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