drivers/md/raid5.c

   1 /*
   2  * raid5.c : Multiple Devices driver for Linux
   3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   4  *         Copyright (C) 1999, 2000 Ingo Molnar
   5  *         Copyright (C) 2002, 2003 H. Peter Anvin
   6  *
   7  * RAID-4/5/6 management functions.
   8  * Thanks to Penguin Computing for making the RAID-6 development possible
   9  * by donating a test server!
  10  *
  11  * This program is free software; you can redistribute it and/or modify
  12  * it under the terms of the GNU General Public License as published by
  13  * the Free Software Foundation; either version 2, or (at your option)
  14  * any later version.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * (for example /usr/src/linux/COPYING); if not, write to the Free
  18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19  */
  20
  21 /*
  22  * BITMAP UNPLUGGING:
  23  *
  24  * The sequencing for updating the bitmap reliably is a little
  25  * subtle (and I got it wrong the first time) so it deserves some
  26  * explanation.
  27  *
  28  * We group bitmap updates into batches.  Each batch has a number.
  29  * We may write out several batches at once, but that isn't very important.
  30  * conf->seq_write is the number of the last batch successfully written.
  31  * conf->seq_flush is the number of the last batch that was closed to
  32  *    new additions.
  33  * When we discover that we will need to write to any block in a stripe
  34  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
  35  * the number of the batch it will be in. This is seq_flush+1.
  36  * When we are ready to do a write, if that batch hasn't been written yet,
  37  *   we plug the array and queue the stripe for later.
  38  * When an unplug happens, we increment bm_flush, thus closing the current
  39  *   batch.
  40  * When we notice that bm_flush > bm_write, we write out all pending updates
  41  * to the bitmap, and advance bm_write to where bm_flush was.
  42  * This may occasionally write a bit out twice, but is sure never to
  43  * miss any bits.
  44  */
  45
  46 #include <linux/blkdev.h>
  47 #include <linux/kthread.h>
  48 #include <linux/raid/pq.h>
  49 #include <linux/async_tx.h>
  50 #include <linux/module.h>
  51 #include <linux/async.h>
  52 #include <linux/seq_file.h>
  53 #include <linux/cpu.h>
  54 #include <linux/slab.h>
  55 #include <linux/ratelimit.h>
  56 #include <linux/nodemask.h>
  57 #include <linux/flex_array.h>
  58 #include <trace/events/block.h>
  59
  60 #include "md.h"
  61 #include "raid5.h"
  62 #include "raid0.h"
  63 #include "bitmap.h"
  64
  65 #define cpu_to_group(cpu) cpu_to_node(cpu)
  66 #define ANY_GROUP NUMA_NO_NODE
  67
  68 static bool devices_handle_discard_safely = false;
  69 module_param(devices_handle_discard_safely, bool, 0644);
  70 MODULE_PARM_DESC(devices_handle_discard_safely,
  71                  "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
  72 static struct workqueue_struct *raid5_wq;
  73 /*
  74  * Stripe cache
  75  */
  76
  77 #define NR_STRIPES              256
  78 #define STRIPE_SIZE             PAGE_SIZE
  79 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
  80 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
  81 #define IO_THRESHOLD            1
  82 #define BYPASS_THRESHOLD        1
  83 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
  84 #define HASH_MASK               (NR_HASH - 1)
  85 #define MAX_STRIPE_BATCH        8
  86
  87 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
  88 {
  89         int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
  90         return &conf->stripe_hashtbl[hash];
  91 }
  92
  93 static inline int stripe_hash_locks_hash(sector_t sect)
  94 {
  95         return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
  96 }
  97
  98 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
  99 {
 100         spin_lock_irq(conf->hash_locks + hash);
 101         spin_lock(&conf->device_lock);
 102 }
 103
 104 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
 105 {
 106         spin_unlock(&conf->device_lock);
 107         spin_unlock_irq(conf->hash_locks + hash);
 108 }
 109
 110 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
 111 {
 112         int i;
 113         local_irq_disable();
 114         spin_lock(conf->hash_locks);
 115         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
 116                 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
 117         spin_lock(&conf->device_lock);
 118 }
 119
 120 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
 121 {
 122         int i;
 123         spin_unlock(&conf->device_lock);
 124         for (i = NR_STRIPE_HASH_LOCKS; i; i--)
 125                 spin_unlock(conf->hash_locks + i - 1);
 126         local_irq_enable();
 127 }
 128
 129 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
 130  * order without overlap.  There may be several bio's per stripe+device, and
 131  * a bio could span several devices.
 132  * When walking this list for a particular stripe+device, we must never proceed
 133  * beyond a bio that extends past this device, as the next bio might no longer
 134  * be valid.
 135  * This function is used to determine the 'next' bio in the list, given the sector
 136  * of the current stripe+device
 137  */
 138 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
 139 {
 140         int sectors = bio_sectors(bio);
 141         if (bio->bi_iter.bi_sector + sectors < sector + STRIPE_SECTORS)
 142                 return bio->bi_next;
 143         else
 144                 return NULL;
 145 }
 146
 147 /*
 148  * We maintain a biased count of active stripes in the bottom 16 bits of
 149  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
 150  */
 151 static inline int raid5_bi_processed_stripes(struct bio *bio)
 152 {
 153         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 154         return (atomic_read(segments) >> 16) & 0xffff;
 155 }
 156
 157 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
 158 {
 159         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 160         return atomic_sub_return(1, segments) & 0xffff;
 161 }
 162
 163 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
 164 {
 165         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 166         atomic_inc(segments);
 167 }
 168
 169 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
 170         unsigned int cnt)
 171 {
 172         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 173         int old, new;
 174
 175         do {
 176                 old = atomic_read(segments);
 177                 new = (old & 0xffff) | (cnt << 16);
 178         } while (atomic_cmpxchg(segments, old, new) != old);
 179 }
 180
 181 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
 182 {
 183         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 184         atomic_set(segments, cnt);
 185 }
 186
 187 /* Find first data disk in a raid6 stripe */
 188 static inline int raid6_d0(struct stripe_head *sh)
 189 {
 190         if (sh->ddf_layout)
 191                 /* ddf always start from first device */
 192                 return 0;
 193         /* md starts just after Q block */
 194         if (sh->qd_idx == sh->disks - 1)
 195                 return 0;
 196         else
 197                 return sh->qd_idx + 1;
 198 }
 199 static inline int raid6_next_disk(int disk, int raid_disks)
 200 {
 201         disk++;
 202         return (disk < raid_disks) ? disk : 0;
 203 }
 204
 205 /* When walking through the disks in a raid5, starting at raid6_d0,
 206  * We need to map each disk to a 'slot', where the data disks are slot
 207  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 208  * is raid_disks-1.  This help does that mapping.
 209  */
 210 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 211                              int *count, int syndrome_disks)
 212 {
 213         int slot = *count;
 214
 215         if (sh->ddf_layout)
 216                 (*count)++;
 217         if (idx == sh->pd_idx)
 218                 return syndrome_disks;
 219         if (idx == sh->qd_idx)
 220                 return syndrome_disks + 1;
 221         if (!sh->ddf_layout)
 222                 (*count)++;
 223         return slot;
 224 }
 225
 226 static void return_io(struct bio_list *return_bi)
 227 {
 228         struct bio *bi;
 229         while ((bi = bio_list_pop(return_bi)) != NULL) {
 230                 bi->bi_iter.bi_size = 0;
 231                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
 232                                          bi, 0);
 233                 bio_endio(bi);
 234         }
 235 }
 236
 237 static void print_raid5_conf (struct r5conf *conf);
 238
 239 static int stripe_operations_active(struct stripe_head *sh)
 240 {
 241         return sh->check_state || sh->reconstruct_state ||
 242                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 243                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 244 }
 245
 246 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
 247 {
 248         struct r5conf *conf = sh->raid_conf;
 249         struct r5worker_group *group;
 250         int thread_cnt;
 251         int i, cpu = sh->cpu;
 252
 253         if (!cpu_online(cpu)) {
 254                 cpu = cpumask_any(cpu_online_mask);
 255                 sh->cpu = cpu;
 256         }
 257
 258         if (list_empty(&sh->lru)) {
 259                 struct r5worker_group *group;
 260                 group = conf->worker_groups + cpu_to_group(cpu);
 261                 list_add_tail(&sh->lru, &group->handle_list);
 262                 group->stripes_cnt++;
 263                 sh->group = group;
 264         }
 265
 266         if (conf->worker_cnt_per_group == 0) {
 267                 md_wakeup_thread(conf->mddev->thread);
 268                 return;
 269         }
 270
 271         group = conf->worker_groups + cpu_to_group(sh->cpu);
 272
 273         group->workers[0].working = true;
 274         /* at least one worker should run to avoid race */
 275         queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
 276
 277         thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
 278         /* wakeup more workers */
 279         for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
 280                 if (group->workers[i].working == false) {
 281                         group->workers[i].working = true;
 282                         queue_work_on(sh->cpu, raid5_wq,
 283                                       &group->workers[i].work);
 284                         thread_cnt--;
 285                 }
 286         }
 287 }
 288
 289 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
 290                               struct list_head *temp_inactive_list)
 291 {
 292         BUG_ON(!list_empty(&sh->lru));
 293         BUG_ON(atomic_read(&conf->active_stripes)==0);
 294         if (test_bit(STRIPE_HANDLE, &sh->state)) {
 295                 if (test_bit(STRIPE_DELAYED, &sh->state) &&
 296                     !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 297                         list_add_tail(&sh->lru, &conf->delayed_list);
 298                 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 299                            sh->bm_seq - conf->seq_write > 0)
 300                         list_add_tail(&sh->lru, &conf->bitmap_list);
 301                 else {
 302                         clear_bit(STRIPE_DELAYED, &sh->state);
 303                         clear_bit(STRIPE_BIT_DELAY, &sh->state);
 304                         if (conf->worker_cnt_per_group == 0) {
 305                                 list_add_tail(&sh->lru, &conf->handle_list);
 306                         } else {
 307                                 raid5_wakeup_stripe_thread(sh);
 308                                 return;
 309                         }
 310                 }
 311                 md_wakeup_thread(conf->mddev->thread);
 312         } else {
 313                 BUG_ON(stripe_operations_active(sh));
 314                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 315                         if (atomic_dec_return(&conf->preread_active_stripes)
 316                             < IO_THRESHOLD)
 317                                 md_wakeup_thread(conf->mddev->thread);
 318                 atomic_dec(&conf->active_stripes);
 319                 if (!test_bit(STRIPE_EXPANDING, &sh->state))
 320                         list_add_tail(&sh->lru, temp_inactive_list);
 321         }
 322 }
 323
 324 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
 325                              struct list_head *temp_inactive_list)
 326 {
 327         if (atomic_dec_and_test(&sh->count))
 328                 do_release_stripe(conf, sh, temp_inactive_list);
 329 }
 330
 331 /*
 332  * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
 333  *
 334  * Be careful: Only one task can add/delete stripes from temp_inactive_list at
 335  * given time. Adding stripes only takes device lock, while deleting stripes
 336  * only takes hash lock.
 337  */
 338 static void release_inactive_stripe_list(struct r5conf *conf,
 339                                          struct list_head *temp_inactive_list,
 340                                          int hash)
 341 {
 342         int size;
 343         unsigned long do_wakeup = 0;
 344         int i = 0;
 345         unsigned long flags;
 346
 347         if (hash == NR_STRIPE_HASH_LOCKS) {
 348                 size = NR_STRIPE_HASH_LOCKS;
 349                 hash = NR_STRIPE_HASH_LOCKS - 1;
 350         } else
 351                 size = 1;
 352         while (size) {
 353                 struct list_head *list = &temp_inactive_list[size - 1];
 354
 355                 /*
 356                  * We don't hold any lock here yet, raid5_get_active_stripe() might
 357                  * remove stripes from the list
 358                  */
 359                 if (!list_empty_careful(list)) {
 360                         spin_lock_irqsave(conf->hash_locks + hash, flags);
 361                         if (list_empty(conf->inactive_list + hash) &&
 362                             !list_empty(list))
 363                                 atomic_dec(&conf->empty_inactive_list_nr);
 364                         list_splice_tail_init(list, conf->inactive_list + hash);
 365                         do_wakeup |= 1 << hash;
 366                         spin_unlock_irqrestore(conf->hash_locks + hash, flags);
 367                 }
 368                 size--;
 369                 hash--;
 370         }
 371
 372         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++) {
 373                 if (do_wakeup & (1 << i))
 374                         wake_up(&conf->wait_for_stripe[i]);
 375         }
 376
 377         if (do_wakeup) {
 378                 if (atomic_read(&conf->active_stripes) == 0)
 379                         wake_up(&conf->wait_for_quiescent);
 380                 if (conf->retry_read_aligned)
 381                         md_wakeup_thread(conf->mddev->thread);
 382         }
 383 }
 384
 385 /* should hold conf->device_lock already */
 386 static int release_stripe_list(struct r5conf *conf,
 387                                struct list_head *temp_inactive_list)
 388 {
 389         struct stripe_head *sh;
 390         int count = 0;
 391         struct llist_node *head;
 392
 393         head = llist_del_all(&conf->released_stripes);
 394         head = llist_reverse_order(head);
 395         while (head) {
 396                 int hash;
 397
 398                 sh = llist_entry(head, struct stripe_head, release_list);
 399                 head = llist_next(head);
 400                 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
 401                 smp_mb();
 402                 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
 403                 /*
 404                  * Don't worry the bit is set here, because if the bit is set
 405                  * again, the count is always > 1. This is true for
 406                  * STRIPE_ON_UNPLUG_LIST bit too.
 407                  */
 408                 hash = sh->hash_lock_index;
 409                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
 410                 count++;
 411         }
 412
 413         return count;
 414 }
 415
 416 void raid5_release_stripe(struct stripe_head *sh)
 417 {
 418         struct r5conf *conf = sh->raid_conf;
 419         unsigned long flags;
 420         struct list_head list;
 421         int hash;
 422         bool wakeup;
 423
 424         /* Avoid release_list until the last reference.
 425          */
 426         if (atomic_add_unless(&sh->count, -1, 1))
 427                 return;
 428
 429         if (unlikely(!conf->mddev->thread) ||
 430                 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
 431                 goto slow_path;
 432         wakeup = llist_add(&sh->release_list, &conf->released_stripes);
 433         if (wakeup)
 434                 md_wakeup_thread(conf->mddev->thread);
 435         return;
 436 slow_path:
 437         local_irq_save(flags);
 438         /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
 439         if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
 440                 INIT_LIST_HEAD(&list);
 441                 hash = sh->hash_lock_index;
 442                 do_release_stripe(conf, sh, &list);
 443                 spin_unlock(&conf->device_lock);
 444                 release_inactive_stripe_list(conf, &list, hash);
 445         }
 446         local_irq_restore(flags);
 447 }
 448
 449 static inline void remove_hash(struct stripe_head *sh)
 450 {
 451         pr_debug("remove_hash(), stripe %llu\n",
 452                 (unsigned long long)sh->sector);
 453
 454         hlist_del_init(&sh->hash);
 455 }
 456
 457 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 458 {
 459         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 460
 461         pr_debug("insert_hash(), stripe %llu\n",
 462                 (unsigned long long)sh->sector);
 463
 464         hlist_add_head(&sh->hash, hp);
 465 }
 466
 467 /* find an idle stripe, make sure it is unhashed, and return it. */
 468 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
 469 {
 470         struct stripe_head *sh = NULL;
 471         struct list_head *first;
 472
 473         if (list_empty(conf->inactive_list + hash))
 474                 goto out;
 475         first = (conf->inactive_list + hash)->next;
 476         sh = list_entry(first, struct stripe_head, lru);
 477         list_del_init(first);
 478         remove_hash(sh);
 479         atomic_inc(&conf->active_stripes);
 480         BUG_ON(hash != sh->hash_lock_index);
 481         if (list_empty(conf->inactive_list + hash))
 482                 atomic_inc(&conf->empty_inactive_list_nr);
 483 out:
 484         return sh;
 485 }
 486
 487 static void shrink_buffers(struct stripe_head *sh)
 488 {
 489         struct page *p;
 490         int i;
 491         int num = sh->raid_conf->pool_size;
 492
 493         for (i = 0; i < num ; i++) {
 494                 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
 495                 p = sh->dev[i].page;
 496                 if (!p)
 497                         continue;
 498                 sh->dev[i].page = NULL;
 499                 put_page(p);
 500         }
 501 }
 502
 503 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
 504 {
 505         int i;
 506         int num = sh->raid_conf->pool_size;
 507
 508         for (i = 0; i < num; i++) {
 509                 struct page *page;
 510
 511                 if (!(page = alloc_page(gfp))) {
 512                         return 1;
 513                 }
 514                 sh->dev[i].page = page;
 515                 sh->dev[i].orig_page = page;
 516         }
 517         return 0;
 518 }
 519
 520 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
 521 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 522                             struct stripe_head *sh);
 523
 524 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 525 {
 526         struct r5conf *conf = sh->raid_conf;
 527         int i, seq;
 528
 529         BUG_ON(atomic_read(&sh->count) != 0);
 530         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 531         BUG_ON(stripe_operations_active(sh));
 532         BUG_ON(sh->batch_head);
 533
 534         pr_debug("init_stripe called, stripe %llu\n",
 535                 (unsigned long long)sector);
 536 retry:
 537         seq = read_seqcount_begin(&conf->gen_lock);
 538         sh->generation = conf->generation - previous;
 539         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 540         sh->sector = sector;
 541         stripe_set_idx(sector, conf, previous, sh);
 542         sh->state = 0;
 543
 544         for (i = sh->disks; i--; ) {
 545                 struct r5dev *dev = &sh->dev[i];
 546
 547                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 548                     test_bit(R5_LOCKED, &dev->flags)) {
 549                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
 550                                (unsigned long long)sh->sector, i, dev->toread,
 551                                dev->read, dev->towrite, dev->written,
 552                                test_bit(R5_LOCKED, &dev->flags));
 553                         WARN_ON(1);
 554                 }
 555                 dev->flags = 0;
 556                 raid5_build_block(sh, i, previous);
 557         }
 558         if (read_seqcount_retry(&conf->gen_lock, seq))
 559                 goto retry;
 560         sh->overwrite_disks = 0;
 561         insert_hash(conf, sh);
 562         sh->cpu = smp_processor_id();
 563         set_bit(STRIPE_BATCH_READY, &sh->state);
 564 }
 565
 566 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 567                                          short generation)
 568 {
 569         struct stripe_head *sh;
 570
 571         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 572         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 573                 if (sh->sector == sector && sh->generation == generation)
 574                         return sh;
 575         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 576         return NULL;
 577 }
 578
 579 /*
 580  * Need to check if array has failed when deciding whether to:
 581  *  - start an array
 582  *  - remove non-faulty devices
 583  *  - add a spare
 584  *  - allow a reshape
 585  * This determination is simple when no reshape is happening.
 586  * However if there is a reshape, we need to carefully check
 587  * both the before and after sections.
 588  * This is because some failed devices may only affect one
 589  * of the two sections, and some non-in_sync devices may
 590  * be insync in the section most affected by failed devices.
 591  */
 592 static int calc_degraded(struct r5conf *conf)
 593 {
 594         int degraded, degraded2;
 595         int i;
 596
 597         rcu_read_lock();
 598         degraded = 0;
 599         for (i = 0; i < conf->previous_raid_disks; i++) {
 600                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 601                 if (rdev && test_bit(Faulty, &rdev->flags))
 602                         rdev = rcu_dereference(conf->disks[i].replacement);
 603                 if (!rdev || test_bit(Faulty, &rdev->flags))
 604                         degraded++;
 605                 else if (test_bit(In_sync, &rdev->flags))
 606                         ;
 607                 else
 608                         /* not in-sync or faulty.
 609                          * If the reshape increases the number of devices,
 610                          * this is being recovered by the reshape, so
 611                          * this 'previous' section is not in_sync.
 612                          * If the number of devices is being reduced however,
 613                          * the device can only be part of the array if
 614                          * we are reverting a reshape, so this section will
 615                          * be in-sync.
 616                          */
 617                         if (conf->raid_disks >= conf->previous_raid_disks)
 618                                 degraded++;
 619         }
 620         rcu_read_unlock();
 621         if (conf->raid_disks == conf->previous_raid_disks)
 622                 return degraded;
 623         rcu_read_lock();
 624         degraded2 = 0;
 625         for (i = 0; i < conf->raid_disks; i++) {
 626                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 627                 if (rdev && test_bit(Faulty, &rdev->flags))
 628                         rdev = rcu_dereference(conf->disks[i].replacement);
 629                 if (!rdev || test_bit(Faulty, &rdev->flags))
 630                         degraded2++;
 631                 else if (test_bit(In_sync, &rdev->flags))
 632                         ;
 633                 else
 634                         /* not in-sync or faulty.
 635                          * If reshape increases the number of devices, this
 636                          * section has already been recovered, else it
 637                          * almost certainly hasn't.
 638                          */
 639                         if (conf->raid_disks <= conf->previous_raid_disks)
 640                                 degraded2++;
 641         }
 642         rcu_read_unlock();
 643         if (degraded2 > degraded)
 644                 return degraded2;
 645         return degraded;
 646 }
 647
 648 static int has_failed(struct r5conf *conf)
 649 {
 650         int degraded;
 651
 652         if (conf->mddev->reshape_position == MaxSector)
 653                 return conf->mddev->degraded > conf->max_degraded;
 654
 655         degraded = calc_degraded(conf);
 656         if (degraded > conf->max_degraded)
 657                 return 1;
 658         return 0;
 659 }
 660
 661 struct stripe_head *
 662 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
 663                         int previous, int noblock, int noquiesce)
 664 {
 665         struct stripe_head *sh;
 666         int hash = stripe_hash_locks_hash(sector);
 667
 668         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 669
 670         spin_lock_irq(conf->hash_locks + hash);
 671
 672         do {
 673                 wait_event_lock_irq(conf->wait_for_quiescent,
 674                                     conf->quiesce == 0 || noquiesce,
 675                                     *(conf->hash_locks + hash));
 676                 sh = __find_stripe(conf, sector, conf->generation - previous);
 677                 if (!sh) {
 678                         if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 679                                 sh = get_free_stripe(conf, hash);
 680                                 if (!sh && !test_bit(R5_DID_ALLOC,
 681                                                      &conf->cache_state))
 682                                         set_bit(R5_ALLOC_MORE,
 683                                                 &conf->cache_state);
 684                         }
 685                         if (noblock && sh == NULL)
 686                                 break;
 687                         if (!sh) {
 688                                 set_bit(R5_INACTIVE_BLOCKED,
 689                                         &conf->cache_state);
 690                                 wait_event_exclusive_cmd(
 691                                         conf->wait_for_stripe[hash],
 692                                         !list_empty(conf->inactive_list + hash) &&
 693                                         (atomic_read(&conf->active_stripes)
 694                                          < (conf->max_nr_stripes * 3 / 4)
 695                                          || !test_bit(R5_INACTIVE_BLOCKED,
 696                                                       &conf->cache_state)),
 697                                         spin_unlock_irq(conf->hash_locks + hash),
 698                                         spin_lock_irq(conf->hash_locks + hash));
 699                                 clear_bit(R5_INACTIVE_BLOCKED,
 700                                           &conf->cache_state);
 701                         } else {
 702                                 init_stripe(sh, sector, previous);
 703                                 atomic_inc(&sh->count);
 704                         }
 705                 } else if (!atomic_inc_not_zero(&sh->count)) {
 706                         spin_lock(&conf->device_lock);
 707                         if (!atomic_read(&sh->count)) {
 708                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 709                                         atomic_inc(&conf->active_stripes);
 710                                 BUG_ON(list_empty(&sh->lru) &&
 711                                        !test_bit(STRIPE_EXPANDING, &sh->state));
 712                                 list_del_init(&sh->lru);
 713                                 if (sh->group) {
 714                                         sh->group->stripes_cnt--;
 715                                         sh->group = NULL;
 716                                 }
 717                         }
 718                         atomic_inc(&sh->count);
 719                         spin_unlock(&conf->device_lock);
 720                 }
 721         } while (sh == NULL);
 722
 723         if (!list_empty(conf->inactive_list + hash))
 724                 wake_up(&conf->wait_for_stripe[hash]);
 725
 726         spin_unlock_irq(conf->hash_locks + hash);
 727         return sh;
 728 }
 729
 730 static bool is_full_stripe_write(struct stripe_head *sh)
 731 {
 732         BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 733         return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 734 }
 735
 736 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 737 {
 738         local_irq_disable();
 739         if (sh1 > sh2) {
 740                 spin_lock(&sh2->stripe_lock);
 741                 spin_lock_nested(&sh1->stripe_lock, 1);
 742         } else {
 743                 spin_lock(&sh1->stripe_lock);
 744                 spin_lock_nested(&sh2->stripe_lock, 1);
 745         }
 746 }
 747
 748 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 749 {
 750         spin_unlock(&sh1->stripe_lock);
 751         spin_unlock(&sh2->stripe_lock);
 752         local_irq_enable();
 753 }
 754
 755 /* Only freshly new full stripe normal write stripe can be added to a batch list */
 756 static bool stripe_can_batch(struct stripe_head *sh)
 757 {
 758         struct r5conf *conf = sh->raid_conf;
 759
 760         if (conf->log)
 761                 return false;
 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         raid5_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         if (r5l_write_stripe(conf->log, sh) == 0)
 903                 return;
 904         for (i = disks; i--; ) {
 905                 int rw;
 906                 int replace_only = 0;
 907                 struct bio *bi, *rbi;
 908                 struct md_rdev *rdev, *rrdev = NULL;
 909
 910                 sh = head_sh;
 911                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
 912                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
 913                                 rw = WRITE_FUA;
 914                         else
 915                                 rw = WRITE;
 916                         if (test_bit(R5_Discard, &sh->dev[i].flags))
 917                                 rw |= REQ_DISCARD;
 918                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
 919                         rw = READ;
 920                 else if (test_and_clear_bit(R5_WantReplace,
 921                                             &sh->dev[i].flags)) {
 922                         rw = WRITE;
 923                         replace_only = 1;
 924                 } else
 925                         continue;
 926                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
 927                         rw |= REQ_SYNC;
 928
 929 again:
 930                 bi = &sh->dev[i].req;
 931                 rbi = &sh->dev[i].rreq; /* For writing to replacement */
 932
 933                 rcu_read_lock();
 934                 rrdev = rcu_dereference(conf->disks[i].replacement);
 935                 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
 936                 rdev = rcu_dereference(conf->disks[i].rdev);
 937                 if (!rdev) {
 938                         rdev = rrdev;
 939                         rrdev = NULL;
 940                 }
 941                 if (rw & WRITE) {
 942                         if (replace_only)
 943                                 rdev = NULL;
 944                         if (rdev == rrdev)
 945                                 /* We raced and saw duplicates */
 946                                 rrdev = NULL;
 947                 } else {
 948                         if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
 949                                 rdev = rrdev;
 950                         rrdev = NULL;
 951                 }
 952
 953                 if (rdev && test_bit(Faulty, &rdev->flags))
 954                         rdev = NULL;
 955                 if (rdev)
 956                         atomic_inc(&rdev->nr_pending);
 957                 if (rrdev && test_bit(Faulty, &rrdev->flags))
 958                         rrdev = NULL;
 959                 if (rrdev)
 960                         atomic_inc(&rrdev->nr_pending);
 961                 rcu_read_unlock();
 962
 963                 /* We have already checked bad blocks for reads.  Now
 964                  * need to check for writes.  We never accept write errors
 965                  * on the replacement, so we don't to check rrdev.
 966                  */
 967                 while ((rw & WRITE) && rdev &&
 968                        test_bit(WriteErrorSeen, &rdev->flags)) {
 969                         sector_t first_bad;
 970                         int bad_sectors;
 971                         int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
 972                                               &first_bad, &bad_sectors);
 973                         if (!bad)
 974                                 break;
 975
 976                         if (bad < 0) {
 977                                 set_bit(BlockedBadBlocks, &rdev->flags);
 978                                 if (!conf->mddev->external &&
 979                                     conf->mddev->flags) {
 980                                         /* It is very unlikely, but we might
 981                                          * still need to write out the
 982                                          * bad block log - better give it
 983                                          * a chance*/
 984                                         md_check_recovery(conf->mddev);
 985                                 }
 986                                 /*
 987                                  * Because md_wait_for_blocked_rdev
 988                                  * will dec nr_pending, we must
 989                                  * increment it first.
 990                                  */
 991                                 atomic_inc(&rdev->nr_pending);
 992                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
 993                         } else {
 994                                 /* Acknowledged bad block - skip the write */
 995                                 rdev_dec_pending(rdev, conf->mddev);
 996                                 rdev = NULL;
 997                         }
 998                 }
 999
1000                 if (rdev) {
1001                         if (s->syncing || s->expanding || s->expanded
1002                             || s->replacing)
1003                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1004
1005                         set_bit(STRIPE_IO_STARTED, &sh->state);
1006
1007                         bio_reset(bi);
1008                         bi->bi_bdev = rdev->bdev;
1009                         bi->bi_rw = rw;
1010                         bi->bi_end_io = (rw & WRITE)
1011                                 ? raid5_end_write_request
1012                                 : raid5_end_read_request;
1013                         bi->bi_private = sh;
1014
1015                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
1016                                 __func__, (unsigned long long)sh->sector,
1017                                 bi->bi_rw, i);
1018                         atomic_inc(&sh->count);
1019                         if (sh != head_sh)
1020                                 atomic_inc(&head_sh->count);
1021                         if (use_new_offset(conf, sh))
1022                                 bi->bi_iter.bi_sector = (sh->sector
1023                                                  + rdev->new_data_offset);
1024                         else
1025                                 bi->bi_iter.bi_sector = (sh->sector
1026                                                  + rdev->data_offset);
1027                         if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1028                                 bi->bi_rw |= REQ_NOMERGE;
1029
1030                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1031                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1032                         sh->dev[i].vec.bv_page = sh->dev[i].page;
1033                         bi->bi_vcnt = 1;
1034                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1035                         bi->bi_io_vec[0].bv_offset = 0;
1036                         bi->bi_iter.bi_size = STRIPE_SIZE;
1037                         /*
1038                          * If this is discard request, set bi_vcnt 0. We don't
1039                          * want to confuse SCSI because SCSI will replace payload
1040                          */
1041                         if (rw & REQ_DISCARD)
1042                                 bi->bi_vcnt = 0;
1043                         if (rrdev)
1044                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1045
1046                         if (conf->mddev->gendisk)
1047                                 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1048                                                       bi, disk_devt(conf->mddev->gendisk),
1049                                                       sh->dev[i].sector);
1050                         generic_make_request(bi);
1051                 }
1052                 if (rrdev) {
1053                         if (s->syncing || s->expanding || s->expanded
1054                             || s->replacing)
1055                                 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1056
1057                         set_bit(STRIPE_IO_STARTED, &sh->state);
1058
1059                         bio_reset(rbi);
1060                         rbi->bi_bdev = rrdev->bdev;
1061                         rbi->bi_rw = rw;
1062                         BUG_ON(!(rw & WRITE));
1063                         rbi->bi_end_io = raid5_end_write_request;
1064                         rbi->bi_private = sh;
1065
1066                         pr_debug("%s: for %llu schedule op %ld on "
1067                                  "replacement disc %d\n",
1068                                 __func__, (unsigned long long)sh->sector,
1069                                 rbi->bi_rw, i);
1070                         atomic_inc(&sh->count);
1071                         if (sh != head_sh)
1072                                 atomic_inc(&head_sh->count);
1073                         if (use_new_offset(conf, sh))
1074                                 rbi->bi_iter.bi_sector = (sh->sector
1075                                                   + rrdev->new_data_offset);
1076                         else
1077                                 rbi->bi_iter.bi_sector = (sh->sector
1078                                                   + rrdev->data_offset);
1079                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1080                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1081                         sh->dev[i].rvec.bv_page = sh->dev[i].page;
1082                         rbi->bi_vcnt = 1;
1083                         rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1084                         rbi->bi_io_vec[0].bv_offset = 0;
1085                         rbi->bi_iter.bi_size = STRIPE_SIZE;
1086                         /*
1087                          * If this is discard request, set bi_vcnt 0. We don't
1088                          * want to confuse SCSI because SCSI will replace payload
1089                          */
1090                         if (rw & REQ_DISCARD)
1091                                 rbi->bi_vcnt = 0;
1092                         if (conf->mddev->gendisk)
1093                                 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1094                                                       rbi, disk_devt(conf->mddev->gendisk),
1095                                                       sh->dev[i].sector);
1096                         generic_make_request(rbi);
1097                 }
1098                 if (!rdev && !rrdev) {
1099                         if (rw & WRITE)
1100                                 set_bit(STRIPE_DEGRADED, &sh->state);
1101                         pr_debug("skip op %ld on disc %d for sector %llu\n",
1102                                 bi->bi_rw, i, (unsigned long long)sh->sector);
1103                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1104                         set_bit(STRIPE_HANDLE, &sh->state);
1105                 }
1106
1107                 if (!head_sh->batch_head)
1108                         continue;
1109                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1110                                       batch_list);
1111                 if (sh != head_sh)
1112                         goto again;
1113         }
1114 }
1115
1116 static struct dma_async_tx_descriptor *
1117 async_copy_data(int frombio, struct bio *bio, struct page **page,
1118         sector_t sector, struct dma_async_tx_descriptor *tx,
1119         struct stripe_head *sh)
1120 {
1121         struct bio_vec bvl;
1122         struct bvec_iter iter;
1123         struct page *bio_page;
1124         int page_offset;
1125         struct async_submit_ctl submit;
1126         enum async_tx_flags flags = 0;
1127
1128         if (bio->bi_iter.bi_sector >= sector)
1129                 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1130         else
1131                 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1132
1133         if (frombio)
1134                 flags |= ASYNC_TX_FENCE;
1135         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1136
1137         bio_for_each_segment(bvl, bio, iter) {
1138                 int len = bvl.bv_len;
1139                 int clen;
1140                 int b_offset = 0;
1141
1142                 if (page_offset < 0) {
1143                         b_offset = -page_offset;
1144                         page_offset += b_offset;
1145                         len -= b_offset;
1146                 }
1147
1148                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1149                         clen = STRIPE_SIZE - page_offset;
1150                 else
1151                         clen = len;
1152
1153                 if (clen > 0) {
1154                         b_offset += bvl.bv_offset;
1155                         bio_page = bvl.bv_page;
1156                         if (frombio) {
1157                                 if (sh->raid_conf->skip_copy &&
1158                                     b_offset == 0 && page_offset == 0 &&
1159                                     clen == STRIPE_SIZE)
1160                                         *page = bio_page;
1161                                 else
1162                                         tx = async_memcpy(*page, bio_page, page_offset,
1163                                                   b_offset, clen, &submit);
1164                         } else
1165                                 tx = async_memcpy(bio_page, *page, b_offset,
1166                                                   page_offset, clen, &submit);
1167                 }
1168                 /* chain the operations */
1169                 submit.depend_tx = tx;
1170
1171                 if (clen < len) /* hit end of page */
1172                         break;
1173                 page_offset +=  len;
1174         }
1175
1176         return tx;
1177 }
1178
1179 static void ops_complete_biofill(void *stripe_head_ref)
1180 {
1181         struct stripe_head *sh = stripe_head_ref;
1182         struct bio_list return_bi = BIO_EMPTY_LIST;
1183         int i;
1184
1185         pr_debug("%s: stripe %llu\n", __func__,
1186                 (unsigned long long)sh->sector);
1187
1188         /* clear completed biofills */
1189         for (i = sh->disks; i--; ) {
1190                 struct r5dev *dev = &sh->dev[i];
1191
1192                 /* acknowledge completion of a biofill operation */
1193                 /* and check if we need to reply to a read request,
1194                  * new R5_Wantfill requests are held off until
1195                  * !STRIPE_BIOFILL_RUN
1196                  */
1197                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1198                         struct bio *rbi, *rbi2;
1199
1200                         BUG_ON(!dev->read);
1201                         rbi = dev->read;
1202                         dev->read = NULL;
1203                         while (rbi && rbi->bi_iter.bi_sector <
1204                                 dev->sector + STRIPE_SECTORS) {
1205                                 rbi2 = r5_next_bio(rbi, dev->sector);
1206                                 if (!raid5_dec_bi_active_stripes(rbi))
1207                                         bio_list_add(&return_bi, rbi);
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         raid5_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         raid5_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         raid5_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         raid5_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         raid5_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                 raid5_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         kmem_cache_destroy(conf->slab_cache);
2281         conf->slab_cache = NULL;
2282 }
2283
2284 static void raid5_end_read_request(struct bio * bi)
2285 {
2286         struct stripe_head *sh = bi->bi_private;
2287         struct r5conf *conf = sh->raid_conf;
2288         int disks = sh->disks, i;
2289         char b[BDEVNAME_SIZE];
2290         struct md_rdev *rdev = NULL;
2291         sector_t s;
2292
2293         for (i=0 ; i<disks; i++)
2294                 if (bi == &sh->dev[i].req)
2295                         break;
2296
2297         pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2298                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2299                 bi->bi_error);
2300         if (i == disks) {
2301                 BUG();
2302                 return;
2303         }
2304         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2305                 /* If replacement finished while this request was outstanding,
2306                  * 'replacement' might be NULL already.
2307                  * In that case it moved down to 'rdev'.
2308                  * rdev is not removed until all requests are finished.
2309                  */
2310                 rdev = conf->disks[i].replacement;
2311         if (!rdev)
2312                 rdev = conf->disks[i].rdev;
2313
2314         if (use_new_offset(conf, sh))
2315                 s = sh->sector + rdev->new_data_offset;
2316         else
2317                 s = sh->sector + rdev->data_offset;
2318         if (!bi->bi_error) {
2319                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2320                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2321                         /* Note that this cannot happen on a
2322                          * replacement device.  We just fail those on
2323                          * any error
2324                          */
2325                         printk_ratelimited(
2326                                 KERN_INFO
2327                                 "md/raid:%s: read error corrected"
2328                                 " (%lu sectors at %llu on %s)\n",
2329                                 mdname(conf->mddev), STRIPE_SECTORS,
2330                                 (unsigned long long)s,
2331                                 bdevname(rdev->bdev, b));
2332                         atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2333                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2334                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2335                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2336                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2337
2338                 if (atomic_read(&rdev->read_errors))
2339                         atomic_set(&rdev->read_errors, 0);
2340         } else {
2341                 const char *bdn = bdevname(rdev->bdev, b);
2342                 int retry = 0;
2343                 int set_bad = 0;
2344
2345                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2346                 atomic_inc(&rdev->read_errors);
2347                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2348                         printk_ratelimited(
2349                                 KERN_WARNING
2350                                 "md/raid:%s: read error on replacement device "
2351                                 "(sector %llu on %s).\n",
2352                                 mdname(conf->mddev),
2353                                 (unsigned long long)s,
2354                                 bdn);
2355                 else if (conf->mddev->degraded >= conf->max_degraded) {
2356                         set_bad = 1;
2357                         printk_ratelimited(
2358                                 KERN_WARNING
2359                                 "md/raid:%s: read error not correctable "
2360                                 "(sector %llu on %s).\n",
2361                                 mdname(conf->mddev),
2362                                 (unsigned long long)s,
2363                                 bdn);
2364                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2365                         /* Oh, no!!! */
2366                         set_bad = 1;
2367                         printk_ratelimited(
2368                                 KERN_WARNING
2369                                 "md/raid:%s: read error NOT corrected!! "
2370                                 "(sector %llu on %s).\n",
2371                                 mdname(conf->mddev),
2372                                 (unsigned long long)s,
2373                                 bdn);
2374                 } else if (atomic_read(&rdev->read_errors)
2375                          > conf->max_nr_stripes)
2376                         printk(KERN_WARNING
2377                                "md/raid:%s: Too many read errors, failing device %s.\n",
2378                                mdname(conf->mddev), bdn);
2379                 else
2380                         retry = 1;
2381                 if (set_bad && test_bit(In_sync, &rdev->flags)
2382                     && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2383                         retry = 1;
2384                 if (retry)
2385                         if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2386                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2387                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2388                         } else
2389                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2390                 else {
2391                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2392                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2393                         if (!(set_bad
2394                               && test_bit(In_sync, &rdev->flags)
2395                               && rdev_set_badblocks(
2396                                       rdev, sh->sector, STRIPE_SECTORS, 0)))
2397                                 md_error(conf->mddev, rdev);
2398                 }
2399         }
2400         rdev_dec_pending(rdev, conf->mddev);
2401         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2402         set_bit(STRIPE_HANDLE, &sh->state);
2403         raid5_release_stripe(sh);
2404 }
2405
2406 static void raid5_end_write_request(struct bio *bi)
2407 {
2408         struct stripe_head *sh = bi->bi_private;
2409         struct r5conf *conf = sh->raid_conf;
2410         int disks = sh->disks, i;
2411         struct md_rdev *uninitialized_var(rdev);
2412         sector_t first_bad;
2413         int bad_sectors;
2414         int replacement = 0;
2415
2416         for (i = 0 ; i < disks; i++) {
2417                 if (bi == &sh->dev[i].req) {
2418                         rdev = conf->disks[i].rdev;
2419                         break;
2420                 }
2421                 if (bi == &sh->dev[i].rreq) {
2422                         rdev = conf->disks[i].replacement;
2423                         if (rdev)
2424                                 replacement = 1;
2425                         else
2426                                 /* rdev was removed and 'replacement'
2427                                  * replaced it.  rdev is not removed
2428                                  * until all requests are finished.
2429                                  */
2430                                 rdev = conf->disks[i].rdev;
2431                         break;
2432                 }
2433         }
2434         pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2435                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2436                 bi->bi_error);
2437         if (i == disks) {
2438                 BUG();
2439                 return;
2440         }
2441
2442         if (replacement) {
2443                 if (bi->bi_error)
2444                         md_error(conf->mddev, rdev);
2445                 else if (is_badblock(rdev, sh->sector,
2446                                      STRIPE_SECTORS,
2447                                      &first_bad, &bad_sectors))
2448                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2449         } else {
2450                 if (bi->bi_error) {
2451                         set_bit(STRIPE_DEGRADED, &sh->state);
2452                         set_bit(WriteErrorSeen, &rdev->flags);
2453                         set_bit(R5_WriteError, &sh->dev[i].flags);
2454                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2455                                 set_bit(MD_RECOVERY_NEEDED,
2456                                         &rdev->mddev->recovery);
2457                 } else if (is_badblock(rdev, sh->sector,
2458                                        STRIPE_SECTORS,
2459                                        &first_bad, &bad_sectors)) {
2460                         set_bit(R5_MadeGood, &sh->dev[i].flags);
2461                         if (test_bit(R5_ReadError, &sh->dev[i].flags))
2462                                 /* That was a successful write so make
2463                                  * sure it looks like we already did
2464                                  * a re-write.
2465                                  */
2466                                 set_bit(R5_ReWrite, &sh->dev[i].flags);
2467                 }
2468         }
2469         rdev_dec_pending(rdev, conf->mddev);
2470
2471         if (sh->batch_head && bi->bi_error && !replacement)
2472                 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2473
2474         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2475                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2476         set_bit(STRIPE_HANDLE, &sh->state);
2477         raid5_release_stripe(sh);
2478
2479         if (sh->batch_head && sh != sh->batch_head)
2480                 raid5_release_stripe(sh->batch_head);
2481 }
2482
2483 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2484 {
2485         struct r5dev *dev = &sh->dev[i];
2486
2487         bio_init(&dev->req);
2488         dev->req.bi_io_vec = &dev->vec;
2489         dev->req.bi_max_vecs = 1;
2490         dev->req.bi_private = sh;
2491
2492         bio_init(&dev->rreq);
2493         dev->rreq.bi_io_vec = &dev->rvec;
2494         dev->rreq.bi_max_vecs = 1;
2495         dev->rreq.bi_private = sh;
2496
2497         dev->flags = 0;
2498         dev->sector = raid5_compute_blocknr(sh, i, previous);
2499 }
2500
2501 static void error(struct mddev *mddev, struct md_rdev *rdev)
2502 {
2503         char b[BDEVNAME_SIZE];
2504         struct r5conf *conf = mddev->private;
2505         unsigned long flags;
2506         pr_debug("raid456: error called\n");
2507
2508         spin_lock_irqsave(&conf->device_lock, flags);
2509         clear_bit(In_sync, &rdev->flags);
2510         mddev->degraded = calc_degraded(conf);
2511         spin_unlock_irqrestore(&conf->device_lock, flags);
2512         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2513
2514         set_bit(Blocked, &rdev->flags);
2515         set_bit(Faulty, &rdev->flags);
2516         set_bit(MD_CHANGE_DEVS, &mddev->flags);
2517         set_bit(MD_CHANGE_PENDING, &mddev->flags);
2518         printk(KERN_ALERT
2519                "md/raid:%s: Disk failure on %s, disabling device.\n"
2520                "md/raid:%s: Operation continuing on %d devices.\n",
2521                mdname(mddev),
2522                bdevname(rdev->bdev, b),
2523                mdname(mddev),
2524                conf->raid_disks - mddev->degraded);
2525 }
2526
2527 /*
2528  * Input: a 'big' sector number,
2529  * Output: index of the data and parity disk, and the sector # in them.
2530  */
2531 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2532                               int previous, int *dd_idx,
2533                               struct stripe_head *sh)
2534 {
2535         sector_t stripe, stripe2;
2536         sector_t chunk_number;
2537         unsigned int chunk_offset;
2538         int pd_idx, qd_idx;
2539         int ddf_layout = 0;
2540         sector_t new_sector;
2541         int algorithm = previous ? conf->prev_algo
2542                                  : conf->algorithm;
2543         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2544                                          : conf->chunk_sectors;
2545         int raid_disks = previous ? conf->previous_raid_disks
2546                                   : conf->raid_disks;
2547         int data_disks = raid_disks - conf->max_degraded;
2548
2549         /* First compute the information on this sector */
2550
2551         /*
2552          * Compute the chunk number and the sector offset inside the chunk
2553          */
2554         chunk_offset = sector_div(r_sector, sectors_per_chunk);
2555         chunk_number = r_sector;
2556
2557         /*
2558          * Compute the stripe number
2559          */
2560         stripe = chunk_number;
2561         *dd_idx = sector_div(stripe, data_disks);
2562         stripe2 = stripe;
2563         /*
2564          * Select the parity disk based on the user selected algorithm.
2565          */
2566         pd_idx = qd_idx = -1;
2567         switch(conf->level) {
2568         case 4:
2569                 pd_idx = data_disks;
2570                 break;
2571         case 5:
2572                 switch (algorithm) {
2573                 case ALGORITHM_LEFT_ASYMMETRIC:
2574                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2575                         if (*dd_idx >= pd_idx)
2576                                 (*dd_idx)++;
2577                         break;
2578                 case ALGORITHM_RIGHT_ASYMMETRIC:
2579                         pd_idx = sector_div(stripe2, raid_disks);
2580                         if (*dd_idx >= pd_idx)
2581                                 (*dd_idx)++;
2582                         break;
2583                 case ALGORITHM_LEFT_SYMMETRIC:
2584                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2585                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2586                         break;
2587                 case ALGORITHM_RIGHT_SYMMETRIC:
2588                         pd_idx = sector_div(stripe2, raid_disks);
2589                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2590                         break;
2591                 case ALGORITHM_PARITY_0:
2592                         pd_idx = 0;
2593                         (*dd_idx)++;
2594                         break;
2595                 case ALGORITHM_PARITY_N:
2596                         pd_idx = data_disks;
2597                         break;
2598                 default:
2599                         BUG();
2600                 }
2601                 break;
2602         case 6:
2603
2604                 switch (algorithm) {
2605                 case ALGORITHM_LEFT_ASYMMETRIC:
2606                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2607                         qd_idx = pd_idx + 1;
2608                         if (pd_idx == raid_disks-1) {
2609                                 (*dd_idx)++;    /* Q D D D P */
2610                                 qd_idx = 0;
2611                         } else if (*dd_idx >= pd_idx)
2612                                 (*dd_idx) += 2; /* D D P Q D */
2613                         break;
2614                 case ALGORITHM_RIGHT_ASYMMETRIC:
2615                         pd_idx = sector_div(stripe2, raid_disks);
2616                         qd_idx = pd_idx + 1;
2617                         if (pd_idx == raid_disks-1) {
2618                                 (*dd_idx)++;    /* Q D D D P */
2619                                 qd_idx = 0;
2620                         } else if (*dd_idx >= pd_idx)
2621                                 (*dd_idx) += 2; /* D D P Q D */
2622                         break;
2623                 case ALGORITHM_LEFT_SYMMETRIC:
2624                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2625                         qd_idx = (pd_idx + 1) % raid_disks;
2626                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2627                         break;
2628                 case ALGORITHM_RIGHT_SYMMETRIC:
2629                         pd_idx = sector_div(stripe2, raid_disks);
2630                         qd_idx = (pd_idx + 1) % raid_disks;
2631                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2632                         break;
2633
2634                 case ALGORITHM_PARITY_0:
2635                         pd_idx = 0;
2636                         qd_idx = 1;
2637                         (*dd_idx) += 2;
2638                         break;
2639                 case ALGORITHM_PARITY_N:
2640                         pd_idx = data_disks;
2641                         qd_idx = data_disks + 1;
2642                         break;
2643
2644                 case ALGORITHM_ROTATING_ZERO_RESTART:
2645                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
2646                          * of blocks for computing Q is different.
2647                          */
2648                         pd_idx = sector_div(stripe2, raid_disks);
2649                         qd_idx = pd_idx + 1;
2650                         if (pd_idx == raid_disks-1) {
2651                                 (*dd_idx)++;    /* Q D D D P */
2652                                 qd_idx = 0;
2653                         } else if (*dd_idx >= pd_idx)
2654                                 (*dd_idx) += 2; /* D D P Q D */
2655                         ddf_layout = 1;
2656                         break;
2657
2658                 case ALGORITHM_ROTATING_N_RESTART:
2659                         /* Same a left_asymmetric, by first stripe is
2660                          * D D D P Q  rather than
2661                          * Q D D D P
2662                          */
2663                         stripe2 += 1;
2664                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2665                         qd_idx = pd_idx + 1;
2666                         if (pd_idx == raid_disks-1) {
2667                                 (*dd_idx)++;    /* Q D D D P */
2668                                 qd_idx = 0;
2669                         } else if (*dd_idx >= pd_idx)
2670                                 (*dd_idx) += 2; /* D D P Q D */
2671                         ddf_layout = 1;
2672                         break;
2673
2674                 case ALGORITHM_ROTATING_N_CONTINUE:
2675                         /* Same as left_symmetric but Q is before P */
2676                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2677                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2678                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2679                         ddf_layout = 1;
2680                         break;
2681
2682                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2683                         /* RAID5 left_asymmetric, with Q on last device */
2684                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2685                         if (*dd_idx >= pd_idx)
2686                                 (*dd_idx)++;
2687                         qd_idx = raid_disks - 1;
2688                         break;
2689
2690                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2691                         pd_idx = sector_div(stripe2, raid_disks-1);
2692                         if (*dd_idx >= pd_idx)
2693                                 (*dd_idx)++;
2694                         qd_idx = raid_disks - 1;
2695                         break;
2696
2697                 case ALGORITHM_LEFT_SYMMETRIC_6:
2698                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2699                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2700                         qd_idx = raid_disks - 1;
2701                         break;
2702
2703                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2704                         pd_idx = sector_div(stripe2, raid_disks-1);
2705                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2706                         qd_idx = raid_disks - 1;
2707                         break;
2708
2709                 case ALGORITHM_PARITY_0_6:
2710                         pd_idx = 0;
2711                         (*dd_idx)++;
2712                         qd_idx = raid_disks - 1;
2713                         break;
2714
2715                 default:
2716                         BUG();
2717                 }
2718                 break;
2719         }
2720
2721         if (sh) {
2722                 sh->pd_idx = pd_idx;
2723                 sh->qd_idx = qd_idx;
2724                 sh->ddf_layout = ddf_layout;
2725         }
2726         /*
2727          * Finally, compute the new sector number
2728          */
2729         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2730         return new_sector;
2731 }
2732
2733 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2734 {
2735         struct r5conf *conf = sh->raid_conf;
2736         int raid_disks = sh->disks;
2737         int data_disks = raid_disks - conf->max_degraded;
2738         sector_t new_sector = sh->sector, check;
2739         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2740                                          : conf->chunk_sectors;
2741         int algorithm = previous ? conf->prev_algo
2742                                  : conf->algorithm;
2743         sector_t stripe;
2744         int chunk_offset;
2745         sector_t chunk_number;
2746         int dummy1, dd_idx = i;
2747         sector_t r_sector;
2748         struct stripe_head sh2;
2749
2750         chunk_offset = sector_div(new_sector, sectors_per_chunk);
2751         stripe = new_sector;
2752
2753         if (i == sh->pd_idx)
2754                 return 0;
2755         switch(conf->level) {
2756         case 4: break;
2757         case 5:
2758                 switch (algorithm) {
2759                 case ALGORITHM_LEFT_ASYMMETRIC:
2760                 case ALGORITHM_RIGHT_ASYMMETRIC:
2761                         if (i > sh->pd_idx)
2762                                 i--;
2763                         break;
2764                 case ALGORITHM_LEFT_SYMMETRIC:
2765                 case ALGORITHM_RIGHT_SYMMETRIC:
2766                         if (i < sh->pd_idx)
2767                                 i += raid_disks;
2768                         i -= (sh->pd_idx + 1);
2769                         break;
2770                 case ALGORITHM_PARITY_0:
2771                         i -= 1;
2772                         break;
2773                 case ALGORITHM_PARITY_N:
2774                         break;
2775                 default:
2776                         BUG();
2777                 }
2778                 break;
2779         case 6:
2780                 if (i == sh->qd_idx)
2781                         return 0; /* It is the Q disk */
2782                 switch (algorithm) {
2783                 case ALGORITHM_LEFT_ASYMMETRIC:
2784                 case ALGORITHM_RIGHT_ASYMMETRIC:
2785                 case ALGORITHM_ROTATING_ZERO_RESTART:
2786                 case ALGORITHM_ROTATING_N_RESTART:
2787                         if (sh->pd_idx == raid_disks-1)
2788                                 i--;    /* Q D D D P */
2789                         else if (i > sh->pd_idx)
2790                                 i -= 2; /* D D P Q D */
2791                         break;
2792                 case ALGORITHM_LEFT_SYMMETRIC:
2793                 case ALGORITHM_RIGHT_SYMMETRIC:
2794                         if (sh->pd_idx == raid_disks-1)
2795                                 i--; /* Q D D D P */
2796                         else {
2797                                 /* D D P Q D */
2798                                 if (i < sh->pd_idx)
2799                                         i += raid_disks;
2800                                 i -= (sh->pd_idx + 2);
2801                         }
2802                         break;
2803                 case ALGORITHM_PARITY_0:
2804                         i -= 2;
2805                         break;
2806                 case ALGORITHM_PARITY_N:
2807                         break;
2808                 case ALGORITHM_ROTATING_N_CONTINUE:
2809                         /* Like left_symmetric, but P is before Q */
2810                         if (sh->pd_idx == 0)
2811                                 i--;    /* P D D D Q */
2812                         else {
2813                                 /* D D Q P D */
2814                                 if (i < sh->pd_idx)
2815                                         i += raid_disks;
2816                                 i -= (sh->pd_idx + 1);
2817                         }
2818                         break;
2819                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2820                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2821                         if (i > sh->pd_idx)
2822                                 i--;
2823                         break;
2824                 case ALGORITHM_LEFT_SYMMETRIC_6:
2825                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2826                         if (i < sh->pd_idx)
2827                                 i += data_disks + 1;
2828                         i -= (sh->pd_idx + 1);
2829                         break;
2830                 case ALGORITHM_PARITY_0_6:
2831                         i -= 1;
2832                         break;
2833                 default:
2834                         BUG();
2835                 }
2836                 break;
2837         }
2838
2839         chunk_number = stripe * data_disks + i;
2840         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2841
2842         check = raid5_compute_sector(conf, r_sector,
2843                                      previous, &dummy1, &sh2);
2844         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2845                 || sh2.qd_idx != sh->qd_idx) {
2846                 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2847                        mdname(conf->mddev));
2848                 return 0;
2849         }
2850         return r_sector;
2851 }
2852
2853 static void
2854 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2855                          int rcw, int expand)
2856 {
2857         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2858         struct r5conf *conf = sh->raid_conf;
2859         int level = conf->level;
2860
2861         if (rcw) {
2862
2863                 for (i = disks; i--; ) {
2864                         struct r5dev *dev = &sh->dev[i];
2865
2866                         if (dev->towrite) {
2867                                 set_bit(R5_LOCKED, &dev->flags);
2868                                 set_bit(R5_Wantdrain, &dev->flags);
2869                                 if (!expand)
2870                                         clear_bit(R5_UPTODATE, &dev->flags);
2871                                 s->locked++;
2872                         }
2873                 }
2874                 /* if we are not expanding this is a proper write request, and
2875                  * there will be bios with new data to be drained into the
2876                  * stripe cache
2877                  */
2878                 if (!expand) {
2879                         if (!s->locked)
2880                                 /* False alarm, nothing to do */
2881                                 return;
2882                         sh->reconstruct_state = reconstruct_state_drain_run;
2883                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2884                 } else
2885                         sh->reconstruct_state = reconstruct_state_run;
2886
2887                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2888
2889                 if (s->locked + conf->max_degraded == disks)
2890                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2891                                 atomic_inc(&conf->pending_full_writes);
2892         } else {
2893                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2894                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2895                 BUG_ON(level == 6 &&
2896                         (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2897                            test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2898
2899                 for (i = disks; i--; ) {
2900                         struct r5dev *dev = &sh->dev[i];
2901                         if (i == pd_idx || i == qd_idx)
2902                                 continue;
2903
2904                         if (dev->towrite &&
2905                             (test_bit(R5_UPTODATE, &dev->flags) ||
2906                              test_bit(R5_Wantcompute, &dev->flags))) {
2907                                 set_bit(R5_Wantdrain, &dev->flags);
2908                                 set_bit(R5_LOCKED, &dev->flags);
2909                                 clear_bit(R5_UPTODATE, &dev->flags);
2910                                 s->locked++;
2911                         }
2912                 }
2913                 if (!s->locked)
2914                         /* False alarm - nothing to do */
2915                         return;
2916                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2917                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2918                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2919                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2920         }
2921
2922         /* keep the parity disk(s) locked while asynchronous operations
2923          * are in flight
2924          */
2925         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2926         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2927         s->locked++;
2928
2929         if (level == 6) {
2930                 int qd_idx = sh->qd_idx;
2931                 struct r5dev *dev = &sh->dev[qd_idx];
2932
2933                 set_bit(R5_LOCKED, &dev->flags);
2934                 clear_bit(R5_UPTODATE, &dev->flags);
2935                 s->locked++;
2936         }
2937
2938         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2939                 __func__, (unsigned long long)sh->sector,
2940                 s->locked, s->ops_request);
2941 }
2942
2943 /*
2944  * Each stripe/dev can have one or more bion attached.
2945  * toread/towrite point to the first in a chain.
2946  * The bi_next chain must be in order.
2947  */
2948 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2949                           int forwrite, int previous)
2950 {
2951         struct bio **bip;
2952         struct r5conf *conf = sh->raid_conf;
2953         int firstwrite=0;
2954
2955         pr_debug("adding bi b#%llu to stripe s#%llu\n",
2956                 (unsigned long long)bi->bi_iter.bi_sector,
2957                 (unsigned long long)sh->sector);
2958
2959         /*
2960          * If several bio share a stripe. The bio bi_phys_segments acts as a
2961          * reference count to avoid race. The reference count should already be
2962          * increased before this function is called (for example, in
2963          * make_request()), so other bio sharing this stripe will not free the
2964          * stripe. If a stripe is owned by one stripe, the stripe lock will
2965          * protect it.
2966          */
2967         spin_lock_irq(&sh->stripe_lock);
2968         /* Don't allow new IO added to stripes in batch list */
2969         if (sh->batch_head)
2970                 goto overlap;
2971         if (forwrite) {
2972                 bip = &sh->dev[dd_idx].towrite;
2973                 if (*bip == NULL)
2974                         firstwrite = 1;
2975         } else
2976                 bip = &sh->dev[dd_idx].toread;
2977         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2978                 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2979                         goto overlap;
2980                 bip = & (*bip)->bi_next;
2981         }
2982         if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2983                 goto overlap;
2984
2985         if (!forwrite || previous)
2986                 clear_bit(STRIPE_BATCH_READY, &sh->state);
2987
2988         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2989         if (*bip)
2990                 bi->bi_next = *bip;
2991         *bip = bi;
2992         raid5_inc_bi_active_stripes(bi);
2993
2994         if (forwrite) {
2995                 /* check if page is covered */
2996                 sector_t sector = sh->dev[dd_idx].sector;
2997                 for (bi=sh->dev[dd_idx].towrite;
2998                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2999                              bi && bi->bi_iter.bi_sector <= sector;
3000                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3001                         if (bio_end_sector(bi) >= sector)
3002                                 sector = bio_end_sector(bi);
3003                 }
3004                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3005                         if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3006                                 sh->overwrite_disks++;
3007         }
3008
3009         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3010                 (unsigned long long)(*bip)->bi_iter.bi_sector,
3011                 (unsigned long long)sh->sector, dd_idx);
3012
3013         if (conf->mddev->bitmap && firstwrite) {
3014                 /* Cannot hold spinlock over bitmap_startwrite,
3015                  * but must ensure this isn't added to a batch until
3016                  * we have added to the bitmap and set bm_seq.
3017                  * So set STRIPE_BITMAP_PENDING to prevent
3018                  * batching.
3019                  * If multiple add_stripe_bio() calls race here they
3020                  * much all set STRIPE_BITMAP_PENDING.  So only the first one
3021                  * to complete "bitmap_startwrite" gets to set
3022                  * STRIPE_BIT_DELAY.  This is important as once a stripe
3023                  * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3024                  * any more.
3025                  */
3026                 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3027                 spin_unlock_irq(&sh->stripe_lock);
3028                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3029                                   STRIPE_SECTORS, 0);
3030                 spin_lock_irq(&sh->stripe_lock);
3031                 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3032                 if (!sh->batch_head) {
3033                         sh->bm_seq = conf->seq_flush+1;
3034                         set_bit(STRIPE_BIT_DELAY, &sh->state);
3035                 }
3036         }
3037         spin_unlock_irq(&sh->stripe_lock);
3038
3039         if (stripe_can_batch(sh))
3040                 stripe_add_to_batch_list(conf, sh);
3041         return 1;
3042
3043  overlap:
3044         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3045         spin_unlock_irq(&sh->stripe_lock);
3046         return 0;
3047 }
3048
3049 static void end_reshape(struct r5conf *conf);
3050
3051 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3052                             struct stripe_head *sh)
3053 {
3054         int sectors_per_chunk =
3055                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3056         int dd_idx;
3057         int chunk_offset = sector_div(stripe, sectors_per_chunk);
3058         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3059
3060         raid5_compute_sector(conf,
3061                              stripe * (disks - conf->max_degraded)
3062                              *sectors_per_chunk + chunk_offset,
3063                              previous,
3064                              &dd_idx, sh);
3065 }
3066
3067 static void
3068 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3069                                 struct stripe_head_state *s, int disks,
3070                                 struct bio_list *return_bi)
3071 {
3072         int i;
3073         BUG_ON(sh->batch_head);
3074         for (i = disks; i--; ) {
3075                 struct bio *bi;
3076                 int bitmap_end = 0;
3077
3078                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3079                         struct md_rdev *rdev;
3080                         rcu_read_lock();
3081                         rdev = rcu_dereference(conf->disks[i].rdev);
3082                         if (rdev && test_bit(In_sync, &rdev->flags))
3083                                 atomic_inc(&rdev->nr_pending);
3084                         else
3085                                 rdev = NULL;
3086                         rcu_read_unlock();
3087                         if (rdev) {
3088                                 if (!rdev_set_badblocks(
3089                                             rdev,
3090                                             sh->sector,
3091                                             STRIPE_SECTORS, 0))
3092                                         md_error(conf->mddev, rdev);
3093                                 rdev_dec_pending(rdev, conf->mddev);
3094                         }
3095                 }
3096                 spin_lock_irq(&sh->stripe_lock);
3097                 /* fail all writes first */
3098                 bi = sh->dev[i].towrite;
3099                 sh->dev[i].towrite = NULL;
3100                 sh->overwrite_disks = 0;
3101                 spin_unlock_irq(&sh->stripe_lock);
3102                 if (bi)
3103                         bitmap_end = 1;
3104
3105                 r5l_stripe_write_finished(sh);
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                                 bio_list_add(return_bi, bi);
3118                         }
3119                         bi = nextbi;
3120                 }
3121                 if (bitmap_end)
3122                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3123                                 STRIPE_SECTORS, 0, 0);
3124                 bitmap_end = 0;
3125                 /* and fail all 'written' */
3126                 bi = sh->dev[i].written;
3127                 sh->dev[i].written = NULL;
3128                 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3129                         WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3130                         sh->dev[i].page = sh->dev[i].orig_page;
3131                 }
3132
3133                 if (bi) bitmap_end = 1;
3134                 while (bi && bi->bi_iter.bi_sector <
3135                        sh->dev[i].sector + STRIPE_SECTORS) {
3136                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3137
3138                         bi->bi_error = -EIO;
3139                         if (!raid5_dec_bi_active_stripes(bi)) {
3140                                 md_write_end(conf->mddev);
3141                                 bio_list_add(return_bi, bi);
3142                         }
3143                         bi = bi2;
3144                 }
3145
3146                 /* fail any reads if this device is non-operational and
3147                  * the data has not reached the cache yet.
3148                  */
3149                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3150                     s->failed > conf->max_degraded &&
3151                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3152                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
3153                         spin_lock_irq(&sh->stripe_lock);
3154                         bi = sh->dev[i].toread;
3155                         sh->dev[i].toread = NULL;
3156                         spin_unlock_irq(&sh->stripe_lock);
3157                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3158                                 wake_up(&conf->wait_for_overlap);
3159                         if (bi)
3160                                 s->to_read--;
3161                         while (bi && bi->bi_iter.bi_sector <
3162                                sh->dev[i].sector + STRIPE_SECTORS) {
3163                                 struct bio *nextbi =
3164                                         r5_next_bio(bi, sh->dev[i].sector);
3165
3166                                 bi->bi_error = -EIO;
3167                                 if (!raid5_dec_bi_active_stripes(bi))
3168                                         bio_list_add(return_bi, bi);
3169                                 bi = nextbi;
3170                         }
3171                 }
3172                 if (bitmap_end)
3173                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3174                                         STRIPE_SECTORS, 0, 0);
3175                 /* If we were in the middle of a write the parity block might
3176                  * still be locked - so just clear all R5_LOCKED flags
3177                  */
3178                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3179         }
3180         s->to_write = 0;
3181         s->written = 0;
3182
3183         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3184                 if (atomic_dec_and_test(&conf->pending_full_writes))
3185                         md_wakeup_thread(conf->mddev->thread);
3186 }
3187
3188 static void
3189 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3190                    struct stripe_head_state *s)
3191 {
3192         int abort = 0;
3193         int i;
3194
3195         BUG_ON(sh->batch_head);
3196         clear_bit(STRIPE_SYNCING, &sh->state);
3197         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3198                 wake_up(&conf->wait_for_overlap);
3199         s->syncing = 0;
3200         s->replacing = 0;
3201         /* There is nothing more to do for sync/check/repair.
3202          * Don't even need to abort as that is handled elsewhere
3203          * if needed, and not always wanted e.g. if there is a known
3204          * bad block here.
3205          * For recover/replace we need to record a bad block on all
3206          * non-sync devices, or abort the recovery
3207          */
3208         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3209                 /* During recovery devices cannot be removed, so
3210                  * locking and refcounting of rdevs is not needed
3211                  */
3212                 for (i = 0; i < conf->raid_disks; i++) {
3213                         struct md_rdev *rdev = conf->disks[i].rdev;
3214                         if (rdev
3215                             && !test_bit(Faulty, &rdev->flags)
3216                             && !test_bit(In_sync, &rdev->flags)
3217                             && !rdev_set_badblocks(rdev, sh->sector,
3218                                                    STRIPE_SECTORS, 0))
3219                                 abort = 1;
3220                         rdev = conf->disks[i].replacement;
3221                         if (rdev
3222                             && !test_bit(Faulty, &rdev->flags)
3223                             && !test_bit(In_sync, &rdev->flags)
3224                             && !rdev_set_badblocks(rdev, sh->sector,
3225                                                    STRIPE_SECTORS, 0))
3226                                 abort = 1;
3227                 }
3228                 if (abort)
3229                         conf->recovery_disabled =
3230                                 conf->mddev->recovery_disabled;
3231         }
3232         md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3233 }
3234
3235 static int want_replace(struct stripe_head *sh, int disk_idx)
3236 {
3237         struct md_rdev *rdev;
3238         int rv = 0;
3239         /* Doing recovery so rcu locking not required */
3240         rdev = sh->raid_conf->disks[disk_idx].replacement;
3241         if (rdev
3242             && !test_bit(Faulty, &rdev->flags)
3243             && !test_bit(In_sync, &rdev->flags)
3244             && (rdev->recovery_offset <= sh->sector
3245                 || rdev->mddev->recovery_cp <= sh->sector))
3246                 rv = 1;
3247
3248         return rv;
3249 }
3250
3251 /* fetch_block - checks the given member device to see if its data needs
3252  * to be read or computed to satisfy a request.
3253  *
3254  * Returns 1 when no more member devices need to be checked, otherwise returns
3255  * 0 to tell the loop in handle_stripe_fill to continue
3256  */
3257
3258 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3259                            int disk_idx, int disks)
3260 {
3261         struct r5dev *dev = &sh->dev[disk_idx];
3262         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3263                                   &sh->dev[s->failed_num[1]] };
3264         int i;
3265
3266
3267         if (test_bit(R5_LOCKED, &dev->flags) ||
3268             test_bit(R5_UPTODATE, &dev->flags))
3269                 /* No point reading this as we already have it or have
3270                  * decided to get it.
3271                  */
3272                 return 0;
3273
3274         if (dev->toread ||
3275             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3276                 /* We need this block to directly satisfy a request */
3277                 return 1;
3278
3279         if (s->syncing || s->expanding ||
3280             (s->replacing && want_replace(sh, disk_idx)))
3281                 /* When syncing, or expanding we read everything.
3282                  * When replacing, we need the replaced block.
3283                  */
3284                 return 1;
3285
3286         if ((s->failed >= 1 && fdev[0]->toread) ||
3287             (s->failed >= 2 && fdev[1]->toread))
3288                 /* If we want to read from a failed device, then
3289                  * we need to actually read every other device.
3290                  */
3291                 return 1;
3292
3293         /* Sometimes neither read-modify-write nor reconstruct-write
3294          * cycles can work.  In those cases we read every block we
3295          * can.  Then the parity-update is certain to have enough to
3296          * work with.
3297          * This can only be a problem when we need to write something,
3298          * and some device has failed.  If either of those tests
3299          * fail we need look no further.
3300          */
3301         if (!s->failed || !s->to_write)
3302                 return 0;
3303
3304         if (test_bit(R5_Insync, &dev->flags) &&
3305             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3306                 /* Pre-reads at not permitted until after short delay
3307                  * to gather multiple requests.  However if this
3308                  * device is no Insync, the block could only be be computed
3309                  * and there is no need to delay that.
3310                  */
3311                 return 0;
3312
3313         for (i = 0; i < s->failed && i < 2; i++) {
3314                 if (fdev[i]->towrite &&
3315                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3316                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3317                         /* If we have a partial write to a failed
3318                          * device, then we will need to reconstruct
3319                          * the content of that device, so all other
3320                          * devices must be read.
3321                          */
3322                         return 1;
3323         }
3324
3325         /* If we are forced to do a reconstruct-write, either because
3326          * the current RAID6 implementation only supports that, or
3327          * or because parity cannot be trusted and we are currently
3328          * recovering it, there is extra need to be careful.
3329          * If one of the devices that we would need to read, because
3330          * it is not being overwritten (and maybe not written at all)
3331          * is missing/faulty, then we need to read everything we can.
3332          */
3333         if (sh->raid_conf->level != 6 &&
3334             sh->sector < sh->raid_conf->mddev->recovery_cp)
3335                 /* reconstruct-write isn't being forced */
3336                 return 0;
3337         for (i = 0; i < s->failed && i < 2; i++) {
3338                 if (s->failed_num[i] != sh->pd_idx &&
3339                     s->failed_num[i] != sh->qd_idx &&
3340                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3341                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3342                         return 1;
3343         }
3344
3345         return 0;
3346 }
3347
3348 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3349                        int disk_idx, int disks)
3350 {
3351         struct r5dev *dev = &sh->dev[disk_idx];
3352
3353         /* is the data in this block needed, and can we get it? */
3354         if (need_this_block(sh, s, disk_idx, disks)) {
3355                 /* we would like to get this block, possibly by computing it,
3356                  * otherwise read it if the backing disk is insync
3357                  */
3358                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3359                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3360                 BUG_ON(sh->batch_head);
3361                 if ((s->uptodate == disks - 1) &&
3362                     (s->failed && (disk_idx == s->failed_num[0] ||
3363                                    disk_idx == s->failed_num[1]))) {
3364                         /* have disk failed, and we're requested to fetch it;
3365                          * do compute it
3366                          */
3367                         pr_debug("Computing stripe %llu block %d\n",
3368                                (unsigned long long)sh->sector, disk_idx);
3369                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3370                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3371                         set_bit(R5_Wantcompute, &dev->flags);
3372                         sh->ops.target = disk_idx;
3373                         sh->ops.target2 = -1; /* no 2nd target */
3374                         s->req_compute = 1;
3375                         /* Careful: from this point on 'uptodate' is in the eye
3376                          * of raid_run_ops which services 'compute' operations
3377                          * before writes. R5_Wantcompute flags a block that will
3378                          * be R5_UPTODATE by the time it is needed for a
3379                          * subsequent operation.
3380                          */
3381                         s->uptodate++;
3382                         return 1;
3383                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3384                         /* Computing 2-failure is *very* expensive; only
3385                          * do it if failed >= 2
3386                          */
3387                         int other;
3388                         for (other = disks; other--; ) {
3389                                 if (other == disk_idx)
3390                                         continue;
3391                                 if (!test_bit(R5_UPTODATE,
3392                                       &sh->dev[other].flags))
3393                                         break;
3394                         }
3395                         BUG_ON(other < 0);
3396                         pr_debug("Computing stripe %llu blocks %d,%d\n",
3397                                (unsigned long long)sh->sector,
3398                                disk_idx, other);
3399                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3400                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3401                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3402                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
3403                         sh->ops.target = disk_idx;
3404                         sh->ops.target2 = other;
3405                         s->uptodate += 2;
3406                         s->req_compute = 1;
3407                         return 1;
3408                 } else if (test_bit(R5_Insync, &dev->flags)) {
3409                         set_bit(R5_LOCKED, &dev->flags);
3410                         set_bit(R5_Wantread, &dev->flags);
3411                         s->locked++;
3412                         pr_debug("Reading block %d (sync=%d)\n",
3413                                 disk_idx, s->syncing);
3414                 }
3415         }
3416
3417         return 0;
3418 }
3419
3420 /**
3421  * handle_stripe_fill - read or compute data to satisfy pending requests.
3422  */
3423 static void handle_stripe_fill(struct stripe_head *sh,
3424                                struct stripe_head_state *s,
3425                                int disks)
3426 {
3427         int i;
3428
3429         /* look for blocks to read/compute, skip this if a compute
3430          * is already in flight, or if the stripe contents are in the
3431          * midst of changing due to a write
3432          */
3433         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3434             !sh->reconstruct_state)
3435                 for (i = disks; i--; )
3436                         if (fetch_block(sh, s, i, disks))
3437                                 break;
3438         set_bit(STRIPE_HANDLE, &sh->state);
3439 }
3440
3441 static void break_stripe_batch_list(struct stripe_head *head_sh,
3442                                     unsigned long handle_flags);
3443 /* handle_stripe_clean_event
3444  * any written block on an uptodate or failed drive can be returned.
3445  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3446  * never LOCKED, so we don't need to test 'failed' directly.
3447  */
3448 static void handle_stripe_clean_event(struct r5conf *conf,
3449         struct stripe_head *sh, int disks, struct bio_list *return_bi)
3450 {
3451         int i;
3452         struct r5dev *dev;
3453         int discard_pending = 0;
3454         struct stripe_head *head_sh = sh;
3455         bool do_endio = false;
3456
3457         for (i = disks; i--; )
3458                 if (sh->dev[i].written) {
3459                         dev = &sh->dev[i];
3460                         if (!test_bit(R5_LOCKED, &dev->flags) &&
3461                             (test_bit(R5_UPTODATE, &dev->flags) ||
3462                              test_bit(R5_Discard, &dev->flags) ||
3463                              test_bit(R5_SkipCopy, &dev->flags))) {
3464                                 /* We can return any write requests */
3465                                 struct bio *wbi, *wbi2;
3466                                 pr_debug("Return write for disc %d\n", i);
3467                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
3468                                         clear_bit(R5_UPTODATE, &dev->flags);
3469                                 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3470                                         WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3471                                 }
3472                                 do_endio = true;
3473
3474 returnbi:
3475                                 dev->page = dev->orig_page;
3476                                 wbi = dev->written;
3477                                 dev->written = NULL;
3478                                 while (wbi && wbi->bi_iter.bi_sector <
3479                                         dev->sector + STRIPE_SECTORS) {
3480                                         wbi2 = r5_next_bio(wbi, dev->sector);
3481                                         if (!raid5_dec_bi_active_stripes(wbi)) {
3482                                                 md_write_end(conf->mddev);
3483                                                 bio_list_add(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
3508         r5l_stripe_write_finished(sh);
3509
3510         if (!discard_pending &&
3511             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3512                 int hash;
3513                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3514                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3515                 if (sh->qd_idx >= 0) {
3516                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3517                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3518                 }
3519                 /* now that discard is done we can proceed with any sync */
3520                 clear_bit(STRIPE_DISCARD, &sh->state);
3521                 /*
3522                  * SCSI discard will change some bio fields and the stripe has
3523                  * no updated data, so remove it from hash list and the stripe
3524                  * will be reinitialized
3525                  */
3526 unhash:
3527                 hash = sh->hash_lock_index;
3528                 spin_lock_irq(conf->hash_locks + hash);
3529                 remove_hash(sh);
3530                 spin_unlock_irq(conf->hash_locks + hash);
3531                 if (head_sh->batch_head) {
3532                         sh = list_first_entry(&sh->batch_list,
3533                                               struct stripe_head, batch_list);
3534                         if (sh != head_sh)
3535                                         goto unhash;
3536                 }
3537                 sh = head_sh;
3538
3539                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3540                         set_bit(STRIPE_HANDLE, &sh->state);
3541
3542         }
3543
3544         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3545                 if (atomic_dec_and_test(&conf->pending_full_writes))
3546                         md_wakeup_thread(conf->mddev->thread);
3547
3548         if (head_sh->batch_head && do_endio)
3549                 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3550 }
3551
3552 static void handle_stripe_dirtying(struct r5conf *conf,
3553                                    struct stripe_head *sh,
3554                                    struct stripe_head_state *s,
3555                                    int disks)
3556 {
3557         int rmw = 0, rcw = 0, i;
3558         sector_t recovery_cp = conf->mddev->recovery_cp;
3559
3560         /* Check whether resync is now happening or should start.
3561          * If yes, then the array is dirty (after unclean shutdown or
3562          * initial creation), so parity in some stripes might be inconsistent.
3563          * In this case, we need to always do reconstruct-write, to ensure
3564          * that in case of drive failure or read-error correction, we
3565          * generate correct data from the parity.
3566          */
3567         if (conf->rmw_level == PARITY_DISABLE_RMW ||
3568             (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3569              s->failed == 0)) {
3570                 /* Calculate the real rcw later - for now make it
3571                  * look like rcw is cheaper
3572                  */
3573                 rcw = 1; rmw = 2;
3574                 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3575                          conf->rmw_level, (unsigned long long)recovery_cp,
3576                          (unsigned long long)sh->sector);
3577         } else for (i = disks; i--; ) {
3578                 /* would I have to read this buffer for read_modify_write */
3579                 struct r5dev *dev = &sh->dev[i];
3580                 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3581                     !test_bit(R5_LOCKED, &dev->flags) &&
3582                     !(test_bit(R5_UPTODATE, &dev->flags) ||
3583                       test_bit(R5_Wantcompute, &dev->flags))) {
3584                         if (test_bit(R5_Insync, &dev->flags))
3585                                 rmw++;
3586                         else
3587                                 rmw += 2*disks;  /* cannot read it */
3588                 }
3589                 /* Would I have to read this buffer for reconstruct_write */
3590                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3591                     i != sh->pd_idx && i != sh->qd_idx &&
3592                     !test_bit(R5_LOCKED, &dev->flags) &&
3593                     !(test_bit(R5_UPTODATE, &dev->flags) ||
3594                     test_bit(R5_Wantcompute, &dev->flags))) {
3595                         if (test_bit(R5_Insync, &dev->flags))
3596                                 rcw++;
3597                         else
3598                                 rcw += 2*disks;
3599                 }
3600         }
3601         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3602                 (unsigned long long)sh->sector, rmw, rcw);
3603         set_bit(STRIPE_HANDLE, &sh->state);
3604         if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_ENABLE_RMW)) && rmw > 0) {
3605                 /* prefer read-modify-write, but need to get some data */
3606                 if (conf->mddev->queue)
3607                         blk_add_trace_msg(conf->mddev->queue,
3608                                           "raid5 rmw %llu %d",
3609                                           (unsigned long long)sh->sector, rmw);
3610                 for (i = disks; i--; ) {
3611                         struct r5dev *dev = &sh->dev[i];
3612                         if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3613                             !test_bit(R5_LOCKED, &dev->flags) &&
3614                             !(test_bit(R5_UPTODATE, &dev->flags) ||
3615                             test_bit(R5_Wantcompute, &dev->flags)) &&
3616                             test_bit(R5_Insync, &dev->flags)) {
3617                                 if (test_bit(STRIPE_PREREAD_ACTIVE,
3618                                              &sh->state)) {
3619                                         pr_debug("Read_old block %d for r-m-w\n",
3620                                                  i);
3621                                         set_bit(R5_LOCKED, &dev->flags);
3622                                         set_bit(R5_Wantread, &dev->flags);
3623                                         s->locked++;
3624                                 } else {
3625                                         set_bit(STRIPE_DELAYED, &sh->state);
3626                                         set_bit(STRIPE_HANDLE, &sh->state);
3627                                 }
3628                         }
3629                 }
3630         }
3631         if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_ENABLE_RMW)) && rcw > 0) {
3632                 /* want reconstruct write, but need to get some data */
3633                 int qread =0;
3634                 rcw = 0;
3635                 for (i = disks; i--; ) {
3636                         struct r5dev *dev = &sh->dev[i];
3637                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3638                             i != sh->pd_idx && i != sh->qd_idx &&
3639                             !test_bit(R5_LOCKED, &dev->flags) &&
3640                             !(test_bit(R5_UPTODATE, &dev->flags) ||
3641                               test_bit(R5_Wantcompute, &dev->flags))) {
3642                                 rcw++;
3643                                 if (test_bit(R5_Insync, &dev->flags) &&
3644                                     test_bit(STRIPE_PREREAD_ACTIVE,
3645                                              &sh->state)) {
3646                                         pr_debug("Read_old block "
3647                                                 "%d for Reconstruct\n", i);
3648                                         set_bit(R5_LOCKED, &dev->flags);
3649                                         set_bit(R5_Wantread, &dev->flags);
3650                                         s->locked++;
3651                                         qread++;
3652                                 } else {
3653                                         set_bit(STRIPE_DELAYED, &sh->state);
3654                                         set_bit(STRIPE_HANDLE, &sh->state);
3655                                 }
3656                         }
3657                 }
3658                 if (rcw && conf->mddev->queue)
3659                         blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3660                                           (unsigned long long)sh->sector,
3661                                           rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3662         }
3663
3664         if (rcw > disks && rmw > disks &&
3665             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3666                 set_bit(STRIPE_DELAYED, &sh->state);
3667
3668         /* now if nothing is locked, and if we have enough data,
3669          * we can start a write request
3670          */
3671         /* since handle_stripe can be called at any time we need to handle the
3672          * case where a compute block operation has been submitted and then a
3673          * subsequent call wants to start a write request.  raid_run_ops only
3674          * handles the case where compute block and reconstruct are requested
3675          * simultaneously.  If this is not the case then new writes need to be
3676          * held off until the compute completes.
3677          */
3678         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3679             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3680             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3681                 schedule_reconstruction(sh, s, rcw == 0, 0);
3682 }
3683
3684 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3685                                 struct stripe_head_state *s, int disks)
3686 {
3687         struct r5dev *dev = NULL;
3688
3689         BUG_ON(sh->batch_head);
3690         set_bit(STRIPE_HANDLE, &sh->state);
3691
3692         switch (sh->check_state) {
3693         case check_state_idle:
3694                 /* start a new check operation if there are no failures */
3695                 if (s->failed == 0) {
3696                         BUG_ON(s->uptodate != disks);
3697                         sh->check_state = check_state_run;
3698                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
3699                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3700                         s->uptodate--;
3701                         break;
3702                 }
3703                 dev = &sh->dev[s->failed_num[0]];
3704                 /* fall through */
3705         case check_state_compute_result:
3706                 sh->check_state = check_state_idle;
3707                 if (!dev)
3708                         dev = &sh->dev[sh->pd_idx];
3709
3710                 /* check that a write has not made the stripe insync */
3711                 if (test_bit(STRIPE_INSYNC, &sh->state))
3712                         break;
3713
3714                 /* either failed parity check, or recovery is happening */
3715                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3716                 BUG_ON(s->uptodate != disks);
3717
3718                 set_bit(R5_LOCKED, &dev->flags);
3719                 s->locked++;
3720                 set_bit(R5_Wantwrite, &dev->flags);
3721
3722                 clear_bit(STRIPE_DEGRADED, &sh->state);
3723                 set_bit(STRIPE_INSYNC, &sh->state);
3724                 break;
3725         case check_state_run:
3726                 break; /* we will be called again upon completion */
3727         case check_state_check_result:
3728                 sh->check_state = check_state_idle;
3729
3730                 /* if a failure occurred during the check operation, leave
3731                  * STRIPE_INSYNC not set and let the stripe be handled again
3732                  */
3733                 if (s->failed)
3734                         break;
3735
3736                 /* handle a successful check operation, if parity is correct
3737                  * we are done.  Otherwise update the mismatch count and repair
3738                  * parity if !MD_RECOVERY_CHECK
3739                  */
3740                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3741                         /* parity is correct (on disc,
3742                          * not in buffer any more)
3743                          */
3744                         set_bit(STRIPE_INSYNC, &sh->state);
3745                 else {
3746                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3747                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3748                                 /* don't try to repair!! */
3749                                 set_bit(STRIPE_INSYNC, &sh->state);
3750                         else {
3751                                 sh->check_state = check_state_compute_run;
3752                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3753                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3754                                 set_bit(R5_Wantcompute,
3755                                         &sh->dev[sh->pd_idx].flags);
3756                                 sh->ops.target = sh->pd_idx;
3757                                 sh->ops.target2 = -1;
3758                                 s->uptodate++;
3759                         }
3760                 }
3761                 break;
3762         case check_state_compute_run:
3763                 break;
3764         default:
3765                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3766                        __func__, sh->check_state,
3767                        (unsigned long long) sh->sector);
3768                 BUG();
3769         }
3770 }
3771
3772 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3773                                   struct stripe_head_state *s,
3774                                   int disks)
3775 {
3776         int pd_idx = sh->pd_idx;
3777         int qd_idx = sh->qd_idx;
3778         struct r5dev *dev;
3779
3780         BUG_ON(sh->batch_head);
3781         set_bit(STRIPE_HANDLE, &sh->state);
3782
3783         BUG_ON(s->failed > 2);
3784
3785         /* Want to check and possibly repair P and Q.
3786          * However there could be one 'failed' device, in which
3787          * case we can only check one of them, possibly using the
3788          * other to generate missing data
3789          */
3790
3791         switch (sh->check_state) {
3792         case check_state_idle:
3793                 /* start a new check operation if there are < 2 failures */
3794                 if (s->failed == s->q_failed) {
3795                         /* The only possible failed device holds Q, so it
3796                          * makes sense to check P (If anything else were failed,
3797                          * we would have used P to recreate it).
3798                          */
3799                         sh->check_state = check_state_run;
3800                 }
3801                 if (!s->q_failed && s->failed < 2) {
3802                         /* Q is not failed, and we didn't use it to generate
3803                          * anything, so it makes sense to check it
3804                          */
3805                         if (sh->check_state == check_state_run)
3806                                 sh->check_state = check_state_run_pq;
3807                         else
3808                                 sh->check_state = check_state_run_q;
3809                 }
3810
3811                 /* discard potentially stale zero_sum_result */
3812                 sh->ops.zero_sum_result = 0;
3813
3814                 if (sh->check_state == check_state_run) {
3815                         /* async_xor_zero_sum destroys the contents of P */
3816                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3817                         s->uptodate--;
3818                 }
3819                 if (sh->check_state >= check_state_run &&
3820                     sh->check_state <= check_state_run_pq) {
3821                         /* async_syndrome_zero_sum preserves P and Q, so
3822                          * no need to mark them !uptodate here
3823                          */
3824                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
3825                         break;
3826                 }
3827
3828                 /* we have 2-disk failure */
3829                 BUG_ON(s->failed != 2);
3830                 /* fall through */
3831         case check_state_compute_result:
3832                 sh->check_state = check_state_idle;
3833
3834                 /* check that a write has not made the stripe insync */
3835                 if (test_bit(STRIPE_INSYNC, &sh->state))
3836                         break;
3837
3838                 /* now write out any block on a failed drive,
3839                  * or P or Q if they were recomputed
3840                  */
3841                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3842                 if (s->failed == 2) {
3843                         dev = &sh->dev[s->failed_num[1]];
3844                         s->locked++;
3845                         set_bit(R5_LOCKED, &dev->flags);
3846                         set_bit(R5_Wantwrite, &dev->flags);
3847                 }
3848                 if (s->failed >= 1) {
3849                         dev = &sh->dev[s->failed_num[0]];
3850                         s->locked++;
3851                         set_bit(R5_LOCKED, &dev->flags);
3852                         set_bit(R5_Wantwrite, &dev->flags);
3853                 }
3854                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3855                         dev = &sh->dev[pd_idx];
3856                         s->locked++;
3857                         set_bit(R5_LOCKED, &dev->flags);
3858                         set_bit(R5_Wantwrite, &dev->flags);
3859                 }
3860                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3861                         dev = &sh->dev[qd_idx];
3862                         s->locked++;
3863                         set_bit(R5_LOCKED, &dev->flags);
3864                         set_bit(R5_Wantwrite, &dev->flags);
3865                 }
3866                 clear_bit(STRIPE_DEGRADED, &sh->state);
3867
3868                 set_bit(STRIPE_INSYNC, &sh->state);
3869                 break;
3870         case check_state_run:
3871         case check_state_run_q:
3872         case check_state_run_pq:
3873                 break; /* we will be called again upon completion */
3874         case check_state_check_result:
3875                 sh->check_state = check_state_idle;
3876
3877                 /* handle a successful check operation, if parity is correct
3878                  * we are done.  Otherwise update the mismatch count and repair
3879                  * parity if !MD_RECOVERY_CHECK
3880                  */
3881                 if (sh->ops.zero_sum_result == 0) {
3882                         /* both parities are correct */
3883                         if (!s->failed)
3884                                 set_bit(STRIPE_INSYNC, &sh->state);
3885                         else {
3886                                 /* in contrast to the raid5 case we can validate
3887                                  * parity, but still have a failure to write
3888                                  * back
3889                                  */
3890                                 sh->check_state = check_state_compute_result;
3891                                 /* Returning at this point means that we may go
3892                                  * off and bring p and/or q uptodate again so
3893                                  * we make sure to check zero_sum_result again
3894                                  * to verify if p or q need writeback
3895                                  */
3896                         }
3897                 } else {
3898                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3899                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3900                                 /* don't try to repair!! */
3901                                 set_bit(STRIPE_INSYNC, &sh->state);
3902                         else {
3903                                 int *target = &sh->ops.target;
3904
3905                                 sh->ops.target = -1;
3906                                 sh->ops.target2 = -1;
3907                                 sh->check_state = check_state_compute_run;
3908                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3909                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3910                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3911                                         set_bit(R5_Wantcompute,
3912                                                 &sh->dev[pd_idx].flags);
3913                                         *target = pd_idx;
3914                                         target = &sh->ops.target2;
3915                                         s->uptodate++;
3916                                 }
3917                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3918                                         set_bit(R5_Wantcompute,
3919                                                 &sh->dev[qd_idx].flags);
3920                                         *target = qd_idx;
3921                                         s->uptodate++;
3922                                 }
3923                         }
3924                 }
3925                 break;
3926         case check_state_compute_run:
3927                 break;
3928         default:
3929                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3930                        __func__, sh->check_state,
3931                        (unsigned long long) sh->sector);
3932                 BUG();
3933         }
3934 }
3935
3936 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3937 {
3938         int i;
3939
3940         /* We have read all the blocks in this stripe and now we need to
3941          * copy some of them into a target stripe for expand.
3942          */
3943         struct dma_async_tx_descriptor *tx = NULL;
3944         BUG_ON(sh->batch_head);
3945         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3946         for (i = 0; i < sh->disks; i++)
3947                 if (i != sh->pd_idx && i != sh->qd_idx) {
3948                         int dd_idx, j;
3949                         struct stripe_head *sh2;
3950                         struct async_submit_ctl submit;
3951
3952                         sector_t bn = raid5_compute_blocknr(sh, i, 1);
3953                         sector_t s = raid5_compute_sector(conf, bn, 0,
3954                                                           &dd_idx, NULL);
3955                         sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
3956                         if (sh2 == NULL)
3957                                 /* so far only the early blocks of this stripe
3958                                  * have been requested.  When later blocks
3959                                  * get requested, we will try again
3960                                  */
3961                                 continue;
3962                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3963                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3964                                 /* must have already done this block */
3965                                 raid5_release_stripe(sh2);
3966                                 continue;
3967                         }
3968
3969                         /* place all the copies on one channel */
3970                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3971                         tx = async_memcpy(sh2->dev[dd_idx].page,
3972                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
3973                                           &submit);
3974
3975                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3976                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3977                         for (j = 0; j < conf->raid_disks; j++)
3978                                 if (j != sh2->pd_idx &&
3979                                     j != sh2->qd_idx &&
3980                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
3981                                         break;
3982                         if (j == conf->raid_disks) {
3983                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3984                                 set_bit(STRIPE_HANDLE, &sh2->state);
3985                         }
3986                         raid5_release_stripe(sh2);
3987
3988                 }
3989         /* done submitting copies, wait for them to complete */
3990         async_tx_quiesce(&tx);
3991 }
3992
3993 /*
3994  * handle_stripe - do things to a stripe.
3995  *
3996  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3997  * state of various bits to see what needs to be done.
3998  * Possible results:
3999  *    return some read requests which now have data
4000  *    return some write requests which are safely on storage
4001  *    schedule a read on some buffers
4002  *    schedule a write of some buffers
4003  *    return confirmation of parity correctness
4004  *
4005  */
4006
4007 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4008 {
4009         struct r5conf *conf = sh->raid_conf;
4010         int disks = sh->disks;
4011         struct r5dev *dev;
4012         int i;
4013         int do_recovery = 0;
4014
4015         memset(s, 0, sizeof(*s));
4016
4017         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4018         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4019         s->failed_num[0] = -1;
4020         s->failed_num[1] = -1;
4021         s->log_failed = r5l_log_disk_error(conf);
4022
4023         /* Now to look around and see what can be done */
4024         rcu_read_lock();
4025         for (i=disks; i--; ) {
4026                 struct md_rdev *rdev;
4027                 sector_t first_bad;
4028                 int bad_sectors;
4029                 int is_bad = 0;
4030
4031                 dev = &sh->dev[i];
4032
4033                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4034                          i, dev->flags,
4035                          dev->toread, dev->towrite, dev->written);
4036                 /* maybe we can reply to a read
4037                  *
4038                  * new wantfill requests are only permitted while
4039                  * ops_complete_biofill is guaranteed to be inactive
4040                  */
4041                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4042                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4043                         set_bit(R5_Wantfill, &dev->flags);
4044
4045                 /* now count some things */
4046                 if (test_bit(R5_LOCKED, &dev->flags))
4047                         s->locked++;
4048                 if (test_bit(R5_UPTODATE, &dev->flags))
4049                         s->uptodate++;
4050                 if (test_bit(R5_Wantcompute, &dev->flags)) {
4051                         s->compute++;
4052                         BUG_ON(s->compute > 2);
4053                 }
4054
4055                 if (test_bit(R5_Wantfill, &dev->flags))
4056                         s->to_fill++;
4057                 else if (dev->toread)
4058                         s->to_read++;
4059                 if (dev->towrite) {
4060                         s->to_write++;
4061                         if (!test_bit(R5_OVERWRITE, &dev->flags))
4062                                 s->non_overwrite++;
4063                 }
4064                 if (dev->written)
4065                         s->written++;
4066                 /* Prefer to use the replacement for reads, but only
4067                  * if it is recovered enough and has no bad blocks.
4068                  */
4069                 rdev = rcu_dereference(conf->disks[i].replacement);
4070                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4071                     rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4072                     !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4073                                  &first_bad, &bad_sectors))
4074                         set_bit(R5_ReadRepl, &dev->flags);
4075                 else {
4076                         if (rdev && !test_bit(Faulty, &rdev->flags))
4077                                 set_bit(R5_NeedReplace, &dev->flags);
4078                         else
4079                                 clear_bit(R5_NeedReplace, &dev->flags);
4080                         rdev = rcu_dereference(conf->disks[i].rdev);
4081                         clear_bit(R5_ReadRepl, &dev->flags);
4082                 }
4083                 if (rdev && test_bit(Faulty, &rdev->flags))
4084                         rdev = NULL;
4085                 if (rdev) {
4086                         is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4087                                              &first_bad, &bad_sectors);
4088                         if (s->blocked_rdev == NULL
4089                             && (test_bit(Blocked, &rdev->flags)
4090                                 || is_bad < 0)) {
4091                                 if (is_bad < 0)
4092                                         set_bit(BlockedBadBlocks,
4093                                                 &rdev->flags);
4094                                 s->blocked_rdev = rdev;
4095                                 atomic_inc(&rdev->nr_pending);
4096                         }
4097                 }
4098                 clear_bit(R5_Insync, &dev->flags);
4099                 if (!rdev)
4100                         /* Not in-sync */;
4101                 else if (is_bad) {
4102                         /* also not in-sync */
4103                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4104                             test_bit(R5_UPTODATE, &dev->flags)) {
4105                                 /* treat as in-sync, but with a read error
4106                                  * which we can now try to correct
4107                                  */
4108                                 set_bit(R5_Insync, &dev->flags);
4109                                 set_bit(R5_ReadError, &dev->flags);
4110                         }
4111                 } else if (test_bit(In_sync, &rdev->flags))
4112                         set_bit(R5_Insync, &dev->flags);
4113                 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4114                         /* in sync if before recovery_offset */
4115                         set_bit(R5_Insync, &dev->flags);
4116                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4117                          test_bit(R5_Expanded, &dev->flags))
4118                         /* If we've reshaped into here, we assume it is Insync.
4119                          * We will shortly update recovery_offset to make
4120                          * it official.
4121                          */
4122                         set_bit(R5_Insync, &dev->flags);
4123
4124                 if (test_bit(R5_WriteError, &dev->flags)) {
4125                         /* This flag does not apply to '.replacement'
4126                          * only to .rdev, so make sure to check that*/
4127                         struct md_rdev *rdev2 = rcu_dereference(
4128                                 conf->disks[i].rdev);
4129                         if (rdev2 == rdev)
4130                                 clear_bit(R5_Insync, &dev->flags);
4131                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4132                                 s->handle_bad_blocks = 1;
4133                                 atomic_inc(&rdev2->nr_pending);
4134                         } else
4135                                 clear_bit(R5_WriteError, &dev->flags);
4136                 }
4137                 if (test_bit(R5_MadeGood, &dev->flags)) {
4138                         /* This flag does not apply to '.replacement'
4139                          * only to .rdev, so make sure to check that*/
4140                         struct md_rdev *rdev2 = rcu_dereference(
4141                                 conf->disks[i].rdev);
4142                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4143                                 s->handle_bad_blocks = 1;
4144                                 atomic_inc(&rdev2->nr_pending);
4145                         } else
4146                                 clear_bit(R5_MadeGood, &dev->flags);
4147                 }
4148                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4149                         struct md_rdev *rdev2 = rcu_dereference(
4150                                 conf->disks[i].replacement);
4151                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4152                                 s->handle_bad_blocks = 1;
4153                                 atomic_inc(&rdev2->nr_pending);
4154                         } else
4155                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
4156                 }
4157                 if (!test_bit(R5_Insync, &dev->flags)) {
4158                         /* The ReadError flag will just be confusing now */
4159                         clear_bit(R5_ReadError, &dev->flags);
4160                         clear_bit(R5_ReWrite, &dev->flags);
4161                 }
4162                 if (test_bit(R5_ReadError, &dev->flags))
4163                         clear_bit(R5_Insync, &dev->flags);
4164                 if (!test_bit(R5_Insync, &dev->flags)) {
4165                         if (s->failed < 2)
4166                                 s->failed_num[s->failed] = i;
4167                         s->failed++;
4168                         if (rdev && !test_bit(Faulty, &rdev->flags))
4169                                 do_recovery = 1;
4170                 }
4171         }
4172         if (test_bit(STRIPE_SYNCING, &sh->state)) {
4173                 /* If there is a failed device being replaced,
4174                  *     we must be recovering.
4175                  * else if we are after recovery_cp, we must be syncing
4176                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4177                  * else we can only be replacing
4178                  * sync and recovery both need to read all devices, and so
4179                  * use the same flag.
4180                  */
4181                 if (do_recovery ||
4182                     sh->sector >= conf->mddev->recovery_cp ||
4183                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4184                         s->syncing = 1;
4185                 else
4186                         s->replacing = 1;
4187         }
4188         rcu_read_unlock();
4189 }
4190
4191 static int clear_batch_ready(struct stripe_head *sh)
4192 {
4193         /* Return '1' if this is a member of batch, or
4194          * '0' if it is a lone stripe or a head which can now be
4195          * handled.
4196          */
4197         struct stripe_head *tmp;
4198         if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4199                 return (sh->batch_head && sh->batch_head != sh);
4200         spin_lock(&sh->stripe_lock);
4201         if (!sh->batch_head) {
4202                 spin_unlock(&sh->stripe_lock);
4203                 return 0;
4204         }
4205
4206         /*
4207          * this stripe could be added to a batch list before we check
4208          * BATCH_READY, skips it
4209          */
4210         if (sh->batch_head != sh) {
4211                 spin_unlock(&sh->stripe_lock);
4212                 return 1;
4213         }
4214         spin_lock(&sh->batch_lock);
4215         list_for_each_entry(tmp, &sh->batch_list, batch_list)
4216                 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4217         spin_unlock(&sh->batch_lock);
4218         spin_unlock(&sh->stripe_lock);
4219
4220         /*
4221          * BATCH_READY is cleared, no new stripes can be added.
4222          * batch_list can be accessed without lock
4223          */
4224         return 0;
4225 }
4226
4227 static void break_stripe_batch_list(struct stripe_head *head_sh,
4228                                     unsigned long handle_flags)
4229 {
4230         struct stripe_head *sh, *next;
4231         int i;
4232         int do_wakeup = 0;
4233
4234         list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4235
4236                 list_del_init(&sh->batch_list);
4237
4238                 WARN_ON_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4239                                           (1 << STRIPE_SYNCING) |
4240                                           (1 << STRIPE_REPLACED) |
4241                                           (1 << STRIPE_PREREAD_ACTIVE) |
4242                                           (1 << STRIPE_DELAYED) |
4243                                           (1 << STRIPE_BIT_DELAY) |
4244                                           (1 << STRIPE_FULL_WRITE) |
4245                                           (1 << STRIPE_BIOFILL_RUN) |
4246                                           (1 << STRIPE_COMPUTE_RUN)  |
4247                                           (1 << STRIPE_OPS_REQ_PENDING) |
4248                                           (1 << STRIPE_DISCARD) |
4249                                           (1 << STRIPE_BATCH_READY) |
4250                                           (1 << STRIPE_BATCH_ERR) |
4251                                           (1 << STRIPE_BITMAP_PENDING)));
4252                 WARN_ON_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4253                                               (1 << STRIPE_REPLACED)));
4254
4255                 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4256                                             (1 << STRIPE_DEGRADED)),
4257                               head_sh->state & (1 << STRIPE_INSYNC));
4258
4259                 sh->check_state = head_sh->check_state;
4260                 sh->reconstruct_state = head_sh->reconstruct_state;
4261                 for (i = 0; i < sh->disks; i++) {
4262                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4263                                 do_wakeup = 1;
4264                         sh->dev[i].flags = head_sh->dev[i].flags &
4265                                 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4266                 }
4267                 spin_lock_irq(&sh->stripe_lock);
4268                 sh->batch_head = NULL;
4269                 spin_unlock_irq(&sh->stripe_lock);
4270                 if (handle_flags == 0 ||
4271                     sh->state & handle_flags)
4272                         set_bit(STRIPE_HANDLE, &sh->state);
4273                 raid5_release_stripe(sh);
4274         }
4275         spin_lock_irq(&head_sh->stripe_lock);
4276         head_sh->batch_head = NULL;
4277         spin_unlock_irq(&head_sh->stripe_lock);
4278         for (i = 0; i < head_sh->disks; i++)
4279                 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4280                         do_wakeup = 1;
4281         if (head_sh->state & handle_flags)
4282                 set_bit(STRIPE_HANDLE, &head_sh->state);
4283
4284         if (do_wakeup)
4285                 wake_up(&head_sh->raid_conf->wait_for_overlap);
4286 }
4287
4288 static void handle_stripe(struct stripe_head *sh)
4289 {
4290         struct stripe_head_state s;
4291         struct r5conf *conf = sh->raid_conf;
4292         int i;
4293         int prexor;
4294         int disks = sh->disks;
4295         struct r5dev *pdev, *qdev;
4296
4297         clear_bit(STRIPE_HANDLE, &sh->state);
4298         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4299                 /* already being handled, ensure it gets handled
4300                  * again when current action finishes */
4301                 set_bit(STRIPE_HANDLE, &sh->state);
4302                 return;
4303         }
4304
4305         if (clear_batch_ready(sh) ) {
4306                 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4307                 return;
4308         }
4309
4310         if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4311                 break_stripe_batch_list(sh, 0);
4312
4313         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4314                 spin_lock(&sh->stripe_lock);
4315                 /* Cannot process 'sync' concurrently with 'discard' */
4316                 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4317                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4318                         set_bit(STRIPE_SYNCING, &sh->state);
4319                         clear_bit(STRIPE_INSYNC, &sh->state);
4320                         clear_bit(STRIPE_REPLACED, &sh->state);
4321                 }
4322                 spin_unlock(&sh->stripe_lock);
4323         }
4324         clear_bit(STRIPE_DELAYED, &sh->state);
4325
4326         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4327                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4328                (unsigned long long)sh->sector, sh->state,
4329                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4330                sh->check_state, sh->reconstruct_state);
4331
4332         analyse_stripe(sh, &s);
4333
4334         if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4335                 goto finish;
4336
4337         if (s.handle_bad_blocks) {
4338                 set_bit(STRIPE_HANDLE, &sh->state);
4339                 goto finish;
4340         }
4341
4342         if (unlikely(s.blocked_rdev)) {
4343                 if (s.syncing || s.expanding || s.expanded ||
4344                     s.replacing || s.to_write || s.written) {
4345                         set_bit(STRIPE_HANDLE, &sh->state);
4346                         goto finish;
4347                 }
4348                 /* There is nothing for the blocked_rdev to block */
4349                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4350                 s.blocked_rdev = NULL;
4351         }
4352
4353         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4354                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4355                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4356         }
4357
4358         pr_debug("locked=%d uptodate=%d to_read=%d"
4359                " to_write=%d failed=%d failed_num=%d,%d\n",
4360                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4361                s.failed_num[0], s.failed_num[1]);
4362         /* check if the array has lost more than max_degraded devices and,
4363          * if so, some requests might need to be failed.
4364          */
4365         if (s.failed > conf->max_degraded || s.log_failed) {
4366                 sh->check_state = 0;
4367                 sh->reconstruct_state = 0;
4368                 break_stripe_batch_list(sh, 0);
4369                 if (s.to_read+s.to_write+s.written)
4370                         handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4371                 if (s.syncing + s.replacing)
4372                         handle_failed_sync(conf, sh, &s);
4373         }
4374
4375         /* Now we check to see if any write operations have recently
4376          * completed
4377          */
4378         prexor = 0;
4379         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4380                 prexor = 1;
4381         if (sh->reconstruct_state == reconstruct_state_drain_result ||
4382             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4383                 sh->reconstruct_state = reconstruct_state_idle;
4384
4385                 /* All the 'written' buffers and the parity block are ready to
4386                  * be written back to disk
4387                  */
4388                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4389                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4390                 BUG_ON(sh->qd_idx >= 0 &&
4391                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4392                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4393                 for (i = disks; i--; ) {
4394                         struct r5dev *dev = &sh->dev[i];
4395                         if (test_bit(R5_LOCKED, &dev->flags) &&
4396                                 (i == sh->pd_idx || i == sh->qd_idx ||
4397                                  dev->written)) {
4398                                 pr_debug("Writing block %d\n", i);
4399                                 set_bit(R5_Wantwrite, &dev->flags);
4400                                 if (prexor)
4401                                         continue;
4402                                 if (s.failed > 1)
4403                                         continue;
4404                                 if (!test_bit(R5_Insync, &dev->flags) ||
4405                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
4406                                      s.failed == 0))
4407                                         set_bit(STRIPE_INSYNC, &sh->state);
4408                         }
4409                 }
4410                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4411                         s.dec_preread_active = 1;
4412         }
4413
4414         /*
4415          * might be able to return some write requests if the parity blocks
4416          * are safe, or on a failed drive
4417          */
4418         pdev = &sh->dev[sh->pd_idx];
4419         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4420                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4421         qdev = &sh->dev[sh->qd_idx];
4422         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4423                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4424                 || conf->level < 6;
4425
4426         if (s.written &&
4427             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4428                              && !test_bit(R5_LOCKED, &pdev->flags)
4429                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
4430                                  test_bit(R5_Discard, &pdev->flags))))) &&
4431             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4432                              && !test_bit(R5_LOCKED, &qdev->flags)
4433                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
4434                                  test_bit(R5_Discard, &qdev->flags))))))
4435                 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4436
4437         /* Now we might consider reading some blocks, either to check/generate
4438          * parity, or to satisfy requests
4439          * or to load a block that is being partially written.
4440          */
4441         if (s.to_read || s.non_overwrite
4442             || (conf->level == 6 && s.to_write && s.failed)
4443             || (s.syncing && (s.uptodate + s.compute < disks))
4444             || s.replacing
4445             || s.expanding)
4446                 handle_stripe_fill(sh, &s, disks);
4447
4448         /* Now to consider new write requests and what else, if anything
4449          * should be read.  We do not handle new writes when:
4450          * 1/ A 'write' operation (copy+xor) is already in flight.
4451          * 2/ A 'check' operation is in flight, as it may clobber the parity
4452          *    block.
4453          */
4454         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4455                 handle_stripe_dirtying(conf, sh, &s, disks);
4456
4457         /* maybe we need to check and possibly fix the parity for this stripe
4458          * Any reads will already have been scheduled, so we just see if enough
4459          * data is available.  The parity check is held off while parity
4460          * dependent operations are in flight.
4461          */
4462         if (sh->check_state ||
4463             (s.syncing && s.locked == 0 &&
4464              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4465              !test_bit(STRIPE_INSYNC, &sh->state))) {
4466                 if (conf->level == 6)
4467                         handle_parity_checks6(conf, sh, &s, disks);
4468                 else
4469                         handle_parity_checks5(conf, sh, &s, disks);
4470         }
4471
4472         if ((s.replacing || s.syncing) && s.locked == 0
4473             && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4474             && !test_bit(STRIPE_REPLACED, &sh->state)) {
4475                 /* Write out to replacement devices where possible */
4476                 for (i = 0; i < conf->raid_disks; i++)
4477                         if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4478                                 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4479                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
4480                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
4481                                 s.locked++;
4482                         }
4483                 if (s.replacing)
4484                         set_bit(STRIPE_INSYNC, &sh->state);
4485                 set_bit(STRIPE_REPLACED, &sh->state);
4486         }
4487         if ((s.syncing || s.replacing) && s.locked == 0 &&
4488             !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4489             test_bit(STRIPE_INSYNC, &sh->state)) {
4490                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4491                 clear_bit(STRIPE_SYNCING, &sh->state);
4492                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4493                         wake_up(&conf->wait_for_overlap);
4494         }
4495
4496         /* If the failed drives are just a ReadError, then we might need
4497          * to progress the repair/check process
4498          */
4499         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4500                 for (i = 0; i < s.failed; i++) {
4501                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
4502                         if (test_bit(R5_ReadError, &dev->flags)
4503                             && !test_bit(R5_LOCKED, &dev->flags)
4504                             && test_bit(R5_UPTODATE, &dev->flags)
4505                                 ) {
4506                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
4507                                         set_bit(R5_Wantwrite, &dev->flags);
4508                                         set_bit(R5_ReWrite, &dev->flags);
4509                                         set_bit(R5_LOCKED, &dev->flags);
4510                                         s.locked++;
4511                                 } else {
4512                                         /* let's read it back */
4513                                         set_bit(R5_Wantread, &dev->flags);
4514                                         set_bit(R5_LOCKED, &dev->flags);
4515                                         s.locked++;
4516                                 }
4517                         }
4518                 }
4519
4520         /* Finish reconstruct operations initiated by the expansion process */
4521         if (sh->reconstruct_state == reconstruct_state_result) {
4522                 struct stripe_head *sh_src
4523                         = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4524                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4525                         /* sh cannot be written until sh_src has been read.
4526                          * so arrange for sh to be delayed a little
4527                          */
4528                         set_bit(STRIPE_DELAYED, &sh->state);
4529                         set_bit(STRIPE_HANDLE, &sh->state);
4530                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4531                                               &sh_src->state))
4532                                 atomic_inc(&conf->preread_active_stripes);
4533                         raid5_release_stripe(sh_src);
4534                         goto finish;
4535                 }
4536                 if (sh_src)
4537                         raid5_release_stripe(sh_src);
4538
4539                 sh->reconstruct_state = reconstruct_state_idle;
4540                 clear_bit(STRIPE_EXPANDING, &sh->state);
4541                 for (i = conf->raid_disks; i--; ) {
4542                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
4543                         set_bit(R5_LOCKED, &sh->dev[i].flags);
4544                         s.locked++;
4545                 }
4546         }
4547
4548         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4549             !sh->reconstruct_state) {
4550                 /* Need to write out all blocks after computing parity */
4551                 sh->disks = conf->raid_disks;
4552                 stripe_set_idx(sh->sector, conf, 0, sh);
4553                 schedule_reconstruction(sh, &s, 1, 1);
4554         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4555                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4556                 atomic_dec(&conf->reshape_stripes);
4557                 wake_up(&conf->wait_for_overlap);
4558                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4559         }
4560
4561         if (s.expanding && s.locked == 0 &&
4562             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4563                 handle_stripe_expansion(conf, sh);
4564
4565 finish:
4566         /* wait for this device to become unblocked */
4567         if (unlikely(s.blocked_rdev)) {
4568                 if (conf->mddev->external)
4569                         md_wait_for_blocked_rdev(s.blocked_rdev,
4570                                                  conf->mddev);
4571                 else
4572                         /* Internal metadata will immediately
4573                          * be written by raid5d, so we don't
4574                          * need to wait here.
4575                          */
4576                         rdev_dec_pending(s.blocked_rdev,
4577                                          conf->mddev);
4578         }
4579
4580         if (s.handle_bad_blocks)
4581                 for (i = disks; i--; ) {
4582                         struct md_rdev *rdev;
4583                         struct r5dev *dev = &sh->dev[i];
4584                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4585                                 /* We own a safe reference to the rdev */
4586                                 rdev = conf->disks[i].rdev;
4587                                 if (!rdev_set_badblocks(rdev, sh->sector,
4588                                                         STRIPE_SECTORS, 0))
4589                                         md_error(conf->mddev, rdev);
4590                                 rdev_dec_pending(rdev, conf->mddev);
4591                         }
4592                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
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                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4599                                 rdev = conf->disks[i].replacement;
4600                                 if (!rdev)
4601                                         /* rdev have been moved down */
4602                                         rdev = conf->disks[i].rdev;
4603                                 rdev_clear_badblocks(rdev, sh->sector,
4604                                                      STRIPE_SECTORS, 0);
4605                                 rdev_dec_pending(rdev, conf->mddev);
4606                         }
4607                 }
4608
4609         if (s.ops_request)
4610                 raid_run_ops(sh, s.ops_request);
4611
4612         ops_run_io(sh, &s);
4613
4614         if (s.dec_preread_active) {
4615                 /* We delay this until after ops_run_io so that if make_request
4616                  * is waiting on a flush, it won't continue until the writes
4617                  * have actually been submitted.
4618                  */
4619                 atomic_dec(&conf->preread_active_stripes);
4620                 if (atomic_read(&conf->preread_active_stripes) <
4621                     IO_THRESHOLD)
4622                         md_wakeup_thread(conf->mddev->thread);
4623         }
4624
4625         if (!bio_list_empty(&s.return_bi)) {
4626                 if (test_bit(MD_CHANGE_PENDING, &conf->mddev->flags)) {
4627                         spin_lock_irq(&conf->device_lock);
4628                         bio_list_merge(&conf->return_bi, &s.return_bi);
4629                         spin_unlock_irq(&conf->device_lock);
4630                         md_wakeup_thread(conf->mddev->thread);
4631                 } else
4632                         return_io(&s.return_bi);
4633         }
4634
4635         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4636 }
4637
4638 static void raid5_activate_delayed(struct r5conf *conf)
4639 {
4640         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4641                 while (!list_empty(&conf->delayed_list)) {
4642                         struct list_head *l = conf->delayed_list.next;
4643                         struct stripe_head *sh;
4644                         sh = list_entry(l, struct stripe_head, lru);
4645                         list_del_init(l);
4646                         clear_bit(STRIPE_DELAYED, &sh->state);
4647                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4648                                 atomic_inc(&conf->preread_active_stripes);
4649                         list_add_tail(&sh->lru, &conf->hold_list);
4650                         raid5_wakeup_stripe_thread(sh);
4651                 }
4652         }
4653 }
4654
4655 static void activate_bit_delay(struct r5conf *conf,
4656         struct list_head *temp_inactive_list)
4657 {
4658         /* device_lock is held */
4659         struct list_head head;
4660         list_add(&head, &conf->bitmap_list);
4661         list_del_init(&conf->bitmap_list);
4662         while (!list_empty(&head)) {
4663                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4664                 int hash;
4665                 list_del_init(&sh->lru);
4666                 atomic_inc(&sh->count);
4667                 hash = sh->hash_lock_index;
4668                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4669         }
4670 }
4671
4672 static int raid5_congested(struct mddev *mddev, int bits)
4673 {
4674         struct r5conf *conf = mddev->private;
4675
4676         /* No difference between reads and writes.  Just check
4677          * how busy the stripe_cache is
4678          */
4679
4680         if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4681                 return 1;
4682         if (conf->quiesce)
4683                 return 1;
4684         if (atomic_read(&conf->empty_inactive_list_nr))
4685                 return 1;
4686
4687         return 0;
4688 }
4689
4690 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4691 {
4692         struct r5conf *conf = mddev->private;
4693         sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4694         unsigned int chunk_sectors;
4695         unsigned int bio_sectors = bio_sectors(bio);
4696
4697         chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
4698         return  chunk_sectors >=
4699                 ((sector & (chunk_sectors - 1)) + bio_sectors);
4700 }
4701
4702 /*
4703  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
4704  *  later sampled by raid5d.
4705  */
4706 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4707 {
4708         unsigned long flags;
4709
4710         spin_lock_irqsave(&conf->device_lock, flags);
4711
4712         bi->bi_next = conf->retry_read_aligned_list;
4713         conf->retry_read_aligned_list = bi;
4714
4715         spin_unlock_irqrestore(&conf->device_lock, flags);
4716         md_wakeup_thread(conf->mddev->thread);
4717 }
4718
4719 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4720 {
4721         struct bio *bi;
4722
4723         bi = conf->retry_read_aligned;
4724         if (bi) {
4725                 conf->retry_read_aligned = NULL;
4726                 return bi;
4727         }
4728         bi = conf->retry_read_aligned_list;
4729         if(bi) {
4730                 conf->retry_read_aligned_list = bi->bi_next;
4731                 bi->bi_next = NULL;
4732                 /*
4733                  * this sets the active strip count to 1 and the processed
4734                  * strip count to zero (upper 8 bits)
4735                  */
4736                 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4737         }
4738
4739         return bi;
4740 }
4741
4742 /*
4743  *  The "raid5_align_endio" should check if the read succeeded and if it
4744  *  did, call bio_endio on the original bio (having bio_put the new bio
4745  *  first).
4746  *  If the read failed..
4747  */
4748 static void raid5_align_endio(struct bio *bi)
4749 {
4750         struct bio* raid_bi  = bi->bi_private;
4751         struct mddev *mddev;
4752         struct r5conf *conf;
4753         struct md_rdev *rdev;
4754         int error = bi->bi_error;
4755
4756         bio_put(bi);
4757
4758         rdev = (void*)raid_bi->bi_next;
4759         raid_bi->bi_next = NULL;
4760         mddev = rdev->mddev;
4761         conf = mddev->private;
4762
4763         rdev_dec_pending(rdev, conf->mddev);
4764
4765         if (!error) {
4766                 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4767                                          raid_bi, 0);
4768                 bio_endio(raid_bi);
4769                 if (atomic_dec_and_test(&conf->active_aligned_reads))
4770                         wake_up(&conf->wait_for_quiescent);
4771                 return;
4772         }
4773
4774         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4775
4776         add_bio_to_retry(raid_bi, conf);
4777 }
4778
4779 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
4780 {
4781         struct r5conf *conf = mddev->private;
4782         int dd_idx;
4783         struct bio* align_bi;
4784         struct md_rdev *rdev;
4785         sector_t end_sector;
4786
4787         if (!in_chunk_boundary(mddev, raid_bio)) {
4788                 pr_debug("%s: non aligned\n", __func__);
4789                 return 0;
4790         }
4791         /*
4792          * use bio_clone_mddev to make a copy of the bio
4793          */
4794         align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4795         if (!align_bi)
4796                 return 0;
4797         /*
4798          *   set bi_end_io to a new function, and set bi_private to the
4799          *     original bio.
4800          */
4801         align_bi->bi_end_io  = raid5_align_endio;
4802         align_bi->bi_private = raid_bio;
4803         /*
4804          *      compute position
4805          */
4806         align_bi->bi_iter.bi_sector =
4807                 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4808                                      0, &dd_idx, NULL);
4809
4810         end_sector = bio_end_sector(align_bi);
4811         rcu_read_lock();
4812         rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4813         if (!rdev || test_bit(Faulty, &rdev->flags) ||
4814             rdev->recovery_offset < end_sector) {
4815                 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4816                 if (rdev &&
4817                     (test_bit(Faulty, &rdev->flags) ||
4818                     !(test_bit(In_sync, &rdev->flags) ||
4819                       rdev->recovery_offset >= end_sector)))
4820                         rdev = NULL;
4821         }
4822         if (rdev) {
4823                 sector_t first_bad;
4824                 int bad_sectors;
4825
4826                 atomic_inc(&rdev->nr_pending);
4827                 rcu_read_unlock();
4828                 raid_bio->bi_next = (void*)rdev;
4829                 align_bi->bi_bdev =  rdev->bdev;
4830                 bio_clear_flag(align_bi, BIO_SEG_VALID);
4831
4832                 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
4833                                 bio_sectors(align_bi),
4834                                 &first_bad, &bad_sectors)) {
4835                         bio_put(align_bi);
4836                         rdev_dec_pending(rdev, mddev);
4837                         return 0;
4838                 }
4839
4840                 /* No reshape active, so we can trust rdev->data_offset */
4841                 align_bi->bi_iter.bi_sector += rdev->data_offset;
4842
4843                 spin_lock_irq(&conf->device_lock);
4844                 wait_event_lock_irq(conf->wait_for_quiescent,
4845                                     conf->quiesce == 0,
4846                                     conf->device_lock);
4847                 atomic_inc(&conf->active_aligned_reads);
4848                 spin_unlock_irq(&conf->device_lock);
4849
4850                 if (mddev->gendisk)
4851                         trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4852                                               align_bi, disk_devt(mddev->gendisk),
4853                                               raid_bio->bi_iter.bi_sector);
4854                 generic_make_request(align_bi);
4855                 return 1;
4856         } else {
4857                 rcu_read_unlock();
4858                 bio_put(align_bi);
4859                 return 0;
4860         }
4861 }
4862
4863 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
4864 {
4865         struct bio *split;
4866
4867         do {
4868                 sector_t sector = raid_bio->bi_iter.bi_sector;
4869                 unsigned chunk_sects = mddev->chunk_sectors;
4870                 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
4871
4872                 if (sectors < bio_sectors(raid_bio)) {
4873                         split = bio_split(raid_bio, sectors, GFP_NOIO, fs_bio_set);
4874                         bio_chain(split, raid_bio);
4875                 } else
4876                         split = raid_bio;
4877
4878                 if (!raid5_read_one_chunk(mddev, split)) {
4879                         if (split != raid_bio)
4880                                 generic_make_request(raid_bio);
4881                         return split;
4882                 }
4883         } while (split != raid_bio);
4884
4885         return NULL;
4886 }
4887
4888 /* __get_priority_stripe - get the next stripe to process
4889  *
4890  * Full stripe writes are allowed to pass preread active stripes up until
4891  * the bypass_threshold is exceeded.  In general the bypass_count
4892  * increments when the handle_list is handled before the hold_list; however, it
4893  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4894  * stripe with in flight i/o.  The bypass_count will be reset when the
4895  * head of the hold_list has changed, i.e. the head was promoted to the
4896  * handle_list.
4897  */
4898 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4899 {
4900         struct stripe_head *sh = NULL, *tmp;
4901         struct list_head *handle_list = NULL;
4902         struct r5worker_group *wg = NULL;
4903
4904         if (conf->worker_cnt_per_group == 0) {
4905                 handle_list = &conf->handle_list;
4906         } else if (group != ANY_GROUP) {
4907                 handle_list = &conf->worker_groups[group].handle_list;
4908                 wg = &conf->worker_groups[group];
4909         } else {
4910                 int i;
4911                 for (i = 0; i < conf->group_cnt; i++) {
4912                         handle_list = &conf->worker_groups[i].handle_list;
4913                         wg = &conf->worker_groups[i];
4914                         if (!list_empty(handle_list))
4915                                 break;
4916                 }
4917         }
4918
4919         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4920                   __func__,
4921                   list_empty(handle_list) ? "empty" : "busy",
4922                   list_empty(&conf->hold_list) ? "empty" : "busy",
4923                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
4924
4925         if (!list_empty(handle_list)) {
4926                 sh = list_entry(handle_list->next, typeof(*sh), lru);
4927
4928                 if (list_empty(&conf->hold_list))
4929                         conf->bypass_count = 0;
4930                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4931                         if (conf->hold_list.next == conf->last_hold)
4932                                 conf->bypass_count++;
4933                         else {
4934                                 conf->last_hold = conf->hold_list.next;
4935                                 conf->bypass_count -= conf->bypass_threshold;
4936                                 if (conf->bypass_count < 0)
4937                                         conf->bypass_count = 0;
4938                         }
4939                 }
4940         } else if (!list_empty(&conf->hold_list) &&
4941                    ((conf->bypass_threshold &&
4942                      conf->bypass_count > conf->bypass_threshold) ||
4943                     atomic_read(&conf->pending_full_writes) == 0)) {
4944
4945                 list_for_each_entry(tmp, &conf->hold_list,  lru) {
4946                         if (conf->worker_cnt_per_group == 0 ||
4947                             group == ANY_GROUP ||
4948                             !cpu_online(tmp->cpu) ||
4949                             cpu_to_group(tmp->cpu) == group) {
4950                                 sh = tmp;
4951                                 break;
4952                         }
4953                 }
4954
4955                 if (sh) {
4956                         conf->bypass_count -= conf->bypass_threshold;
4957                         if (conf->bypass_count < 0)
4958                                 conf->bypass_count = 0;
4959                 }
4960                 wg = NULL;
4961         }
4962
4963         if (!sh)
4964                 return NULL;
4965
4966         if (wg) {
4967                 wg->stripes_cnt--;
4968                 sh->group = NULL;
4969         }
4970         list_del_init(&sh->lru);
4971         BUG_ON(atomic_inc_return(&sh->count) != 1);
4972         return sh;
4973 }
4974
4975 struct raid5_plug_cb {
4976         struct blk_plug_cb      cb;
4977         struct list_head        list;
4978         struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4979 };
4980
4981 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4982 {
4983         struct raid5_plug_cb *cb = container_of(
4984                 blk_cb, struct raid5_plug_cb, cb);
4985         struct stripe_head *sh;
4986         struct mddev *mddev = cb->cb.data;
4987         struct r5conf *conf = mddev->private;
4988         int cnt = 0;
4989         int hash;
4990
4991         if (cb->list.next && !list_empty(&cb->list)) {
4992                 spin_lock_irq(&conf->device_lock);
4993                 while (!list_empty(&cb->list)) {
4994                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
4995                         list_del_init(&sh->lru);
4996                         /*
4997                          * avoid race release_stripe_plug() sees
4998                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
4999                          * is still in our list
5000                          */
5001                         smp_mb__before_atomic();
5002                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5003                         /*
5004                          * STRIPE_ON_RELEASE_LIST could be set here. In that
5005                          * case, the count is always > 1 here
5006                          */
5007                         hash = sh->hash_lock_index;
5008                         __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5009                         cnt++;
5010                 }
5011                 spin_unlock_irq(&conf->device_lock);
5012         }
5013         release_inactive_stripe_list(conf, cb->temp_inactive_list,
5014                                      NR_STRIPE_HASH_LOCKS);
5015         if (mddev->queue)
5016                 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5017         kfree(cb);
5018 }
5019
5020 static void release_stripe_plug(struct mddev *mddev,
5021                                 struct stripe_head *sh)
5022 {
5023         struct blk_plug_cb *blk_cb = blk_check_plugged(
5024                 raid5_unplug, mddev,
5025                 sizeof(struct raid5_plug_cb));
5026         struct raid5_plug_cb *cb;
5027
5028         if (!blk_cb) {
5029                 raid5_release_stripe(sh);
5030                 return;
5031         }
5032
5033         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5034
5035         if (cb->list.next == NULL) {
5036                 int i;
5037                 INIT_LIST_HEAD(&cb->list);
5038                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5039                         INIT_LIST_HEAD(cb->temp_inactive_list + i);
5040         }
5041
5042         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5043                 list_add_tail(&sh->lru, &cb->list);
5044         else
5045                 raid5_release_stripe(sh);
5046 }
5047
5048 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5049 {
5050         struct r5conf *conf = mddev->private;
5051         sector_t logical_sector, last_sector;
5052         struct stripe_head *sh;
5053         int remaining;
5054         int stripe_sectors;
5055
5056         if (mddev->reshape_position != MaxSector)
5057                 /* Skip discard while reshape is happening */
5058                 return;
5059
5060         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5061         last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5062
5063         bi->bi_next = NULL;
5064         bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5065
5066         stripe_sectors = conf->chunk_sectors *
5067                 (conf->raid_disks - conf->max_degraded);
5068         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5069                                                stripe_sectors);
5070         sector_div(last_sector, stripe_sectors);
5071
5072         logical_sector *= conf->chunk_sectors;
5073         last_sector *= conf->chunk_sectors;
5074
5075         for (; logical_sector < last_sector;
5076              logical_sector += STRIPE_SECTORS) {
5077                 DEFINE_WAIT(w);
5078                 int d;
5079         again:
5080                 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5081                 prepare_to_wait(&conf->wait_for_overlap, &w,
5082                                 TASK_UNINTERRUPTIBLE);
5083                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5084                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5085                         raid5_release_stripe(sh);
5086                         schedule();
5087                         goto again;
5088                 }
5089                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5090                 spin_lock_irq(&sh->stripe_lock);
5091                 for (d = 0; d < conf->raid_disks; d++) {
5092                         if (d == sh->pd_idx || d == sh->qd_idx)
5093                                 continue;
5094                         if (sh->dev[d].towrite || sh->dev[d].toread) {
5095                                 set_bit(R5_Overlap, &sh->dev[d].flags);
5096                                 spin_unlock_irq(&sh->stripe_lock);
5097                                 raid5_release_stripe(sh);
5098                                 schedule();
5099                                 goto again;
5100                         }
5101                 }
5102                 set_bit(STRIPE_DISCARD, &sh->state);
5103                 finish_wait(&conf->wait_for_overlap, &w);
5104                 sh->overwrite_disks = 0;
5105                 for (d = 0; d < conf->raid_disks; d++) {
5106                         if (d == sh->pd_idx || d == sh->qd_idx)
5107                                 continue;
5108                         sh->dev[d].towrite = bi;
5109                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5110                         raid5_inc_bi_active_stripes(bi);
5111                         sh->overwrite_disks++;
5112                 }
5113                 spin_unlock_irq(&sh->stripe_lock);
5114                 if (conf->mddev->bitmap) {
5115                         for (d = 0;
5116                              d < conf->raid_disks - conf->max_degraded;
5117                              d++)
5118                                 bitmap_startwrite(mddev->bitmap,
5119                                                   sh->sector,
5120                                                   STRIPE_SECTORS,
5121                                                   0);
5122                         sh->bm_seq = conf->seq_flush + 1;
5123                         set_bit(STRIPE_BIT_DELAY, &sh->state);
5124                 }
5125
5126                 set_bit(STRIPE_HANDLE, &sh->state);
5127                 clear_bit(STRIPE_DELAYED, &sh->state);
5128                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5129                         atomic_inc(&conf->preread_active_stripes);
5130                 release_stripe_plug(mddev, sh);
5131         }
5132
5133         remaining = raid5_dec_bi_active_stripes(bi);
5134         if (remaining == 0) {
5135                 md_write_end(mddev);
5136                 bio_endio(bi);
5137         }
5138 }
5139
5140 static void make_request(struct mddev *mddev, struct bio * bi)
5141 {
5142         struct r5conf *conf = mddev->private;
5143         int dd_idx;
5144         sector_t new_sector;
5145         sector_t logical_sector, last_sector;
5146         struct stripe_head *sh;
5147         const int rw = bio_data_dir(bi);
5148         int remaining;
5149         DEFINE_WAIT(w);
5150         bool do_prepare;
5151
5152         if (unlikely(bi->bi_rw & REQ_FLUSH)) {
5153                 int ret = r5l_handle_flush_request(conf->log, bi);
5154
5155                 if (ret == 0)
5156                         return;
5157                 if (ret == -ENODEV) {
5158                         md_flush_request(mddev, bi);
5159                         return;
5160                 }
5161                 /* ret == -EAGAIN, fallback */
5162         }
5163
5164         md_write_start(mddev, bi);
5165
5166         /*
5167          * If array is degraded, better not do chunk aligned read because
5168          * later we might have to read it again in order to reconstruct
5169          * data on failed drives.
5170          */
5171         if (rw == READ && mddev->degraded == 0 &&
5172             mddev->reshape_position == MaxSector) {
5173                 bi = chunk_aligned_read(mddev, bi);
5174                 if (!bi)
5175                         return;
5176         }
5177
5178         if (unlikely(bi->bi_rw & REQ_DISCARD)) {
5179                 make_discard_request(mddev, bi);
5180                 return;
5181         }
5182
5183         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5184         last_sector = bio_end_sector(bi);
5185         bi->bi_next = NULL;
5186         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
5187
5188         prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5189         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5190                 int previous;
5191                 int seq;
5192
5193                 do_prepare = false;
5194         retry:
5195                 seq = read_seqcount_begin(&conf->gen_lock);
5196                 previous = 0;
5197                 if (do_prepare)
5198                         prepare_to_wait(&conf->wait_for_overlap, &w,
5199                                 TASK_UNINTERRUPTIBLE);
5200                 if (unlikely(conf->reshape_progress != MaxSector)) {
5201                         /* spinlock is needed as reshape_progress may be
5202                          * 64bit on a 32bit platform, and so it might be
5203                          * possible to see a half-updated value
5204                          * Of course reshape_progress could change after
5205                          * the lock is dropped, so once we get a reference
5206                          * to the stripe that we think it is, we will have
5207                          * to check again.
5208                          */
5209                         spin_lock_irq(&conf->device_lock);
5210                         if (mddev->reshape_backwards
5211                             ? logical_sector < conf->reshape_progress
5212                             : logical_sector >= conf->reshape_progress) {
5213                                 previous = 1;
5214                         } else {
5215                                 if (mddev->reshape_backwards
5216                                     ? logical_sector < conf->reshape_safe
5217                                     : logical_sector >= conf->reshape_safe) {
5218                                         spin_unlock_irq(&conf->device_lock);
5219                                         schedule();
5220                                         do_prepare = true;
5221                                         goto retry;
5222                                 }
5223                         }
5224                         spin_unlock_irq(&conf->device_lock);
5225                 }
5226
5227                 new_sector = raid5_compute_sector(conf, logical_sector,
5228                                                   previous,
5229                                                   &dd_idx, NULL);
5230                 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5231                         (unsigned long long)new_sector,
5232                         (unsigned long long)logical_sector);
5233
5234                 sh = raid5_get_active_stripe(conf, new_sector, previous,
5235                                        (bi->bi_rw&RWA_MASK), 0);
5236                 if (sh) {
5237                         if (unlikely(previous)) {
5238                                 /* expansion might have moved on while waiting for a
5239                                  * stripe, so we must do the range check again.
5240                                  * Expansion could still move past after this
5241                                  * test, but as we are holding a reference to
5242                                  * 'sh', we know that if that happens,
5243                                  *  STRIPE_EXPANDING will get set and the expansion
5244                                  * won't proceed until we finish with the stripe.
5245                                  */
5246                                 int must_retry = 0;
5247                                 spin_lock_irq(&conf->device_lock);
5248                                 if (mddev->reshape_backwards
5249                                     ? logical_sector >= conf->reshape_progress
5250                                     : logical_sector < conf->reshape_progress)
5251                                         /* mismatch, need to try again */
5252                                         must_retry = 1;
5253                                 spin_unlock_irq(&conf->device_lock);
5254                                 if (must_retry) {
5255                                         raid5_release_stripe(sh);
5256                                         schedule();
5257                                         do_prepare = true;
5258                                         goto retry;
5259                                 }
5260                         }
5261                         if (read_seqcount_retry(&conf->gen_lock, seq)) {
5262                                 /* Might have got the wrong stripe_head
5263                                  * by accident
5264                                  */
5265                                 raid5_release_stripe(sh);
5266                                 goto retry;
5267                         }
5268
5269                         if (rw == WRITE &&
5270                             logical_sector >= mddev->suspend_lo &&
5271                             logical_sector < mddev->suspend_hi) {
5272                                 raid5_release_stripe(sh);
5273                                 /* As the suspend_* range is controlled by
5274                                  * userspace, we want an interruptible
5275                                  * wait.
5276                                  */
5277                                 flush_signals(current);
5278                                 prepare_to_wait(&conf->wait_for_overlap,
5279                                                 &w, TASK_INTERRUPTIBLE);
5280                                 if (logical_sector >= mddev->suspend_lo &&
5281                                     logical_sector < mddev->suspend_hi) {
5282                                         schedule();
5283                                         do_prepare = true;
5284                                 }
5285                                 goto retry;
5286                         }
5287
5288                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5289                             !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5290                                 /* Stripe is busy expanding or
5291                                  * add failed due to overlap.  Flush everything
5292                                  * and wait a while
5293                                  */
5294                                 md_wakeup_thread(mddev->thread);
5295                                 raid5_release_stripe(sh);
5296                                 schedule();
5297                                 do_prepare = true;
5298                                 goto retry;
5299                         }
5300                         set_bit(STRIPE_HANDLE, &sh->state);
5301                         clear_bit(STRIPE_DELAYED, &sh->state);
5302                         if ((!sh->batch_head || sh == sh->batch_head) &&
5303                             (bi->bi_rw & REQ_SYNC) &&
5304                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5305                                 atomic_inc(&conf->preread_active_stripes);
5306                         release_stripe_plug(mddev, sh);
5307                 } else {
5308                         /* cannot get stripe for read-ahead, just give-up */
5309                         bi->bi_error = -EIO;
5310                         break;
5311                 }
5312         }
5313         finish_wait(&conf->wait_for_overlap, &w);
5314
5315         remaining = raid5_dec_bi_active_stripes(bi);
5316         if (remaining == 0) {
5317
5318                 if ( rw == WRITE )
5319                         md_write_end(mddev);
5320
5321                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5322                                          bi, 0);
5323                 bio_endio(bi);
5324         }
5325 }
5326
5327 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5328
5329 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5330 {
5331         /* reshaping is quite different to recovery/resync so it is
5332          * handled quite separately ... here.
5333          *
5334          * On each call to sync_request, we gather one chunk worth of
5335          * destination stripes and flag them as expanding.
5336          * Then we find all the source stripes and request reads.
5337          * As the reads complete, handle_stripe will copy the data
5338          * into the destination stripe and release that stripe.
5339          */
5340         struct r5conf *conf = mddev->private;
5341         struct stripe_head *sh;
5342         sector_t first_sector, last_sector;
5343         int raid_disks = conf->previous_raid_disks;
5344         int data_disks = raid_disks - conf->max_degraded;
5345         int new_data_disks = conf->raid_disks - conf->max_degraded;
5346         int i;
5347         int dd_idx;
5348         sector_t writepos, readpos, safepos;
5349         sector_t stripe_addr;
5350         int reshape_sectors;
5351         struct list_head stripes;
5352         sector_t retn;
5353
5354         if (sector_nr == 0) {
5355                 /* If restarting in the middle, skip the initial sectors */
5356                 if (mddev->reshape_backwards &&
5357                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5358                         sector_nr = raid5_size(mddev, 0, 0)
5359                                 - conf->reshape_progress;
5360                 } else if (mddev->reshape_backwards &&
5361                            conf->reshape_progress == MaxSector) {
5362                         /* shouldn't happen, but just in case, finish up.*/
5363                         sector_nr = MaxSector;
5364                 } else if (!mddev->reshape_backwards &&
5365                            conf->reshape_progress > 0)
5366                         sector_nr = conf->reshape_progress;
5367                 sector_div(sector_nr, new_data_disks);
5368                 if (sector_nr) {
5369                         mddev->curr_resync_completed = sector_nr;
5370                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5371                         *skipped = 1;
5372                         retn = sector_nr;
5373                         goto finish;
5374                 }
5375         }
5376
5377         /* We need to process a full chunk at a time.
5378          * If old and new chunk sizes differ, we need to process the
5379          * largest of these
5380          */
5381
5382         reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5383
5384         /* We update the metadata at least every 10 seconds, or when
5385          * the data about to be copied would over-write the source of
5386          * the data at the front of the range.  i.e. one new_stripe
5387          * along from reshape_progress new_maps to after where
5388          * reshape_safe old_maps to
5389          */
5390         writepos = conf->reshape_progress;
5391         sector_div(writepos, new_data_disks);
5392         readpos = conf->reshape_progress;
5393         sector_div(readpos, data_disks);
5394         safepos = conf->reshape_safe;
5395         sector_div(safepos, data_disks);
5396         if (mddev->reshape_backwards) {
5397                 BUG_ON(writepos < reshape_sectors);
5398                 writepos -= reshape_sectors;
5399                 readpos += reshape_sectors;
5400                 safepos += reshape_sectors;
5401         } else {
5402                 writepos += reshape_sectors;
5403                 /* readpos and safepos are worst-case calculations.
5404                  * A negative number is overly pessimistic, and causes
5405                  * obvious problems for unsigned storage.  So clip to 0.
5406                  */
5407                 readpos -= min_t(sector_t, reshape_sectors, readpos);
5408                 safepos -= min_t(sector_t, reshape_sectors, safepos);
5409         }
5410
5411         /* Having calculated the 'writepos' possibly use it
5412          * to set 'stripe_addr' which is where we will write to.
5413          */
5414         if (mddev->reshape_backwards) {
5415                 BUG_ON(conf->reshape_progress == 0);
5416                 stripe_addr = writepos;
5417                 BUG_ON((mddev->dev_sectors &
5418                         ~((sector_t)reshape_sectors - 1))
5419                        - reshape_sectors - stripe_addr
5420                        != sector_nr);
5421         } else {
5422                 BUG_ON(writepos != sector_nr + reshape_sectors);
5423                 stripe_addr = sector_nr;
5424         }
5425
5426         /* 'writepos' is the most advanced device address we might write.
5427          * 'readpos' is the least advanced device address we might read.
5428          * 'safepos' is the least address recorded in the metadata as having
5429          *     been reshaped.
5430          * If there is a min_offset_diff, these are adjusted either by
5431          * increasing the safepos/readpos if diff is negative, or
5432          * increasing writepos if diff is positive.
5433          * If 'readpos' is then behind 'writepos', there is no way that we can
5434          * ensure safety in the face of a crash - that must be done by userspace
5435          * making a backup of the data.  So in that case there is no particular
5436          * rush to update metadata.
5437          * Otherwise if 'safepos' is behind 'writepos', then we really need to
5438          * update the metadata to advance 'safepos' to match 'readpos' so that
5439          * we can be safe in the event of a crash.
5440          * So we insist on updating metadata if safepos is behind writepos and
5441          * readpos is beyond writepos.
5442          * In any case, update the metadata every 10 seconds.
5443          * Maybe that number should be configurable, but I'm not sure it is
5444          * worth it.... maybe it could be a multiple of safemode_delay???
5445          */
5446         if (conf->min_offset_diff < 0) {
5447                 safepos += -conf->min_offset_diff;
5448                 readpos += -conf->min_offset_diff;
5449         } else
5450                 writepos += conf->min_offset_diff;
5451
5452         if ((mddev->reshape_backwards
5453              ? (safepos > writepos && readpos < writepos)
5454              : (safepos < writepos && readpos > writepos)) ||
5455             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5456                 /* Cannot proceed until we've updated the superblock... */
5457                 wait_event(conf->wait_for_overlap,
5458                            atomic_read(&conf->reshape_stripes)==0
5459                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5460                 if (atomic_read(&conf->reshape_stripes) != 0)
5461                         return 0;
5462                 mddev->reshape_position = conf->reshape_progress;
5463                 mddev->curr_resync_completed = sector_nr;
5464                 conf->reshape_checkpoint = jiffies;
5465                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5466                 md_wakeup_thread(mddev->thread);
5467                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5468                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5469                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5470                         return 0;
5471                 spin_lock_irq(&conf->device_lock);
5472                 conf->reshape_safe = mddev->reshape_position;
5473                 spin_unlock_irq(&conf->device_lock);
5474                 wake_up(&conf->wait_for_overlap);
5475                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5476         }
5477
5478         INIT_LIST_HEAD(&stripes);
5479         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5480                 int j;
5481                 int skipped_disk = 0;
5482                 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5483                 set_bit(STRIPE_EXPANDING, &sh->state);
5484                 atomic_inc(&conf->reshape_stripes);
5485                 /* If any of this stripe is beyond the end of the old
5486                  * array, then we need to zero those blocks
5487                  */
5488                 for (j=sh->disks; j--;) {
5489                         sector_t s;
5490                         if (j == sh->pd_idx)
5491                                 continue;
5492                         if (conf->level == 6 &&
5493                             j == sh->qd_idx)
5494                                 continue;
5495                         s = raid5_compute_blocknr(sh, j, 0);
5496                         if (s < raid5_size(mddev, 0, 0)) {
5497                                 skipped_disk = 1;
5498                                 continue;
5499                         }
5500                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5501                         set_bit(R5_Expanded, &sh->dev[j].flags);
5502                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
5503                 }
5504                 if (!skipped_disk) {
5505                         set_bit(STRIPE_EXPAND_READY, &sh->state);
5506                         set_bit(STRIPE_HANDLE, &sh->state);
5507                 }
5508                 list_add(&sh->lru, &stripes);
5509         }
5510         spin_lock_irq(&conf->device_lock);
5511         if (mddev->reshape_backwards)
5512                 conf->reshape_progress -= reshape_sectors * new_data_disks;
5513         else
5514                 conf->reshape_progress += reshape_sectors * new_data_disks;
5515         spin_unlock_irq(&conf->device_lock);
5516         /* Ok, those stripe are ready. We can start scheduling
5517          * reads on the source stripes.
5518          * The source stripes are determined by mapping the first and last
5519          * block on the destination stripes.
5520          */
5521         first_sector =
5522                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5523                                      1, &dd_idx, NULL);
5524         last_sector =
5525                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5526                                             * new_data_disks - 1),
5527                                      1, &dd_idx, NULL);
5528         if (last_sector >= mddev->dev_sectors)
5529                 last_sector = mddev->dev_sectors - 1;
5530         while (first_sector <= last_sector) {
5531                 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5532                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5533                 set_bit(STRIPE_HANDLE, &sh->state);
5534                 raid5_release_stripe(sh);
5535                 first_sector += STRIPE_SECTORS;
5536         }
5537         /* Now that the sources are clearly marked, we can release
5538          * the destination stripes
5539          */
5540         while (!list_empty(&stripes)) {
5541                 sh = list_entry(stripes.next, struct stripe_head, lru);
5542                 list_del_init(&sh->lru);
5543                 raid5_release_stripe(sh);
5544         }
5545         /* If this takes us to the resync_max point where we have to pause,
5546          * then we need to write out the superblock.
5547          */
5548         sector_nr += reshape_sectors;
5549         retn = reshape_sectors;
5550 finish:
5551         if (mddev->curr_resync_completed > mddev->resync_max ||
5552             (sector_nr - mddev->curr_resync_completed) * 2
5553             >= mddev->resync_max - mddev->curr_resync_completed) {
5554                 /* Cannot proceed until we've updated the superblock... */
5555                 wait_event(conf->wait_for_overlap,
5556                            atomic_read(&conf->reshape_stripes) == 0
5557                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5558                 if (atomic_read(&conf->reshape_stripes) != 0)
5559                         goto ret;
5560                 mddev->reshape_position = conf->reshape_progress;
5561                 mddev->curr_resync_completed = sector_nr;
5562                 conf->reshape_checkpoint = jiffies;
5563                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5564                 md_wakeup_thread(mddev->thread);
5565                 wait_event(mddev->sb_wait,
5566                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5567                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5568                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5569                         goto ret;
5570                 spin_lock_irq(&conf->device_lock);
5571                 conf->reshape_safe = mddev->reshape_position;
5572                 spin_unlock_irq(&conf->device_lock);
5573                 wake_up(&conf->wait_for_overlap);
5574                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5575         }
5576 ret:
5577         return retn;
5578 }
5579
5580 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5581 {
5582         struct r5conf *conf = mddev->private;
5583         struct stripe_head *sh;
5584         sector_t max_sector = mddev->dev_sectors;
5585         sector_t sync_blocks;
5586         int still_degraded = 0;
5587         int i;
5588
5589         if (sector_nr >= max_sector) {
5590                 /* just being told to finish up .. nothing much to do */
5591
5592                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5593                         end_reshape(conf);
5594                         return 0;
5595                 }
5596
5597                 if (mddev->curr_resync < max_sector) /* aborted */
5598                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5599                                         &sync_blocks, 1);
5600                 else /* completed sync */
5601                         conf->fullsync = 0;
5602                 bitmap_close_sync(mddev->bitmap);
5603
5604                 return 0;
5605         }
5606
5607         /* Allow raid5_quiesce to complete */
5608         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5609
5610         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5611                 return reshape_request(mddev, sector_nr, skipped);
5612
5613         /* No need to check resync_max as we never do more than one
5614          * stripe, and as resync_max will always be on a chunk boundary,
5615          * if the check in md_do_sync didn't fire, there is no chance
5616          * of overstepping resync_max here
5617          */
5618
5619         /* if there is too many failed drives and we are trying
5620          * to resync, then assert that we are finished, because there is
5621          * nothing we can do.
5622          */
5623         if (mddev->degraded >= conf->max_degraded &&
5624             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5625                 sector_t rv = mddev->dev_sectors - sector_nr;
5626                 *skipped = 1;
5627                 return rv;
5628         }
5629         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5630             !conf->fullsync &&
5631             !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5632             sync_blocks >= STRIPE_SECTORS) {
5633                 /* we can skip this block, and probably more */
5634                 sync_blocks /= STRIPE_SECTORS;
5635                 *skipped = 1;
5636                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5637         }
5638
5639         bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
5640
5641         sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
5642         if (sh == NULL) {
5643                 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
5644                 /* make sure we don't swamp the stripe cache if someone else
5645                  * is trying to get access
5646                  */
5647                 schedule_timeout_uninterruptible(1);
5648         }
5649         /* Need to check if array will still be degraded after recovery/resync
5650          * Note in case of > 1 drive failures it's possible we're rebuilding
5651          * one drive while leaving another faulty drive in array.
5652          */
5653         rcu_read_lock();
5654         for (i = 0; i < conf->raid_disks; i++) {
5655                 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5656
5657                 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5658                         still_degraded = 1;
5659         }
5660         rcu_read_unlock();
5661
5662         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5663
5664         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5665         set_bit(STRIPE_HANDLE, &sh->state);
5666
5667         raid5_release_stripe(sh);
5668
5669         return STRIPE_SECTORS;
5670 }
5671
5672 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5673 {
5674         /* We may not be able to submit a whole bio at once as there
5675          * may not be enough stripe_heads available.
5676          * We cannot pre-allocate enough stripe_heads as we may need
5677          * more than exist in the cache (if we allow ever large chunks).
5678          * So we do one stripe head at a time and record in
5679          * ->bi_hw_segments how many have been done.
5680          *
5681          * We *know* that this entire raid_bio is in one chunk, so
5682          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5683          */
5684         struct stripe_head *sh;
5685         int dd_idx;
5686         sector_t sector, logical_sector, last_sector;
5687         int scnt = 0;
5688         int remaining;
5689         int handled = 0;
5690
5691         logical_sector = raid_bio->bi_iter.bi_sector &
5692                 ~((sector_t)STRIPE_SECTORS-1);
5693         sector = raid5_compute_sector(conf, logical_sector,
5694                                       0, &dd_idx, NULL);
5695         last_sector = bio_end_sector(raid_bio);
5696
5697         for (; logical_sector < last_sector;
5698              logical_sector += STRIPE_SECTORS,
5699                      sector += STRIPE_SECTORS,
5700                      scnt++) {
5701
5702                 if (scnt < raid5_bi_processed_stripes(raid_bio))
5703                         /* already done this stripe */
5704                         continue;
5705
5706                 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
5707
5708                 if (!sh) {
5709                         /* failed to get a stripe - must wait */
5710                         raid5_set_bi_processed_stripes(raid_bio, scnt);
5711                         conf->retry_read_aligned = raid_bio;
5712                         return handled;
5713                 }
5714
5715                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5716                         raid5_release_stripe(sh);
5717                         raid5_set_bi_processed_stripes(raid_bio, scnt);
5718                         conf->retry_read_aligned = raid_bio;
5719                         return handled;
5720                 }
5721
5722                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5723                 handle_stripe(sh);
5724                 raid5_release_stripe(sh);
5725                 handled++;
5726         }
5727         remaining = raid5_dec_bi_active_stripes(raid_bio);
5728         if (remaining == 0) {
5729                 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5730                                          raid_bio, 0);
5731                 bio_endio(raid_bio);
5732         }
5733         if (atomic_dec_and_test(&conf->active_aligned_reads))
5734                 wake_up(&conf->wait_for_quiescent);
5735         return handled;
5736 }
5737
5738 static int handle_active_stripes(struct r5conf *conf, int group,
5739                                  struct r5worker *worker,
5740                                  struct list_head *temp_inactive_list)
5741 {
5742         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5743         int i, batch_size = 0, hash;
5744         bool release_inactive = false;
5745
5746         while (batch_size < MAX_STRIPE_BATCH &&
5747                         (sh = __get_priority_stripe(conf, group)) != NULL)
5748                 batch[batch_size++] = sh;
5749
5750         if (batch_size == 0) {
5751                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5752                         if (!list_empty(temp_inactive_list + i))
5753                                 break;
5754                 if (i == NR_STRIPE_HASH_LOCKS) {
5755                         spin_unlock_irq(&conf->device_lock);
5756                         r5l_flush_stripe_to_raid(conf->log);
5757                         spin_lock_irq(&conf->device_lock);
5758                         return batch_size;
5759                 }
5760                 release_inactive = true;
5761         }
5762         spin_unlock_irq(&conf->device_lock);
5763
5764         release_inactive_stripe_list(conf, temp_inactive_list,
5765                                      NR_STRIPE_HASH_LOCKS);
5766
5767         r5l_flush_stripe_to_raid(conf->log);
5768         if (release_inactive) {
5769                 spin_lock_irq(&conf->device_lock);
5770                 return 0;
5771         }
5772
5773         for (i = 0; i < batch_size; i++)
5774                 handle_stripe(batch[i]);
5775         r5l_write_stripe_run(conf->log);
5776
5777         cond_resched();
5778
5779         spin_lock_irq(&conf->device_lock);
5780         for (i = 0; i < batch_size; i++) {
5781                 hash = batch[i]->hash_lock_index;
5782                 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5783         }
5784         return batch_size;
5785 }
5786
5787 static void raid5_do_work(struct work_struct *work)
5788 {
5789         struct r5worker *worker = container_of(work, struct r5worker, work);
5790         struct r5worker_group *group = worker->group;
5791         struct r5conf *conf = group->conf;
5792         int group_id = group - conf->worker_groups;
5793         int handled;
5794         struct blk_plug plug;
5795
5796         pr_debug("+++ raid5worker active\n");
5797
5798         blk_start_plug(&plug);
5799         handled = 0;
5800         spin_lock_irq(&conf->device_lock);
5801         while (1) {
5802                 int batch_size, released;
5803
5804                 released = release_stripe_list(conf, worker->temp_inactive_list);
5805
5806                 batch_size = handle_active_stripes(conf, group_id, worker,
5807                                                    worker->temp_inactive_list);
5808                 worker->working = false;
5809                 if (!batch_size && !released)
5810                         break;
5811                 handled += batch_size;
5812         }
5813         pr_debug("%d stripes handled\n", handled);
5814
5815         spin_unlock_irq(&conf->device_lock);
5816         blk_finish_plug(&plug);
5817
5818         pr_debug("--- raid5worker inactive\n");
5819 }
5820
5821 /*
5822  * This is our raid5 kernel thread.
5823  *
5824  * We scan the hash table for stripes which can be handled now.
5825  * During the scan, completed stripes are saved for us by the interrupt
5826  * handler, so that they will not have to wait for our next wakeup.
5827  */
5828 static void raid5d(struct md_thread *thread)
5829 {
5830         struct mddev *mddev = thread->mddev;
5831         struct r5conf *conf = mddev->private;
5832         int handled;
5833         struct blk_plug plug;
5834
5835         pr_debug("+++ raid5d active\n");
5836
5837         md_check_recovery(mddev);
5838
5839         if (!bio_list_empty(&conf->return_bi) &&
5840             !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5841                 struct bio_list tmp = BIO_EMPTY_LIST;
5842                 spin_lock_irq(&conf->device_lock);
5843                 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5844                         bio_list_merge(&tmp, &conf->return_bi);
5845                         bio_list_init(&conf->return_bi);
5846                 }
5847                 spin_unlock_irq(&conf->device_lock);
5848                 return_io(&tmp);
5849         }
5850
5851         blk_start_plug(&plug);
5852         handled = 0;
5853         spin_lock_irq(&conf->device_lock);
5854         while (1) {
5855                 struct bio *bio;
5856                 int batch_size, released;
5857
5858                 released = release_stripe_list(conf, conf->temp_inactive_list);
5859                 if (released)
5860                         clear_bit(R5_DID_ALLOC, &conf->cache_state);
5861
5862                 if (
5863                     !list_empty(&conf->bitmap_list)) {
5864                         /* Now is a good time to flush some bitmap updates */
5865                         conf->seq_flush++;
5866                         spin_unlock_irq(&conf->device_lock);
5867                         bitmap_unplug(mddev->bitmap);
5868                         spin_lock_irq(&conf->device_lock);
5869                         conf->seq_write = conf->seq_flush;
5870                         activate_bit_delay(conf, conf->temp_inactive_list);
5871                 }
5872                 raid5_activate_delayed(conf);
5873
5874                 while ((bio = remove_bio_from_retry(conf))) {
5875                         int ok;
5876                         spin_unlock_irq(&conf->device_lock);
5877                         ok = retry_aligned_read(conf, bio);
5878                         spin_lock_irq(&conf->device_lock);
5879                         if (!ok)
5880                                 break;
5881                         handled++;
5882                 }
5883
5884                 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5885                                                    conf->temp_inactive_list);
5886                 if (!batch_size && !released)
5887                         break;
5888                 handled += batch_size;
5889
5890                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5891                         spin_unlock_irq(&conf->device_lock);
5892                         md_check_recovery(mddev);
5893                         spin_lock_irq(&conf->device_lock);
5894                 }
5895         }
5896         pr_debug("%d stripes handled\n", handled);
5897
5898         spin_unlock_irq(&conf->device_lock);
5899         if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
5900             mutex_trylock(&conf->cache_size_mutex)) {
5901                 grow_one_stripe(conf, __GFP_NOWARN);
5902                 /* Set flag even if allocation failed.  This helps
5903                  * slow down allocation requests when mem is short
5904                  */
5905                 set_bit(R5_DID_ALLOC, &conf->cache_state);
5906                 mutex_unlock(&conf->cache_size_mutex);
5907         }
5908
5909         r5l_flush_stripe_to_raid(conf->log);
5910
5911         async_tx_issue_pending_all();
5912         blk_finish_plug(&plug);
5913
5914         pr_debug("--- raid5d inactive\n");
5915 }
5916
5917 static ssize_t
5918 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5919 {
5920         struct r5conf *conf;
5921         int ret = 0;
5922         spin_lock(&mddev->lock);
5923         conf = mddev->private;
5924         if (conf)
5925                 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
5926         spin_unlock(&mddev->lock);
5927         return ret;
5928 }
5929
5930 int
5931 raid5_set_cache_size(struct mddev *mddev, int size)
5932 {
5933         struct r5conf *conf = mddev->private;
5934         int err;
5935
5936         if (size <= 16 || size > 32768)
5937                 return -EINVAL;
5938
5939         conf->min_nr_stripes = size;
5940         mutex_lock(&conf->cache_size_mutex);
5941         while (size < conf->max_nr_stripes &&
5942                drop_one_stripe(conf))
5943                 ;
5944         mutex_unlock(&conf->cache_size_mutex);
5945
5946
5947         err = md_allow_write(mddev);
5948         if (err)
5949                 return err;
5950
5951         mutex_lock(&conf->cache_size_mutex);
5952         while (size > conf->max_nr_stripes)
5953                 if (!grow_one_stripe(conf, GFP_KERNEL))
5954                         break;
5955         mutex_unlock(&conf->cache_size_mutex);
5956
5957         return 0;
5958 }
5959 EXPORT_SYMBOL(raid5_set_cache_size);
5960
5961 static ssize_t
5962 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5963 {
5964         struct r5conf *conf;
5965         unsigned long new;
5966         int err;
5967
5968         if (len >= PAGE_SIZE)
5969                 return -EINVAL;
5970         if (kstrtoul(page, 10, &new))
5971                 return -EINVAL;
5972         err = mddev_lock(mddev);
5973         if (err)
5974                 return err;
5975         conf = mddev->private;
5976         if (!conf)
5977                 err = -ENODEV;
5978         else
5979                 err = raid5_set_cache_size(mddev, new);
5980         mddev_unlock(mddev);
5981
5982         return err ?: len;
5983 }
5984
5985 static struct md_sysfs_entry
5986 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5987                                 raid5_show_stripe_cache_size,
5988                                 raid5_store_stripe_cache_size);
5989
5990 static ssize_t
5991 raid5_show_rmw_level(struct mddev  *mddev, char *page)
5992 {
5993         struct r5conf *conf = mddev->private;
5994         if (conf)
5995                 return sprintf(page, "%d\n", conf->rmw_level);
5996         else
5997                 return 0;
5998 }
5999
6000 static ssize_t
6001 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6002 {
6003         struct r5conf *conf = mddev->private;
6004         unsigned long new;
6005
6006         if (!conf)
6007                 return -ENODEV;
6008
6009         if (len >= PAGE_SIZE)
6010                 return -EINVAL;
6011
6012         if (kstrtoul(page, 10, &new))
6013                 return -EINVAL;
6014
6015         if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6016                 return -EINVAL;
6017
6018         if (new != PARITY_DISABLE_RMW &&
6019             new != PARITY_ENABLE_RMW &&
6020             new != PARITY_PREFER_RMW)
6021                 return -EINVAL;
6022
6023         conf->rmw_level = new;
6024         return len;
6025 }
6026
6027 static struct md_sysfs_entry
6028 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6029                          raid5_show_rmw_level,
6030                          raid5_store_rmw_level);
6031
6032
6033 static ssize_t
6034 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6035 {
6036         struct r5conf *conf;
6037         int ret = 0;
6038         spin_lock(&mddev->lock);
6039         conf = mddev->private;
6040         if (conf)
6041                 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6042         spin_unlock(&mddev->lock);
6043         return ret;
6044 }
6045
6046 static ssize_t
6047 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6048 {
6049         struct r5conf *conf;
6050         unsigned long new;
6051         int err;
6052
6053         if (len >= PAGE_SIZE)
6054                 return -EINVAL;
6055         if (kstrtoul(page, 10, &new))
6056                 return -EINVAL;
6057
6058         err = mddev_lock(mddev);
6059         if (err)
6060                 return err;
6061         conf = mddev->private;
6062         if (!conf)
6063                 err = -ENODEV;
6064         else if (new > conf->min_nr_stripes)
6065                 err = -EINVAL;
6066         else
6067                 conf->bypass_threshold = new;
6068         mddev_unlock(mddev);
6069         return err ?: len;
6070 }
6071
6072 static struct md_sysfs_entry
6073 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6074                                         S_IRUGO | S_IWUSR,
6075                                         raid5_show_preread_threshold,
6076                                         raid5_store_preread_threshold);
6077
6078 static ssize_t
6079 raid5_show_skip_copy(struct mddev *mddev, char *page)
6080 {
6081         struct r5conf *conf;
6082         int ret = 0;
6083         spin_lock(&mddev->lock);
6084         conf = mddev->private;
6085         if (conf)
6086                 ret = sprintf(page, "%d\n", conf->skip_copy);
6087         spin_unlock(&mddev->lock);
6088         return ret;
6089 }
6090
6091 static ssize_t
6092 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6093 {
6094         struct r5conf *conf;
6095         unsigned long new;
6096         int err;
6097
6098         if (len >= PAGE_SIZE)
6099                 return -EINVAL;
6100         if (kstrtoul(page, 10, &new))
6101                 return -EINVAL;
6102         new = !!new;
6103
6104         err = mddev_lock(mddev);
6105         if (err)
6106                 return err;
6107         conf = mddev->private;
6108         if (!conf)
6109                 err = -ENODEV;
6110         else if (new != conf->skip_copy) {
6111                 mddev_suspend(mddev);
6112                 conf->skip_copy = new;
6113                 if (new)
6114                         mddev->queue->backing_dev_info.capabilities |=
6115                                 BDI_CAP_STABLE_WRITES;
6116                 else
6117                         mddev->queue->backing_dev_info.capabilities &=
6118                                 ~BDI_CAP_STABLE_WRITES;
6119                 mddev_resume(mddev);
6120         }
6121         mddev_unlock(mddev);
6122         return err ?: len;
6123 }
6124
6125 static struct md_sysfs_entry
6126 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6127                                         raid5_show_skip_copy,
6128                                         raid5_store_skip_copy);
6129
6130 static ssize_t
6131 stripe_cache_active_show(struct mddev *mddev, char *page)
6132 {
6133         struct r5conf *conf = mddev->private;
6134         if (conf)
6135                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6136         else
6137                 return 0;
6138 }
6139
6140 static struct md_sysfs_entry
6141 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6142
6143 static ssize_t
6144 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6145 {
6146         struct r5conf *conf;
6147         int ret = 0;
6148         spin_lock(&mddev->lock);
6149         conf = mddev->private;
6150         if (conf)
6151                 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6152         spin_unlock(&mddev->lock);
6153         return ret;
6154 }
6155
6156 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6157                                int *group_cnt,
6158                                int *worker_cnt_per_group,
6159                                struct r5worker_group **worker_groups);
6160 static ssize_t
6161 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6162 {
6163         struct r5conf *conf;
6164         unsigned long new;
6165         int err;
6166         struct r5worker_group *new_groups, *old_groups;
6167         int group_cnt, worker_cnt_per_group;
6168
6169         if (len >= PAGE_SIZE)
6170                 return -EINVAL;
6171         if (kstrtoul(page, 10, &new))
6172                 return -EINVAL;
6173
6174         err = mddev_lock(mddev);
6175         if (err)
6176                 return err;
6177         conf = mddev->private;
6178         if (!conf)
6179                 err = -ENODEV;
6180         else if (new != conf->worker_cnt_per_group) {
6181                 mddev_suspend(mddev);
6182
6183                 old_groups = conf->worker_groups;
6184                 if (old_groups)
6185                         flush_workqueue(raid5_wq);
6186
6187                 err = alloc_thread_groups(conf, new,
6188                                           &group_cnt, &worker_cnt_per_group,
6189                                           &new_groups);
6190                 if (!err) {
6191                         spin_lock_irq(&conf->device_lock);
6192                         conf->group_cnt = group_cnt;
6193                         conf->worker_cnt_per_group = worker_cnt_per_group;
6194                         conf->worker_groups = new_groups;
6195                         spin_unlock_irq(&conf->device_lock);
6196
6197                         if (old_groups)
6198                                 kfree(old_groups[0].workers);
6199                         kfree(old_groups);
6200                 }
6201                 mddev_resume(mddev);
6202         }
6203         mddev_unlock(mddev);
6204
6205         return err ?: len;
6206 }
6207
6208 static struct md_sysfs_entry
6209 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6210                                 raid5_show_group_thread_cnt,
6211                                 raid5_store_group_thread_cnt);
6212
6213 static struct attribute *raid5_attrs[] =  {
6214         &raid5_stripecache_size.attr,
6215         &raid5_stripecache_active.attr,
6216         &raid5_preread_bypass_threshold.attr,
6217         &raid5_group_thread_cnt.attr,
6218         &raid5_skip_copy.attr,
6219         &raid5_rmw_level.attr,
6220         NULL,
6221 };
6222 static struct attribute_group raid5_attrs_group = {
6223         .name = NULL,
6224         .attrs = raid5_attrs,
6225 };
6226
6227 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6228                                int *group_cnt,
6229                                int *worker_cnt_per_group,
6230                                struct r5worker_group **worker_groups)
6231 {
6232         int i, j, k;
6233         ssize_t size;
6234         struct r5worker *workers;
6235
6236         *worker_cnt_per_group = cnt;
6237         if (cnt == 0) {
6238                 *group_cnt = 0;
6239                 *worker_groups = NULL;
6240                 return 0;
6241         }
6242         *group_cnt = num_possible_nodes();
6243         size = sizeof(struct r5worker) * cnt;
6244         workers = kzalloc(size * *group_cnt, GFP_NOIO);
6245         *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6246                                 *group_cnt, GFP_NOIO);
6247         if (!*worker_groups || !workers) {
6248                 kfree(workers);
6249                 kfree(*worker_groups);
6250                 return -ENOMEM;
6251         }
6252
6253         for (i = 0; i < *group_cnt; i++) {
6254                 struct r5worker_group *group;
6255
6256                 group = &(*worker_groups)[i];
6257                 INIT_LIST_HEAD(&group->handle_list);
6258                 group->conf = conf;
6259                 group->workers = workers + i * cnt;
6260
6261                 for (j = 0; j < cnt; j++) {
6262                         struct r5worker *worker = group->workers + j;
6263                         worker->group = group;
6264                         INIT_WORK(&worker->work, raid5_do_work);
6265
6266                         for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6267                                 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6268                 }
6269         }
6270
6271         return 0;
6272 }
6273
6274 static void free_thread_groups(struct r5conf *conf)
6275 {
6276         if (conf->worker_groups)
6277                 kfree(conf->worker_groups[0].workers);
6278         kfree(conf->worker_groups);
6279         conf->worker_groups = NULL;
6280 }
6281
6282 static sector_t
6283 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6284 {
6285         struct r5conf *conf = mddev->private;
6286
6287         if (!sectors)
6288                 sectors = mddev->dev_sectors;
6289         if (!raid_disks)
6290                 /* size is defined by the smallest of previous and new size */
6291                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6292
6293         sectors &= ~((sector_t)conf->chunk_sectors - 1);
6294         sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6295         return sectors * (raid_disks - conf->max_degraded);
6296 }
6297
6298 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6299 {
6300         safe_put_page(percpu->spare_page);
6301         if (percpu->scribble)
6302                 flex_array_free(percpu->scribble);
6303         percpu->spare_page = NULL;
6304         percpu->scribble = NULL;
6305 }
6306
6307 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6308 {
6309         if (conf->level == 6 && !percpu->spare_page)
6310                 percpu->spare_page = alloc_page(GFP_KERNEL);
6311         if (!percpu->scribble)
6312                 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6313                                                       conf->previous_raid_disks),
6314                                                   max(conf->chunk_sectors,
6315                                                       conf->prev_chunk_sectors)
6316                                                    / STRIPE_SECTORS,
6317                                                   GFP_KERNEL);
6318
6319         if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6320                 free_scratch_buffer(conf, percpu);
6321                 return -ENOMEM;
6322         }
6323
6324         return 0;
6325 }
6326
6327 static void raid5_free_percpu(struct r5conf *conf)
6328 {
6329         unsigned long cpu;
6330
6331         if (!conf->percpu)
6332                 return;
6333
6334 #ifdef CONFIG_HOTPLUG_CPU
6335         unregister_cpu_notifier(&conf->cpu_notify);
6336 #endif
6337
6338         get_online_cpus();
6339         for_each_possible_cpu(cpu)
6340                 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6341         put_online_cpus();
6342
6343         free_percpu(conf->percpu);
6344 }
6345
6346 static void free_conf(struct r5conf *conf)
6347 {
6348         if (conf->log)
6349                 r5l_exit_log(conf->log);
6350         if (conf->shrinker.seeks)
6351                 unregister_shrinker(&conf->shrinker);
6352
6353         free_thread_groups(conf);
6354         shrink_stripes(conf);
6355         raid5_free_percpu(conf);
6356         kfree(conf->disks);
6357         kfree(conf->stripe_hashtbl);
6358         kfree(conf);
6359 }
6360
6361 #ifdef CONFIG_HOTPLUG_CPU
6362 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
6363                               void *hcpu)
6364 {
6365         struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
6366         long cpu = (long)hcpu;
6367         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6368
6369         switch (action) {
6370         case CPU_UP_PREPARE:
6371         case CPU_UP_PREPARE_FROZEN:
6372                 if (alloc_scratch_buffer(conf, percpu)) {
6373                         pr_err("%s: failed memory allocation for cpu%ld\n",
6374                                __func__, cpu);
6375                         return notifier_from_errno(-ENOMEM);
6376                 }
6377                 break;
6378         case CPU_DEAD:
6379         case CPU_DEAD_FROZEN:
6380                 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6381                 break;
6382         default:
6383                 break;
6384         }
6385         return NOTIFY_OK;
6386 }
6387 #endif
6388
6389 static int raid5_alloc_percpu(struct r5conf *conf)
6390 {
6391         unsigned long cpu;
6392         int err = 0;
6393
6394         conf->percpu = alloc_percpu(struct raid5_percpu);
6395         if (!conf->percpu)
6396                 return -ENOMEM;
6397
6398 #ifdef CONFIG_HOTPLUG_CPU
6399         conf->cpu_notify.notifier_call = raid456_cpu_notify;
6400         conf->cpu_notify.priority = 0;
6401         err = register_cpu_notifier(&conf->cpu_notify);
6402         if (err)
6403                 return err;
6404 #endif
6405
6406         get_online_cpus();
6407         for_each_present_cpu(cpu) {
6408                 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6409                 if (err) {
6410                         pr_err("%s: failed memory allocation for cpu%ld\n",
6411                                __func__, cpu);
6412                         break;
6413                 }
6414         }
6415         put_online_cpus();
6416
6417         return err;
6418 }
6419
6420 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6421                                       struct shrink_control *sc)
6422 {
6423         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6424         unsigned long ret = SHRINK_STOP;
6425
6426         if (mutex_trylock(&conf->cache_size_mutex)) {
6427                 ret= 0;
6428                 while (ret < sc->nr_to_scan &&
6429                        conf->max_nr_stripes > conf->min_nr_stripes) {
6430                         if (drop_one_stripe(conf) == 0) {
6431                                 ret = SHRINK_STOP;
6432                                 break;
6433                         }
6434                         ret++;
6435                 }
6436                 mutex_unlock(&conf->cache_size_mutex);
6437         }
6438         return ret;
6439 }
6440
6441 static unsigned long raid5_cache_count(struct shrinker *shrink,
6442                                        struct shrink_control *sc)
6443 {
6444         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6445
6446         if (conf->max_nr_stripes < conf->min_nr_stripes)
6447                 /* unlikely, but not impossible */
6448                 return 0;
6449         return conf->max_nr_stripes - conf->min_nr_stripes;
6450 }
6451
6452 static struct r5conf *setup_conf(struct mddev *mddev)
6453 {
6454         struct r5conf *conf;
6455         int raid_disk, memory, max_disks;
6456         struct md_rdev *rdev;
6457         struct disk_info *disk;
6458         char pers_name[6];
6459         int i;
6460         int group_cnt, worker_cnt_per_group;
6461         struct r5worker_group *new_group;
6462
6463         if (mddev->new_level != 5
6464             && mddev->new_level != 4
6465             && mddev->new_level != 6) {
6466                 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6467                        mdname(mddev), mddev->new_level);
6468                 return ERR_PTR(-EIO);
6469         }
6470         if ((mddev->new_level == 5
6471              && !algorithm_valid_raid5(mddev->new_layout)) ||
6472             (mddev->new_level == 6
6473              && !algorithm_valid_raid6(mddev->new_layout))) {
6474                 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6475                        mdname(mddev), mddev->new_layout);
6476                 return ERR_PTR(-EIO);
6477         }
6478         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6479                 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6480                        mdname(mddev), mddev->raid_disks);
6481                 return ERR_PTR(-EINVAL);
6482         }
6483
6484         if (!mddev->new_chunk_sectors ||
6485             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6486             !is_power_of_2(mddev->new_chunk_sectors)) {
6487                 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6488                        mdname(mddev), mddev->new_chunk_sectors << 9);
6489                 return ERR_PTR(-EINVAL);
6490         }
6491
6492         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6493         if (conf == NULL)
6494                 goto abort;
6495         /* Don't enable multi-threading by default*/
6496         if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6497                                  &new_group)) {
6498                 conf->group_cnt = group_cnt;
6499                 conf->worker_cnt_per_group = worker_cnt_per_group;
6500                 conf->worker_groups = new_group;
6501         } else
6502                 goto abort;
6503         spin_lock_init(&conf->device_lock);
6504         seqcount_init(&conf->gen_lock);
6505         mutex_init(&conf->cache_size_mutex);
6506         init_waitqueue_head(&conf->wait_for_quiescent);
6507         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++) {
6508                 init_waitqueue_head(&conf->wait_for_stripe[i]);
6509         }
6510         init_waitqueue_head(&conf->wait_for_overlap);
6511         INIT_LIST_HEAD(&conf->handle_list);
6512         INIT_LIST_HEAD(&conf->hold_list);
6513         INIT_LIST_HEAD(&conf->delayed_list);
6514         INIT_LIST_HEAD(&conf->bitmap_list);
6515         bio_list_init(&conf->return_bi);
6516         init_llist_head(&conf->released_stripes);
6517         atomic_set(&conf->active_stripes, 0);
6518         atomic_set(&conf->preread_active_stripes, 0);
6519         atomic_set(&conf->active_aligned_reads, 0);
6520         conf->bypass_threshold = BYPASS_THRESHOLD;
6521         conf->recovery_disabled = mddev->recovery_disabled - 1;
6522
6523         conf->raid_disks = mddev->raid_disks;
6524         if (mddev->reshape_position == MaxSector)
6525                 conf->previous_raid_disks = mddev->raid_disks;
6526         else
6527                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6528         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6529
6530         conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6531                               GFP_KERNEL);
6532         if (!conf->disks)
6533                 goto abort;
6534
6535         conf->mddev = mddev;
6536
6537         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6538                 goto abort;
6539
6540         /* We init hash_locks[0] separately to that it can be used
6541          * as the reference lock in the spin_lock_nest_lock() call
6542          * in lock_all_device_hash_locks_irq in order to convince
6543          * lockdep that we know what we are doing.
6544          */
6545         spin_lock_init(conf->hash_locks);
6546         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6547                 spin_lock_init(conf->hash_locks + i);
6548
6549         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6550                 INIT_LIST_HEAD(conf->inactive_list + i);
6551
6552         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6553                 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6554
6555         conf->level = mddev->new_level;
6556         conf->chunk_sectors = mddev->new_chunk_sectors;
6557         if (raid5_alloc_percpu(conf) != 0)
6558                 goto abort;
6559
6560         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6561
6562         rdev_for_each(rdev, mddev) {
6563                 raid_disk = rdev->raid_disk;
6564                 if (raid_disk >= max_disks
6565                     || raid_disk < 0 || test_bit(Journal, &rdev->flags))
6566                         continue;
6567                 disk = conf->disks + raid_disk;
6568
6569                 if (test_bit(Replacement, &rdev->flags)) {
6570                         if (disk->replacement)
6571                                 goto abort;
6572                         disk->replacement = rdev;
6573                 } else {
6574                         if (disk->rdev)
6575                                 goto abort;
6576                         disk->rdev = rdev;
6577                 }
6578
6579                 if (test_bit(In_sync, &rdev->flags)) {
6580                         char b[BDEVNAME_SIZE];
6581                         printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6582                                " disk %d\n",
6583                                mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6584                 } else if (rdev->saved_raid_disk != raid_disk)
6585                         /* Cannot rely on bitmap to complete recovery */
6586                         conf->fullsync = 1;
6587         }
6588
6589         conf->level = mddev->new_level;
6590         if (conf->level == 6) {
6591                 conf->max_degraded = 2;
6592                 if (raid6_call.xor_syndrome)
6593                         conf->rmw_level = PARITY_ENABLE_RMW;
6594                 else
6595                         conf->rmw_level = PARITY_DISABLE_RMW;
6596         } else {
6597                 conf->max_degraded = 1;
6598                 conf->rmw_level = PARITY_ENABLE_RMW;
6599         }
6600         conf->algorithm = mddev->new_layout;
6601         conf->reshape_progress = mddev->reshape_position;
6602         if (conf->reshape_progress != MaxSector) {
6603                 conf->prev_chunk_sectors = mddev->chunk_sectors;
6604                 conf->prev_algo = mddev->layout;
6605         } else {
6606                 conf->prev_chunk_sectors = conf->chunk_sectors;
6607                 conf->prev_algo = conf->algorithm;
6608         }
6609
6610         conf->min_nr_stripes = NR_STRIPES;
6611         memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6612                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6613         atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6614         if (grow_stripes(conf, conf->min_nr_stripes)) {
6615                 printk(KERN_ERR
6616                        "md/raid:%s: couldn't allocate %dkB for buffers\n",
6617                        mdname(mddev), memory);
6618                 goto abort;
6619         } else
6620                 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6621                        mdname(mddev), memory);
6622         /*
6623          * Losing a stripe head costs more than the time to refill it,
6624          * it reduces the queue depth and so can hurt throughput.
6625          * So set it rather large, scaled by number of devices.
6626          */
6627         conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6628         conf->shrinker.scan_objects = raid5_cache_scan;
6629         conf->shrinker.count_objects = raid5_cache_count;
6630         conf->shrinker.batch = 128;
6631         conf->shrinker.flags = 0;
6632         register_shrinker(&conf->shrinker);
6633
6634         sprintf(pers_name, "raid%d", mddev->new_level);
6635         conf->thread = md_register_thread(raid5d, mddev, pers_name);
6636         if (!conf->thread) {
6637                 printk(KERN_ERR
6638                        "md/raid:%s: couldn't allocate thread.\n",
6639                        mdname(mddev));
6640                 goto abort;
6641         }
6642
6643         return conf;
6644
6645  abort:
6646         if (conf) {
6647                 free_conf(conf);
6648                 return ERR_PTR(-EIO);
6649         } else
6650                 return ERR_PTR(-ENOMEM);
6651 }
6652
6653 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6654 {
6655         switch (algo) {
6656         case ALGORITHM_PARITY_0:
6657                 if (raid_disk < max_degraded)
6658                         return 1;
6659                 break;
6660         case ALGORITHM_PARITY_N:
6661                 if (raid_disk >= raid_disks - max_degraded)
6662                         return 1;
6663                 break;
6664         case ALGORITHM_PARITY_0_6:
6665                 if (raid_disk == 0 ||
6666                     raid_disk == raid_disks - 1)
6667                         return 1;
6668                 break;
6669         case ALGORITHM_LEFT_ASYMMETRIC_6:
6670         case ALGORITHM_RIGHT_ASYMMETRIC_6:
6671         case ALGORITHM_LEFT_SYMMETRIC_6:
6672         case ALGORITHM_RIGHT_SYMMETRIC_6:
6673                 if (raid_disk == raid_disks - 1)
6674                         return 1;
6675         }
6676         return 0;
6677 }
6678
6679 static int run(struct mddev *mddev)
6680 {
6681         struct r5conf *conf;
6682         int working_disks = 0;
6683         int dirty_parity_disks = 0;
6684         struct md_rdev *rdev;
6685         struct md_rdev *journal_dev = NULL;
6686         sector_t reshape_offset = 0;
6687         int i;
6688         long long min_offset_diff = 0;
6689         int first = 1;
6690
6691         if (mddev->recovery_cp != MaxSector)
6692                 printk(KERN_NOTICE "md/raid:%s: not clean"
6693                        " -- starting background reconstruction\n",
6694                        mdname(mddev));
6695
6696         rdev_for_each(rdev, mddev) {
6697                 long long diff;
6698
6699                 if (test_bit(Journal, &rdev->flags)) {
6700                         journal_dev = rdev;
6701                         continue;
6702                 }
6703                 if (rdev->raid_disk < 0)
6704                         continue;
6705                 diff = (rdev->new_data_offset - rdev->data_offset);
6706                 if (first) {
6707                         min_offset_diff = diff;
6708                         first = 0;
6709                 } else if (mddev->reshape_backwards &&
6710                          diff < min_offset_diff)
6711                         min_offset_diff = diff;
6712                 else if (!mddev->reshape_backwards &&
6713                          diff > min_offset_diff)
6714                         min_offset_diff = diff;
6715         }
6716
6717         if (mddev->reshape_position != MaxSector) {
6718                 /* Check that we can continue the reshape.
6719                  * Difficulties arise if the stripe we would write to
6720                  * next is at or after the stripe we would read from next.
6721                  * For a reshape that changes the number of devices, this
6722                  * is only possible for a very short time, and mdadm makes
6723                  * sure that time appears to have past before assembling
6724                  * the array.  So we fail if that time hasn't passed.
6725                  * For a reshape that keeps the number of devices the same
6726                  * mdadm must be monitoring the reshape can keeping the
6727                  * critical areas read-only and backed up.  It will start
6728                  * the array in read-only mode, so we check for that.
6729                  */
6730                 sector_t here_new, here_old;
6731                 int old_disks;
6732                 int max_degraded = (mddev->level == 6 ? 2 : 1);
6733                 int chunk_sectors;
6734                 int new_data_disks;
6735
6736                 if (journal_dev) {
6737                         printk(KERN_ERR "md/raid:%s: don't support reshape with journal - aborting.\n",
6738                                mdname(mddev));
6739                         return -EINVAL;
6740                 }
6741
6742                 if (mddev->new_level != mddev->level) {
6743                         printk(KERN_ERR "md/raid:%s: unsupported reshape "
6744                                "required - aborting.\n",
6745                                mdname(mddev));
6746                         return -EINVAL;
6747                 }
6748                 old_disks = mddev->raid_disks - mddev->delta_disks;
6749                 /* reshape_position must be on a new-stripe boundary, and one
6750                  * further up in new geometry must map after here in old
6751                  * geometry.
6752                  * If the chunk sizes are different, then as we perform reshape
6753                  * in units of the largest of the two, reshape_position needs
6754                  * be a multiple of the largest chunk size times new data disks.
6755                  */
6756                 here_new = mddev->reshape_position;
6757                 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
6758                 new_data_disks = mddev->raid_disks - max_degraded;
6759                 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
6760                         printk(KERN_ERR "md/raid:%s: reshape_position not "
6761                                "on a stripe boundary\n", mdname(mddev));
6762                         return -EINVAL;
6763                 }
6764                 reshape_offset = here_new * chunk_sectors;
6765                 /* here_new is the stripe we will write to */
6766                 here_old = mddev->reshape_position;
6767                 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
6768                 /* here_old is the first stripe that we might need to read
6769                  * from */
6770                 if (mddev->delta_disks == 0) {
6771                         /* We cannot be sure it is safe to start an in-place
6772                          * reshape.  It is only safe if user-space is monitoring
6773                          * and taking constant backups.
6774                          * mdadm always starts a situation like this in
6775                          * readonly mode so it can take control before
6776                          * allowing any writes.  So just check for that.
6777                          */
6778                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6779                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
6780                                 /* not really in-place - so OK */;
6781                         else if (mddev->ro == 0) {
6782                                 printk(KERN_ERR "md/raid:%s: in-place reshape "
6783                                        "must be started in read-only mode "
6784                                        "- aborting\n",
6785                                        mdname(mddev));
6786                                 return -EINVAL;
6787                         }
6788                 } else if (mddev->reshape_backwards
6789                     ? (here_new * chunk_sectors + min_offset_diff <=
6790                        here_old * chunk_sectors)
6791                     : (here_new * chunk_sectors >=
6792                        here_old * chunk_sectors + (-min_offset_diff))) {
6793                         /* Reading from the same stripe as writing to - bad */
6794                         printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6795                                "auto-recovery - aborting.\n",
6796                                mdname(mddev));
6797                         return -EINVAL;
6798                 }
6799                 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6800                        mdname(mddev));
6801                 /* OK, we should be able to continue; */
6802         } else {
6803                 BUG_ON(mddev->level != mddev->new_level);
6804                 BUG_ON(mddev->layout != mddev->new_layout);
6805                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6806                 BUG_ON(mddev->delta_disks != 0);
6807         }
6808
6809         if (mddev->private == NULL)
6810                 conf = setup_conf(mddev);
6811         else
6812                 conf = mddev->private;
6813
6814         if (IS_ERR(conf))
6815                 return PTR_ERR(conf);
6816
6817         if (test_bit(MD_HAS_JOURNAL, &mddev->flags) && !journal_dev) {
6818                 printk(KERN_ERR "md/raid:%s: journal disk is missing, force array readonly\n",
6819                        mdname(mddev));
6820                 mddev->ro = 1;
6821                 set_disk_ro(mddev->gendisk, 1);
6822         }
6823
6824         conf->min_offset_diff = min_offset_diff;
6825         mddev->thread = conf->thread;
6826         conf->thread = NULL;
6827         mddev->private = conf;
6828
6829         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6830              i++) {
6831                 rdev = conf->disks[i].rdev;
6832                 if (!rdev && conf->disks[i].replacement) {
6833                         /* The replacement is all we have yet */
6834                         rdev = conf->disks[i].replacement;
6835                         conf->disks[i].replacement = NULL;
6836                         clear_bit(Replacement, &rdev->flags);
6837                         conf->disks[i].rdev = rdev;
6838                 }
6839                 if (!rdev)
6840                         continue;
6841                 if (conf->disks[i].replacement &&
6842                     conf->reshape_progress != MaxSector) {
6843                         /* replacements and reshape simply do not mix. */
6844                         printk(KERN_ERR "md: cannot handle concurrent "
6845                                "replacement and reshape.\n");
6846                         goto abort;
6847                 }
6848                 if (test_bit(In_sync, &rdev->flags)) {
6849                         working_disks++;
6850                         continue;
6851                 }
6852                 /* This disc is not fully in-sync.  However if it
6853                  * just stored parity (beyond the recovery_offset),
6854                  * when we don't need to be concerned about the
6855                  * array being dirty.
6856                  * When reshape goes 'backwards', we never have
6857                  * partially completed devices, so we only need
6858                  * to worry about reshape going forwards.
6859                  */
6860                 /* Hack because v0.91 doesn't store recovery_offset properly. */
6861                 if (mddev->major_version == 0 &&
6862                     mddev->minor_version > 90)
6863                         rdev->recovery_offset = reshape_offset;
6864
6865                 if (rdev->recovery_offset < reshape_offset) {
6866                         /* We need to check old and new layout */
6867                         if (!only_parity(rdev->raid_disk,
6868                                          conf->algorithm,
6869                                          conf->raid_disks,
6870                                          conf->max_degraded))
6871                                 continue;
6872                 }
6873                 if (!only_parity(rdev->raid_disk,
6874                                  conf->prev_algo,
6875                                  conf->previous_raid_disks,
6876                                  conf->max_degraded))
6877                         continue;
6878                 dirty_parity_disks++;
6879         }
6880
6881         /*
6882          * 0 for a fully functional array, 1 or 2 for a degraded array.
6883          */
6884         mddev->degraded = calc_degraded(conf);
6885
6886         if (has_failed(conf)) {
6887                 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6888                         " (%d/%d failed)\n",
6889                         mdname(mddev), mddev->degraded, conf->raid_disks);
6890                 goto abort;
6891         }
6892
6893         /* device size must be a multiple of chunk size */
6894         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6895         mddev->resync_max_sectors = mddev->dev_sectors;
6896
6897         if (mddev->degraded > dirty_parity_disks &&
6898             mddev->recovery_cp != MaxSector) {
6899                 if (mddev->ok_start_degraded)
6900                         printk(KERN_WARNING
6901                                "md/raid:%s: starting dirty degraded array"
6902                                " - data corruption possible.\n",
6903                                mdname(mddev));
6904                 else {
6905                         printk(KERN_ERR
6906                                "md/raid:%s: cannot start dirty degraded array.\n",
6907                                mdname(mddev));
6908                         goto abort;
6909                 }
6910         }
6911
6912         if (mddev->degraded == 0)
6913                 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6914                        " devices, algorithm %d\n", mdname(mddev), conf->level,
6915                        mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6916                        mddev->new_layout);
6917         else
6918                 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6919                        " out of %d devices, algorithm %d\n",
6920                        mdname(mddev), conf->level,
6921                        mddev->raid_disks - mddev->degraded,
6922                        mddev->raid_disks, mddev->new_layout);
6923
6924         print_raid5_conf(conf);
6925
6926         if (conf->reshape_progress != MaxSector) {
6927                 conf->reshape_safe = conf->reshape_progress;
6928                 atomic_set(&conf->reshape_stripes, 0);
6929                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6930                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6931                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6932                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6933                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6934                                                         "reshape");
6935         }
6936
6937         /* Ok, everything is just fine now */
6938         if (mddev->to_remove == &raid5_attrs_group)
6939                 mddev->to_remove = NULL;
6940         else if (mddev->kobj.sd &&
6941             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6942                 printk(KERN_WARNING
6943                        "raid5: failed to create sysfs attributes for %s\n",
6944                        mdname(mddev));
6945         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6946
6947         if (mddev->queue) {
6948                 int chunk_size;
6949                 bool discard_supported = true;
6950                 /* read-ahead size must cover two whole stripes, which
6951                  * is 2 * (datadisks) * chunksize where 'n' is the
6952                  * number of raid devices
6953                  */
6954                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6955                 int stripe = data_disks *
6956                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6957                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6958                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6959
6960                 chunk_size = mddev->chunk_sectors << 9;
6961                 blk_queue_io_min(mddev->queue, chunk_size);
6962                 blk_queue_io_opt(mddev->queue, chunk_size *
6963                                  (conf->raid_disks - conf->max_degraded));
6964                 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6965                 /*
6966                  * We can only discard a whole stripe. It doesn't make sense to
6967                  * discard data disk but write parity disk
6968                  */
6969                 stripe = stripe * PAGE_SIZE;
6970                 /* Round up to power of 2, as discard handling
6971                  * currently assumes that */
6972                 while ((stripe-1) & stripe)
6973                         stripe = (stripe | (stripe-1)) + 1;
6974                 mddev->queue->limits.discard_alignment = stripe;
6975                 mddev->queue->limits.discard_granularity = stripe;
6976                 /*
6977                  * unaligned part of discard request will be ignored, so can't
6978                  * guarantee discard_zeroes_data
6979                  */
6980                 mddev->queue->limits.discard_zeroes_data = 0;
6981
6982                 blk_queue_max_write_same_sectors(mddev->queue, 0);
6983
6984                 rdev_for_each(rdev, mddev) {
6985                         disk_stack_limits(mddev->gendisk, rdev->bdev,
6986                                           rdev->data_offset << 9);
6987                         disk_stack_limits(mddev->gendisk, rdev->bdev,
6988                                           rdev->new_data_offset << 9);
6989                         /*
6990                          * discard_zeroes_data is required, otherwise data
6991                          * could be lost. Consider a scenario: discard a stripe
6992                          * (the stripe could be inconsistent if
6993                          * discard_zeroes_data is 0); write one disk of the
6994                          * stripe (the stripe could be inconsistent again
6995                          * depending on which disks are used to calculate
6996                          * parity); the disk is broken; The stripe data of this
6997                          * disk is lost.
6998                          */
6999                         if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
7000                             !bdev_get_queue(rdev->bdev)->
7001                                                 limits.discard_zeroes_data)
7002                                 discard_supported = false;
7003                         /* Unfortunately, discard_zeroes_data is not currently
7004                          * a guarantee - just a hint.  So we only allow DISCARD
7005                          * if the sysadmin has confirmed that only safe devices
7006                          * are in use by setting a module parameter.
7007                          */
7008                         if (!devices_handle_discard_safely) {
7009                                 if (discard_supported) {
7010                                         pr_info("md/raid456: discard support disabled due to uncertainty.\n");
7011                                         pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
7012                                 }
7013                                 discard_supported = false;
7014                         }
7015                 }
7016
7017                 if (discard_supported &&
7018                    mddev->queue->limits.max_discard_sectors >= stripe &&
7019                    mddev->queue->limits.discard_granularity >= stripe)
7020                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
7021                                                 mddev->queue);
7022                 else
7023                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
7024                                                 mddev->queue);
7025         }
7026
7027         if (journal_dev) {
7028                 char b[BDEVNAME_SIZE];
7029
7030                 printk(KERN_INFO"md/raid:%s: using device %s as journal\n",
7031                        mdname(mddev), bdevname(journal_dev->bdev, b));
7032                 r5l_init_log(conf, journal_dev);
7033         }
7034
7035         return 0;
7036 abort:
7037         md_unregister_thread(&mddev->thread);
7038         print_raid5_conf(conf);
7039         free_conf(conf);
7040         mddev->private = NULL;
7041         printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
7042         return -EIO;
7043 }
7044
7045 static void raid5_free(struct mddev *mddev, void *priv)
7046 {
7047         struct r5conf *conf = priv;
7048
7049         free_conf(conf);
7050         mddev->to_remove = &raid5_attrs_group;
7051 }
7052
7053 static void status(struct seq_file *seq, struct mddev *mddev)
7054 {
7055         struct r5conf *conf = mddev->private;
7056         int i;
7057
7058         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7059                 conf->chunk_sectors / 2, mddev->layout);
7060         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7061         for (i = 0; i < conf->raid_disks; i++)
7062                 seq_printf (seq, "%s",
7063                                conf->disks[i].rdev &&
7064                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
7065         seq_printf (seq, "]");
7066 }
7067
7068 static void print_raid5_conf (struct r5conf *conf)
7069 {
7070         int i;
7071         struct disk_info *tmp;
7072
7073         printk(KERN_DEBUG "RAID conf printout:\n");
7074         if (!conf) {
7075                 printk("(conf==NULL)\n");
7076                 return;
7077         }
7078         printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
7079                conf->raid_disks,
7080                conf->raid_disks - conf->mddev->degraded);
7081
7082         for (i = 0; i < conf->raid_disks; i++) {
7083                 char b[BDEVNAME_SIZE];
7084                 tmp = conf->disks + i;
7085                 if (tmp->rdev)
7086                         printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
7087                                i, !test_bit(Faulty, &tmp->rdev->flags),
7088                                bdevname(tmp->rdev->bdev, b));
7089         }
7090 }
7091
7092 static int raid5_spare_active(struct mddev *mddev)
7093 {
7094         int i;
7095         struct r5conf *conf = mddev->private;
7096         struct disk_info *tmp;
7097         int count = 0;
7098         unsigned long flags;
7099
7100         for (i = 0; i < conf->raid_disks; i++) {
7101                 tmp = conf->disks + i;
7102                 if (tmp->replacement
7103                     && tmp->replacement->recovery_offset == MaxSector
7104                     && !test_bit(Faulty, &tmp->replacement->flags)
7105                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7106                         /* Replacement has just become active. */
7107                         if (!tmp->rdev
7108                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7109                                 count++;
7110                         if (tmp->rdev) {
7111                                 /* Replaced device not technically faulty,
7112                                  * but we need to be sure it gets removed
7113                                  * and never re-added.
7114                                  */
7115                                 set_bit(Faulty, &tmp->rdev->flags);
7116                                 sysfs_notify_dirent_safe(
7117                                         tmp->rdev->sysfs_state);
7118                         }
7119                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7120                 } else if (tmp->rdev
7121                     && tmp->rdev->recovery_offset == MaxSector
7122                     && !test_bit(Faulty, &tmp->rdev->flags)
7123                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7124                         count++;
7125                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7126                 }
7127         }
7128         spin_lock_irqsave(&conf->device_lock, flags);
7129         mddev->degraded = calc_degraded(conf);
7130         spin_unlock_irqrestore(&conf->device_lock, flags);
7131         print_raid5_conf(conf);
7132         return count;
7133 }
7134
7135 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7136 {
7137         struct r5conf *conf = mddev->private;
7138         int err = 0;
7139         int number = rdev->raid_disk;
7140         struct md_rdev **rdevp;
7141         struct disk_info *p = conf->disks + number;
7142
7143         print_raid5_conf(conf);
7144         if (test_bit(Journal, &rdev->flags)) {
7145                 /*
7146                  * journal disk is not removable, but we need give a chance to
7147                  * update superblock of other disks. Otherwise journal disk
7148                  * will be considered as 'fresh'
7149                  */
7150                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7151                 return -EINVAL;
7152         }
7153         if (rdev == p->rdev)
7154                 rdevp = &p->rdev;
7155         else if (rdev == p->replacement)
7156                 rdevp = &p->replacement;
7157         else
7158                 return 0;
7159
7160         if (number >= conf->raid_disks &&
7161             conf->reshape_progress == MaxSector)
7162                 clear_bit(In_sync, &rdev->flags);
7163
7164         if (test_bit(In_sync, &rdev->flags) ||
7165             atomic_read(&rdev->nr_pending)) {
7166                 err = -EBUSY;
7167                 goto abort;
7168         }
7169         /* Only remove non-faulty devices if recovery
7170          * isn't possible.
7171          */
7172         if (!test_bit(Faulty, &rdev->flags) &&
7173             mddev->recovery_disabled != conf->recovery_disabled &&
7174             !has_failed(conf) &&
7175             (!p->replacement || p->replacement == rdev) &&
7176             number < conf->raid_disks) {
7177                 err = -EBUSY;
7178                 goto abort;
7179         }
7180         *rdevp = NULL;
7181         synchronize_rcu();
7182         if (atomic_read(&rdev->nr_pending)) {
7183                 /* lost the race, try later */
7184                 err = -EBUSY;
7185                 *rdevp = rdev;
7186         } else if (p->replacement) {
7187                 /* We must have just cleared 'rdev' */
7188                 p->rdev = p->replacement;
7189                 clear_bit(Replacement, &p->replacement->flags);
7190                 smp_mb(); /* Make sure other CPUs may see both as identical
7191                            * but will never see neither - if they are careful
7192                            */
7193                 p->replacement = NULL;
7194                 clear_bit(WantReplacement, &rdev->flags);
7195         } else
7196                 /* We might have just removed the Replacement as faulty-
7197                  * clear the bit just in case
7198                  */
7199                 clear_bit(WantReplacement, &rdev->flags);
7200 abort:
7201
7202         print_raid5_conf(conf);
7203         return err;
7204 }
7205
7206 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7207 {
7208         struct r5conf *conf = mddev->private;
7209         int err = -EEXIST;
7210         int disk;
7211         struct disk_info *p;
7212         int first = 0;
7213         int last = conf->raid_disks - 1;
7214
7215         if (test_bit(Journal, &rdev->flags))
7216                 return -EINVAL;
7217         if (mddev->recovery_disabled == conf->recovery_disabled)
7218                 return -EBUSY;
7219
7220         if (rdev->saved_raid_disk < 0 && has_failed(conf))
7221                 /* no point adding a device */
7222                 return -EINVAL;
7223
7224         if (rdev->raid_disk >= 0)
7225                 first = last = rdev->raid_disk;
7226
7227         /*
7228          * find the disk ... but prefer rdev->saved_raid_disk
7229          * if possible.
7230          */
7231         if (rdev->saved_raid_disk >= 0 &&
7232             rdev->saved_raid_disk >= first &&
7233             conf->disks[rdev->saved_raid_disk].rdev == NULL)
7234                 first = rdev->saved_raid_disk;
7235
7236         for (disk = first; disk <= last; disk++) {
7237                 p = conf->disks + disk;
7238                 if (p->rdev == NULL) {
7239                         clear_bit(In_sync, &rdev->flags);
7240                         rdev->raid_disk = disk;
7241                         err = 0;
7242                         if (rdev->saved_raid_disk != disk)
7243                                 conf->fullsync = 1;
7244                         rcu_assign_pointer(p->rdev, rdev);
7245                         goto out;
7246                 }
7247         }
7248         for (disk = first; disk <= last; disk++) {
7249                 p = conf->disks + disk;
7250                 if (test_bit(WantReplacement, &p->rdev->flags) &&
7251                     p->replacement == NULL) {
7252                         clear_bit(In_sync, &rdev->flags);
7253                         set_bit(Replacement, &rdev->flags);
7254                         rdev->raid_disk = disk;
7255                         err = 0;
7256                         conf->fullsync = 1;
7257                         rcu_assign_pointer(p->replacement, rdev);
7258                         break;
7259                 }
7260         }
7261 out:
7262         print_raid5_conf(conf);
7263         return err;
7264 }
7265
7266 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7267 {
7268         /* no resync is happening, and there is enough space
7269          * on all devices, so we can resize.
7270          * We need to make sure resync covers any new space.
7271          * If the array is shrinking we should possibly wait until
7272          * any io in the removed space completes, but it hardly seems
7273          * worth it.
7274          */
7275         sector_t newsize;
7276         struct r5conf *conf = mddev->private;
7277
7278         if (conf->log)
7279                 return -EINVAL;
7280         sectors &= ~((sector_t)conf->chunk_sectors - 1);
7281         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7282         if (mddev->external_size &&
7283             mddev->array_sectors > newsize)
7284                 return -EINVAL;
7285         if (mddev->bitmap) {
7286                 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7287                 if (ret)
7288                         return ret;
7289         }
7290         md_set_array_sectors(mddev, newsize);
7291         set_capacity(mddev->gendisk, mddev->array_sectors);
7292         revalidate_disk(mddev->gendisk);
7293         if (sectors > mddev->dev_sectors &&
7294             mddev->recovery_cp > mddev->dev_sectors) {
7295                 mddev->recovery_cp = mddev->dev_sectors;
7296                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7297         }
7298         mddev->dev_sectors = sectors;
7299         mddev->resync_max_sectors = sectors;
7300         return 0;
7301 }
7302
7303 static int check_stripe_cache(struct mddev *mddev)
7304 {
7305         /* Can only proceed if there are plenty of stripe_heads.
7306          * We need a minimum of one full stripe,, and for sensible progress
7307          * it is best to have about 4 times that.
7308          * If we require 4 times, then the default 256 4K stripe_heads will
7309          * allow for chunk sizes up to 256K, which is probably OK.
7310          * If the chunk size is greater, user-space should request more
7311          * stripe_heads first.
7312          */
7313         struct r5conf *conf = mddev->private;
7314         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7315             > conf->min_nr_stripes ||
7316             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7317             > conf->min_nr_stripes) {
7318                 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
7319                        mdname(mddev),
7320                        ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7321                         / STRIPE_SIZE)*4);
7322                 return 0;
7323         }
7324         return 1;
7325 }
7326
7327 static int check_reshape(struct mddev *mddev)
7328 {
7329         struct r5conf *conf = mddev->private;
7330
7331         if (conf->log)
7332                 return -EINVAL;
7333         if (mddev->delta_disks == 0 &&
7334             mddev->new_layout == mddev->layout &&
7335             mddev->new_chunk_sectors == mddev->chunk_sectors)
7336                 return 0; /* nothing to do */
7337         if (has_failed(conf))
7338                 return -EINVAL;
7339         if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7340                 /* We might be able to shrink, but the devices must
7341                  * be made bigger first.
7342                  * For raid6, 4 is the minimum size.
7343                  * Otherwise 2 is the minimum
7344                  */
7345                 int min = 2;
7346                 if (mddev->level == 6)
7347                         min = 4;
7348                 if (mddev->raid_disks + mddev->delta_disks < min)
7349                         return -EINVAL;
7350         }
7351
7352         if (!check_stripe_cache(mddev))
7353                 return -ENOSPC;
7354
7355         if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7356             mddev->delta_disks > 0)
7357                 if (resize_chunks(conf,
7358                                   conf->previous_raid_disks
7359                                   + max(0, mddev->delta_disks),
7360                                   max(mddev->new_chunk_sectors,
7361                                       mddev->chunk_sectors)
7362                             ) < 0)
7363                         return -ENOMEM;
7364         return resize_stripes(conf, (conf->previous_raid_disks
7365                                      + mddev->delta_disks));
7366 }
7367
7368 static int raid5_start_reshape(struct mddev *mddev)
7369 {
7370         struct r5conf *conf = mddev->private;
7371         struct md_rdev *rdev;
7372         int spares = 0;
7373         unsigned long flags;
7374
7375         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7376                 return -EBUSY;
7377
7378         if (!check_stripe_cache(mddev))
7379                 return -ENOSPC;
7380
7381         if (has_failed(conf))
7382                 return -EINVAL;
7383
7384         rdev_for_each(rdev, mddev) {
7385                 if (!test_bit(In_sync, &rdev->flags)
7386                     && !test_bit(Faulty, &rdev->flags))
7387                         spares++;
7388         }
7389
7390         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7391                 /* Not enough devices even to make a degraded array
7392                  * of that size
7393                  */
7394                 return -EINVAL;
7395
7396         /* Refuse to reduce size of the array.  Any reductions in
7397          * array size must be through explicit setting of array_size
7398          * attribute.
7399          */
7400         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7401             < mddev->array_sectors) {
7402                 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7403                        "before number of disks\n", mdname(mddev));
7404                 return -EINVAL;
7405         }
7406
7407         atomic_set(&conf->reshape_stripes, 0);
7408         spin_lock_irq(&conf->device_lock);
7409         write_seqcount_begin(&conf->gen_lock);
7410         conf->previous_raid_disks = conf->raid_disks;
7411         conf->raid_disks += mddev->delta_disks;
7412         conf->prev_chunk_sectors = conf->chunk_sectors;
7413         conf->chunk_sectors = mddev->new_chunk_sectors;
7414         conf->prev_algo = conf->algorithm;
7415         conf->algorithm = mddev->new_layout;
7416         conf->generation++;
7417         /* Code that selects data_offset needs to see the generation update
7418          * if reshape_progress has been set - so a memory barrier needed.
7419          */
7420         smp_mb();
7421         if (mddev->reshape_backwards)
7422                 conf->reshape_progress = raid5_size(mddev, 0, 0);
7423         else
7424                 conf->reshape_progress = 0;
7425         conf->reshape_safe = conf->reshape_progress;
7426         write_seqcount_end(&conf->gen_lock);
7427         spin_unlock_irq(&conf->device_lock);
7428
7429         /* Now make sure any requests that proceeded on the assumption
7430          * the reshape wasn't running - like Discard or Read - have
7431          * completed.
7432          */
7433         mddev_suspend(mddev);
7434         mddev_resume(mddev);
7435
7436         /* Add some new drives, as many as will fit.
7437          * We know there are enough to make the newly sized array work.
7438          * Don't add devices if we are reducing the number of
7439          * devices in the array.  This is because it is not possible
7440          * to correctly record the "partially reconstructed" state of
7441          * such devices during the reshape and confusion could result.
7442          */
7443         if (mddev->delta_disks >= 0) {
7444                 rdev_for_each(rdev, mddev)
7445                         if (rdev->raid_disk < 0 &&
7446                             !test_bit(Faulty, &rdev->flags)) {
7447                                 if (raid5_add_disk(mddev, rdev) == 0) {
7448                                         if (rdev->raid_disk
7449                                             >= conf->previous_raid_disks)
7450                                                 set_bit(In_sync, &rdev->flags);
7451                                         else
7452                                                 rdev->recovery_offset = 0;
7453
7454                                         if (sysfs_link_rdev(mddev, rdev))
7455                                                 /* Failure here is OK */;
7456                                 }
7457                         } else if (rdev->raid_disk >= conf->previous_raid_disks
7458                                    && !test_bit(Faulty, &rdev->flags)) {
7459                                 /* This is a spare that was manually added */
7460                                 set_bit(In_sync, &rdev->flags);
7461                         }
7462
7463                 /* When a reshape changes the number of devices,
7464                  * ->degraded is measured against the larger of the
7465                  * pre and post number of devices.
7466                  */
7467                 spin_lock_irqsave(&conf->device_lock, flags);
7468                 mddev->degraded = calc_degraded(conf);
7469                 spin_unlock_irqrestore(&conf->device_lock, flags);
7470         }
7471         mddev->raid_disks = conf->raid_disks;
7472         mddev->reshape_position = conf->reshape_progress;
7473         set_bit(MD_CHANGE_DEVS, &mddev->flags);
7474
7475         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7476         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7477         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7478         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7479         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7480         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7481                                                 "reshape");
7482         if (!mddev->sync_thread) {
7483                 mddev->recovery = 0;
7484                 spin_lock_irq(&conf->device_lock);
7485                 write_seqcount_begin(&conf->gen_lock);
7486                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7487                 mddev->new_chunk_sectors =
7488                         conf->chunk_sectors = conf->prev_chunk_sectors;
7489                 mddev->new_layout = conf->algorithm = conf->prev_algo;
7490                 rdev_for_each(rdev, mddev)
7491                         rdev->new_data_offset = rdev->data_offset;
7492                 smp_wmb();
7493                 conf->generation --;
7494                 conf->reshape_progress = MaxSector;
7495                 mddev->reshape_position = MaxSector;
7496                 write_seqcount_end(&conf->gen_lock);
7497                 spin_unlock_irq(&conf->device_lock);
7498                 return -EAGAIN;
7499         }
7500         conf->reshape_checkpoint = jiffies;
7501         md_wakeup_thread(mddev->sync_thread);
7502         md_new_event(mddev);
7503         return 0;
7504 }
7505
7506 /* This is called from the reshape thread and should make any
7507  * changes needed in 'conf'
7508  */
7509 static void end_reshape(struct r5conf *conf)
7510 {
7511
7512         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7513                 struct md_rdev *rdev;
7514
7515                 spin_lock_irq(&conf->device_lock);
7516                 conf->previous_raid_disks = conf->raid_disks;
7517                 rdev_for_each(rdev, conf->mddev)
7518                         rdev->data_offset = rdev->new_data_offset;
7519                 smp_wmb();
7520                 conf->reshape_progress = MaxSector;
7521                 conf->mddev->reshape_position = MaxSector;
7522                 spin_unlock_irq(&conf->device_lock);
7523                 wake_up(&conf->wait_for_overlap);
7524
7525                 /* read-ahead size must cover two whole stripes, which is
7526                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7527                  */
7528                 if (conf->mddev->queue) {
7529                         int data_disks = conf->raid_disks - conf->max_degraded;
7530                         int stripe = data_disks * ((conf->chunk_sectors << 9)
7531                                                    / PAGE_SIZE);
7532                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7533                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7534                 }
7535         }
7536 }
7537
7538 /* This is called from the raid5d thread with mddev_lock held.
7539  * It makes config changes to the device.
7540  */
7541 static void raid5_finish_reshape(struct mddev *mddev)
7542 {
7543         struct r5conf *conf = mddev->private;
7544
7545         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7546
7547                 if (mddev->delta_disks > 0) {
7548                         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7549                         set_capacity(mddev->gendisk, mddev->array_sectors);
7550                         revalidate_disk(mddev->gendisk);
7551                 } else {
7552                         int d;
7553                         spin_lock_irq(&conf->device_lock);
7554                         mddev->degraded = calc_degraded(conf);
7555                         spin_unlock_irq(&conf->device_lock);
7556                         for (d = conf->raid_disks ;
7557                              d < conf->raid_disks - mddev->delta_disks;
7558                              d++) {
7559                                 struct md_rdev *rdev = conf->disks[d].rdev;
7560                                 if (rdev)
7561                                         clear_bit(In_sync, &rdev->flags);
7562                                 rdev = conf->disks[d].replacement;
7563                                 if (rdev)
7564                                         clear_bit(In_sync, &rdev->flags);
7565                         }
7566                 }
7567                 mddev->layout = conf->algorithm;
7568                 mddev->chunk_sectors = conf->chunk_sectors;
7569                 mddev->reshape_position = MaxSector;
7570                 mddev->delta_disks = 0;
7571                 mddev->reshape_backwards = 0;
7572         }
7573 }
7574
7575 static void raid5_quiesce(struct mddev *mddev, int state)
7576 {
7577         struct r5conf *conf = mddev->private;
7578
7579         switch(state) {
7580         case 2: /* resume for a suspend */
7581                 wake_up(&conf->wait_for_overlap);
7582                 break;
7583
7584         case 1: /* stop all writes */
7585                 lock_all_device_hash_locks_irq(conf);
7586                 /* '2' tells resync/reshape to pause so that all
7587                  * active stripes can drain
7588                  */
7589                 conf->quiesce = 2;
7590                 wait_event_cmd(conf->wait_for_quiescent,
7591                                     atomic_read(&conf->active_stripes) == 0 &&
7592                                     atomic_read(&conf->active_aligned_reads) == 0,
7593                                     unlock_all_device_hash_locks_irq(conf),
7594                                     lock_all_device_hash_locks_irq(conf));
7595                 conf->quiesce = 1;
7596                 unlock_all_device_hash_locks_irq(conf);
7597                 /* allow reshape to continue */
7598                 wake_up(&conf->wait_for_overlap);
7599                 break;
7600
7601         case 0: /* re-enable writes */
7602                 lock_all_device_hash_locks_irq(conf);
7603                 conf->quiesce = 0;
7604                 wake_up(&conf->wait_for_quiescent);
7605                 wake_up(&conf->wait_for_overlap);
7606                 unlock_all_device_hash_locks_irq(conf);
7607                 break;
7608         }
7609         r5l_quiesce(conf->log, state);
7610 }
7611
7612 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7613 {
7614         struct r0conf *raid0_conf = mddev->private;
7615         sector_t sectors;
7616
7617         /* for raid0 takeover only one zone is supported */
7618         if (raid0_conf->nr_strip_zones > 1) {
7619                 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7620                        mdname(mddev));
7621                 return ERR_PTR(-EINVAL);
7622         }
7623
7624         sectors = raid0_conf->strip_zone[0].zone_end;
7625         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7626         mddev->dev_sectors = sectors;
7627         mddev->new_level = level;
7628         mddev->new_layout = ALGORITHM_PARITY_N;
7629         mddev->new_chunk_sectors = mddev->chunk_sectors;
7630         mddev->raid_disks += 1;
7631         mddev->delta_disks = 1;
7632         /* make sure it will be not marked as dirty */
7633         mddev->recovery_cp = MaxSector;
7634
7635         return setup_conf(mddev);
7636 }
7637
7638 static void *raid5_takeover_raid1(struct mddev *mddev)
7639 {
7640         int chunksect;
7641
7642         if (mddev->raid_disks != 2 ||
7643             mddev->degraded > 1)
7644                 return ERR_PTR(-EINVAL);
7645
7646         /* Should check if there are write-behind devices? */
7647
7648         chunksect = 64*2; /* 64K by default */
7649
7650         /* The array must be an exact multiple of chunksize */
7651         while (chunksect && (mddev->array_sectors & (chunksect-1)))
7652                 chunksect >>= 1;
7653
7654         if ((chunksect<<9) < STRIPE_SIZE)
7655                 /* array size does not allow a suitable chunk size */
7656                 return ERR_PTR(-EINVAL);
7657
7658         mddev->new_level = 5;
7659         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7660         mddev->new_chunk_sectors = chunksect;
7661
7662         return setup_conf(mddev);
7663 }
7664
7665 static void *raid5_takeover_raid6(struct mddev *mddev)
7666 {
7667         int new_layout;
7668
7669         switch (mddev->layout) {
7670         case ALGORITHM_LEFT_ASYMMETRIC_6:
7671                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7672                 break;
7673         case ALGORITHM_RIGHT_ASYMMETRIC_6:
7674                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7675                 break;
7676         case ALGORITHM_LEFT_SYMMETRIC_6:
7677                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7678                 break;
7679         case ALGORITHM_RIGHT_SYMMETRIC_6:
7680                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7681                 break;
7682         case ALGORITHM_PARITY_0_6:
7683                 new_layout = ALGORITHM_PARITY_0;
7684                 break;
7685         case ALGORITHM_PARITY_N:
7686                 new_layout = ALGORITHM_PARITY_N;
7687                 break;
7688         default:
7689                 return ERR_PTR(-EINVAL);
7690         }
7691         mddev->new_level = 5;
7692         mddev->new_layout = new_layout;
7693         mddev->delta_disks = -1;
7694         mddev->raid_disks -= 1;
7695         return setup_conf(mddev);
7696 }
7697
7698 static int raid5_check_reshape(struct mddev *mddev)
7699 {
7700         /* For a 2-drive array, the layout and chunk size can be changed
7701          * immediately as not restriping is needed.
7702          * For larger arrays we record the new value - after validation
7703          * to be used by a reshape pass.
7704          */
7705         struct r5conf *conf = mddev->private;
7706         int new_chunk = mddev->new_chunk_sectors;
7707
7708         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7709                 return -EINVAL;
7710         if (new_chunk > 0) {
7711                 if (!is_power_of_2(new_chunk))
7712                         return -EINVAL;
7713                 if (new_chunk < (PAGE_SIZE>>9))
7714                         return -EINVAL;
7715                 if (mddev->array_sectors & (new_chunk-1))
7716                         /* not factor of array size */
7717                         return -EINVAL;
7718         }
7719
7720         /* They look valid */
7721
7722         if (mddev->raid_disks == 2) {
7723                 /* can make the change immediately */
7724                 if (mddev->new_layout >= 0) {
7725                         conf->algorithm = mddev->new_layout;
7726                         mddev->layout = mddev->new_layout;
7727                 }
7728                 if (new_chunk > 0) {
7729                         conf->chunk_sectors = new_chunk ;
7730                         mddev->chunk_sectors = new_chunk;
7731                 }
7732                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7733                 md_wakeup_thread(mddev->thread);
7734         }
7735         return check_reshape(mddev);
7736 }
7737
7738 static int raid6_check_reshape(struct mddev *mddev)
7739 {
7740         int new_chunk = mddev->new_chunk_sectors;
7741
7742         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7743                 return -EINVAL;
7744         if (new_chunk > 0) {
7745                 if (!is_power_of_2(new_chunk))
7746                         return -EINVAL;
7747                 if (new_chunk < (PAGE_SIZE >> 9))
7748                         return -EINVAL;
7749                 if (mddev->array_sectors & (new_chunk-1))
7750                         /* not factor of array size */
7751                         return -EINVAL;
7752         }
7753
7754         /* They look valid */
7755         return check_reshape(mddev);
7756 }
7757
7758 static void *raid5_takeover(struct mddev *mddev)
7759 {
7760         /* raid5 can take over:
7761          *  raid0 - if there is only one strip zone - make it a raid4 layout
7762          *  raid1 - if there are two drives.  We need to know the chunk size
7763          *  raid4 - trivial - just use a raid4 layout.
7764          *  raid6 - Providing it is a *_6 layout
7765          */
7766         if (mddev->level == 0)
7767                 return raid45_takeover_raid0(mddev, 5);
7768         if (mddev->level == 1)
7769                 return raid5_takeover_raid1(mddev);
7770         if (mddev->level == 4) {
7771                 mddev->new_layout = ALGORITHM_PARITY_N;
7772                 mddev->new_level = 5;
7773                 return setup_conf(mddev);
7774         }
7775         if (mddev->level == 6)
7776                 return raid5_takeover_raid6(mddev);
7777
7778         return ERR_PTR(-EINVAL);
7779 }
7780
7781 static void *raid4_takeover(struct mddev *mddev)
7782 {
7783         /* raid4 can take over:
7784          *  raid0 - if there is only one strip zone
7785          *  raid5 - if layout is right
7786          */
7787         if (mddev->level == 0)
7788                 return raid45_takeover_raid0(mddev, 4);
7789         if (mddev->level == 5 &&
7790             mddev->layout == ALGORITHM_PARITY_N) {
7791                 mddev->new_layout = 0;
7792                 mddev->new_level = 4;
7793                 return setup_conf(mddev);
7794         }
7795         return ERR_PTR(-EINVAL);
7796 }
7797
7798 static struct md_personality raid5_personality;
7799
7800 static void *raid6_takeover(struct mddev *mddev)
7801 {
7802         /* Currently can only take over a raid5.  We map the
7803          * personality to an equivalent raid6 personality
7804          * with the Q block at the end.
7805          */
7806         int new_layout;
7807
7808         if (mddev->pers != &raid5_personality)
7809                 return ERR_PTR(-EINVAL);
7810         if (mddev->degraded > 1)
7811                 return ERR_PTR(-EINVAL);
7812         if (mddev->raid_disks > 253)
7813                 return ERR_PTR(-EINVAL);
7814         if (mddev->raid_disks < 3)
7815                 return ERR_PTR(-EINVAL);
7816
7817         switch (mddev->layout) {
7818         case ALGORITHM_LEFT_ASYMMETRIC:
7819                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7820                 break;
7821         case ALGORITHM_RIGHT_ASYMMETRIC:
7822                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7823                 break;
7824         case ALGORITHM_LEFT_SYMMETRIC:
7825                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7826                 break;
7827         case ALGORITHM_RIGHT_SYMMETRIC:
7828                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7829                 break;
7830         case ALGORITHM_PARITY_0:
7831                 new_layout = ALGORITHM_PARITY_0_6;
7832                 break;
7833         case ALGORITHM_PARITY_N:
7834                 new_layout = ALGORITHM_PARITY_N;
7835                 break;
7836         default:
7837                 return ERR_PTR(-EINVAL);
7838         }
7839         mddev->new_level = 6;
7840         mddev->new_layout = new_layout;
7841         mddev->delta_disks = 1;
7842         mddev->raid_disks += 1;
7843         return setup_conf(mddev);
7844 }
7845
7846 static struct md_personality raid6_personality =
7847 {
7848         .name           = "raid6",
7849         .level          = 6,
7850         .owner          = THIS_MODULE,
7851         .make_request   = make_request,
7852         .run            = run,
7853         .free           = raid5_free,
7854         .status         = status,
7855         .error_handler  = error,
7856         .hot_add_disk   = raid5_add_disk,
7857         .hot_remove_disk= raid5_remove_disk,
7858         .spare_active   = raid5_spare_active,
7859         .sync_request   = sync_request,
7860         .resize         = raid5_resize,
7861         .size           = raid5_size,
7862         .check_reshape  = raid6_check_reshape,
7863         .start_reshape  = raid5_start_reshape,
7864         .finish_reshape = raid5_finish_reshape,
7865         .quiesce        = raid5_quiesce,
7866         .takeover       = raid6_takeover,
7867         .congested      = raid5_congested,
7868 };
7869 static struct md_personality raid5_personality =
7870 {
7871         .name           = "raid5",
7872         .level          = 5,
7873         .owner          = THIS_MODULE,
7874         .make_request   = make_request,
7875         .run            = run,
7876         .free           = raid5_free,
7877         .status         = status,
7878         .error_handler  = error,
7879         .hot_add_disk   = raid5_add_disk,
7880         .hot_remove_disk= raid5_remove_disk,
7881         .spare_active   = raid5_spare_active,
7882         .sync_request   = sync_request,
7883         .resize         = raid5_resize,
7884         .size           = raid5_size,
7885         .check_reshape  = raid5_check_reshape,
7886         .start_reshape  = raid5_start_reshape,
7887         .finish_reshape = raid5_finish_reshape,
7888         .quiesce        = raid5_quiesce,
7889         .takeover       = raid5_takeover,
7890         .congested      = raid5_congested,
7891 };
7892
7893 static struct md_personality raid4_personality =
7894 {
7895         .name           = "raid4",
7896         .level          = 4,
7897         .owner          = THIS_MODULE,
7898         .make_request   = make_request,
7899         .run            = run,
7900         .free           = raid5_free,
7901         .status         = status,
7902         .error_handler  = error,
7903         .hot_add_disk   = raid5_add_disk,
7904         .hot_remove_disk= raid5_remove_disk,
7905         .spare_active   = raid5_spare_active,
7906         .sync_request   = sync_request,
7907         .resize         = raid5_resize,
7908         .size           = raid5_size,
7909         .check_reshape  = raid5_check_reshape,
7910         .start_reshape  = raid5_start_reshape,
7911         .finish_reshape = raid5_finish_reshape,
7912         .quiesce        = raid5_quiesce,
7913         .takeover       = raid4_takeover,
7914         .congested      = raid5_congested,
7915 };
7916
7917 static int __init raid5_init(void)
7918 {
7919         raid5_wq = alloc_workqueue("raid5wq",
7920                 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7921         if (!raid5_wq)
7922                 return -ENOMEM;
7923         register_md_personality(&raid6_personality);
7924         register_md_personality(&raid5_personality);
7925         register_md_personality(&raid4_personality);
7926         return 0;
7927 }
7928
7929 static void raid5_exit(void)
7930 {
7931         unregister_md_personality(&raid6_personality);
7932         unregister_md_personality(&raid5_personality);
7933         unregister_md_personality(&raid4_personality);
7934         destroy_workqueue(raid5_wq);
7935 }
7936
7937 module_init(raid5_init);
7938 module_exit(raid5_exit);
7939 MODULE_LICENSE("GPL");
7940 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7941 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7942 MODULE_ALIAS("md-raid5");
7943 MODULE_ALIAS("md-raid4");
7944 MODULE_ALIAS("md-level-5");
7945 MODULE_ALIAS("md-level-4");
7946 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7947 MODULE_ALIAS("md-raid6");
7948 MODULE_ALIAS("md-level-6");
7949
7950 /* This used to be two separate modules, they were: */
7951 MODULE_ALIAS("raid5");
7952 MODULE_ALIAS("raid6");