread_active) {
4616                 /* We delay this until after ops_run_io so that if make_request
4617                  * is waiting on a flush, it won't continue until the writes
4618                  * have actually been submitted.
4619                  */
4620                 atomic_dec(&conf->preread_active_stripes);
4621                 if (atomic_read(&conf->preread_active_stripes) <
4622                     IO_THRESHOLD)
4623                         md_wakeup_thread(conf->mddev->thread);
4624         }
4625
4626         if (!bio_list_empty(&s.return_bi)) {
4627                 if (test_bit(MD_CHANGE_PENDING, &conf->mddev->flags)) {
4628                         spin_lock_irq(&conf->device_lock);
4629                         bio_list_merge(&conf->return_bi, &s.return_bi);
4630                         spin_unlock_irq(&conf->device_lock);
4631                         md_wakeup_thread(conf->mddev->thread);
4632                 } else
4633                         return_io(&s.return_bi);
4634         }
4635
4636         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4637 }
4638
4639 static void raid5_activate_delayed(struct r5conf *conf)
4640 {
4641         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4642                 while (!list_empty(&conf->delayed_list)) {
4643                         struct list_head *l = conf->delayed_list.next;
4644                         struct stripe_head *sh;
4645                         sh = list_entry(l, struct stripe_head, lru);
4646                         list_del_init(l);
4647                         clear_bit(STRIPE_DELAYED, &sh->state);
4648                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4649                                 atomic_inc(&conf->preread_active_stripes);
4650                         list_add_tail(&sh->lru, &conf->hold_list);
4651                         raid5_wakeup_stripe_thread(sh);
4652                 }
4653         }
4654 }
4655
4656 static void activate_bit_delay(struct r5conf *conf,
4657         struct list_head *temp_inactive_list)
4658 {
4659         /* device_lock is held */
4660         struct list_head head;
4661         list_add(&head, &conf->bitmap_list);
4662         list_del_init(&conf->bitmap_list);
4663         while (!list_empty(&head)) {
4664                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4665                 int hash;
4666                 list_del_init(&sh->lru);
4667                 atomic_inc(&sh->count);
4668                 hash = sh->hash_lock_index;
4669                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4670         }
4671 }
4672
4673 static int raid5_congested(struct mddev *mddev, int bits)
4674 {
4675         struct r5conf *conf = mddev->private;
4676
4677         /* No difference between reads and writes.  Just check
4678          * how busy the stripe_cache is
4679          */
4680
4681         if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4682                 return 1;
4683         if (conf->quiesce)
4684                 return 1;
4685         if (atomic_read(&conf->empty_inactive_list_nr))
4686                 return 1;
4687
4688         return 0;
4689 }
4690
4691 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4692 {
4693         struct r5conf *conf = mddev->private;
4694         sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4695         unsigned int chunk_sectors;
4696         unsigned int bio_sectors = bio_sectors(bio);
4697
4698         chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
4699         return  chunk_sectors >=
4700                 ((sector & (chunk_sectors - 1)) + bio_sectors);
4701 }
4702
4703 /*
4704  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
4705  *  later sampled by raid5d.
4706  */
4707 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4708 {
4709         unsigned long flags;
4710
4711         spin_lock_irqsave(&conf->device_lock, flags);
4712
4713         bi->bi_next = conf->retry_read_aligned_list;
4714         conf->retry_read_aligned_list = bi;
4715
4716         spin_unlock_irqrestore(&conf->device_lock, flags);
4717         md_wakeup_thread(conf->mddev->thread);
4718 }
4719
4720 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4721 {
4722         struct bio *bi;
4723
4724         bi = conf->retry_read_aligned;
4725         if (bi) {
4726                 conf->retry_read_aligned = NULL;
4727                 return bi;
4728         }
4729         bi = conf->retry_read_aligned_list;
4730         if(bi) {
4731                 conf->retry_read_aligned_list = bi->bi_next;
4732                 bi->bi_next = NULL;
4733                 /*
4734                  * this sets the active strip count to 1 and the processed
4735                  * strip count to zero (upper 8 bits)
4736                  */
4737                 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4738         }
4739
4740         return bi;
4741 }
4742
4743 /*
4744  *  The "raid5_align_endio" should check if the read succeeded and if it
4745  *  did, call bio_endio on the original bio (having bio_put the new bio
4746  *  first).
4747  *  If the read failed..
4748  */
4749 static void raid5_align_endio(struct bio *bi)
4750 {
4751         struct bio* raid_bi  = bi->bi_private;
4752         struct mddev *mddev;
4753         struct r5conf *conf;
4754         struct md_rdev *rdev;
4755         int error = bi->bi_error;
4756
4757         bio_put(bi);
4758
4759         rdev = (void*)raid_bi->bi_next;
4760         raid_bi->bi_next = NULL;
4761         mddev = rdev->mddev;
4762         conf = mddev->private;
4763
4764         rdev_dec_pending(rdev, conf->mddev);
4765
4766         if (!error) {
4767                 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4768                                          raid_bi, 0);
4769                 bio_endio(raid_bi);
4770                 if (atomic_dec_and_test(&conf->active_aligned_reads))
4771                         wake_up(&conf->wait_for_quiescent);
4772                 return;
4773         }
4774
4775         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4776
4777         add_bio_to_retry(raid_bi, conf);
4778 }
4779
4780 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
4781 {
4782         struct r5conf *conf = mddev->private;
4783         int dd_idx;
4784         struct bio* align_bi;
4785         struct md_rdev *rdev;
4786         sector_t end_sector;
4787
4788         if (!in_chunk_boundary(mddev, raid_bio)) {
4789                 pr_debug("%s: non aligned\n", __func__);
4790                 return 0;
4791         }
4792         /*
4793          * use bio_clone_mddev to make a copy of the bio
4794          */
4795         align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4796         if (!align_bi)
4797                 return 0;
4798         /*
4799          *   set bi_end_io to a new function, and set bi_private to the
4800          *     original bio.
4801          */
4802         align_bi->bi_end_io  = raid5_align_endio;
4803         align_bi->bi_private = raid_bio;
4804         /*
4805          *      compute position
4806          */
4807         align_bi->bi_iter.bi_sector =
4808                 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4809                                      0, &dd_idx, NULL);
4810
4811         end_sector = bio_end_sector(align_bi);
4812         rcu_read_lock();
4813         rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4814         if (!rdev || test_bit(Faulty, &rdev->flags) ||
4815             rdev->recovery_offset < end_sector) {
4816                 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4817                 if (rdev &&
4818                     (test_bit(Faulty, &rdev->flags) ||
4819                     !(test_bit(In_sync, &rdev->flags) ||
4820                       rdev->recovery_offset >= end_sector)))
4821                         rdev = NULL;
4822         }
4823         if (rdev) {
4824                 sector_t first_bad;
4825                 int bad_sectors;
4826
4827                 atomic_inc(&rdev->nr_pending);
4828                 rcu_read_unlock();
4829                 raid_bio->bi_next = (void*)rdev;
4830                 align_bi->bi_bdev =  rdev->bdev;
4831                 bio_clear_flag(align_bi, BIO_SEG_VALID);
4832
4833                 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
4834                                 bio_sectors(align_bi),
4835                                 &first_bad, &bad_sectors)) {
4836                         bio_put(align_bi);
4837                         rdev_dec_pending(rdev, mddev);
4838                         return 0;
4839                 }
4840
4841                 /* No reshape active, so we can trust rdev->data_offset */
4842                 align_bi->bi_iter.bi_sector += rdev->data_offset;
4843
4844                 spin_lock_irq(&conf->device_lock);
4845                 wait_event_lock_irq(conf->wait_for_quiescent,
4846                                     conf->quiesce == 0,
4847                                     conf->device_lock);
4848                 atomic_inc(&conf->active_aligned_reads);
4849                 spin_unlock_irq(&conf->device_lock);
4850
4851                 if (mddev->gendisk)
4852                         trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4853                                               align_bi, disk_devt(mddev->gendisk),
4854                                               raid_bio->bi_iter.bi_sector);
4855                 generic_make_request(align_bi);
4856                 return 1;
4857         } else {
4858                 rcu_read_unlock();
4859                 bio_put(align_bi);
4860                 return 0;
4861         }
4862 }
4863
4864 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
4865 {
4866         struct bio *split;
4867
4868         do {
4869                 sector_t sector = raid_bio->bi_iter.bi_sector;
4870                 unsigned chunk_sects = mddev->chunk_sectors;
4871                 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
4872
4873                 if (sectors < bio_sectors(raid_bio)) {
4874                         split = bio_split(raid_bio, sectors, GFP_NOIO, fs_bio_set);
4875                         bio_chain(split, raid_bio);
4876                 } else
4877                         split = raid_bio;
4878
4879                 if (!raid5_read_one_chunk(mddev, split)) {
4880                         if (split != raid_bio)
4881                                 generic_make_request(raid_bio);
4882                         return split;
4883                 }
4884         } while (split != raid_bio);
4885
4886         return NULL;
4887 }
4888
4889 /* __get_priority_stripe - get the next stripe to process
4890  *
4891  * Full stripe writes are allowed to pass preread active stripes up until
4892  * the bypass_threshold is exceeded.  In general the bypass_count
4893  * increments when the handle_list is handled before the hold_list; however, it
4894  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4895  * stripe with in flight i/o.  The bypass_count will be reset when the
4896  * head of the hold_list has changed, i.e. the head was promoted to the
4897  * handle_list.
4898  */
4899 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4900 {
4901         struct stripe_head *sh = NULL, *tmp;
4902         struct list_head *handle_list = NULL;
4903         struct r5worker_group *wg = NULL;
4904
4905         if (conf->worker_cnt_per_group == 0) {
4906                 handle_list = &conf->handle_list;
4907         } else if (group != ANY_GROUP) {
4908                 handle_list = &conf->worker_groups[group].handle_list;
4909                 wg = &conf->worker_groups[group];
4910         } else {
4911                 int i;
4912                 for (i = 0; i < conf->group_cnt; i++) {
4913                         handle_list = &conf->worker_groups[i].handle_list;
4914                         wg = &conf->worker_groups[i];
4915                         if (!list_empty(handle_list))
4916                                 break;
4917                 }
4918         }
4919
4920         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4921                   __func__,
4922                   list_empty(handle_list) ? "empty" : "busy",
4923                   list_empty(&conf->hold_list) ? "empty" : "busy",
4924                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
4925
4926         if (!list_empty(handle_list)) {
4927                 sh = list_entry(handle_list->next, typeof(*sh), lru);
4928
4929                 if (list_empty(&conf->hold_list))
4930                         conf->bypass_count = 0;
4931                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4932                         if (conf->hold_list.next == conf->last_hold)
4933                                 conf->bypass_count++;
4934                         else {
4935                                 conf->last_hold = conf->hold_list.next;
4936                                 conf->bypass_count -= conf->bypass_threshold;
4937                                 if (conf->bypass_count < 0)
4938                                         conf->bypass_count = 0;
4939                         }
4940                 }
4941         } else if (!list_empty(&conf->hold_list) &&
4942                    ((conf->bypass_threshold &&
4943                      conf->bypass_count > conf->bypass_threshold) ||
4944                     atomic_read(&conf->pending_full_writes) == 0)) {
4945
4946                 list_for_each_entry(tmp, &conf->hold_list,  lru) {
4947                         if (conf->worker_cnt_per_group == 0 ||
4948                             group == ANY_GROUP ||
4949                             !cpu_online(tmp->cpu) ||
4950                             cpu_to_group(tmp->cpu) == group) {
4951                                 sh = tmp;
4952                                 break;
4953                         }
4954                 }
4955
4956                 if (sh) {
4957                         conf->bypass_count -= conf->bypass_threshold;
4958                         if (conf->bypass_count < 0)
4959                                 conf->bypass_count = 0;
4960                 }
4961                 wg = NULL;
4962         }
4963
4964         if (!sh)
4965                 return NULL;
4966
4967         if (wg) {
4968                 wg->stripes_cnt--;
4969                 sh->group = NULL;
4970         }
4971         list_del_init(&sh->lru);
4972         BUG_ON(atomic_inc_return(&sh->count) != 1);
4973         return sh;
4974 }
4975
4976 struct raid5_plug_cb {
4977         struct blk_plug_cb      cb;
4978         struct list_head        list;
4979         struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4980 };
4981
4982 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4983 {
4984         struct raid5_plug_cb *cb = container_of(
4985                 blk_cb, struct raid5_plug_cb, cb);
4986         struct stripe_head *sh;
4987         struct mddev *mddev = cb->cb.data;
4988         struct r5conf *conf = mddev->private;
4989         int cnt = 0;
4990         int hash;
4991
4992         if (cb->list.next && !list_empty(&cb->list)) {
4993                 spin_lock_irq(&conf->device_lock);
4994                 while (!list_empty(&cb->list)) {
4995                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
4996                         list_del_init(&sh->lru);
4997                         /*
4998                          * avoid race release_stripe_plug() sees
4999                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
5000                          * is still in our list
5001                          */
5002                         smp_mb__before_atomic();
5003                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5004                         /*
5005                          * STRIPE_ON_RELEASE_LIST could be set here. In that
5006                          * case, the count is always > 1 here
5007                          */
5008                         hash = sh->hash_lock_index;
5009                         __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5010                         cnt++;
5011                 }
5012                 spin_unlock_irq(&conf->device_lock);
5013         }
5014         release_inactive_stripe_list(conf, cb->temp_inactive_list,
5015                                      NR_STRIPE_HASH_LOCKS);
5016         if (mddev->queue)
5017                 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5018         kfree(cb);
5019 }
5020
5021 static void release_stripe_plug(struct mddev *mddev,
5022                                 struct stripe_head *sh)
5023 {
5024         struct blk_plug_cb *blk_cb = blk_check_plugged(
5025                 raid5_unplug, mddev,
5026                 sizeof(struct raid5_plug_cb));
5027         struct raid5_plug_cb *cb;
5028
5029         if (!blk_cb) {
5030                 raid5_release_stripe(sh);
5031                 return;
5032         }
5033
5034         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5035
5036         if (cb->list.next == NULL) {
5037                 int i;
5038                 INIT_LIST_HEAD(&cb->list);
5039                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5040                         INIT_LIST_HEAD(cb->temp_inactive_list + i);
5041         }
5042
5043         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5044                 list_add_tail(&sh->lru, &cb->list);
5045         else
5046                 raid5_release_stripe(sh);
5047 }
5048
5049 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5050 {
5051         struct r5conf *conf = mddev->private;
5052         sector_t logical_sector, last_sector;
5053         struct stripe_head *sh;
5054         int remaining;
5055         int stripe_sectors;
5056
5057         if (mddev->reshape_position != MaxSector)
5058                 /* Skip discard while reshape is happening */
5059                 return;
5060
5061         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5062         last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5063
5064         bi->bi_next = NULL;
5065         bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5066
5067         stripe_sectors = conf->chunk_sectors *
5068                 (conf->raid_disks - conf->max_degraded);
5069         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5070                                                stripe_sectors);
5071         sector_div(last_sector, stripe_sectors);
5072
5073         logical_sector *= conf->chunk_sectors;
5074         last_sector *= conf->chunk_sectors;
5075
5076         for (; logical_sector < last_sector;
5077              logical_sector += STRIPE_SECTORS) {
5078                 DEFINE_WAIT(w);
5079                 int d;
5080         again:
5081                 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5082                 prepare_to_wait(&conf->wait_for_overlap, &w,
5083                                 TASK_UNINTERRUPTIBLE);
5084                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5085                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5086                         raid5_release_stripe(sh);
5087                         schedule();
5088                         goto again;
5089                 }
5090                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5091                 spin_lock_irq(&sh->stripe_lock);
5092                 for (d = 0; d < conf->raid_disks; d++) {
5093                         if (d == sh->pd_idx || d == sh->qd_idx)
5094                                 continue;
5095                         if (sh->dev[d].towrite || sh->dev[d].toread) {
5096                                 set_bit(R5_Overlap, &sh->dev[d].flags);
5097                                 spin_unlock_irq(&sh->stripe_lock);
5098                                 raid5_release_stripe(sh);
5099                                 schedule();
5100                                 goto again;
5101                         }
5102                 }
5103                 set_bit(STRIPE_DISCARD, &sh->state);
5104                 finish_wait(&conf->wait_for_overlap, &w);
5105                 sh->overwrite_disks = 0;
5106                 for (d = 0; d < conf->raid_disks; d++) {
5107                         if (d == sh->pd_idx || d == sh->qd_idx)
5108                                 continue;
5109                         sh->dev[d].towrite = bi;
5110                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5111                         raid5_inc_bi_active_stripes(bi);
5112                         sh->overwrite_disks++;
5113                 }
5114                 spin_unlock_irq(&sh->stripe_lock);
5115                 if (conf->mddev->bitmap) {
5116                         for (d = 0;
5117                              d < conf->raid_disks - conf->max_degraded;
5118                              d++)
5119                                 bitmap_startwrite(mddev->bitmap,
5120                                                   sh->sector,
5121                                                   STRIPE_SECTORS,
5122                                                   0);
5123                         sh->bm_seq = conf->seq_flush + 1;
5124                         set_bit(STRIPE_BIT_DELAY, &sh->state);
5125                 }
5126
5127                 set_bit(STRIPE_HANDLE, &sh->state);
5128                 clear_bit(STRIPE_DELAYED, &sh->state);
5129                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5130                         atomic_inc(&conf->preread_active_stripes);
5131                 release_stripe_plug(mddev, sh);
5132         }
5133
5134         remaining = raid5_dec_bi_active_stripes(bi);
5135         if (remaining == 0) {
5136                 md_write_end(mddev);
5137                 bio_endio(bi);
5138         }
5139 }
5140
5141 static void raid5_make_request(struct mddev *mddev, struct bio * bi)
5142 {
5143         struct r5conf *conf = mddev->private;
5144         int dd_idx;
5145         sector_t new_sector;
5146         sector_t logical_sector, last_sector;
5147         struct stripe_head *sh;
5148         const int rw = bio_data_dir(bi);
5149         int remaining;
5150         DEFINE_WAIT(w);
5151         bool do_prepare;
5152
5153         if (unlikely(bi->bi_rw & REQ_PREFLUSH)) {
5154                 int ret = r5l_handle_flush_request(conf->log, bi);
5155
5156                 if (ret == 0)
5157                         return;
5158                 if (ret == -ENODEV) {
5159                         md_flush_request(mddev, bi);
5160                         return;
5161                 }
5162                 /* ret == -EAGAIN, fallback */
5163         }
5164
5165         md_write_start(mddev, bi);
5166
5167         /*
5168          * If array is degraded, better not do chunk aligned read because
5169          * later we might have to read it again in order to reconstruct
5170          * data on failed drives.
5171          */
5172         if (rw == READ && mddev->degraded == 0 &&
5173             mddev->reshape_position == MaxSector) {
5174                 bi = chunk_aligned_read(mddev, bi);
5175                 if (!bi)
5176                         return;
5177         }
5178
5179         if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5180                 make_discard_request(mddev, bi);
5181                 return;
5182         }
5183
5184         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5185         last_sector = bio_end_sector(bi);
5186         bi->bi_next = NULL;
5187         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
5188
5189         prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5190         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5191                 int previous;
5192                 int seq;
5193
5194                 do_prepare = false;
5195         retry:
5196                 seq = read_seqcount_begin(&conf->gen_lock);
5197                 previous = 0;
5198                 if (do_prepare)
5199                         prepare_to_wait(&conf->wait_for_overlap, &w,
5200                                 TASK_UNINTERRUPTIBLE);
5201                 if (unlikely(conf->reshape_progress != MaxSector)) {
5202                         /* spinlock is needed as reshape_progress may be
5203                          * 64bit on a 32bit platform, and so it might be
5204                          * possible to see a half-updated value
5205                          * Of course reshape_progress could change after
5206                          * the lock is dropped, so once we get a reference
5207                          * to the stripe that we think it is, we will have
5208                          * to check again.
5209                          */
5210                         spin_lock_irq(&conf->device_lock);
5211                         if (mddev->reshape_backwards
5212                             ? logical_sector < conf->reshape_progress
5213                             : logical_sector >= conf->reshape_progress) {
5214                                 previous = 1;
5215                         } else {
5216                                 if (mddev->reshape_backwards
5217                                     ? logical_sector < conf->reshape_safe
5218                                     : logical_sector >= conf->reshape_safe) {
5219                                         spin_unlock_irq(&conf->device_lock);
5220                                         schedule();
5221                                         do_prepare = true;
5222                                         goto retry;
5223                                 }
5224                         }
5225                         spin_unlock_irq(&conf->device_lock);
5226                 }
5227
5228                 new_sector = raid5_compute_sector(conf, logical_sector,
5229                                                   previous,
5230                                                   &dd_idx, NULL);
5231                 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5232                         (unsigned long long)new_sector,
5233                         (unsigned long long)logical_sector);
5234
5235                 sh = raid5_get_active_stripe(conf, new_sector, previous,
5236                                        (bi->bi_rw & REQ_RAHEAD), 0);
5237                 if (sh) {
5238                         if (unlikely(previous)) {
5239                                 /* expansion might have moved on while waiting for a
5240                                  * stripe, so we must do the range check again.
5241                                  * Expansion could still move past after this
5242                                  * test, but as we are holding a reference to
5243                                  * 'sh', we know that if that happens,
5244                                  *  STRIPE_EXPANDING will get set and the expansion
5245                                  * won't proceed until we finish with the stripe.
5246                                  */
5247                                 int must_retry = 0;
5248                                 spin_lock_irq(&conf->device_lock);
5249                                 if (mddev->reshape_backwards
5250                                     ? logical_sector >= conf->reshape_progress
5251                                     : logical_sector < conf->reshape_progress)
5252                                         /* mismatch, need to try again */
5253                                         must_retry = 1;
5254                                 spin_unlock_irq(&conf->device_lock);
5255                                 if (must_retry) {
5256                                         raid5_release_stripe(sh);
5257                                         schedule();
5258                                         do_prepare = true;
5259                                         goto retry;
5260                                 }
5261                         }
5262                         if (read_seqcount_retry(&conf->gen_lock, seq)) {
5263                                 /* Might have got the wrong stripe_head
5264                                  * by accident
5265                                  */
5266                                 raid5_release_stripe(sh);
5267                                 goto retry;
5268                         }
5269
5270                         if (rw == WRITE &&
5271                             logical_sector >= mddev->suspend_lo &&
5272                             logical_sector < mddev->suspend_hi) {
5273                                 raid5_release_stripe(sh);
5274                                 /* As the suspend_* range is controlled by
5275                                  * userspace, we want an interruptible
5276                                  * wait.
5277                                  */
5278                                 flush_signals(current);
5279                                 prepare_to_wait(&conf->wait_for_overlap,
5280                                                 &w, TASK_INTERRUPTIBLE);
5281                                 if (logical_sector >= mddev->suspend_lo &&
5282                                     logical_sector < mddev->suspend_hi) {
5283                                         schedule();
5284                                         do_prepare = true;
5285                                 }
5286                                 goto retry;
5287                         }
5288
5289                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5290                             !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5291                                 /* Stripe is busy expanding or
5292                                  * add failed due to overlap.  Flush everything
5293                                  * and wait a while
5294                                  */
5295                                 md_wakeup_thread(mddev->thread);
5296                                 raid5_release_stripe(sh);
5297                                 schedule();
5298                                 do_prepare = true;
5299                                 goto retry;
5300                         }
5301                         set_bit(STRIPE_HANDLE, &sh->state);
5302                         clear_bit(STRIPE_DELAYED, &sh->state);
5303                         if ((!sh->batch_head || sh == sh->batch_head) &&
5304                             (bi->bi_rw & REQ_SYNC) &&
5305                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5306                                 atomic_inc(&conf->preread_active_stripes);
5307                         release_stripe_plug(mddev, sh);
5308                 } else {
5309                         /* cannot get stripe for read-ahead, just give-up */
5310                         bi->bi_error = -EIO;
5311                         break;
5312                 }
5313         }
5314         finish_wait(&conf->wait_for_overlap, &w);
5315
5316         remaining = raid5_dec_bi_active_stripes(bi);
5317         if (remaining == 0) {
5318
5319                 if ( rw == WRITE )
5320                         md_write_end(mddev);
5321
5322                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5323                                          bi, 0);
5324                 bio_endio(bi);
5325         }
5326 }
5327
5328 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5329
5330 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5331 {
5332         /* reshaping is quite different to recovery/resync so it is
5333          * handled quite separately ... here.
5334          *
5335          * On each call to sync_request, we gather one chunk worth of
5336          * destination stripes and flag them as expanding.
5337          * Then we find all the source stripes and request reads.
5338          * As the reads complete, handle_stripe will copy the data
5339          * into the destination stripe and release that stripe.
5340          */
5341         struct r5conf *conf = mddev->private;
5342         struct stripe_head *sh;
5343         sector_t first_sector, last_sector;
5344         int raid_disks = conf->previous_raid_disks;
5345         int data_disks = raid_disks - conf->max_degraded;
5346         int new_data_disks = conf->raid_disks - conf->max_degraded;
5347         int i;
5348         int dd_idx;
5349         sector_t writepos, readpos, safepos;
5350         sector_t stripe_addr;
5351         int reshape_sectors;
5352         struct list_head stripes;
5353         sector_t retn;
5354
5355         if (sector_nr == 0) {
5356                 /* If restarting in the middle, skip the initial sectors */
5357                 if (mddev->reshape_backwards &&
5358                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5359                         sector_nr = raid5_size(mddev, 0, 0)
5360                                 - conf->reshape_progress;
5361                 } else if (mddev->reshape_backwards &&
5362                            conf->reshape_progress == MaxSector) {
5363                         /* shouldn't happen, but just in case, finish up.*/
5364                         sector_nr = MaxSector;
5365                 } else if (!mddev->reshape_backwards &&
5366                            conf->reshape_progress > 0)
5367                         sector_nr = conf->reshape_progress;
5368                 sector_div(sector_nr, new_data_disks);
5369                 if (sector_nr) {
5370                         mddev->curr_resync_completed = sector_nr;
5371                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5372                         *skipped = 1;
5373                         retn = sector_nr;
5374                         goto finish;
5375                 }
5376         }
5377
5378         /* We need to process a full chunk at a time.
5379          * If old and new chunk sizes differ, we need to process the
5380          * largest of these
5381          */
5382
5383         reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5384
5385         /* We update the metadata at least every 10 seconds, or when
5386          * the data about to be copied would over-write the source of
5387          * the data at the front of the range.  i.e. one new_stripe
5388          * along from reshape_progress new_maps to after where
5389          * reshape_safe old_maps to
5390          */
5391         writepos = conf->reshape_progress;
5392         sector_div(writepos, new_data_disks);
5393         readpos = conf->reshape_progress;
5394         sector_div(readpos, data_disks);
5395         safepos = conf->reshape_safe;
5396         sector_div(safepos, data_disks);
5397         if (mddev->reshape_backwards) {
5398                 BUG_ON(writepos < reshape_sectors);
5399                 writepos -= reshape_sectors;
5400                 readpos += reshape_sectors;
5401                 safepos += reshape_sectors;
5402         } else {
5403                 writepos += reshape_sectors;
5404                 /* readpos and safepos are worst-case calculations.
5405                  * A negative number is overly pessimistic, and causes
5406                  * obvious problems for unsigned storage.  So clip to 0.
5407                  */
5408                 readpos -= min_t(sector_t, reshape_sectors, readpos);
5409                 safepos -= min_t(sector_t, reshape_sectors, safepos);
5410         }
5411
5412         /* Having calculated the 'writepos' possibly use it
5413          * to set 'stripe_addr' which is where we will write to.
5414          */
5415         if (mddev->reshape_backwards) {
5416                 BUG_ON(conf->reshape_progress == 0);
5417                 stripe_addr = writepos;
5418                 BUG_ON((mddev->dev_sectors &
5419                         ~((sector_t)reshape_sectors - 1))
5420                        - reshape_sectors - stripe_addr
5421                        != sector_nr);
5422         } else {
5423                 BUG_ON(writepos != sector_nr + reshape_sectors);
5424                 stripe_addr = sector_nr;
5425         }
5426
5427         /* 'writepos' is the most advanced device address we might write.
5428          * 'readpos' is the least advanced device address we might read.
5429          * 'safepos' is the least address recorded in the metadata as having
5430          *     been reshaped.
5431          * If there is a min_offset_diff, these are adjusted either by
5432          * increasing the safepos/readpos if diff is negative, or
5433          * increasing writepos if diff is positive.
5434          * If 'readpos' is then behind 'writepos', there is no way that we can
5435          * ensure safety in the face of a crash - that must be done by userspace
5436          * making a backup of the data.  So in that case there is no particular
5437          * rush to update metadata.
5438          * Otherwise if 'safepos' is behind 'writepos', then we really need to
5439          * update the metadata to advance 'safepos' to match 'readpos' so that
5440          * we can be safe in the event of a crash.
5441          * So we insist on updating metadata if safepos is behind writepos and
5442          * readpos is beyond writepos.
5443          * In any case, update the metadata every 10 seconds.
5444          * Maybe that number should be configurable, but I'm not sure it is
5445          * worth it.... maybe it could be a multiple of safemode_delay???
5446          */
5447         if (conf->min_offset_diff < 0) {
5448                 safepos += -conf->min_offset_diff;
5449                 readpos += -conf->min_offset_diff;
5450         } else
5451                 writepos += conf->min_offset_diff;
5452
5453         if ((mddev->reshape_backwards
5454              ? (safepos > writepos && readpos < writepos)
5455              : (safepos < writepos && readpos > writepos)) ||
5456             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5457                 /* Cannot proceed until we've updated the superblock... */
5458                 wait_event(conf->wait_for_overlap,
5459                            atomic_read(&conf->reshape_stripes)==0
5460                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5461                 if (atomic_read(&conf->reshape_stripes) != 0)
5462                         return 0;
5463                 mddev->reshape_position = conf->reshape_progress;
5464                 mddev->curr_resync_completed = sector_nr;
5465                 conf->reshape_checkpoint = jiffies;
5466                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5467                 md_wakeup_thread(mddev->thread);
5468                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5469                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5470                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5471                         return 0;
5472                 spin_lock_irq(&conf->device_lock);
5473                 conf->reshape_safe = mddev->reshape_position;
5474                 spin_unlock_irq(&conf->device_lock);
5475                 wake_up(&conf->wait_for_overlap);
5476                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5477         }
5478
5479         INIT_LIST_HEAD(&stripes);
5480         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5481                 int j;
5482                 int skipped_disk = 0;
5483                 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5484                 set_bit(STRIPE_EXPANDING, &sh->state);
5485                 atomic_inc(&conf->reshape_stripes);
5486                 /* If any of this stripe is beyond the end of the old
5487                  * array, then we need to zero those blocks
5488                  */
5489                 for (j=sh->disks; j--;) {
5490                         sector_t s;
5491                         if (j == sh->pd_idx)
5492                                 continue;
5493                         if (conf->level == 6 &&
5494                             j == sh->qd_idx)
5495                                 continue;
5496                         s = raid5_compute_blocknr(sh, j, 0);
5497                         if (s < raid5_size(mddev, 0, 0)) {
5498                                 skipped_disk = 1;
5499                                 continue;
5500                         }
5501                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5502                         set_bit(R5_Expanded, &sh->dev[j].flags);
5503                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
5504                 }
5505                 if (!skipped_disk) {
5506                         set_bit(STRIPE_EXPAND_READY, &sh->state);
5507                         set_bit(STRIPE_HANDLE, &sh->state);
5508                 }
5509                 list_add(&sh->lru, &stripes);
5510         }
5511         spin_lock_irq(&conf->device_lock);
5512         if (mddev->reshape_backwards)
5513                 conf->reshape_progress -= reshape_sectors * new_data_disks;
5514         else
5515                 conf->reshape_progress += reshape_sectors * new_data_disks;
5516         spin_unlock_irq(&conf->device_lock);
5517         /* Ok, those stripe are ready. We can start scheduling
5518          * reads on the source stripes.
5519          * The source stripes are determined by mapping the first and last
5520          * block on the destination stripes.
5521          */
5522         first_sector =
5523                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5524                                      1, &dd_idx, NULL);
5525         last_sector =
5526                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5527                                             * new_data_disks - 1),
5528                                      1, &dd_idx, NULL);
5529         if (last_sector >= mddev->dev_sectors)
5530                 last_sector = mddev->dev_sectors - 1;
5531         while (first_sector <= last_sector) {
5532                 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5533                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5534                 set_bit(STRIPE_HANDLE, &sh->state);
5535                 raid5_release_stripe(sh);
5536                 first_sector += STRIPE_SECTORS;
5537         }
5538         /* Now that the sources are clearly marked, we can release
5539          * the destination stripes
5540          */
5541         while (!list_empty(&stripes)) {
5542                 sh = list_entry(stripes.next, struct stripe_head, lru);
5543                 list_del_init(&sh->lru);
5544                 raid5_release_stripe(sh);
5545         }
5546         /* If this takes us to the resync_max point where we have to pause,
5547          * then we need to write out the superblock.
5548          */
5549         sector_nr += reshape_sectors;
5550         retn = reshape_sectors;
5551 finish:
5552         if (mddev->curr_resync_completed > mddev->resync_max ||
5553             (sector_nr - mddev->curr_resync_completed) * 2
5554             >= mddev->resync_max - mddev->curr_resync_completed) {
5555                 /* Cannot proceed until we've updated the superblock... */
5556                 wait_event(conf->wait_for_overlap,
5557                            atomic_read(&conf->reshape_stripes) == 0
5558                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5559                 if (atomic_read(&conf->reshape_stripes) != 0)
5560                         goto ret;
5561                 mddev->reshape_position = conf->reshape_progress;
5562                 mddev->curr_resync_completed = sector_nr;
5563                 conf->reshape_checkpoint = jiffies;
5564                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5565                 md_wakeup_thread(mddev->thread);
5566                 wait_event(mddev->sb_wait,
5567                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5568                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5569                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5570                         goto ret;
5571                 spin_lock_irq(&conf->device_lock);
5572                 conf->reshape_safe = mddev->reshape_position;
5573                 spin_unlock_irq(&conf->device_lock);
5574                 wake_up(&conf->wait_for_overlap);
5575                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5576         }
5577 ret:
5578         return retn;
5579 }
5580
5581 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
5582                                           int *skipped)
5583 {
5584         struct r5conf *conf = mddev->private;
5585         struct stripe_head *sh;
5586         sector_t max_sector = mddev->dev_sectors;
5587         sector_t sync_blocks;
5588         int still_degraded = 0;
5589         int i;
5590
5591         if (sector_nr >= max_sector) {
5592                 /* just being told to finish up .. nothing much to do */
5593
5594                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5595                         end_reshape(conf);
5596                         return 0;
5597                 }
5598
5599                 if (mddev->curr_resync < max_sector) /* aborted */
5600                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5601                                         &sync_blocks, 1);
5602                 else /* completed sync */
5603                         conf->fullsync = 0;
5604                 bitmap_close_sync(mddev->bitmap);
5605
5606                 return 0;
5607         }
5608
5609         /* Allow raid5_quiesce to complete */
5610         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5611
5612         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5613                 return reshape_request(mddev, sector_nr, skipped);
5614
5615         /* No need to check resync_max as we never do more than one
5616          * stripe, and as resync_max will always be on a chunk boundary,
5617          * if the check in md_do_sync didn't fire, there is no chance
5618          * of overstepping resync_max here
5619          */
5620
5621         /* if there is too many failed drives and we are trying
5622          * to resync, then assert that we are finished, because there is
5623          * nothing we can do.
5624          */
5625         if (mddev->degraded >= conf->max_degraded &&
5626             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5627                 sector_t rv = mddev->dev_sectors - sector_nr;
5628                 *skipped = 1;
5629                 return rv;
5630         }
5631         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5632             !conf->fullsync &&
5633             !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5634             sync_blocks >= STRIPE_SECTORS) {
5635                 /* we can skip this block, and probably more */
5636                 sync_blocks /= STRIPE_SECTORS;
5637                 *skipped = 1;
5638                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5639         }
5640
5641         bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
5642
5643         sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
5644         if (sh == NULL) {
5645                 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
5646                 /* make sure we don't swamp the stripe cache if someone else
5647                  * is trying to get access
5648                  */
5649                 schedule_timeout_uninterruptible(1);
5650         }
5651         /* Need to check if array will still be degraded after recovery/resync
5652          * Note in case of > 1 drive failures it's possible we're rebuilding
5653          * one drive while leaving another faulty drive in array.
5654          */
5655         rcu_read_lock();
5656         for (i = 0; i < conf->raid_disks; i++) {
5657                 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5658
5659                 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5660                         still_degraded = 1;
5661         }
5662         rcu_read_unlock();
5663
5664         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5665
5666         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5667         set_bit(STRIPE_HANDLE, &sh->state);
5668
5669         raid5_release_stripe(sh);
5670
5671         return STRIPE_SECTORS;
5672 }
5673
5674 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5675 {
5676         /* We may not be able to submit a whole bio at once as there
5677          * may not be enough stripe_heads available.
5678          * We cannot pre-allocate enough stripe_heads as we may need
5679          * more than exist in the cache (if we allow ever large chunks).
5680          * So we do one stripe head at a time and record in
5681          * ->bi_hw_segments how many have been done.
5682          *
5683          * We *know* that this entire raid_bio is in one chunk, so
5684          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5685          */
5686         struct stripe_head *sh;
5687         int dd_idx;
5688         sector_t sector, logical_sector, last_sector;
5689         int scnt = 0;
5690         int remaining;
5691         int handled = 0;
5692
5693         logical_sector = raid_bio->bi_iter.bi_sector &
5694                 ~((sector_t)STRIPE_SECTORS-1);
5695         sector = raid5_compute_sector(conf, logical_sector,
5696                                       0, &dd_idx, NULL);
5697         last_sector = bio_end_sector(raid_bio);
5698
5699         for (; logical_sector < last_sector;
5700              logical_sector += STRIPE_SECTORS,
5701                      sector += STRIPE_SECTORS,
5702                      scnt++) {
5703
5704                 if (scnt < raid5_bi_processed_stripes(raid_bio))
5705                         /* already done this stripe */
5706                         continue;
5707
5708                 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
5709
5710                 if (!sh) {
5711                         /* failed to get a stripe - must wait */
5712                         raid5_set_bi_processed_stripes(raid_bio, scnt);
5713                         conf->retry_read_aligned = raid_bio;
5714                         return handled;
5715                 }
5716
5717                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5718                         raid5_release_stripe(sh);
5719                         raid5_set_bi_processed_stripes(raid_bio, scnt);
5720                         conf->retry_read_aligned = raid_bio;
5721                         return handled;
5722                 }
5723
5724                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5725                 handle_stripe(sh);
5726                 raid5_release_stripe(sh);
5727                 handled++;
5728         }
5729         remaining = raid5_dec_bi_active_stripes(raid_bio);
5730         if (remaining == 0) {
5731                 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5732                                          raid_bio, 0);
5733                 bio_endio(raid_bio);
5734         }
5735         if (atomic_dec_and_test(&conf->active_aligned_reads))
5736                 wake_up(&conf->wait_for_quiescent);
5737         return handled;
5738 }
5739
5740 static int handle_active_stripes(struct r5conf *conf, int group,
5741                                  struct r5worker *worker,
5742                                  struct list_head *temp_inactive_list)
5743 {
5744         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5745         int i, batch_size = 0, hash;
5746         bool release_inactive = false;
5747
5748         while (batch_size < MAX_STRIPE_BATCH &&
5749                         (sh = __get_priority_stripe(conf, group)) != NULL)
5750                 batch[batch_size++] = sh;
5751
5752         if (batch_size == 0) {
5753                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5754                         if (!list_empty(temp_inactive_list + i))
5755                                 break;
5756                 if (i == NR_STRIPE_HASH_LOCKS) {
5757                         spin_unlock_irq(&conf->device_lock);
5758                         r5l_flush_stripe_to_raid(conf->log);
5759                         spin_lock_irq(&conf->device_lock);
5760                         return batch_size;
5761                 }
5762                 release_inactive = true;
5763         }
5764         spin_unlock_irq(&conf->device_lock);
5765
5766         release_inactive_stripe_list(conf, temp_inactive_list,
5767                                      NR_STRIPE_HASH_LOCKS);
5768
5769         r5l_flush_stripe_to_raid(conf->log);
5770         if (release_inactive) {
5771                 spin_lock_irq(&conf->device_lock);
5772                 return 0;
5773         }
5774
5775         for (i = 0; i < batch_size; i++)
5776                 handle_stripe(batch[i]);
5777         r5l_write_stripe_run(conf->log);
5778
5779         cond_resched();
5780
5781         spin_lock_irq(&conf->device_lock);
5782         for (i = 0; i < batch_size; i++) {
5783                 hash = batch[i]->hash_lock_index;
5784                 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5785         }
5786         return batch_size;
5787 }
5788
5789 static void raid5_do_work(struct work_struct *work)
5790 {
5791         struct r5worker *worker = container_of(work, struct r5worker, work);
5792         struct r5worker_group *group = worker->group;
5793         struct r5conf *conf = group->conf;
5794         int group_id = group - conf->worker_groups;
5795         int handled;
5796         struct blk_plug plug;
5797
5798         pr_debug("+++ raid5worker active\n");
5799
5800         blk_start_plug(&plug);
5801         handled = 0;
5802         spin_lock_irq(&conf->device_lock);
5803         while (1) {
5804                 int batch_size, released;
5805
5806                 released = release_stripe_list(conf, worker->temp_inactive_list);
5807
5808                 batch_size = handle_active_stripes(conf, group_id, worker,
5809                                                    worker->temp_inactive_list);
5810                 worker->working = false;
5811                 if (!batch_size && !released)
5812                         break;
5813                 handled += batch_size;
5814         }
5815         pr_debug("%d stripes handled\n", handled);
5816
5817         spin_unlock_irq(&conf->device_lock);
5818         blk_finish_plug(&plug);
5819
5820         pr_debug("--- raid5worker inactive\n");
5821 }
5822
5823 /*
5824  * This is our raid5 kernel thread.
5825  *
5826  * We scan the hash table for stripes which can be handled now.
5827  * During the scan, completed stripes are saved for us by the interrupt
5828  * handler, so that they will not have to wait for our next wakeup.
5829  */
5830 static void raid5d(struct md_thread *thread)
5831 {
5832         struct mddev *mddev = thread->mddev;
5833         struct r5conf *conf = mddev->private;
5834         int handled;
5835         struct blk_plug plug;
5836
5837         pr_debug("+++ raid5d active\n");
5838
5839         md_check_recovery(mddev);
5840
5841         if (!bio_list_empty(&conf->return_bi) &&
5842             !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5843                 struct bio_list tmp = BIO_EMPTY_LIST;
5844                 spin_lock_irq(&conf->device_lock);
5845                 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5846                         bio_list_merge(&tmp, &conf->return_bi);
5847                         bio_list_init(&conf->return_bi);
5848                 }
5849                 spin_unlock_irq(&conf->device_lock);
5850                 return_io(&tmp);
5851         }
5852
5853         blk_start_plug(&plug);
5854         handled = 0;
5855         spin_lock_irq(&conf->device_lock);
5856         while (1) {
5857                 struct bio *bio;
5858                 int batch_size, released;
5859
5860                 released = release_stripe_list(conf, conf->temp_inactive_list);
5861                 if (released)
5862                         clear_bit(R5_DID_ALLOC, &conf->cache_state);
5863
5864                 if (
5865                     !list_empty(&conf->bitmap_list)) {
5866                         /* Now is a good time to flush some bitmap updates */
5867                         conf->seq_flush++;
5868                         spin_unlock_irq(&conf->device_lock);
5869                         bitmap_unplug(mddev->bitmap);
5870                         spin_lock_irq(&conf->device_lock);
5871                         conf->seq_write = conf->seq_flush;
5872                         activate_bit_delay(conf, conf->temp_inactive_list);
5873                 }
5874                 raid5_activate_delayed(conf);
5875
5876                 while ((bio = remove_bio_from_retry(conf))) {
5877                         int ok;
5878                         spin_unlock_irq(&conf->device_lock);
5879                         ok = retry_aligned_read(conf, bio);
5880                         spin_lock_irq(&conf->device_lock);
5881                         if (!ok)
5882                                 break;
5883                         handled++;
5884                 }
5885
5886                 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5887                                                    conf->temp_inactive_list);
5888                 if (!batch_size && !released)
5889                         break;
5890                 handled += batch_size;
5891
5892                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5893                         spin_unlock_irq(&conf->device_lock);
5894                         md_check_recovery(mddev);
5895                         spin_lock_irq(&conf->device_lock);
5896                 }
5897         }
5898         pr_debug("%d stripes handled\n", handled);
5899
5900         spin_unlock_irq(&conf->device_lock);
5901         if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
5902             mutex_trylock(&conf->cache_size_mutex)) {
5903                 grow_one_stripe(conf, __GFP_NOWARN);
5904                 /* Set flag even if allocation failed.  This helps
5905                  * slow down allocation requests when mem is short
5906                  */
5907                 set_bit(R5_DID_ALLOC, &conf->cache_state);
5908                 mutex_unlock(&conf->cache_size_mutex);
5909         }
5910
5911         r5l_flush_stripe_to_raid(conf->log);
5912
5913         async_tx_issue_pending_all();
5914         blk_finish_plug(&plug);
5915
5916         pr_debug("--- raid5d inactive\n");
5917 }
5918
5919 static ssize_t
5920 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5921 {
5922         struct r5conf *conf;
5923         int ret = 0;
5924         spin_lock(&mddev->lock);
5925         conf = mddev->private;
5926         if (conf)
5927                 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
5928         spin_unlock(&mddev->lock);
5929         return ret;
5930 }
5931
5932 int
5933 raid5_set_cache_size(struct mddev *mddev, int size)
5934 {
5935         struct r5conf *conf = mddev->private;
5936         int err;
5937
5938         if (size <= 16 || size > 32768)
5939                 return -EINVAL;
5940
5941         conf->min_nr_stripes = size;
5942         mutex_lock(&conf->cache_size_mutex);
5943         while (size < conf->max_nr_stripes &&
5944                drop_one_stripe(conf))
5945                 ;
5946         mutex_unlock(&conf->cache_size_mutex);
5947
5948
5949         err = md_allow_write(mddev);
5950         if (err)
5951                 return err;
5952
5953         mutex_lock(&conf->cache_size_mutex);
5954         while (size > conf->max_nr_stripes)
5955                 if (!grow_one_stripe(conf, GFP_KERNEL))
5956                         break;
5957         mutex_unlock(&conf->cache_size_mutex);
5958
5959         return 0;
5960 }
5961 EXPORT_SYMBOL(raid5_set_cache_size);
5962
5963 static ssize_t
5964 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5965 {
5966         struct r5conf *conf;
5967         unsigned long new;
5968         int err;
5969
5970         if (len >= PAGE_SIZE)
5971                 return -EINVAL;
5972         if (kstrtoul(page, 10, &new))
5973                 return -EINVAL;
5974         err = mddev_lock(mddev);
5975         if (err)
5976                 return err;
5977         conf = mddev->private;
5978         if (!conf)
5979                 err = -ENODEV;
5980         else
5981                 err = raid5_set_cache_size(mddev, new);
5982         mddev_unlock(mddev);
5983
5984         return err ?: len;
5985 }
5986
5987 static struct md_sysfs_entry
5988 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5989                                 raid5_show_stripe_cache_size,
5990                                 raid5_store_stripe_cache_size);
5991
5992 static ssize_t
5993 raid5_show_rmw_level(struct mddev  *mddev, char *page)
5994 {
5995         struct r5conf *conf = mddev->private;
5996         if (conf)
5997                 return sprintf(page, "%d\n", conf->rmw_level);
5998         else
5999                 return 0;
6000 }
6001
6002 static ssize_t
6003 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6004 {
6005         struct r5conf *conf = mddev->private;
6006         unsigned long new;
6007
6008         if (!conf)
6009                 return -ENODEV;
6010
6011         if (len >= PAGE_SIZE)
6012                 return -EINVAL;
6013
6014         if (kstrtoul(page, 10, &new))
6015                 return -EINVAL;
6016
6017         if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6018                 return -EINVAL;
6019
6020         if (new != PARITY_DISABLE_RMW &&
6021             new != PARITY_ENABLE_RMW &&
6022             new != PARITY_PREFER_RMW)
6023                 return -EINVAL;
6024
6025         conf->rmw_level = new;
6026         return len;
6027 }
6028
6029 static struct md_sysfs_entry
6030 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6031                          raid5_show_rmw_level,
6032                          raid5_store_rmw_level);
6033
6034
6035 static ssize_t
6036 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6037 {
6038         struct r5conf *conf;
6039         int ret = 0;
6040         spin_lock(&mddev->lock);
6041         conf = mddev->private;
6042         if (conf)
6043                 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6044         spin_unlock(&mddev->lock);
6045         return ret;
6046 }
6047
6048 static ssize_t
6049 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6050 {
6051         struct r5conf *conf;
6052         unsigned long new;
6053         int err;
6054
6055         if (len >= PAGE_SIZE)
6056                 return -EINVAL;
6057         if (kstrtoul(page, 10, &new))
6058                 return -EINVAL;
6059
6060         err = mddev_lock(mddev);
6061         if (err)
6062                 return err;
6063         conf = mddev->private;
6064         if (!conf)
6065                 err = -ENODEV;
6066         else if (new > conf->min_nr_stripes)
6067                 err = -EINVAL;
6068         else
6069                 conf->bypass_threshold = new;
6070         mddev_unlock(mddev);
6071         return err ?: len;
6072 }
6073
6074 static struct md_sysfs_entry
6075 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6076                                         S_IRUGO | S_IWUSR,
6077                                         raid5_show_preread_threshold,
6078                                         raid5_store_preread_threshold);
6079
6080 static ssize_t
6081 raid5_show_skip_copy(struct mddev *mddev, char *page)
6082 {
6083         struct r5conf *conf;
6084         int ret = 0;
6085         spin_lock(&mddev->lock);
6086         conf = mddev->private;
6087         if (conf)
6088                 ret = sprintf(page, "%d\n", conf->skip_copy);
6089         spin_unlock(&mddev->lock);
6090         return ret;
6091 }
6092
6093 static ssize_t
6094 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6095 {
6096         struct r5conf *conf;
6097         unsigned long new;
6098         int err;
6099
6100         if (len >= PAGE_SIZE)
6101                 return -EINVAL;
6102         if (kstrtoul(page, 10, &new))
6103                 return -EINVAL;
6104         new = !!new;
6105
6106         err = mddev_lock(mddev);
6107         if (err)
6108                 return err;
6109         conf = mddev->private;
6110         if (!conf)
6111                 err = -ENODEV;
6112         else if (new != conf->skip_copy) {
6113                 mddev_suspend(mddev);
6114                 conf->skip_copy = new;
6115                 if (new)
6116                         mddev->queue->backing_dev_info.capabilities |=
6117                                 BDI_CAP_STABLE_WRITES;
6118                 else
6119                         mddev->queue->backing_dev_info.capabilities &=
6120                                 ~BDI_CAP_STABLE_WRITES;
6121                 mddev_resume(mddev);
6122         }
6123         mddev_unlock(mddev);
6124         return err ?: len;
6125 }
6126
6127 static struct md_sysfs_entry
6128 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6129                                         raid5_show_skip_copy,
6130                                         raid5_store_skip_copy);
6131
6132 static ssize_t
6133 stripe_cache_active_show(struct mddev *mddev, char *page)
6134 {
6135         struct r5conf *conf = mddev->private;
6136         if (conf)
6137                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6138         else
6139                 return 0;
6140 }
6141
6142 static struct md_sysfs_entry
6143 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6144
6145 static ssize_t
6146 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6147 {
6148         struct r5conf *conf;
6149         int ret = 0;
6150         spin_lock(&mddev->lock);
6151         conf = mddev->private;
6152         if (conf)
6153                 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6154         spin_unlock(&mddev->lock);
6155         return ret;
6156 }
6157
6158 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6159                                int *group_cnt,
6160                                int *worker_cnt_per_group,
6161                                struct r5worker_group **worker_groups);
6162 static ssize_t
6163 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6164 {
6165         struct r5conf *conf;
6166         unsigned long new;
6167         int err;
6168         struct r5worker_group *new_groups, *old_groups;
6169         int group_cnt, worker_cnt_per_group;
6170
6171         if (len >= PAGE_SIZE)
6172                 return -EINVAL;
6173         if (kstrtoul(page, 10, &new))
6174                 return -EINVAL;
6175
6176         err = mddev_lock(mddev);
6177         if (err)
6178                 return err;
6179         conf = mddev->private;
6180         if (!conf)
6181                 err = -ENODEV;
6182         else if (new != conf->worker_cnt_per_group) {
6183                 mddev_suspend(mddev);
6184
6185                 old_groups = conf->worker_groups;
6186                 if (old_groups)
6187                         flush_workqueue(raid5_wq);
6188
6189                 err = alloc_thread_groups(conf, new,
6190                                           &group_cnt, &worker_cnt_per_group,
6191                                           &new_groups);
6192                 if (!err) {
6193                         spin_lock_irq(&conf->device_lock);
6194                         conf->group_cnt = group_cnt;
6195                         conf->worker_cnt_per_group = worker_cnt_per_group;
6196                         conf->worker_groups = new_groups;
6197                         spin_unlock_irq(&conf->device_lock);
6198
6199                         if (old_groups)
6200                                 kfree(old_groups[0].workers);
6201                         kfree(old_groups);
6202                 }
6203                 mddev_resume(mddev);
6204         }
6205         mddev_unlock(mddev);
6206
6207         return err ?: len;
6208 }
6209
6210 static struct md_sysfs_entry
6211 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6212                                 raid5_show_group_thread_cnt,
6213                                 raid5_store_group_thread_cnt);
6214
6215 static struct attribute *raid5_attrs[] =  {
6216         &raid5_stripecache_size.attr,
6217         &raid5_stripecache_active.attr,
6218         &raid5_preread_bypass_threshold.attr,
6219         &raid5_group_thread_cnt.attr,
6220         &raid5_skip_copy.attr,
6221         &raid5_rmw_level.attr,
6222         NULL,
6223 };
6224 static struct attribute_group raid5_attrs_group = {
6225         .name = NULL,
6226         .attrs = raid5_attrs,
6227 };
6228
6229 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6230                                int *group_cnt,
6231                                int *worker_cnt_per_group,
6232                                struct r5worker_group **worker_groups)
6233 {
6234         int i, j, k;
6235         ssize_t size;
6236         struct r5worker *workers;
6237
6238         *worker_cnt_per_group = cnt;
6239         if (cnt == 0) {
6240                 *group_cnt = 0;
6241                 *worker_groups = NULL;
6242                 return 0;
6243         }
6244         *group_cnt = num_possible_nodes();
6245         size = sizeof(struct r5worker) * cnt;
6246         workers = kzalloc(size * *group_cnt, GFP_NOIO);
6247         *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6248                                 *group_cnt, GFP_NOIO);
6249         if (!*worker_groups || !workers) {
6250                 kfree(workers);
6251                 kfree(*worker_groups);
6252                 return -ENOMEM;
6253         }
6254
6255         for (i = 0; i < *group_cnt; i++) {
6256                 struct r5worker_group *group;
6257
6258                 group = &(*worker_groups)[i];
6259                 INIT_LIST_HEAD(&group->handle_list);
6260                 group->conf = conf;
6261                 group->workers = workers + i * cnt;
6262
6263                 for (j = 0; j < cnt; j++) {
6264                         struct r5worker *worker = group->workers + j;
6265                         worker->group = group;
6266                         INIT_WORK(&worker->work, raid5_do_work);
6267
6268                         for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6269                                 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6270                 }
6271         }
6272
6273         return 0;
6274 }
6275
6276 static void free_thread_groups(struct r5conf *conf)
6277 {
6278         if (conf->worker_groups)
6279                 kfree(conf->worker_groups[0].workers);
6280         kfree(conf->worker_groups);
6281         conf->worker_groups = NULL;
6282 }
6283
6284 static sector_t
6285 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6286 {
6287         struct r5conf *conf = mddev->private;
6288
6289         if (!sectors)
6290                 sectors = mddev->dev_sectors;
6291         if (!raid_disks)
6292                 /* size is defined by the smallest of previous and new size */
6293                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6294
6295         sectors &= ~((sector_t)conf->chunk_sectors - 1);
6296         sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6297         return sectors * (raid_disks - conf->max_degraded);
6298 }
6299
6300 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6301 {
6302         safe_put_page(percpu->spare_page);
6303         if (percpu->scribble)
6304                 flex_array_free(percpu->scribble);
6305         percpu->spare_page = NULL;
6306         percpu->scribble = NULL;
6307 }
6308
6309 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6310 {
6311         if (conf->level == 6 && !percpu->spare_page)
6312                 percpu->spare_page = alloc_page(GFP_KERNEL);
6313         if (!percpu->scribble)
6314                 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6315                                                       conf->previous_raid_disks),
6316                                                   max(conf->chunk_sectors,
6317                                                       conf->prev_chunk_sectors)
6318                                                    / STRIPE_SECTORS,
6319                                                   GFP_KERNEL);
6320
6321         if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6322                 free_scratch_buffer(conf, percpu);
6323                 return -ENOMEM;
6324         }
6325
6326         return 0;
6327 }
6328
6329 static void raid5_free_percpu(struct r5conf *conf)
6330 {
6331         unsigned long cpu;
6332
6333         if (!conf->percpu)
6334                 return;
6335
6336 #ifdef CONFIG_HOTPLUG_CPU
6337         unregister_cpu_notifier(&conf->cpu_notify);
6338 #endif
6339
6340         get_online_cpus();
6341         for_each_possible_cpu(cpu)
6342                 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6343         put_online_cpus();
6344
6345         free_percpu(conf->percpu);
6346 }
6347
6348 static void free_conf(struct r5conf *conf)
6349 {
6350         if (conf->log)
6351                 r5l_exit_log(conf->log);
6352         if (conf->shrinker.seeks)
6353                 unregister_shrinker(&conf->shrinker);
6354
6355         free_thread_groups(conf);
6356         shrink_stripes(conf);
6357         raid5_free_percpu(conf);
6358         kfree(conf->disks);
6359         kfree(conf->stripe_hashtbl);
6360         kfree(conf);
6361 }
6362
6363 #ifdef CONFIG_HOTPLUG_CPU
6364 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
6365                               void *hcpu)
6366 {
6367         struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
6368         long cpu = (long)hcpu;
6369         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6370
6371         switch (action) {
6372         case CPU_UP_PREPARE:
6373         case CPU_UP_PREPARE_FROZEN:
6374                 if (alloc_scratch_buffer(conf, percpu)) {
6375                         pr_err("%s: failed memory allocation for cpu%ld\n",
6376                                __func__, cpu);
6377                         return notifier_from_errno(-ENOMEM);
6378                 }
6379                 break;
6380         case CPU_DEAD:
6381         case CPU_DEAD_FROZEN:
6382         case CPU_UP_CANCELED:
6383         case CPU_UP_CANCELED_FROZEN:
6384                 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6385                 break;
6386         default:
6387                 break;
6388         }
6389         return NOTIFY_OK;
6390 }
6391 #endif
6392
6393 static int raid5_alloc_percpu(struct r5conf *conf)
6394 {
6395         unsigned long cpu;
6396         int err = 0;
6397
6398         conf->percpu = alloc_percpu(struct raid5_percpu);
6399         if (!conf->percpu)
6400                 return -ENOMEM;
6401
6402 #ifdef CONFIG_HOTPLUG_CPU
6403         conf->cpu_notify.notifier_call = raid456_cpu_notify;
6404         conf->cpu_notify.priority = 0;
6405         err = register_cpu_notifier(&conf->cpu_notify);
6406         if (err)
6407                 return err;
6408 #endif
6409
6410         get_online_cpus();
6411         for_each_present_cpu(cpu) {
6412                 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6413                 if (err) {
6414                         pr_err("%s: failed memory allocation for cpu%ld\n",
6415                                __func__, cpu);
6416                         break;
6417                 }
6418         }
6419         put_online_cpus();
6420
6421         if (!err) {
6422                 conf->scribble_disks = max(conf->raid_disks,
6423                         conf->previous_raid_disks);
6424                 conf->scribble_sectors = max(conf->chunk_sectors,
6425                         conf->prev_chunk_sectors);
6426         }
6427         return err;
6428 }
6429
6430 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6431                                       struct shrink_control *sc)
6432 {
6433         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6434         unsigned long ret = SHRINK_STOP;
6435
6436         if (mutex_trylock(&conf->cache_size_mutex)) {
6437                 ret= 0;
6438                 while (ret < sc->nr_to_scan &&
6439                        conf->max_nr_stripes > conf->min_nr_stripes) {
6440                         if (drop_one_stripe(conf) == 0) {
6441                                 ret = SHRINK_STOP;
6442                                 break;
6443                         }
6444                         ret++;
6445                 }
6446                 mutex_unlock(&conf->cache_size_mutex);
6447         }
6448         return ret;
6449 }
6450
6451 static unsigned long raid5_cache_count(struct shrinker *shrink,
6452                                        struct shrink_control *sc)
6453 {
6454         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6455
6456         if (conf->max_nr_stripes < conf->min_nr_stripes)
6457                 /* unlikely, but not impossible */
6458                 return 0;
6459         return conf->max_nr_stripes - conf->min_nr_stripes;
6460 }
6461
6462 static struct r5conf *setup_conf(struct mddev *mddev)
6463 {
6464         struct r5conf *conf;
6465         int raid_disk, memory, max_disks;
6466         struct md_rdev *rdev;
6467         struct disk_info *disk;
6468         char pers_name[6];
6469         int i;
6470         int group_cnt, worker_cnt_per_group;
6471         struct r5worker_group *new_group;
6472
6473         if (mddev->new_level != 5
6474             && mddev->new_level != 4
6475             && mddev->new_level != 6) {
6476                 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6477                        mdname(mddev), mddev->new_level);
6478                 return ERR_PTR(-EIO);
6479         }
6480         if ((mddev->new_level == 5
6481              && !algorithm_valid_raid5(mddev->new_layout)) ||
6482             (mddev->new_level == 6
6483              && !algorithm_valid_raid6(mddev->new_layout))) {
6484                 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6485                        mdname(mddev), mddev->new_layout);
6486                 return ERR_PTR(-EIO);
6487         }
6488         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6489                 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6490                        mdname(mddev), mddev->raid_disks);
6491                 return ERR_PTR(-EINVAL);
6492         }
6493
6494         if (!mddev->new_chunk_sectors ||
6495             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6496             !is_power_of_2(mddev->new_chunk_sectors)) {
6497                 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6498                        mdname(mddev), mddev->new_chunk_sectors << 9);
6499                 return ERR_PTR(-EINVAL);
6500         }
6501
6502         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6503         if (conf == NULL)
6504                 goto abort;
6505         /* Don't enable multi-threading by default*/
6506         if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6507                                  &new_group)) {
6508                 conf->group_cnt = group_cnt;
6509                 conf->worker_cnt_per_group = worker_cnt_per_group;
6510                 conf->worker_groups = new_group;
6511         } else
6512                 goto abort;
6513         spin_lock_init(&conf->device_lock);
6514         seqcount_init(&conf->gen_lock);
6515         mutex_init(&conf->cache_size_mutex);
6516         init_waitqueue_head(&conf->wait_for_quiescent);
6517         init_waitqueue_head(&conf->wait_for_stripe);
6518         init_waitqueue_head(&conf->wait_for_overlap);
6519         INIT_LIST_HEAD(&conf->handle_list);
6520         INIT_LIST_HEAD(&conf->hold_list);
6521         INIT_LIST_HEAD(&conf->delayed_list);
6522         INIT_LIST_HEAD(&conf->bitmap_list);
6523         bio_list_init(&conf->return_bi);
6524         init_llist_head(&conf->released_stripes);
6525         atomic_set(&conf->active_stripes, 0);
6526         atomic_set(&conf->preread_active_stripes, 0);
6527         atomic_set(&conf->active_aligned_reads, 0);
6528         conf->bypass_threshold = BYPASS_THRESHOLD;
6529         conf->recovery_disabled = mddev->recovery_disabled - 1;
6530
6531         conf->raid_disks = mddev->raid_disks;
6532         if (mddev->reshape_position == MaxSector)
6533                 conf->previous_raid_disks = mddev->raid_disks;
6534         else
6535                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6536         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6537
6538         conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6539                               GFP_KERNEL);
6540         if (!conf->disks)
6541                 goto abort;
6542
6543         conf->mddev = mddev;
6544
6545         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6546                 goto abort;
6547
6548         /* We init hash_locks[0] separately to that it can be used
6549          * as the reference lock in the spin_lock_nest_lock() call
6550          * in lock_all_device_hash_locks_irq in order to convince
6551          * lockdep that we know what we are doing.
6552          */
6553         spin_lock_init(conf->hash_locks);
6554         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6555                 spin_lock_init(conf->hash_locks + i);
6556
6557         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6558                 INIT_LIST_HEAD(conf->inactive_list + i);
6559
6560         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6561                 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6562
6563         conf->level = mddev->new_level;
6564         conf->chunk_sectors = mddev->new_chunk_sectors;
6565         if (raid5_alloc_percpu(conf) != 0)
6566                 goto abort;
6567
6568         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6569
6570         rdev_for_each(rdev, mddev) {
6571                 raid_disk = rdev->raid_disk;
6572                 if (raid_disk >= max_disks
6573                     || raid_disk < 0 || test_bit(Journal, &rdev->flags))
6574                         continue;
6575                 disk = conf->disks + raid_disk;
6576
6577                 if (test_bit(Replacement, &rdev->flags)) {
6578                         if (disk->replacement)
6579                                 goto abort;
6580                         disk->replacement = rdev;
6581                 } else {
6582                         if (disk->rdev)
6583                                 goto abort;
6584                         disk->rdev = rdev;
6585                 }
6586
6587                 if (test_bit(In_sync, &rdev->flags)) {
6588                         char b[BDEVNAME_SIZE];
6589                         printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6590                                " disk %d\n",
6591                                mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6592                 } else if (rdev->saved_raid_disk != raid_disk)
6593                         /* Cannot rely on bitmap to complete recovery */
6594                         conf->fullsync = 1;
6595         }
6596
6597         conf->level = mddev->new_level;
6598         if (conf->level == 6) {
6599                 conf->max_degraded = 2;
6600                 if (raid6_call.xor_syndrome)
6601                         conf->rmw_level = PARITY_ENABLE_RMW;
6602                 else
6603                         conf->rmw_level = PARITY_DISABLE_RMW;
6604         } else {
6605                 conf->max_degraded = 1;
6606                 conf->rmw_level = PARITY_ENABLE_RMW;
6607         }
6608         conf->algorithm = mddev->new_layout;
6609         conf->reshape_progress = mddev->reshape_position;
6610         if (conf->reshape_progress != MaxSector) {
6611                 conf->prev_chunk_sectors = mddev->chunk_sectors;
6612                 conf->prev_algo = mddev->layout;
6613         } else {
6614                 conf->prev_chunk_sectors = conf->chunk_sectors;
6615                 conf->prev_algo = conf->algorithm;
6616         }
6617
6618         conf->min_nr_stripes = NR_STRIPES;
6619         memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6620                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6621         atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6622         if (grow_stripes(conf, conf->min_nr_stripes)) {
6623                 printk(KERN_ERR
6624                        "md/raid:%s: couldn't allocate %dkB for buffers\n",
6625                        mdname(mddev), memory);
6626                 goto abort;
6627         } else
6628                 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6629                        mdname(mddev), memory);
6630         /*
6631          * Losing a stripe head costs more than the time to refill it,
6632          * it reduces the queue depth and so can hurt throughput.
6633          * So set it rather large, scaled by number of devices.
6634          */
6635         conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6636         conf->shrinker.scan_objects = raid5_cache_scan;
6637         conf->shrinker.count_objects = raid5_cache_count;
6638         conf->shrinker.batch = 128;
6639         conf->shrinker.flags = 0;
6640         register_shrinker(&conf->shrinker);
6641
6642         sprintf(pers_name, "raid%d", mddev->new_level);
6643         conf->thread = md_register_thread(raid5d, mddev, pers_name);
6644         if (!conf->thread) {
6645                 printk(KERN_ERR
6646                        "md/raid:%s: couldn't allocate thread.\n",
6647                        mdname(mddev));
6648                 goto abort;
6649         }
6650
6651         return conf;
6652
6653  abort:
6654         if (conf) {
6655                 free_conf(conf);
6656                 return ERR_PTR(-EIO);
6657         } else
6658                 return ERR_PTR(-ENOMEM);
6659 }
6660
6661 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6662 {
6663         switch (algo) {
6664         case ALGORITHM_PARITY_0:
6665                 if (raid_disk < max_degraded)
6666                         return 1;
6667                 break;
6668         case ALGORITHM_PARITY_N:
6669                 if (raid_disk >= raid_disks - max_degraded)
6670                         return 1;
6671                 break;
6672         case ALGORITHM_PARITY_0_6:
6673                 if (raid_disk == 0 ||
6674                     raid_disk == raid_disks - 1)
6675                         return 1;
6676                 break;
6677         case ALGORITHM_LEFT_ASYMMETRIC_6:
6678         case ALGORITHM_RIGHT_ASYMMETRIC_6:
6679         case ALGORITHM_LEFT_SYMMETRIC_6:
6680         case ALGORITHM_RIGHT_SYMMETRIC_6:
6681                 if (raid_disk == raid_disks - 1)
6682                         return 1;
6683         }
6684         return 0;
6685 }
6686
6687 static int raid5_run(struct mddev *mddev)
6688 {
6689         struct r5conf *conf;
6690         int working_disks = 0;
6691         int dirty_parity_disks = 0;
6692         struct md_rdev *rdev;
6693         struct md_rdev *journal_dev = NULL;
6694         sector_t reshape_offset = 0;
6695         int i;
6696         long long min_offset_diff = 0;
6697         int first = 1;
6698
6699         if (mddev->recovery_cp != MaxSector)
6700                 printk(KERN_NOTICE "md/raid:%s: not clean"
6701                        " -- starting background reconstruction\n",
6702                        mdname(mddev));
6703
6704         rdev_for_each(rdev, mddev) {
6705                 long long diff;
6706
6707                 if (test_bit(Journal, &rdev->flags)) {
6708                         journal_dev = rdev;
6709                         continue;
6710                 }
6711                 if (rdev->raid_disk < 0)
6712                         continue;
6713                 diff = (rdev->new_data_offset - rdev->data_offset);
6714                 if (first) {
6715                         min_offset_diff = diff;
6716                         first = 0;
6717                 } else if (mddev->reshape_backwards &&
6718                          diff < min_offset_diff)
6719                         min_offset_diff = diff;
6720                 else if (!mddev->reshape_backwards &&
6721                          diff > min_offset_diff)
6722                         min_offset_diff = diff;
6723         }
6724
6725         if (mddev->reshape_position != MaxSector) {
6726                 /* Check that we can continue the reshape.
6727                  * Difficulties arise if the stripe we would write to
6728                  * next is at or after the stripe we would read from next.
6729                  * For a reshape that changes the number of devices, this
6730                  * is only possible for a very short time, and mdadm makes
6731                  * sure that time appears to have past before assembling
6732                  * the array.  So we fail if that time hasn't passed.
6733                  * For a reshape that keeps the number of devices the same
6734                  * mdadm must be monitoring the reshape can keeping the
6735                  * critical areas read-only and backed up.  It will start
6736                  * the array in read-only mode, so we check for that.
6737                  */
6738                 sector_t here_new, here_old;
6739                 int old_disks;
6740                 int max_degraded = (mddev->level == 6 ? 2 : 1);
6741                 int chunk_sectors;
6742                 int new_data_disks;
6743
6744                 if (journal_dev) {
6745                         printk(KERN_ERR "md/raid:%s: don't support reshape with journal - aborting.\n",
6746                                mdname(mddev));
6747                         return -EINVAL;
6748                 }
6749
6750                 if (mddev->new_level != mddev->level) {
6751                         printk(KERN_ERR "md/raid:%s: unsupported reshape "
6752                                "required - aborting.\n",
6753                                mdname(mddev));
6754                         return -EINVAL;
6755                 }
6756                 old_disks = mddev->raid_disks - mddev->delta_disks;
6757                 /* reshape_position must be on a new-stripe boundary, and one
6758                  * further up in new geometry must map after here in old
6759                  * geometry.
6760                  * If the chunk sizes are different, then as we perform reshape
6761                  * in units of the largest of the two, reshape_position needs
6762                  * be a multiple of the largest chunk size times new data disks.
6763                  */
6764                 here_new = mddev->reshape_position;
6765                 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
6766                 new_data_disks = mddev->raid_disks - max_degraded;
6767                 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
6768                         printk(KERN_ERR "md/raid:%s: reshape_position not "
6769                                "on a stripe boundary\n", mdname(mddev));
6770                         return -EINVAL;
6771                 }
6772                 reshape_offset = here_new * chunk_sectors;
6773                 /* here_new is the stripe we will write to */
6774                 here_old = mddev->reshape_position;
6775                 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
6776                 /* here_old is the first stripe that we might need to read
6777                  * from */
6778                 if (mddev->delta_disks == 0) {
6779                         /* We cannot be sure it is safe to start an in-place
6780                          * reshape.  It is only safe if user-space is monitoring
6781                          * and taking constant backups.
6782                          * mdadm always starts a situation like this in
6783                          * readonly mode so it can take control before
6784                          * allowing any writes.  So just check for that.
6785                          */
6786                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6787                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
6788                                 /* not really in-place - so OK */;
6789                         else if (mddev->ro == 0) {
6790                                 printk(KERN_ERR "md/raid:%s: in-place reshape "
6791                                        "must be started in read-only mode "
6792                                        "- aborting\n",
6793                                        mdname(mddev));
6794                                 return -EINVAL;
6795                         }
6796                 } else if (mddev->reshape_backwards
6797                     ? (here_new * chunk_sectors + min_offset_diff <=
6798                        here_old * chunk_sectors)
6799                     : (here_new * chunk_sectors >=
6800                        here_old * chunk_sectors + (-min_offset_diff))) {
6801                         /* Reading from the same stripe as writing to - bad */
6802                         printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6803                                "auto-recovery - aborting.\n",
6804                                mdname(mddev));
6805                         return -EINVAL;
6806                 }
6807                 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6808                        mdname(mddev));
6809                 /* OK, we should be able to continue; */
6810         } else {
6811                 BUG_ON(mddev->level != mddev->new_level);
6812                 BUG_ON(mddev->layout != mddev->new_layout);
6813                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6814                 BUG_ON(mddev->delta_disks != 0);
6815         }
6816
6817         if (mddev->private == NULL)
6818                 conf = setup_conf(mddev);
6819         else
6820                 conf = mddev->private;
6821
6822         if (IS_ERR(conf))
6823                 return PTR_ERR(conf);
6824
6825         if (test_bit(MD_HAS_JOURNAL, &mddev->flags) && !journal_dev) {
6826                 printk(KERN_ERR "md/raid:%s: journal disk is missing, force array readonly\n",
6827                        mdname(mddev));
6828                 mddev->ro = 1;
6829                 set_disk_ro(mddev->gendisk, 1);
6830         }
6831
6832         conf->min_offset_diff = min_offset_diff;
6833         mddev->thread = conf->thread;
6834         conf->thread = NULL;
6835         mddev->private = conf;
6836
6837         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6838              i++) {
6839                 rdev = conf->disks[i].rdev;
6840                 if (!rdev && conf->disks[i].replacement) {
6841                         /* The replacement is all we have yet */
6842                         rdev = conf->disks[i].replacement;
6843                         conf->disks[i].replacement = NULL;
6844                         clear_bit(Replacement, &rdev->flags);
6845                         conf->disks[i].rdev = rdev;
6846                 }
6847                 if (!rdev)
6848                         continue;
6849                 if (conf->disks[i].replacement &&
6850                     conf->reshape_progress != MaxSector) {
6851                         /* replacements and reshape simply do not mix. */
6852                         printk(KERN_ERR "md: cannot handle concurrent "
6853                                "replacement and reshape.\n");
6854                         goto abort;
6855                 }
6856                 if (test_bit(In_sync, &rdev->flags)) {
6857                         working_disks++;
6858                         continue;
6859                 }
6860                 /* This disc is not fully in-sync.  However if it
6861                  * just stored parity (beyond the recovery_offset),
6862                  * when we don't need to be concerned about the
6863                  * array being dirty.
6864                  * When reshape goes 'backwards', we never have
6865                  * partially completed devices, so we only need
6866                  * to worry about reshape going forwards.
6867                  */
6868                 /* Hack because v0.91 doesn't store recovery_offset properly. */
6869                 if (mddev->major_version == 0 &&
6870                     mddev->minor_version > 90)
6871                         rdev->recovery_offset = reshape_offset;
6872
6873                 if (rdev->recovery_offset < reshape_offset) {
6874                         /* We need to check old and new layout */
6875                         if (!only_parity(rdev->raid_disk,
6876                                          conf->algorithm,
6877                                          conf->raid_disks,
6878                                          conf->max_degraded))
6879                                 continue;
6880                 }
6881                 if (!only_parity(rdev->raid_disk,
6882                                  conf->prev_algo,
6883                                  conf->previous_raid_disks,
6884                                  conf->max_degraded))
6885                         continue;
6886                 dirty_parity_disks++;
6887         }
6888
6889         /*
6890          * 0 for a fully functional array, 1 or 2 for a degraded array.
6891          */
6892         mddev->degraded = calc_degraded(conf);
6893
6894         if (has_failed(conf)) {
6895                 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6896                         " (%d/%d failed)\n",
6897                         mdname(mddev), mddev->degraded, conf->raid_disks);
6898                 goto abort;
6899         }
6900
6901         /* device size must be a multiple of chunk size */
6902         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6903         mddev->resync_max_sectors = mddev->dev_sectors;
6904
6905         if (mddev->degraded > dirty_parity_disks &&
6906             mddev->recovery_cp != MaxSector) {
6907                 if (mddev->ok_start_degraded)
6908                         printk(KERN_WARNING
6909                                "md/raid:%s: starting dirty degraded array"
6910                                " - data corruption possible.\n",
6911                                mdname(mddev));
6912                 else {
6913                         printk(KERN_ERR
6914                                "md/raid:%s: cannot start dirty degraded array.\n",
6915                                mdname(mddev));
6916                         goto abort;
6917                 }
6918         }
6919
6920         if (mddev->degraded == 0)
6921                 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6922                        " devices, algorithm %d\n", mdname(mddev), conf->level,
6923                        mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6924                        mddev->new_layout);
6925         else
6926                 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6927                        " out of %d devices, algorithm %d\n",
6928                        mdname(mddev), conf->level,
6929                        mddev->raid_disks - mddev->degraded,
6930                        mddev->raid_disks, mddev->new_layout);
6931
6932         print_raid5_conf(conf);
6933
6934         if (conf->reshape_progress != MaxSector) {
6935                 conf->reshape_safe = conf->reshape_progress;
6936                 atomic_set(&conf->reshape_stripes, 0);
6937                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6938                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6939                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6940                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6941                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6942                                                         "reshape");
6943         }
6944
6945         /* Ok, everything is just fine now */
6946         if (mddev->to_remove == &raid5_attrs_group)
6947                 mddev->to_remove = NULL;
6948         else if (mddev->kobj.sd &&
6949             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6950                 printk(KERN_WARNING
6951                        "raid5: failed to create sysfs attributes for %s\n",
6952                        mdname(mddev));
6953         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6954
6955         if (mddev->queue) {
6956                 int chunk_size;
6957                 bool discard_supported = true;
6958                 /* read-ahead size must cover two whole stripes, which
6959                  * is 2 * (datadisks) * chunksize where 'n' is the
6960                  * number of raid devices
6961                  */
6962                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6963                 int stripe = data_disks *
6964                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6965                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6966                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6967
6968                 chunk_size = mddev->chunk_sectors << 9;
6969                 blk_queue_io_min(mddev->queue, chunk_size);
6970                 blk_queue_io_opt(mddev->queue, chunk_size *
6971                                  (conf->raid_disks - conf->max_degraded));
6972                 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6973                 /*
6974                  * We can only discard a whole stripe. It doesn't make sense to
6975                  * discard data disk but write parity disk
6976                  */
6977                 stripe = stripe * PAGE_SIZE;
6978                 /* Round up to power of 2, as discard handling
6979                  * currently assumes that */
6980                 while ((stripe-1) & stripe)
6981                         stripe = (stripe | (stripe-1)) + 1;
6982                 mddev->queue->limits.discard_alignment = stripe;
6983                 mddev->queue->limits.discard_granularity = stripe;
6984                 /*
6985                  * unaligned part of discard request will be ignored, so can't
6986                  * guarantee discard_zeroes_data
6987                  */
6988                 mddev->queue->limits.discard_zeroes_data = 0;
6989
6990                 blk_queue_max_write_same_sectors(mddev->queue, 0);
6991
6992                 rdev_for_each(rdev, mddev) {
6993                         disk_stack_limits(mddev->gendisk, rdev->bdev,
6994                                           rdev->data_offset << 9);
6995                         disk_stack_limits(mddev->gendisk, rdev->bdev,
6996                                           rdev->new_data_offset << 9);
6997                         /*
6998                          * discard_zeroes_data is required, otherwise data
6999                          * could be lost. Consider a scenario: discard a stripe
7000                          * (the stripe could be inconsistent if
7001                          * discard_zeroes_data is 0); write one disk of the
7002                          * stripe (the stripe could be inconsistent again
7003                          * depending on which disks are used to calculate
7004                          * parity); the disk is broken; The stripe data of this
7005                          * disk is lost.
7006                          */
7007                         if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
7008                             !bdev_get_queue(rdev->bdev)->
7009                                                 limits.discard_zeroes_data)
7010                                 discard_supported = false;
7011                         /* Unfortunately, discard_zeroes_data is not currently
7012                          * a guarantee - just a hint.  So we only allow DISCARD
7013                          * if the sysadmin has confirmed that only safe devices
7014                          * are in use by setting a module parameter.
7015                          */
7016                         if (!devices_handle_discard_safely) {
7017                                 if (discard_supported) {
7018                                         pr_info("md/raid456: discard support disabled due to uncertainty.\n");
7019                                         pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
7020                                 }
7021                                 discard_supported = false;
7022                         }
7023                 }
7024
7025                 if (discard_supported &&
7026                     mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7027                     mddev->queue->limits.discard_granularity >= stripe)
7028                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
7029                                                 mddev->queue);
7030                 else
7031                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
7032                                                 mddev->queue);
7033         }
7034
7035         if (journal_dev) {
7036                 char b[BDEVNAME_SIZE];
7037
7038                 printk(KERN_INFO"md/raid:%s: using device %s as journal\n",
7039                        mdname(mddev), bdevname(journal_dev->bdev, b));
7040                 r5l_init_log(conf, journal_dev);
7041         }
7042
7043         return 0;
7044 abort:
7045         md_unregister_thread(&mddev->thread);
7046         print_raid5_conf(conf);
7047         free_conf(conf);
7048         mddev->private = NULL;
7049         printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
7050         return -EIO;
7051 }
7052
7053 static void raid5_free(struct mddev *mddev, void *priv)
7054 {
7055         struct r5conf *conf = priv;
7056
7057         free_conf(conf);
7058         mddev->to_remove = &raid5_attrs_group;
7059 }
7060
7061 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7062 {
7063         struct r5conf *conf = mddev->private;
7064         int i;
7065
7066         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7067                 conf->chunk_sectors / 2, mddev->layout);
7068         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7069         for (i = 0; i < conf->raid_disks; i++)
7070                 seq_printf (seq, "%s",
7071                                conf->disks[i].rdev &&
7072                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
7073         seq_printf (seq, "]");
7074 }
7075
7076 static void print_raid5_conf (struct r5conf *conf)
7077 {
7078         int i;
7079         struct disk_info *tmp;
7080
7081         printk(KERN_DEBUG "RAID conf printout:\n");
7082         if (!conf) {
7083                 printk("(conf==NULL)\n");
7084                 return;
7085         }
7086         printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
7087                conf->raid_disks,
7088                conf->raid_disks - conf->mddev->degraded);
7089
7090         for (i = 0; i < conf->raid_disks; i++) {
7091                 char b[BDEVNAME_SIZE];
7092                 tmp = conf->disks + i;
7093                 if (tmp->rdev)
7094                         printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
7095                                i, !test_bit(Faulty, &tmp->rdev->flags),
7096                                bdevname(tmp->rdev->bdev, b));
7097         }
7098 }
7099
7100 static int raid5_spare_active(struct mddev *mddev)
7101 {
7102         int i;
7103         struct r5conf *conf = mddev->private;
7104         struct disk_info *tmp;
7105         int count = 0;
7106         unsigned long flags;
7107
7108         for (i = 0; i < conf->raid_disks; i++) {
7109                 tmp = conf->disks + i;
7110                 if (tmp->replacement
7111                     && tmp->replacement->recovery_offset == MaxSector
7112                     && !test_bit(Faulty, &tmp->replacement->flags)
7113                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7114                         /* Replacement has just become active. */
7115                         if (!tmp->rdev
7116                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7117                                 count++;
7118                         if (tmp->rdev) {
7119                                 /* Replaced device not technically faulty,
7120                                  * but we need to be sure it gets removed
7121                                  * and never re-added.
7122                                  */
7123                                 set_bit(Faulty, &tmp->rdev->flags);
7124                                 sysfs_notify_dirent_safe(
7125                                         tmp->rdev->sysfs_state);
7126                         }
7127                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7128                 } else if (tmp->rdev
7129                     && tmp->rdev->recovery_offset == MaxSector
7130                     && !test_bit(Faulty, &tmp->rdev->flags)
7131                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7132                         count++;
7133                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7134                 }
7135         }
7136         spin_lock_irqsave(&conf->device_lock, flags);
7137         mddev->degraded = calc_degraded(conf);
7138         spin_unlock_irqrestore(&conf->device_lock, flags);
7139         print_raid5_conf(conf);
7140         return count;
7141 }
7142
7143 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7144 {
7145         struct r5conf *conf = mddev->private;
7146         int err = 0;
7147         int number = rdev->raid_disk;
7148         struct md_rdev **rdevp;
7149         struct disk_info *p = conf->disks + number;
7150
7151         print_raid5_conf(conf);
7152         if (test_bit(Journal, &rdev->flags) && conf->log) {
7153                 struct r5l_log *log;
7154                 /*
7155                  * we can't wait pending write here, as this is called in
7156                  * raid5d, wait will deadlock.
7157                  */
7158                 if (atomic_read(&mddev->writes_pending))
7159                         return -EBUSY;
7160                 log = conf->log;
7161                 conf->log = NULL;
7162                 synchronize_rcu();
7163                 r5l_exit_log(log);
7164                 return 0;
7165         }
7166         if (rdev == p->rdev)
7167                 rdevp = &p->rdev;
7168         else if (rdev == p->replacement)
7169                 rdevp = &p->replacement;
7170         else
7171                 return 0;
7172
7173         if (number >= conf->raid_disks &&
7174             conf->reshape_progress == MaxSector)
7175                 clear_bit(In_sync, &rdev->flags);
7176
7177         if (test_bit(In_sync, &rdev->flags) ||
7178             atomic_read(&rdev->nr_pending)) {
7179                 err = -EBUSY;
7180                 goto abort;
7181         }
7182         /* Only remove non-faulty devices if recovery
7183          * isn't possible.
7184          */
7185         if (!test_bit(Faulty, &rdev->flags) &&
7186             mddev->recovery_disabled != conf->recovery_disabled &&
7187             !has_failed(conf) &&
7188             (!p->replacement || p->replacement == rdev) &&
7189             number < conf->raid_disks) {
7190                 err = -EBUSY;
7191                 goto abort;
7192         }
7193         *rdevp = NULL;
7194         synchronize_rcu();
7195         if (atomic_read(&rdev->nr_pending)) {
7196                 /* lost the race, try later */
7197                 err = -EBUSY;
7198                 *rdevp = rdev;
7199         } else if (p->replacement) {
7200                 /* We must have just cleared 'rdev' */
7201                 p->rdev = p->replacement;
7202                 clear_bit(Replacement, &p->replacement->flags);
7203                 smp_mb(); /* Make sure other CPUs may see both as identical
7204                            * but will never see neither - if they are careful
7205                            */
7206                 p->replacement = NULL;
7207                 clear_bit(WantReplacement, &rdev->flags);
7208         } else
7209                 /* We might have just removed the Replacement as faulty-
7210                  * clear the bit just in case
7211                  */
7212                 clear_bit(WantReplacement, &rdev->flags);
7213 abort:
7214
7215         print_raid5_conf(conf);
7216         return err;
7217 }
7218
7219 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7220 {
7221         struct r5conf *conf = mddev->private;
7222         int err = -EEXIST;
7223         int disk;
7224         struct disk_info *p;
7225         int first = 0;
7226         int last = conf->raid_disks - 1;
7227
7228         if (test_bit(Journal, &rdev->flags)) {
7229                 char b[BDEVNAME_SIZE];
7230                 if (conf->log)
7231                         return -EBUSY;
7232
7233                 rdev->raid_disk = 0;
7234                 /*
7235                  * The array is in readonly mode if journal is missing, so no
7236                  * write requests running. We should be safe
7237                  */
7238                 r5l_init_log(conf, rdev);
7239                 printk(KERN_INFO"md/raid:%s: using device %s as journal\n",
7240                        mdname(mddev), bdevname(rdev->bdev, b));
7241                 return 0;
7242         }
7243         if (mddev->recovery_disabled == conf->recovery_disabled)
7244                 return -EBUSY;
7245
7246         if (rdev->saved_raid_disk < 0 && has_failed(conf))
7247                 /* no point adding a device */
7248                 return -EINVAL;
7249
7250         if (rdev->raid_disk >= 0)
7251                 first = last = rdev->raid_disk;
7252
7253         /*
7254          * find the disk ... but prefer rdev->saved_raid_disk
7255          * if possible.
7256          */
7257         if (rdev->saved_raid_disk >= 0 &&
7258             rdev->saved_raid_disk >= first &&
7259             conf->disks[rdev->saved_raid_disk].rdev == NULL)
7260                 first = rdev->saved_raid_disk;
7261
7262         for (disk = first; disk <= last; disk++) {
7263                 p = conf->disks + disk;
7264                 if (p->rdev == NULL) {
7265                         clear_bit(In_sync, &rdev->flags);
7266                         rdev->raid_disk = disk;
7267                         err = 0;
7268                         if (rdev->saved_raid_disk != disk)
7269                                 conf->fullsync = 1;
7270                         rcu_assign_pointer(p->rdev, rdev);
7271                         goto out;
7272                 }
7273         }
7274         for (disk = first; disk <= last; disk++) {
7275                 p = conf->disks + disk;
7276                 if (test_bit(WantReplacement, &p->rdev->flags) &&
7277                     p->replacement == NULL) {
7278                         clear_bit(In_sync, &rdev->flags);
7279                         set_bit(Replacement, &rdev->flags);
7280                         rdev->raid_disk = disk;
7281                         err = 0;
7282                         conf->fullsync = 1;
7283                         rcu_assign_pointer(p->replacement, rdev);
7284                         break;
7285                 }
7286         }
7287 out:
7288         print_raid5_conf(conf);
7289         return err;
7290 }
7291
7292 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7293 {
7294         /* no resync is happening, and there is enough space
7295          * on all devices, so we can resize.
7296          * We need to make sure resync covers any new space.
7297          * If the array is shrinking we should possibly wait until
7298          * any io in the removed space completes, but it hardly seems
7299          * worth it.
7300          */
7301         sector_t newsize;
7302         struct r5conf *conf = mddev->private;
7303
7304         if (conf->log)
7305                 return -EINVAL;
7306         sectors &= ~((sector_t)conf->chunk_sectors - 1);
7307         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7308         if (mddev->external_size &&
7309             mddev->array_sectors > newsize)
7310                 return -EINVAL;
7311         if (mddev->bitmap) {
7312                 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7313                 if (ret)
7314                         return ret;
7315         }
7316         md_set_array_sectors(mddev, newsize);
7317         set_capacity(mddev->gendisk, mddev->array_sectors);
7318         revalidate_disk(mddev->gendisk);
7319         if (sectors > mddev->dev_sectors &&
7320             mddev->recovery_cp > mddev->dev_sectors) {
7321                 mddev->recovery_cp = mddev->dev_sectors;
7322                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7323         }
7324         mddev->dev_sectors = sectors;
7325         mddev->resync_max_sectors = sectors;
7326         return 0;
7327 }
7328
7329 static int check_stripe_cache(struct mddev *mddev)
7330 {
7331         /* Can only proceed if there are plenty of stripe_heads.
7332          * We need a minimum of one full stripe,, and for sensible progress
7333          * it is best to have about 4 times that.
7334          * If we require 4 times, then the default 256 4K stripe_heads will
7335          * allow for chunk sizes up to 256K, which is probably OK.
7336          * If the chunk size is greater, user-space should request more
7337          * stripe_heads first.
7338          */
7339         struct r5conf *conf = mddev->private;
7340         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7341             > conf->min_nr_stripes ||
7342             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7343             > conf->min_nr_stripes) {
7344                 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
7345                        mdname(mddev),
7346                        ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7347                         / STRIPE_SIZE)*4);
7348                 return 0;
7349         }
7350         return 1;
7351 }
7352
7353 static int check_reshape(struct mddev *mddev)
7354 {
7355         struct r5conf *conf = mddev->private;
7356
7357         if (conf->log)
7358                 return -EINVAL;
7359         if (mddev->delta_disks == 0 &&
7360             mddev->new_layout == mddev->layout &&
7361             mddev->new_chunk_sectors == mddev->chunk_sectors)
7362                 return 0; /* nothing to do */
7363         if (has_failed(conf))
7364                 return -EINVAL;
7365         if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7366                 /* We might be able to shrink, but the devices must
7367                  * be made bigger first.
7368                  * For raid6, 4 is the minimum size.
7369                  * Otherwise 2 is the minimum
7370                  */
7371                 int min = 2;
7372                 if (mddev->level == 6)
7373                         min = 4;
7374                 if (mddev->raid_disks + mddev->delta_disks < min)
7375                         return -EINVAL;
7376         }
7377
7378         if (!check_stripe_cache(mddev))
7379                 return -ENOSPC;
7380
7381         if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7382             mddev->delta_disks > 0)
7383                 if (resize_chunks(conf,
7384                                   conf->previous_raid_disks
7385                                   + max(0, mddev->delta_disks),
7386                                   max(mddev->new_chunk_sectors,
7387                                       mddev->chunk_sectors)
7388                             ) < 0)
7389                         return -ENOMEM;
7390         return resize_stripes(conf, (conf->previous_raid_disks
7391                                      + mddev->delta_disks));
7392 }
7393
7394 static int raid5_start_reshape(struct mddev *mddev)
7395 {
7396         struct r5conf *conf = mddev->private;
7397         struct md_rdev *rdev;
7398         int spares = 0;
7399         unsigned long flags;
7400
7401         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7402                 return -EBUSY;
7403
7404         if (!check_stripe_cache(mddev))
7405                 return -ENOSPC;
7406
7407         if (has_failed(conf))
7408                 return -EINVAL;
7409
7410         rdev_for_each(rdev, mddev) {
7411                 if (!test_bit(In_sync, &rdev->flags)
7412                     && !test_bit(Faulty, &rdev->flags))
7413                         spares++;
7414         }
7415
7416         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7417                 /* Not enough devices even to make a degraded array
7418                  * of that size
7419                  */
7420                 return -EINVAL;
7421
7422         /* Refuse to reduce size of the array.  Any reductions in
7423          * array size must be through explicit setting of array_size
7424          * attribute.
7425          */
7426         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7427             < mddev->array_sectors) {
7428                 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7429                        "before number of disks\n", mdname(mddev));
7430                 return -EINVAL;
7431         }
7432
7433         atomic_set(&conf->reshape_stripes, 0);
7434         spin_lock_irq(&conf->device_lock);
7435         write_seqcount_begin(&conf->gen_lock);
7436         conf->previous_raid_disks = conf->raid_disks;
7437         conf->raid_disks += mddev->delta_disks;
7438         conf->prev_chunk_sectors = conf->chunk_sectors;
7439         conf->chunk_sectors = mddev->new_chunk_sectors;
7440         conf->prev_algo = conf->algorithm;
7441         conf->algorithm = mddev->new_layout;
7442         conf->generation++;
7443         /* Code that selects data_offset needs to see the generation update
7444          * if reshape_progress has been set - so a memory barrier needed.
7445          */
7446         smp_mb();
7447         if (mddev->reshape_backwards)
7448                 conf->reshape_progress = raid5_size(mddev, 0, 0);
7449         else
7450                 conf->reshape_progress = 0;
7451         conf->reshape_safe = conf->reshape_progress;
7452         write_seqcount_end(&conf->gen_lock);
7453         spin_unlock_irq(&conf->device_lock);
7454
7455         /* Now make sure any requests that proceeded on the assumption
7456          * the reshape wasn't running - like Discard or Read - have
7457          * completed.
7458          */
7459         mddev_suspend(mddev);
7460         mddev_resume(mddev);
7461
7462         /* Add some new drives, as many as will fit.
7463          * We know there are enough to make the newly sized array work.
7464          * Don't add devices if we are reducing the number of
7465          * devices in the array.  This is because it is not possible
7466          * to correctly record the "partially reconstructed" state of
7467          * such devices during the reshape and confusion could result.
7468          */
7469         if (mddev->delta_disks >= 0) {
7470                 rdev_for_each(rdev, mddev)
7471                         if (rdev->raid_disk < 0 &&
7472                             !test_bit(Faulty, &rdev->flags)) {
7473                                 if (raid5_add_disk(mddev, rdev) == 0) {
7474                                         if (rdev->raid_disk
7475                                             >= conf->previous_raid_disks)
7476                                                 set_bit(In_sync, &rdev->flags);
7477                                         else
7478                                                 rdev->recovery_offset = 0;
7479
7480                                         if (sysfs_link_rdev(mddev, rdev))
7481                                                 /* Failure here is OK */;
7482                                 }
7483                         } else if (rdev->raid_disk >= conf->previous_raid_disks
7484                                    && !test_bit(Faulty, &rdev->flags)) {
7485                                 /* This is a spare that was manually added */
7486                                 set_bit(In_sync, &rdev->flags);
7487                         }
7488
7489                 /* When a reshape changes the number of devices,
7490                  * ->degraded is measured against the larger of the
7491                  * pre and post number of devices.
7492                  */
7493                 spin_lock_irqsave(&conf->device_lock, flags);
7494                 mddev->degraded = calc_degraded(conf);
7495                 spin_unlock_irqrestore(&conf->device_lock, flags);
7496         }
7497         mddev->raid_disks = conf->raid_disks;
7498         mddev->reshape_position = conf->reshape_progress;
7499         set_bit(MD_CHANGE_DEVS, &mddev->flags);
7500
7501         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7502         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7503         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7504         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7505         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7506         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7507                                                 "reshape");
7508         if (!mddev->sync_thread) {
7509                 mddev->recovery = 0;
7510                 spin_lock_irq(&conf->device_lock);
7511                 write_seqcount_begin(&conf->gen_lock);
7512                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7513                 mddev->new_chunk_sectors =
7514                         conf->chunk_sectors = conf->prev_chunk_sectors;
7515                 mddev->new_layout = conf->algorithm = conf->prev_algo;
7516                 rdev_for_each(rdev, mddev)
7517                         rdev->new_data_offset = rdev->data_offset;
7518                 smp_wmb();
7519                 conf->generation --;
7520                 conf->reshape_progress = MaxSector;
7521                 mddev->reshape_position = MaxSector;
7522                 write_seqcount_end(&conf->gen_lock);
7523                 spin_unlock_irq(&conf->device_lock);
7524                 return -EAGAIN;
7525         }
7526         conf->reshape_checkpoint = jiffies;
7527         md_wakeup_thread(mddev->sync_thread);
7528         md_new_event(mddev);
7529         return 0;
7530 }
7531
7532 /* This is called from the reshape thread and should make any
7533  * changes needed in 'conf'
7534  */
7535 static void end_reshape(struct r5conf *conf)
7536 {
7537
7538         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7539                 struct md_rdev *rdev;
7540
7541                 spin_lock_irq(&conf->device_lock);
7542                 conf->previous_raid_disks = conf->raid_disks;
7543                 rdev_for_each(rdev, conf->mddev)
7544                         rdev->data_offset = rdev->new_data_offset;
7545                 smp_wmb();
7546                 conf->reshape_progress = MaxSector;
7547                 conf->mddev->reshape_position = MaxSector;
7548                 spin_unlock_irq(&conf->device_lock);
7549                 wake_up(&conf->wait_for_overlap);
7550
7551                 /* read-ahead size must cover two whole stripes, which is
7552                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7553                  */
7554                 if (conf->mddev->queue) {
7555                         int data_disks = conf->raid_disks - conf->max_degraded;
7556                         int stripe = data_disks * ((conf->chunk_sectors << 9)
7557                                                    / PAGE_SIZE);
7558                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7559                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7560                 }
7561         }
7562 }
7563
7564 /* This is called from the raid5d thread with mddev_lock held.
7565  * It makes config changes to the device.
7566  */
7567 static void raid5_finish_reshape(struct mddev *mddev)
7568 {
7569         struct r5conf *conf = mddev->private;
7570
7571         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7572
7573                 if (mddev->delta_disks > 0) {
7574                         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7575                         if (mddev->queue) {
7576                                 set_capacity(mddev->gendisk, mddev->array_sectors);
7577                                 revalidate_disk(mddev->gendisk);
7578                         }
7579                 } else {
7580                         int d;
7581                         spin_lock_irq(&conf->device_lock);
7582                         mddev->degraded = calc_degraded(conf);
7583                         spin_unlock_irq(&conf->device_lock);
7584                         for (d = conf->raid_disks ;
7585                              d < conf->raid_disks - mddev->delta_disks;
7586                              d++) {
7587                                 struct md_rdev *rdev = conf->disks[d].rdev;
7588                                 if (rdev)
7589                                         clear_bit(In_sync, &rdev->flags);
7590                                 rdev = conf->disks[d].replacement;
7591                                 if (rdev)
7592                                         clear_bit(In_sync, &rdev->flags);
7593                         }
7594                 }
7595                 mddev->layout = conf->algorithm;
7596                 mddev->chunk_sectors = conf->chunk_sectors;
7597                 mddev->reshape_position = MaxSector;
7598                 mddev->delta_disks = 0;
7599                 mddev->reshape_backwards = 0;
7600         }
7601 }
7602
7603 static void raid5_quiesce(struct mddev *mddev, int state)
7604 {
7605         struct r5conf *conf = mddev->private;
7606
7607         switch(state) {
7608         case 2: /* resume for a suspend */
7609                 wake_up(&conf->wait_for_overlap);
7610                 break;
7611
7612         case 1: /* stop all writes */
7613                 lock_all_device_hash_locks_irq(conf);
7614                 /* '2' tells resync/reshape to pause so that all
7615                  * active stripes can drain
7616                  */
7617                 conf->quiesce = 2;
7618                 wait_event_cmd(conf->wait_for_quiescent,
7619                                     atomic_read(&conf->active_stripes) == 0 &&
7620                                     atomic_read(&conf->active_aligned_reads) == 0,
7621                                     unlock_all_device_hash_locks_irq(conf),
7622                                     lock_all_device_hash_locks_irq(conf));
7623                 conf->quiesce = 1;
7624                 unlock_all_device_hash_locks_irq(conf);
7625                 /* allow reshape to continue */
7626                 wake_up(&conf->wait_for_overlap);
7627                 break;
7628
7629         case 0: /* re-enable writes */
7630                 lock_all_device_hash_locks_irq(conf);
7631                 conf->quiesce = 0;
7632                 wake_up(&conf->wait_for_quiescent);
7633                 wake_up(&conf->wait_for_overlap);
7634                 unlock_all_device_hash_locks_irq(conf);
7635                 break;
7636         }
7637         r5l_quiesce(conf->log, state);
7638 }
7639
7640 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7641 {
7642         struct r0conf *raid0_conf = mddev->private;
7643         sector_t sectors;
7644
7645         /* for raid0 takeover only one zone is supported */
7646         if (raid0_conf->nr_strip_zones > 1) {
7647                 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7648                        mdname(mddev));
7649                 return ERR_PTR(-EINVAL);
7650         }
7651
7652         sectors = raid0_conf->strip_zone[0].zone_end;
7653         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7654         mddev->dev_sectors = sectors;
7655         mddev->new_level = level;
7656         mddev->new_layout = ALGORITHM_PARITY_N;
7657         mddev->new_chunk_sectors = mddev->chunk_sectors;
7658         mddev->raid_disks += 1;
7659         mddev->delta_disks = 1;
7660         /* make sure it will be not marked as dirty */
7661         mddev->recovery_cp = MaxSector;
7662
7663         return setup_conf(mddev);
7664 }
7665
7666 static void *raid5_takeover_raid1(struct mddev *mddev)
7667 {
7668         int chunksect;
7669
7670         if (mddev->raid_disks != 2 ||
7671             mddev->degraded > 1)
7672                 return ERR_PTR(-EINVAL);
7673
7674         /* Should check if there are write-behind devices? */
7675
7676         chunksect = 64*2; /* 64K by default */
7677
7678         /* The array must be an exact multiple of chunksize */
7679         while (chunksect && (mddev->array_sectors & (chunksect-1)))
7680                 chunksect >>= 1;
7681
7682         if ((chunksect<<9) < STRIPE_SIZE)
7683                 /* array size does not allow a suitable chunk size */
7684                 return ERR_PTR(-EINVAL);
7685
7686         mddev->new_level = 5;
7687         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7688         mddev->new_chunk_sectors = chunksect;
7689
7690         return setup_conf(mddev);
7691 }
7692
7693 static void *raid5_takeover_raid6(struct mddev *mddev)
7694 {
7695         int new_layout;
7696
7697         switch (mddev->layout) {
7698         case ALGORITHM_LEFT_ASYMMETRIC_6:
7699                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7700                 break;
7701         case ALGORITHM_RIGHT_ASYMMETRIC_6:
7702                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7703                 break;
7704         case ALGORITHM_LEFT_SYMMETRIC_6:
7705                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7706                 break;
7707         case ALGORITHM_RIGHT_SYMMETRIC_6:
7708                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7709                 break;
7710         case ALGORITHM_PARITY_0_6:
7711                 new_layout = ALGORITHM_PARITY_0;
7712                 break;
7713         case ALGORITHM_PARITY_N:
7714                 new_layout = ALGORITHM_PARITY_N;
7715                 break;
7716         default:
7717                 return ERR_PTR(-EINVAL);
7718         }
7719         mddev->new_level = 5;
7720         mddev->new_layout = new_layout;
7721         mddev->delta_disks = -1;
7722         mddev->raid_disks -= 1;
7723         return setup_conf(mddev);
7724 }
7725
7726 static int raid5_check_reshape(struct mddev *mddev)
7727 {
7728         /* For a 2-drive array, the layout and chunk size can be changed
7729          * immediately as not restriping is needed.
7730          * For larger arrays we record the new value - after validation
7731          * to be used by a reshape pass.
7732          */
7733         struct r5conf *conf = mddev->private;
7734         int new_chunk = mddev->new_chunk_sectors;
7735
7736         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7737                 return -EINVAL;
7738         if (new_chunk > 0) {
7739                 if (!is_power_of_2(new_chunk))
7740                         return -EINVAL;
7741                 if (new_chunk < (PAGE_SIZE>>9))
7742                         return -EINVAL;
7743                 if (mddev->array_sectors & (new_chunk-1))
7744                         /* not factor of array size */
7745                         return -EINVAL;
7746         }
7747
7748         /* They look valid */
7749
7750         if (mddev->raid_disks == 2) {
7751                 /* can make the change immediately */
7752                 if (mddev->new_layout >= 0) {
7753                         conf->algorithm = mddev->new_layout;
7754                         mddev->layout = mddev->new_layout;
7755                 }
7756                 if (new_chunk > 0) {
7757                         conf->chunk_sectors = new_chunk ;
7758                         mddev->chunk_sectors = new_chunk;
7759                 }
7760                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7761                 md_wakeup_thread(mddev->thread);
7762         }
7763         return check_reshape(mddev);
7764 }
7765
7766 static int raid6_check_reshape(struct mddev *mddev)
7767 {
7768         int new_chunk = mddev->new_chunk_sectors;
7769
7770         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7771                 return -EINVAL;
7772         if (new_chunk > 0) {
7773                 if (!is_power_of_2(new_chunk))
7774                         return -EINVAL;
7775                 if (new_chunk < (PAGE_SIZE >> 9))
7776                         return -EINVAL;
7777                 if (mddev->array_sectors & (new_chunk-1))
7778                         /* not factor of array size */
7779                         return -EINVAL;
7780         }
7781
7782         /* They look valid */
7783         return check_reshape(mddev);
7784 }
7785
7786 static void *raid5_takeover(struct mddev *mddev)
7787 {
7788         /* raid5 can take over:
7789          *  raid0 - if there is only one strip zone - make it a raid4 layout
7790          *  raid1 - if there are two drives.  We need to know the chunk size
7791          *  raid4 - trivial - just use a raid4 layout.
7792          *  raid6 - Providing it is a *_6 layout
7793          */
7794         if (mddev->level == 0)
7795                 return raid45_takeover_raid0(mddev, 5);
7796         if (mddev->level == 1)
7797                 return raid5_takeover_raid1(mddev);
7798         if (mddev->level == 4) {
7799                 mddev->new_layout = ALGORITHM_PARITY_N;
7800                 mddev->new_level = 5;
7801                 return setup_conf(mddev);
7802         }
7803         if (mddev->level == 6)
7804                 return raid5_takeover_raid6(mddev);
7805
7806         return ERR_PTR(-EINVAL);
7807 }
7808
7809 static void *raid4_takeover(struct mddev *mddev)
7810 {
7811         /* raid4 can take over:
7812          *  raid0 - if there is only one strip zone
7813          *  raid5 - if layout is right
7814          */
7815         if (mddev->level == 0)
7816                 return raid45_takeover_raid0(mddev, 4);
7817         if (mddev->level == 5 &&
7818             mddev->layout == ALGORITHM_PARITY_N) {
7819                 mddev->new_layout = 0;
7820                 mddev->new_level = 4;
7821                 return setup_conf(mddev);
7822         }
7823         return ERR_PTR(-EINVAL);
7824 }
7825
7826 static struct md_personality raid5_personality;
7827
7828 static void *raid6_takeover(struct mddev *mddev)
7829 {
7830         /* Currently can only take over a raid5.  We map the
7831          * personality to an equivalent raid6 personality
7832          * with the Q block at the end.
7833          */
7834         int new_layout;
7835
7836         if (mddev->pers != &raid5_personality)
7837                 return ERR_PTR(-EINVAL);
7838         if (mddev->degraded > 1)
7839                 return ERR_PTR(-EINVAL);
7840         if (mddev->raid_disks > 253)
7841                 return ERR_PTR(-EINVAL);
7842         if (mddev->raid_disks < 3)
7843                 return ERR_PTR(-EINVAL);
7844
7845         switch (mddev->layout) {
7846         case ALGORITHM_LEFT_ASYMMETRIC:
7847                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7848                 break;
7849         case ALGORITHM_RIGHT_ASYMMETRIC:
7850                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7851                 break;
7852         case ALGORITHM_LEFT_SYMMETRIC:
7853                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7854                 break;
7855         case ALGORITHM_RIGHT_SYMMETRIC:
7856                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7857                 break;
7858         case ALGORITHM_PARITY_0:
7859                 new_layout = ALGORITHM_PARITY_0_6;
7860                 break;
7861         case ALGORITHM_PARITY_N:
7862                 new_layout = ALGORITHM_PARITY_N;
7863                 break;
7864         default:
7865                 return ERR_PTR(-EINVAL);
7866         }
7867         mddev->new_level = 6;
7868         mddev->new_layout = new_layout;
7869         mddev->delta_disks = 1;
7870         mddev->raid_disks += 1;
7871         return setup_conf(mddev);
7872 }
7873
7874 static struct md_personality raid6_personality =
7875 {
7876         .name           = "raid6",
7877         .level          = 6,
7878         .owner          = THIS_MODULE,
7879         .make_request   = raid5_make_request,
7880         .run            = raid5_run,
7881         .free           = raid5_free,
7882         .status         = raid5_status,
7883         .error_handler  = raid5_error,
7884         .hot_add_disk   = raid5_add_disk,
7885         .hot_remove_disk= raid5_remove_disk,
7886         .spare_active   = raid5_spare_active,
7887         .sync_request   = raid5_sync_request,
7888         .resize         = raid5_resize,
7889         .size           = raid5_size,
7890         .check_reshape  = raid6_check_reshape,
7891         .start_reshape  = raid5_start_reshape,
7892         .finish_reshape = raid5_finish_reshape,
7893         .quiesce        = raid5_quiesce,
7894         .takeover       = raid6_takeover,
7895         .congested      = raid5_congested,
7896 };
7897 static struct md_personality raid5_personality =
7898 {
7899         .name           = "raid5",
7900         .level          = 5,
7901         .owner          = THIS_MODULE,
7902         .make_request   = raid5_make_request,
7903         .run            = raid5_run,
7904         .free           = raid5_free,
7905         .status         = raid5_status,
7906         .error_handler  = raid5_error,
7907         .hot_add_disk   = raid5_add_disk,
7908         .hot_remove_disk= raid5_remove_disk,
7909         .spare_active   = raid5_spare_active,
7910         .sync_request   = raid5_sync_request,
7911         .resize         = raid5_resize,
7912         .size           = raid5_size,
7913         .check_reshape  = raid5_check_reshape,
7914         .start_reshape  = raid5_start_reshape,
7915         .finish_reshape = raid5_finish_reshape,
7916         .quiesce        = raid5_quiesce,
7917         .takeover       = raid5_takeover,
7918         .congested      = raid5_congested,
7919 };
7920
7921 static struct md_personality raid4_personality =
7922 {
7923         .name           = "raid4",
7924         .level          = 4,
7925         .owner          = THIS_MODULE,
7926         .make_request   = raid5_make_request,
7927         .run            = raid5_run,
7928         .free           = raid5_free,
7929         .status         = raid5_status,
7930         .error_handler  = raid5_error,
7931         .hot_add_disk   = raid5_add_disk,
7932         .hot_remove_disk= raid5_remove_disk,
7933         .spare_active   = raid5_spare_active,
7934         .sync_request   = raid5_sync_request,
7935         .resize         = raid5_resize,
7936         .size           = raid5_size,
7937         .check_reshape  = raid5_check_reshape,
7938         .start_reshape  = raid5_start_reshape,
7939         .finish_reshape = raid5_finish_reshape,
7940         .quiesce        = raid5_quiesce,
7941         .takeover       = raid4_takeover,
7942         .congested      = raid5_congested,
7943 };
7944
7945 static int __init raid5_init(void)
7946 {
7947         raid5_wq = alloc_workqueue("raid5wq",
7948                 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7949         if (!raid5_wq)
7950                 return -ENOMEM;
7951         register_md_personality(&raid6_personality);
7952         register_md_personality(&raid5_personality);
7953         register_md_personality(&raid4_personality);
7954         return 0;
7955 }
7956
7957 static void raid5_exit(void)
7958 {
7959         unregister_md_personality(&raid6_personality);
7960         unregister_md_personality(&raid5_personality);
7961         unregister_md_personality(&raid4_personality);
7962         destroy_workqueue(raid5_wq);
7963 }
7964
7965 module_init(raid5_init);
7966 module_exit(raid5_exit);
7967 MODULE_LICENSE("GPL");
7968 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7969 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7970 MODULE_ALIAS("md-raid5");
7971 MODULE_ALIAS("md-raid4");
7972 MODULE_ALIAS("md-level-5");
7973 MODULE_ALIAS("md-level-4");
7974 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7975 MODULE_ALIAS("md-raid6");
7976 MODULE_ALIAS("md-level-6");
7977
7978 /* This used to be two separate modules, they were: */
7979 MODULE_ALIAS("raid5");
7980 MODULE_ALIAS("raid6");