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 <trace/events/block.h>
  57
  58 #include "md.h"
  59 #include "raid5.h"
  60 #include "raid0.h"
  61 #include "bitmap.h"
  62
  63 /*
  64  * Stripe cache
  65  */
  66
  67 #define NR_STRIPES              256
  68 #define STRIPE_SIZE             PAGE_SIZE
  69 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
  70 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
  71 #define IO_THRESHOLD            1
  72 #define BYPASS_THRESHOLD        1
  73 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
  74 #define HASH_MASK               (NR_HASH - 1)
  75
  76 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
  77 {
  78         int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
  79         return &conf->stripe_hashtbl[hash];
  80 }
  81
  82 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
  83  * order without overlap.  There may be several bio's per stripe+device, and
  84  * a bio could span several devices.
  85  * When walking this list for a particular stripe+device, we must never proceed
  86  * beyond a bio that extends past this device, as the next bio might no longer
  87  * be valid.
  88  * This function is used to determine the 'next' bio in the list, given the sector
  89  * of the current stripe+device
  90  */
  91 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
  92 {
  93         int sectors = bio_sectors(bio);
  94         if (bio->bi_sector + sectors < sector + STRIPE_SECTORS)
  95                 return bio->bi_next;
  96         else
  97                 return NULL;
  98 }
  99
 100 /*
 101  * We maintain a biased count of active stripes in the bottom 16 bits of
 102  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
 103  */
 104 static inline int raid5_bi_processed_stripes(struct bio *bio)
 105 {
 106         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 107         return (atomic_read(segments) >> 16) & 0xffff;
 108 }
 109
 110 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
 111 {
 112         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 113         return atomic_sub_return(1, segments) & 0xffff;
 114 }
 115
 116 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
 117 {
 118         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 119         atomic_inc(segments);
 120 }
 121
 122 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
 123         unsigned int cnt)
 124 {
 125         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 126         int old, new;
 127
 128         do {
 129                 old = atomic_read(segments);
 130                 new = (old & 0xffff) | (cnt << 16);
 131         } while (atomic_cmpxchg(segments, old, new) != old);
 132 }
 133
 134 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
 135 {
 136         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 137         atomic_set(segments, cnt);
 138 }
 139
 140 /* Find first data disk in a raid6 stripe */
 141 static inline int raid6_d0(struct stripe_head *sh)
 142 {
 143         if (sh->ddf_layout)
 144                 /* ddf always start from first device */
 145                 return 0;
 146         /* md starts just after Q block */
 147         if (sh->qd_idx == sh->disks - 1)
 148                 return 0;
 149         else
 150                 return sh->qd_idx + 1;
 151 }
 152 static inline int raid6_next_disk(int disk, int raid_disks)
 153 {
 154         disk++;
 155         return (disk < raid_disks) ? disk : 0;
 156 }
 157
 158 /* When walking through the disks in a raid5, starting at raid6_d0,
 159  * We need to map each disk to a 'slot', where the data disks are slot
 160  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 161  * is raid_disks-1.  This help does that mapping.
 162  */
 163 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 164                              int *count, int syndrome_disks)
 165 {
 166         int slot = *count;
 167
 168         if (sh->ddf_layout)
 169                 (*count)++;
 170         if (idx == sh->pd_idx)
 171                 return syndrome_disks;
 172         if (idx == sh->qd_idx)
 173                 return syndrome_disks + 1;
 174         if (!sh->ddf_layout)
 175                 (*count)++;
 176         return slot;
 177 }
 178
 179 static void return_io(struct bio *return_bi)
 180 {
 181         struct bio *bi = return_bi;
 182         while (bi) {
 183
 184                 return_bi = bi->bi_next;
 185                 bi->bi_next = NULL;
 186                 bi->bi_size = 0;
 187                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
 188                                          bi, 0);
 189                 bio_endio(bi, 0);
 190                 bi = return_bi;
 191         }
 192 }
 193
 194 static void print_raid5_conf (struct r5conf *conf);
 195
 196 static int stripe_operations_active(struct stripe_head *sh)
 197 {
 198         return sh->check_state || sh->reconstruct_state ||
 199                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 200                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 201 }
 202
 203 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh)
 204 {
 205         BUG_ON(!list_empty(&sh->lru));
 206         BUG_ON(atomic_read(&conf->active_stripes)==0);
 207         if (test_bit(STRIPE_HANDLE, &sh->state)) {
 208                 if (test_bit(STRIPE_DELAYED, &sh->state) &&
 209                     !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 210                         list_add_tail(&sh->lru, &conf->delayed_list);
 211                 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 212                            sh->bm_seq - conf->seq_write > 0)
 213                         list_add_tail(&sh->lru, &conf->bitmap_list);
 214                 else {
 215                         clear_bit(STRIPE_DELAYED, &sh->state);
 216                         clear_bit(STRIPE_BIT_DELAY, &sh->state);
 217                         list_add_tail(&sh->lru, &conf->handle_list);
 218                 }
 219                 md_wakeup_thread(conf->mddev->thread);
 220         } else {
 221                 BUG_ON(stripe_operations_active(sh));
 222                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 223                         if (atomic_dec_return(&conf->preread_active_stripes)
 224                             < IO_THRESHOLD)
 225                                 md_wakeup_thread(conf->mddev->thread);
 226                 atomic_dec(&conf->active_stripes);
 227                 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
 228                         list_add_tail(&sh->lru, &conf->inactive_list);
 229                         wake_up(&conf->wait_for_stripe);
 230                         if (conf->retry_read_aligned)
 231                                 md_wakeup_thread(conf->mddev->thread);
 232                 }
 233         }
 234 }
 235
 236 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
 237 {
 238         if (atomic_dec_and_test(&sh->count))
 239                 do_release_stripe(conf, sh);
 240 }
 241
 242 static void release_stripe(struct stripe_head *sh)
 243 {
 244         struct r5conf *conf = sh->raid_conf;
 245         unsigned long flags;
 246
 247         local_irq_save(flags);
 248         if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
 249                 do_release_stripe(conf, sh);
 250                 spin_unlock(&conf->device_lock);
 251         }
 252         local_irq_restore(flags);
 253 }
 254
 255 static inline void remove_hash(struct stripe_head *sh)
 256 {
 257         pr_debug("remove_hash(), stripe %llu\n",
 258                 (unsigned long long)sh->sector);
 259
 260         hlist_del_init(&sh->hash);
 261 }
 262
 263 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 264 {
 265         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 266
 267         pr_debug("insert_hash(), stripe %llu\n",
 268                 (unsigned long long)sh->sector);
 269
 270         hlist_add_head(&sh->hash, hp);
 271 }
 272
 273
 274 /* find an idle stripe, make sure it is unhashed, and return it. */
 275 static struct stripe_head *get_free_stripe(struct r5conf *conf)
 276 {
 277         struct stripe_head *sh = NULL;
 278         struct list_head *first;
 279
 280         if (list_empty(&conf->inactive_list))
 281                 goto out;
 282         first = conf->inactive_list.next;
 283         sh = list_entry(first, struct stripe_head, lru);
 284         list_del_init(first);
 285         remove_hash(sh);
 286         atomic_inc(&conf->active_stripes);
 287 out:
 288         return sh;
 289 }
 290
 291 static void shrink_buffers(struct stripe_head *sh)
 292 {
 293         struct page *p;
 294         int i;
 295         int num = sh->raid_conf->pool_size;
 296
 297         for (i = 0; i < num ; i++) {
 298                 p = sh->dev[i].page;
 299                 if (!p)
 300                         continue;
 301                 sh->dev[i].page = NULL;
 302                 put_page(p);
 303         }
 304 }
 305
 306 static int grow_buffers(struct stripe_head *sh)
 307 {
 308         int i;
 309         int num = sh->raid_conf->pool_size;
 310
 311         for (i = 0; i < num; i++) {
 312                 struct page *page;
 313
 314                 if (!(page = alloc_page(GFP_KERNEL))) {
 315                         return 1;
 316                 }
 317                 sh->dev[i].page = page;
 318         }
 319         return 0;
 320 }
 321
 322 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
 323 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 324                             struct stripe_head *sh);
 325
 326 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 327 {
 328         struct r5conf *conf = sh->raid_conf;
 329         int i;
 330
 331         BUG_ON(atomic_read(&sh->count) != 0);
 332         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 333         BUG_ON(stripe_operations_active(sh));
 334
 335         pr_debug("init_stripe called, stripe %llu\n",
 336                 (unsigned long long)sh->sector);
 337
 338         remove_hash(sh);
 339
 340         sh->generation = conf->generation - previous;
 341         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 342         sh->sector = sector;
 343         stripe_set_idx(sector, conf, previous, sh);
 344         sh->state = 0;
 345
 346
 347         for (i = sh->disks; i--; ) {
 348                 struct r5dev *dev = &sh->dev[i];
 349
 350                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 351                     test_bit(R5_LOCKED, &dev->flags)) {
 352                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
 353                                (unsigned long long)sh->sector, i, dev->toread,
 354                                dev->read, dev->towrite, dev->written,
 355                                test_bit(R5_LOCKED, &dev->flags));
 356                         WARN_ON(1);
 357                 }
 358                 dev->flags = 0;
 359                 raid5_build_block(sh, i, previous);
 360         }
 361         insert_hash(conf, sh);
 362 }
 363
 364 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 365                                          short generation)
 366 {
 367         struct stripe_head *sh;
 368
 369         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 370         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 371                 if (sh->sector == sector && sh->generation == generation)
 372                         return sh;
 373         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 374         return NULL;
 375 }
 376
 377 /*
 378  * Need to check if array has failed when deciding whether to:
 379  *  - start an array
 380  *  - remove non-faulty devices
 381  *  - add a spare
 382  *  - allow a reshape
 383  * This determination is simple when no reshape is happening.
 384  * However if there is a reshape, we need to carefully check
 385  * both the before and after sections.
 386  * This is because some failed devices may only affect one
 387  * of the two sections, and some non-in_sync devices may
 388  * be insync in the section most affected by failed devices.
 389  */
 390 static int calc_degraded(struct r5conf *conf)
 391 {
 392         int degraded, degraded2;
 393         int i;
 394
 395         rcu_read_lock();
 396         degraded = 0;
 397         for (i = 0; i < conf->previous_raid_disks; i++) {
 398                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 399                 if (rdev && test_bit(Faulty, &rdev->flags))
 400                         rdev = rcu_dereference(conf->disks[i].replacement);
 401                 if (!rdev || test_bit(Faulty, &rdev->flags))
 402                         degraded++;
 403                 else if (test_bit(In_sync, &rdev->flags))
 404                         ;
 405                 else
 406                         /* not in-sync or faulty.
 407                          * If the reshape increases the number of devices,
 408                          * this is being recovered by the reshape, so
 409                          * this 'previous' section is not in_sync.
 410                          * If the number of devices is being reduced however,
 411                          * the device can only be part of the array if
 412                          * we are reverting a reshape, so this section will
 413                          * be in-sync.
 414                          */
 415                         if (conf->raid_disks >= conf->previous_raid_disks)
 416                                 degraded++;
 417         }
 418         rcu_read_unlock();
 419         if (conf->raid_disks == conf->previous_raid_disks)
 420                 return degraded;
 421         rcu_read_lock();
 422         degraded2 = 0;
 423         for (i = 0; i < conf->raid_disks; i++) {
 424                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 425                 if (rdev && test_bit(Faulty, &rdev->flags))
 426                         rdev = rcu_dereference(conf->disks[i].replacement);
 427                 if (!rdev || test_bit(Faulty, &rdev->flags))
 428                         degraded2++;
 429                 else if (test_bit(In_sync, &rdev->flags))
 430                         ;
 431                 else
 432                         /* not in-sync or faulty.
 433                          * If reshape increases the number of devices, this
 434                          * section has already been recovered, else it
 435                          * almost certainly hasn't.
 436                          */
 437                         if (conf->raid_disks <= conf->previous_raid_disks)
 438                                 degraded2++;
 439         }
 440         rcu_read_unlock();
 441         if (degraded2 > degraded)
 442                 return degraded2;
 443         return degraded;
 444 }
 445
 446 static int has_failed(struct r5conf *conf)
 447 {
 448         int degraded;
 449
 450         if (conf->mddev->reshape_position == MaxSector)
 451                 return conf->mddev->degraded > conf->max_degraded;
 452
 453         degraded = calc_degraded(conf);
 454         if (degraded > conf->max_degraded)
 455                 return 1;
 456         return 0;
 457 }
 458
 459 static struct stripe_head *
 460 get_active_stripe(struct r5conf *conf, sector_t sector,
 461                   int previous, int noblock, int noquiesce)
 462 {
 463         struct stripe_head *sh;
 464
 465         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 466
 467         spin_lock_irq(&conf->device_lock);
 468
 469         do {
 470                 wait_event_lock_irq(conf->wait_for_stripe,
 471                                     conf->quiesce == 0 || noquiesce,
 472                                     conf->device_lock);
 473                 sh = __find_stripe(conf, sector, conf->generation - previous);
 474                 if (!sh) {
 475                         if (!conf->inactive_blocked)
 476                                 sh = get_free_stripe(conf);
 477                         if (noblock && sh == NULL)
 478                                 break;
 479                         if (!sh) {
 480                                 conf->inactive_blocked = 1;
 481                                 wait_event_lock_irq(conf->wait_for_stripe,
 482                                                     !list_empty(&conf->inactive_list) &&
 483                                                     (atomic_read(&conf->active_stripes)
 484                                                      < (conf->max_nr_stripes *3/4)
 485                                                      || !conf->inactive_blocked),
 486                                                     conf->device_lock);
 487                                 conf->inactive_blocked = 0;
 488                         } else
 489                                 init_stripe(sh, sector, previous);
 490                 } else {
 491                         if (atomic_read(&sh->count)) {
 492                                 BUG_ON(!list_empty(&sh->lru)
 493                                     && !test_bit(STRIPE_EXPANDING, &sh->state)
 494                                     && !test_bit(STRIPE_ON_UNPLUG_LIST, &sh->state));
 495                         } else {
 496                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 497                                         atomic_inc(&conf->active_stripes);
 498                                 if (list_empty(&sh->lru) &&
 499                                     !test_bit(STRIPE_EXPANDING, &sh->state))
 500                                         BUG();
 501                                 list_del_init(&sh->lru);
 502                         }
 503                 }
 504         } while (sh == NULL);
 505
 506         if (sh)
 507                 atomic_inc(&sh->count);
 508
 509         spin_unlock_irq(&conf->device_lock);
 510         return sh;
 511 }
 512
 513 /* Determine if 'data_offset' or 'new_data_offset' should be used
 514  * in this stripe_head.
 515  */
 516 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
 517 {
 518         sector_t progress = conf->reshape_progress;
 519         /* Need a memory barrier to make sure we see the value
 520          * of conf->generation, or ->data_offset that was set before
 521          * reshape_progress was updated.
 522          */
 523         smp_rmb();
 524         if (progress == MaxSector)
 525                 return 0;
 526         if (sh->generation == conf->generation - 1)
 527                 return 0;
 528         /* We are in a reshape, and this is a new-generation stripe,
 529          * so use new_data_offset.
 530          */
 531         return 1;
 532 }
 533
 534 static void
 535 raid5_end_read_request(struct bio *bi, int error);
 536 static void
 537 raid5_end_write_request(struct bio *bi, int error);
 538
 539 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
 540 {
 541         struct r5conf *conf = sh->raid_conf;
 542         int i, disks = sh->disks;
 543
 544         might_sleep();
 545
 546         for (i = disks; i--; ) {
 547                 int rw;
 548                 int replace_only = 0;
 549                 struct bio *bi, *rbi;
 550                 struct md_rdev *rdev, *rrdev = NULL;
 551                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
 552                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
 553                                 rw = WRITE_FUA;
 554                         else
 555                                 rw = WRITE;
 556                         if (test_bit(R5_Discard, &sh->dev[i].flags))
 557                                 rw |= REQ_DISCARD;
 558                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
 559                         rw = READ;
 560                 else if (test_and_clear_bit(R5_WantReplace,
 561                                             &sh->dev[i].flags)) {
 562                         rw = WRITE;
 563                         replace_only = 1;
 564                 } else
 565                         continue;
 566                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
 567                         rw |= REQ_SYNC;
 568
 569                 bi = &sh->dev[i].req;
 570                 rbi = &sh->dev[i].rreq; /* For writing to replacement */
 571
 572                 rcu_read_lock();
 573                 rrdev = rcu_dereference(conf->disks[i].replacement);
 574                 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
 575                 rdev = rcu_dereference(conf->disks[i].rdev);
 576                 if (!rdev) {
 577                         rdev = rrdev;
 578                         rrdev = NULL;
 579                 }
 580                 if (rw & WRITE) {
 581                         if (replace_only)
 582                                 rdev = NULL;
 583                         if (rdev == rrdev)
 584                                 /* We raced and saw duplicates */
 585                                 rrdev = NULL;
 586                 } else {
 587                         if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
 588                                 rdev = rrdev;
 589                         rrdev = NULL;
 590                 }
 591
 592                 if (rdev && test_bit(Faulty, &rdev->flags))
 593                         rdev = NULL;
 594                 if (rdev)
 595                         atomic_inc(&rdev->nr_pending);
 596                 if (rrdev && test_bit(Faulty, &rrdev->flags))
 597                         rrdev = NULL;
 598                 if (rrdev)
 599                         atomic_inc(&rrdev->nr_pending);
 600                 rcu_read_unlock();
 601
 602                 /* We have already checked bad blocks for reads.  Now
 603                  * need to check for writes.  We never accept write errors
 604                  * on the replacement, so we don't to check rrdev.
 605                  */
 606                 while ((rw & WRITE) && rdev &&
 607                        test_bit(WriteErrorSeen, &rdev->flags)) {
 608                         sector_t first_bad;
 609                         int bad_sectors;
 610                         int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
 611                                               &first_bad, &bad_sectors);
 612                         if (!bad)
 613                                 break;
 614
 615                         if (bad < 0) {
 616                                 set_bit(BlockedBadBlocks, &rdev->flags);
 617                                 if (!conf->mddev->external &&
 618                                     conf->mddev->flags) {
 619                                         /* It is very unlikely, but we might
 620                                          * still need to write out the
 621                                          * bad block log - better give it
 622                                          * a chance*/
 623                                         md_check_recovery(conf->mddev);
 624                                 }
 625                                 /*
 626                                  * Because md_wait_for_blocked_rdev
 627                                  * will dec nr_pending, we must
 628                                  * increment it first.
 629                                  */
 630                                 atomic_inc(&rdev->nr_pending);
 631                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
 632                         } else {
 633                                 /* Acknowledged bad block - skip the write */
 634                                 rdev_dec_pending(rdev, conf->mddev);
 635                                 rdev = NULL;
 636                         }
 637                 }
 638
 639                 if (rdev) {
 640                         if (s->syncing || s->expanding || s->expanded
 641                             || s->replacing)
 642                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
 643
 644                         set_bit(STRIPE_IO_STARTED, &sh->state);
 645
 646                         bio_reset(bi);
 647                         bi->bi_bdev = rdev->bdev;
 648                         bi->bi_rw = rw;
 649                         bi->bi_end_io = (rw & WRITE)
 650                                 ? raid5_end_write_request
 651                                 : raid5_end_read_request;
 652                         bi->bi_private = sh;
 653
 654                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
 655                                 __func__, (unsigned long long)sh->sector,
 656                                 bi->bi_rw, i);
 657                         atomic_inc(&sh->count);
 658                         if (use_new_offset(conf, sh))
 659                                 bi->bi_sector = (sh->sector
 660                                                  + rdev->new_data_offset);
 661                         else
 662                                 bi->bi_sector = (sh->sector
 663                                                  + rdev->data_offset);
 664                         if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
 665                                 bi->bi_rw |= REQ_FLUSH;
 666
 667                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
 668                         bi->bi_io_vec[0].bv_offset = 0;
 669                         bi->bi_size = STRIPE_SIZE;
 670                         if (rrdev)
 671                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
 672
 673                         if (conf->mddev->gendisk)
 674                                 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
 675                                                       bi, disk_devt(conf->mddev->gendisk),
 676                                                       sh->dev[i].sector);
 677                         generic_make_request(bi);
 678                 }
 679                 if (rrdev) {
 680                         if (s->syncing || s->expanding || s->expanded
 681                             || s->replacing)
 682                                 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
 683
 684                         set_bit(STRIPE_IO_STARTED, &sh->state);
 685
 686                         bio_reset(rbi);
 687                         rbi->bi_bdev = rrdev->bdev;
 688                         rbi->bi_rw = rw;
 689                         BUG_ON(!(rw & WRITE));
 690                         rbi->bi_end_io = raid5_end_write_request;
 691                         rbi->bi_private = sh;
 692
 693                         pr_debug("%s: for %llu schedule op %ld on "
 694                                  "replacement disc %d\n",
 695                                 __func__, (unsigned long long)sh->sector,
 696                                 rbi->bi_rw, i);
 697                         atomic_inc(&sh->count);
 698                         if (use_new_offset(conf, sh))
 699                                 rbi->bi_sector = (sh->sector
 700                                                   + rrdev->new_data_offset);
 701                         else
 702                                 rbi->bi_sector = (sh->sector
 703                                                   + rrdev->data_offset);
 704                         rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
 705                         rbi->bi_io_vec[0].bv_offset = 0;
 706                         rbi->bi_size = STRIPE_SIZE;
 707                         if (conf->mddev->gendisk)
 708                                 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
 709                                                       rbi, disk_devt(conf->mddev->gendisk),
 710                                                       sh->dev[i].sector);
 711                         generic_make_request(rbi);
 712                 }
 713                 if (!rdev && !rrdev) {
 714                         if (rw & WRITE)
 715                                 set_bit(STRIPE_DEGRADED, &sh->state);
 716                         pr_debug("skip op %ld on disc %d for sector %llu\n",
 717                                 bi->bi_rw, i, (unsigned long long)sh->sector);
 718                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 719                         set_bit(STRIPE_HANDLE, &sh->state);
 720                 }
 721         }
 722 }
 723
 724 static struct dma_async_tx_descriptor *
 725 async_copy_data(int frombio, struct bio *bio, struct page *page,
 726         sector_t sector, struct dma_async_tx_descriptor *tx)
 727 {
 728         struct bio_vec *bvl;
 729         struct page *bio_page;
 730         int i;
 731         int page_offset;
 732         struct async_submit_ctl submit;
 733         enum async_tx_flags flags = 0;
 734
 735         if (bio->bi_sector >= sector)
 736                 page_offset = (signed)(bio->bi_sector - sector) * 512;
 737         else
 738                 page_offset = (signed)(sector - bio->bi_sector) * -512;
 739
 740         if (frombio)
 741                 flags |= ASYNC_TX_FENCE;
 742         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
 743
 744         bio_for_each_segment(bvl, bio, i) {
 745                 int len = bvl->bv_len;
 746                 int clen;
 747                 int b_offset = 0;
 748
 749                 if (page_offset < 0) {
 750                         b_offset = -page_offset;
 751                         page_offset += b_offset;
 752                         len -= b_offset;
 753                 }
 754
 755                 if (len > 0 && page_offset + len > STRIPE_SIZE)
 756                         clen = STRIPE_SIZE - page_offset;
 757                 else
 758                         clen = len;
 759
 760                 if (clen > 0) {
 761                         b_offset += bvl->bv_offset;
 762                         bio_page = bvl->bv_page;
 763                         if (frombio)
 764                                 tx = async_memcpy(page, bio_page, page_offset,
 765                                                   b_offset, clen, &submit);
 766                         else
 767                                 tx = async_memcpy(bio_page, page, b_offset,
 768                                                   page_offset, clen, &submit);
 769                 }
 770                 /* chain the operations */
 771                 submit.depend_tx = tx;
 772
 773                 if (clen < len) /* hit end of page */
 774                         break;
 775                 page_offset +=  len;
 776         }
 777
 778         return tx;
 779 }
 780
 781 static void ops_complete_biofill(void *stripe_head_ref)
 782 {
 783         struct stripe_head *sh = stripe_head_ref;
 784         struct bio *return_bi = NULL;
 785         int i;
 786
 787         pr_debug("%s: stripe %llu\n", __func__,
 788                 (unsigned long long)sh->sector);
 789
 790         /* clear completed biofills */
 791         for (i = sh->disks; i--; ) {
 792                 struct r5dev *dev = &sh->dev[i];
 793
 794                 /* acknowledge completion of a biofill operation */
 795                 /* and check if we need to reply to a read request,
 796                  * new R5_Wantfill requests are held off until
 797                  * !STRIPE_BIOFILL_RUN
 798                  */
 799                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
 800                         struct bio *rbi, *rbi2;
 801
 802                         BUG_ON(!dev->read);
 803                         rbi = dev->read;
 804                         dev->read = NULL;
 805                         while (rbi && rbi->bi_sector <
 806                                 dev->sector + STRIPE_SECTORS) {
 807                                 rbi2 = r5_next_bio(rbi, dev->sector);
 808                                 if (!raid5_dec_bi_active_stripes(rbi)) {
 809                                         rbi->bi_next = return_bi;
 810                                         return_bi = rbi;
 811                                 }
 812                                 rbi = rbi2;
 813                         }
 814                 }
 815         }
 816         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
 817
 818         return_io(return_bi);
 819
 820         set_bit(STRIPE_HANDLE, &sh->state);
 821         release_stripe(sh);
 822 }
 823
 824 static void ops_run_biofill(struct stripe_head *sh)
 825 {
 826         struct dma_async_tx_descriptor *tx = NULL;
 827         struct async_submit_ctl submit;
 828         int i;
 829
 830         pr_debug("%s: stripe %llu\n", __func__,
 831                 (unsigned long long)sh->sector);
 832
 833         for (i = sh->disks; i--; ) {
 834                 struct r5dev *dev = &sh->dev[i];
 835                 if (test_bit(R5_Wantfill, &dev->flags)) {
 836                         struct bio *rbi;
 837                         spin_lock_irq(&sh->stripe_lock);
 838                         dev->read = rbi = dev->toread;
 839                         dev->toread = NULL;
 840                         spin_unlock_irq(&sh->stripe_lock);
 841                         while (rbi && rbi->bi_sector <
 842                                 dev->sector + STRIPE_SECTORS) {
 843                                 tx = async_copy_data(0, rbi, dev->page,
 844                                         dev->sector, tx);
 845                                 rbi = r5_next_bio(rbi, dev->sector);
 846                         }
 847                 }
 848         }
 849
 850         atomic_inc(&sh->count);
 851         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
 852         async_trigger_callback(&submit);
 853 }
 854
 855 static void mark_target_uptodate(struct stripe_head *sh, int target)
 856 {
 857         struct r5dev *tgt;
 858
 859         if (target < 0)
 860                 return;
 861
 862         tgt = &sh->dev[target];
 863         set_bit(R5_UPTODATE, &tgt->flags);
 864         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
 865         clear_bit(R5_Wantcompute, &tgt->flags);
 866 }
 867
 868 static void ops_complete_compute(void *stripe_head_ref)
 869 {
 870         struct stripe_head *sh = stripe_head_ref;
 871
 872         pr_debug("%s: stripe %llu\n", __func__,
 873                 (unsigned long long)sh->sector);
 874
 875         /* mark the computed target(s) as uptodate */
 876         mark_target_uptodate(sh, sh->ops.target);
 877         mark_target_uptodate(sh, sh->ops.target2);
 878
 879         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
 880         if (sh->check_state == check_state_compute_run)
 881                 sh->check_state = check_state_compute_result;
 882         set_bit(STRIPE_HANDLE, &sh->state);
 883         release_stripe(sh);
 884 }
 885
 886 /* return a pointer to the address conversion region of the scribble buffer */
 887 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
 888                                  struct raid5_percpu *percpu)
 889 {
 890         return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
 891 }
 892
 893 static struct dma_async_tx_descriptor *
 894 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
 895 {
 896         int disks = sh->disks;
 897         struct page **xor_srcs = percpu->scribble;
 898         int target = sh->ops.target;
 899         struct r5dev *tgt = &sh->dev[target];
 900         struct page *xor_dest = tgt->page;
 901         int count = 0;
 902         struct dma_async_tx_descriptor *tx;
 903         struct async_submit_ctl submit;
 904         int i;
 905
 906         pr_debug("%s: stripe %llu block: %d\n",
 907                 __func__, (unsigned long long)sh->sector, target);
 908         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
 909
 910         for (i = disks; i--; )
 911                 if (i != target)
 912                         xor_srcs[count++] = sh->dev[i].page;
 913
 914         atomic_inc(&sh->count);
 915
 916         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
 917                           ops_complete_compute, sh, to_addr_conv(sh, percpu));
 918         if (unlikely(count == 1))
 919                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
 920         else
 921                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
 922
 923         return tx;
 924 }
 925
 926 /* set_syndrome_sources - populate source buffers for gen_syndrome
 927  * @srcs - (struct page *) array of size sh->disks
 928  * @sh - stripe_head to parse
 929  *
 930  * Populates srcs in proper layout order for the stripe and returns the
 931  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
 932  * destination buffer is recorded in srcs[count] and the Q destination
 933  * is recorded in srcs[count+1]].
 934  */
 935 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
 936 {
 937         int disks = sh->disks;
 938         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
 939         int d0_idx = raid6_d0(sh);
 940         int count;
 941         int i;
 942
 943         for (i = 0; i < disks; i++)
 944                 srcs[i] = NULL;
 945
 946         count = 0;
 947         i = d0_idx;
 948         do {
 949                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
 950
 951                 srcs[slot] = sh->dev[i].page;
 952                 i = raid6_next_disk(i, disks);
 953         } while (i != d0_idx);
 954
 955         return syndrome_disks;
 956 }
 957
 958 static struct dma_async_tx_descriptor *
 959 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
 960 {
 961         int disks = sh->disks;
 962         struct page **blocks = percpu->scribble;
 963         int target;
 964         int qd_idx = sh->qd_idx;
 965         struct dma_async_tx_descriptor *tx;
 966         struct async_submit_ctl submit;
 967         struct r5dev *tgt;
 968         struct page *dest;
 969         int i;
 970         int count;
 971
 972         if (sh->ops.target < 0)
 973                 target = sh->ops.target2;
 974         else if (sh->ops.target2 < 0)
 975                 target = sh->ops.target;
 976         else
 977                 /* we should only have one valid target */
 978                 BUG();
 979         BUG_ON(target < 0);
 980         pr_debug("%s: stripe %llu block: %d\n",
 981                 __func__, (unsigned long long)sh->sector, target);
 982
 983         tgt = &sh->dev[target];
 984         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
 985         dest = tgt->page;
 986
 987         atomic_inc(&sh->count);
 988
 989         if (target == qd_idx) {
 990                 count = set_syndrome_sources(blocks, sh);
 991                 blocks[count] = NULL; /* regenerating p is not necessary */
 992                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
 993                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
 994                                   ops_complete_compute, sh,
 995                                   to_addr_conv(sh, percpu));
 996                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
 997         } else {
 998                 /* Compute any data- or p-drive using XOR */
 999                 count = 0;
1000                 for (i = disks; i-- ; ) {
1001                         if (i == target || i == qd_idx)
1002                                 continue;
1003                         blocks[count++] = sh->dev[i].page;
1004                 }
1005
1006                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1007                                   NULL, ops_complete_compute, sh,
1008                                   to_addr_conv(sh, percpu));
1009                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1010         }
1011
1012         return tx;
1013 }
1014
1015 static struct dma_async_tx_descriptor *
1016 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1017 {
1018         int i, count, disks = sh->disks;
1019         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1020         int d0_idx = raid6_d0(sh);
1021         int faila = -1, failb = -1;
1022         int target = sh->ops.target;
1023         int target2 = sh->ops.target2;
1024         struct r5dev *tgt = &sh->dev[target];
1025         struct r5dev *tgt2 = &sh->dev[target2];
1026         struct dma_async_tx_descriptor *tx;
1027         struct page **blocks = percpu->scribble;
1028         struct async_submit_ctl submit;
1029
1030         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1031                  __func__, (unsigned long long)sh->sector, target, target2);
1032         BUG_ON(target < 0 || target2 < 0);
1033         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1034         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1035
1036         /* we need to open-code set_syndrome_sources to handle the
1037          * slot number conversion for 'faila' and 'failb'
1038          */
1039         for (i = 0; i < disks ; i++)
1040                 blocks[i] = NULL;
1041         count = 0;
1042         i = d0_idx;
1043         do {
1044                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1045
1046                 blocks[slot] = sh->dev[i].page;
1047
1048                 if (i == target)
1049                         faila = slot;
1050                 if (i == target2)
1051                         failb = slot;
1052                 i = raid6_next_disk(i, disks);
1053         } while (i != d0_idx);
1054
1055         BUG_ON(faila == failb);
1056         if (failb < faila)
1057                 swap(faila, failb);
1058         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1059                  __func__, (unsigned long long)sh->sector, faila, failb);
1060
1061         atomic_inc(&sh->count);
1062
1063         if (failb == syndrome_disks+1) {
1064                 /* Q disk is one of the missing disks */
1065                 if (faila == syndrome_disks) {
1066                         /* Missing P+Q, just recompute */
1067                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1068                                           ops_complete_compute, sh,
1069                                           to_addr_conv(sh, percpu));
1070                         return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1071                                                   STRIPE_SIZE, &submit);
1072                 } else {
1073                         struct page *dest;
1074                         int data_target;
1075                         int qd_idx = sh->qd_idx;
1076
1077                         /* Missing D+Q: recompute D from P, then recompute Q */
1078                         if (target == qd_idx)
1079                                 data_target = target2;
1080                         else
1081                                 data_target = target;
1082
1083                         count = 0;
1084                         for (i = disks; i-- ; ) {
1085                                 if (i == data_target || i == qd_idx)
1086                                         continue;
1087                                 blocks[count++] = sh->dev[i].page;
1088                         }
1089                         dest = sh->dev[data_target].page;
1090                         init_async_submit(&submit,
1091                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1092                                           NULL, NULL, NULL,
1093                                           to_addr_conv(sh, percpu));
1094                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1095                                        &submit);
1096
1097                         count = set_syndrome_sources(blocks, sh);
1098                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1099                                           ops_complete_compute, sh,
1100                                           to_addr_conv(sh, percpu));
1101                         return async_gen_syndrome(blocks, 0, count+2,
1102                                                   STRIPE_SIZE, &submit);
1103                 }
1104         } else {
1105                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1106                                   ops_complete_compute, sh,
1107                                   to_addr_conv(sh, percpu));
1108                 if (failb == syndrome_disks) {
1109                         /* We're missing D+P. */
1110                         return async_raid6_datap_recov(syndrome_disks+2,
1111                                                        STRIPE_SIZE, faila,
1112                                                        blocks, &submit);
1113                 } else {
1114                         /* We're missing D+D. */
1115                         return async_raid6_2data_recov(syndrome_disks+2,
1116                                                        STRIPE_SIZE, faila, failb,
1117                                                        blocks, &submit);
1118                 }
1119         }
1120 }
1121
1122
1123 static void ops_complete_prexor(void *stripe_head_ref)
1124 {
1125         struct stripe_head *sh = stripe_head_ref;
1126
1127         pr_debug("%s: stripe %llu\n", __func__,
1128                 (unsigned long long)sh->sector);
1129 }
1130
1131 static struct dma_async_tx_descriptor *
1132 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1133                struct dma_async_tx_descriptor *tx)
1134 {
1135         int disks = sh->disks;
1136         struct page **xor_srcs = percpu->scribble;
1137         int count = 0, pd_idx = sh->pd_idx, i;
1138         struct async_submit_ctl submit;
1139
1140         /* existing parity data subtracted */
1141         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1142
1143         pr_debug("%s: stripe %llu\n", __func__,
1144                 (unsigned long long)sh->sector);
1145
1146         for (i = disks; i--; ) {
1147                 struct r5dev *dev = &sh->dev[i];
1148                 /* Only process blocks that are known to be uptodate */
1149                 if (test_bit(R5_Wantdrain, &dev->flags))
1150                         xor_srcs[count++] = dev->page;
1151         }
1152
1153         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1154                           ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1155         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1156
1157         return tx;
1158 }
1159
1160 static struct dma_async_tx_descriptor *
1161 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1162 {
1163         int disks = sh->disks;
1164         int i;
1165
1166         pr_debug("%s: stripe %llu\n", __func__,
1167                 (unsigned long long)sh->sector);
1168
1169         for (i = disks; i--; ) {
1170                 struct r5dev *dev = &sh->dev[i];
1171                 struct bio *chosen;
1172
1173                 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1174                         struct bio *wbi;
1175
1176                         spin_lock_irq(&sh->stripe_lock);
1177                         chosen = dev->towrite;
1178                         dev->towrite = NULL;
1179                         BUG_ON(dev->written);
1180                         wbi = dev->written = chosen;
1181                         spin_unlock_irq(&sh->stripe_lock);
1182
1183                         while (wbi && wbi->bi_sector <
1184                                 dev->sector + STRIPE_SECTORS) {
1185                                 if (wbi->bi_rw & REQ_FUA)
1186                                         set_bit(R5_WantFUA, &dev->flags);
1187                                 if (wbi->bi_rw & REQ_SYNC)
1188                                         set_bit(R5_SyncIO, &dev->flags);
1189                                 if (wbi->bi_rw & REQ_DISCARD)
1190                                         set_bit(R5_Discard, &dev->flags);
1191                                 else
1192                                         tx = async_copy_data(1, wbi, dev->page,
1193                                                 dev->sector, tx);
1194                                 wbi = r5_next_bio(wbi, dev->sector);
1195                         }
1196                 }
1197         }
1198
1199         return tx;
1200 }
1201
1202 static void ops_complete_reconstruct(void *stripe_head_ref)
1203 {
1204         struct stripe_head *sh = stripe_head_ref;
1205         int disks = sh->disks;
1206         int pd_idx = sh->pd_idx;
1207         int qd_idx = sh->qd_idx;
1208         int i;
1209         bool fua = false, sync = false, discard = false;
1210
1211         pr_debug("%s: stripe %llu\n", __func__,
1212                 (unsigned long long)sh->sector);
1213
1214         for (i = disks; i--; ) {
1215                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1216                 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1217                 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1218         }
1219
1220         for (i = disks; i--; ) {
1221                 struct r5dev *dev = &sh->dev[i];
1222
1223                 if (dev->written || i == pd_idx || i == qd_idx) {
1224                         if (!discard)
1225                                 set_bit(R5_UPTODATE, &dev->flags);
1226                         if (fua)
1227                                 set_bit(R5_WantFUA, &dev->flags);
1228                         if (sync)
1229                                 set_bit(R5_SyncIO, &dev->flags);
1230                 }
1231         }
1232
1233         if (sh->reconstruct_state == reconstruct_state_drain_run)
1234                 sh->reconstruct_state = reconstruct_state_drain_result;
1235         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1236                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1237         else {
1238                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1239                 sh->reconstruct_state = reconstruct_state_result;
1240         }
1241
1242         set_bit(STRIPE_HANDLE, &sh->state);
1243         release_stripe(sh);
1244 }
1245
1246 static void
1247 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1248                      struct dma_async_tx_descriptor *tx)
1249 {
1250         int disks = sh->disks;
1251         struct page **xor_srcs = percpu->scribble;
1252         struct async_submit_ctl submit;
1253         int count = 0, pd_idx = sh->pd_idx, i;
1254         struct page *xor_dest;
1255         int prexor = 0;
1256         unsigned long flags;
1257
1258         pr_debug("%s: stripe %llu\n", __func__,
1259                 (unsigned long long)sh->sector);
1260
1261         for (i = 0; i < sh->disks; i++) {
1262                 if (pd_idx == i)
1263                         continue;
1264                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1265                         break;
1266         }
1267         if (i >= sh->disks) {
1268                 atomic_inc(&sh->count);
1269                 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1270                 ops_complete_reconstruct(sh);
1271                 return;
1272         }
1273         /* check if prexor is active which means only process blocks
1274          * that are part of a read-modify-write (written)
1275          */
1276         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1277                 prexor = 1;
1278                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1279                 for (i = disks; i--; ) {
1280                         struct r5dev *dev = &sh->dev[i];
1281                         if (dev->written)
1282                                 xor_srcs[count++] = dev->page;
1283                 }
1284         } else {
1285                 xor_dest = sh->dev[pd_idx].page;
1286                 for (i = disks; i--; ) {
1287                         struct r5dev *dev = &sh->dev[i];
1288                         if (i != pd_idx)
1289                                 xor_srcs[count++] = dev->page;
1290                 }
1291         }
1292
1293         /* 1/ if we prexor'd then the dest is reused as a source
1294          * 2/ if we did not prexor then we are redoing the parity
1295          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1296          * for the synchronous xor case
1297          */
1298         flags = ASYNC_TX_ACK |
1299                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1300
1301         atomic_inc(&sh->count);
1302
1303         init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1304                           to_addr_conv(sh, percpu));
1305         if (unlikely(count == 1))
1306                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1307         else
1308                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1309 }
1310
1311 static void
1312 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1313                      struct dma_async_tx_descriptor *tx)
1314 {
1315         struct async_submit_ctl submit;
1316         struct page **blocks = percpu->scribble;
1317         int count, i;
1318
1319         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1320
1321         for (i = 0; i < sh->disks; i++) {
1322                 if (sh->pd_idx == i || sh->qd_idx == i)
1323                         continue;
1324                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1325                         break;
1326         }
1327         if (i >= sh->disks) {
1328                 atomic_inc(&sh->count);
1329                 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1330                 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1331                 ops_complete_reconstruct(sh);
1332                 return;
1333         }
1334
1335         count = set_syndrome_sources(blocks, sh);
1336
1337         atomic_inc(&sh->count);
1338
1339         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1340                           sh, to_addr_conv(sh, percpu));
1341         async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1342 }
1343
1344 static void ops_complete_check(void *stripe_head_ref)
1345 {
1346         struct stripe_head *sh = stripe_head_ref;
1347
1348         pr_debug("%s: stripe %llu\n", __func__,
1349                 (unsigned long long)sh->sector);
1350
1351         sh->check_state = check_state_check_result;
1352         set_bit(STRIPE_HANDLE, &sh->state);
1353         release_stripe(sh);
1354 }
1355
1356 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1357 {
1358         int disks = sh->disks;
1359         int pd_idx = sh->pd_idx;
1360         int qd_idx = sh->qd_idx;
1361         struct page *xor_dest;
1362         struct page **xor_srcs = percpu->scribble;
1363         struct dma_async_tx_descriptor *tx;
1364         struct async_submit_ctl submit;
1365         int count;
1366         int i;
1367
1368         pr_debug("%s: stripe %llu\n", __func__,
1369                 (unsigned long long)sh->sector);
1370
1371         count = 0;
1372         xor_dest = sh->dev[pd_idx].page;
1373         xor_srcs[count++] = xor_dest;
1374         for (i = disks; i--; ) {
1375                 if (i == pd_idx || i == qd_idx)
1376                         continue;
1377                 xor_srcs[count++] = sh->dev[i].page;
1378         }
1379
1380         init_async_submit(&submit, 0, NULL, NULL, NULL,
1381                           to_addr_conv(sh, percpu));
1382         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1383                            &sh->ops.zero_sum_result, &submit);
1384
1385         atomic_inc(&sh->count);
1386         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1387         tx = async_trigger_callback(&submit);
1388 }
1389
1390 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1391 {
1392         struct page **srcs = percpu->scribble;
1393         struct async_submit_ctl submit;
1394         int count;
1395
1396         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1397                 (unsigned long long)sh->sector, checkp);
1398
1399         count = set_syndrome_sources(srcs, sh);
1400         if (!checkp)
1401                 srcs[count] = NULL;
1402
1403         atomic_inc(&sh->count);
1404         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1405                           sh, to_addr_conv(sh, percpu));
1406         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1407                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1408 }
1409
1410 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1411 {
1412         int overlap_clear = 0, i, disks = sh->disks;
1413         struct dma_async_tx_descriptor *tx = NULL;
1414         struct r5conf *conf = sh->raid_conf;
1415         int level = conf->level;
1416         struct raid5_percpu *percpu;
1417         unsigned long cpu;
1418
1419         cpu = get_cpu();
1420         percpu = per_cpu_ptr(conf->percpu, cpu);
1421         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1422                 ops_run_biofill(sh);
1423                 overlap_clear++;
1424         }
1425
1426         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1427                 if (level < 6)
1428                         tx = ops_run_compute5(sh, percpu);
1429                 else {
1430                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
1431                                 tx = ops_run_compute6_1(sh, percpu);
1432                         else
1433                                 tx = ops_run_compute6_2(sh, percpu);
1434                 }
1435                 /* terminate the chain if reconstruct is not set to be run */
1436                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1437                         async_tx_ack(tx);
1438         }
1439
1440         if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1441                 tx = ops_run_prexor(sh, percpu, tx);
1442
1443         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1444                 tx = ops_run_biodrain(sh, tx);
1445                 overlap_clear++;
1446         }
1447
1448         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1449                 if (level < 6)
1450                         ops_run_reconstruct5(sh, percpu, tx);
1451                 else
1452                         ops_run_reconstruct6(sh, percpu, tx);
1453         }
1454
1455         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1456                 if (sh->check_state == check_state_run)
1457                         ops_run_check_p(sh, percpu);
1458                 else if (sh->check_state == check_state_run_q)
1459                         ops_run_check_pq(sh, percpu, 0);
1460                 else if (sh->check_state == check_state_run_pq)
1461                         ops_run_check_pq(sh, percpu, 1);
1462                 else
1463                         BUG();
1464         }
1465
1466         if (overlap_clear)
1467                 for (i = disks; i--; ) {
1468                         struct r5dev *dev = &sh->dev[i];
1469                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1470                                 wake_up(&sh->raid_conf->wait_for_overlap);
1471                 }
1472         put_cpu();
1473 }
1474
1475 static int grow_one_stripe(struct r5conf *conf)
1476 {
1477         struct stripe_head *sh;
1478         sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1479         if (!sh)
1480                 return 0;
1481
1482         sh->raid_conf = conf;
1483
1484         spin_lock_init(&sh->stripe_lock);
1485
1486         if (grow_buffers(sh)) {
1487                 shrink_buffers(sh);
1488                 kmem_cache_free(conf->slab_cache, sh);
1489                 return 0;
1490         }
1491         /* we just created an active stripe so... */
1492         atomic_set(&sh->count, 1);
1493         atomic_inc(&conf->active_stripes);
1494         INIT_LIST_HEAD(&sh->lru);
1495         release_stripe(sh);
1496         return 1;
1497 }
1498
1499 static int grow_stripes(struct r5conf *conf, int num)
1500 {
1501         struct kmem_cache *sc;
1502         int devs = max(conf->raid_disks, conf->previous_raid_disks);
1503
1504         if (conf->mddev->gendisk)
1505                 sprintf(conf->cache_name[0],
1506                         "raid%d-%s", conf->level, mdname(conf->mddev));
1507         else
1508                 sprintf(conf->cache_name[0],
1509                         "raid%d-%p", conf->level, conf->mddev);
1510         sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1511
1512         conf->active_name = 0;
1513         sc = kmem_cache_create(conf->cache_name[conf->active_name],
1514                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1515                                0, 0, NULL);
1516         if (!sc)
1517                 return 1;
1518         conf->slab_cache = sc;
1519         conf->pool_size = devs;
1520         while (num--)
1521                 if (!grow_one_stripe(conf))
1522                         return 1;
1523         return 0;
1524 }
1525
1526 /**
1527  * scribble_len - return the required size of the scribble region
1528  * @num - total number of disks in the array
1529  *
1530  * The size must be enough to contain:
1531  * 1/ a struct page pointer for each device in the array +2
1532  * 2/ room to convert each entry in (1) to its corresponding dma
1533  *    (dma_map_page()) or page (page_address()) address.
1534  *
1535  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1536  * calculate over all devices (not just the data blocks), using zeros in place
1537  * of the P and Q blocks.
1538  */
1539 static size_t scribble_len(int num)
1540 {
1541         size_t len;
1542
1543         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1544
1545         return len;
1546 }
1547
1548 static int resize_stripes(struct r5conf *conf, int newsize)
1549 {
1550         /* Make all the stripes able to hold 'newsize' devices.
1551          * New slots in each stripe get 'page' set to a new page.
1552          *
1553          * This happens in stages:
1554          * 1/ create a new kmem_cache and allocate the required number of
1555          *    stripe_heads.
1556          * 2/ gather all the old stripe_heads and transfer the pages across
1557          *    to the new stripe_heads.  This will have the side effect of
1558          *    freezing the array as once all stripe_heads have been collected,
1559          *    no IO will be possible.  Old stripe heads are freed once their
1560          *    pages have been transferred over, and the old kmem_cache is
1561          *    freed when all stripes are done.
1562          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
1563          *    we simple return a failre status - no need to clean anything up.
1564          * 4/ allocate new pages for the new slots in the new stripe_heads.
1565          *    If this fails, we don't bother trying the shrink the
1566          *    stripe_heads down again, we just leave them as they are.
1567          *    As each stripe_head is processed the new one is released into
1568          *    active service.
1569          *
1570          * Once step2 is started, we cannot afford to wait for a write,
1571          * so we use GFP_NOIO allocations.
1572          */
1573         struct stripe_head *osh, *nsh;
1574         LIST_HEAD(newstripes);
1575         struct disk_info *ndisks;
1576         unsigned long cpu;
1577         int err;
1578         struct kmem_cache *sc;
1579         int i;
1580
1581         if (newsize <= conf->pool_size)
1582                 return 0; /* never bother to shrink */
1583
1584         err = md_allow_write(conf->mddev);
1585         if (err)
1586                 return err;
1587
1588         /* Step 1 */
1589         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1590                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1591                                0, 0, NULL);
1592         if (!sc)
1593                 return -ENOMEM;
1594
1595         for (i = conf->max_nr_stripes; i; i--) {
1596                 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1597                 if (!nsh)
1598                         break;
1599
1600                 nsh->raid_conf = conf;
1601                 spin_lock_init(&nsh->stripe_lock);
1602
1603                 list_add(&nsh->lru, &newstripes);
1604         }
1605         if (i) {
1606                 /* didn't get enough, give up */
1607                 while (!list_empty(&newstripes)) {
1608                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1609                         list_del(&nsh->lru);
1610                         kmem_cache_free(sc, nsh);
1611                 }
1612                 kmem_cache_destroy(sc);
1613                 return -ENOMEM;
1614         }
1615         /* Step 2 - Must use GFP_NOIO now.
1616          * OK, we have enough stripes, start collecting inactive
1617          * stripes and copying them over
1618          */
1619         list_for_each_entry(nsh, &newstripes, lru) {
1620                 spin_lock_irq(&conf->device_lock);
1621                 wait_event_lock_irq(conf->wait_for_stripe,
1622                                     !list_empty(&conf->inactive_list),
1623                                     conf->device_lock);
1624                 osh = get_free_stripe(conf);
1625                 spin_unlock_irq(&conf->device_lock);
1626                 atomic_set(&nsh->count, 1);
1627                 for(i=0; i<conf->pool_size; i++)
1628                         nsh->dev[i].page = osh->dev[i].page;
1629                 for( ; i<newsize; i++)
1630                         nsh->dev[i].page = NULL;
1631                 kmem_cache_free(conf->slab_cache, osh);
1632         }
1633         kmem_cache_destroy(conf->slab_cache);
1634
1635         /* Step 3.
1636          * At this point, we are holding all the stripes so the array
1637          * is completely stalled, so now is a good time to resize
1638          * conf->disks and the scribble region
1639          */
1640         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1641         if (ndisks) {
1642                 for (i=0; i<conf->raid_disks; i++)
1643                         ndisks[i] = conf->disks[i];
1644                 kfree(conf->disks);
1645                 conf->disks = ndisks;
1646         } else
1647                 err = -ENOMEM;
1648
1649         get_online_cpus();
1650         conf->scribble_len = scribble_len(newsize);
1651         for_each_present_cpu(cpu) {
1652                 struct raid5_percpu *percpu;
1653                 void *scribble;
1654
1655                 percpu = per_cpu_ptr(conf->percpu, cpu);
1656                 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1657
1658                 if (scribble) {
1659                         kfree(percpu->scribble);
1660                         percpu->scribble = scribble;
1661                 } else {
1662                         err = -ENOMEM;
1663                         break;
1664                 }
1665         }
1666         put_online_cpus();
1667
1668         /* Step 4, return new stripes to service */
1669         while(!list_empty(&newstripes)) {
1670                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1671                 list_del_init(&nsh->lru);
1672
1673                 for (i=conf->raid_disks; i < newsize; i++)
1674                         if (nsh->dev[i].page == NULL) {
1675                                 struct page *p = alloc_page(GFP_NOIO);
1676                                 nsh->dev[i].page = p;
1677                                 if (!p)
1678                                         err = -ENOMEM;
1679                         }
1680                 release_stripe(nsh);
1681         }
1682         /* critical section pass, GFP_NOIO no longer needed */
1683
1684         conf->slab_cache = sc;
1685         conf->active_name = 1-conf->active_name;
1686         conf->pool_size = newsize;
1687         return err;
1688 }
1689
1690 static int drop_one_stripe(struct r5conf *conf)
1691 {
1692         struct stripe_head *sh;
1693
1694         spin_lock_irq(&conf->device_lock);
1695         sh = get_free_stripe(conf);
1696         spin_unlock_irq(&conf->device_lock);
1697         if (!sh)
1698                 return 0;
1699         BUG_ON(atomic_read(&sh->count));
1700         shrink_buffers(sh);
1701         kmem_cache_free(conf->slab_cache, sh);
1702         atomic_dec(&conf->active_stripes);
1703         return 1;
1704 }
1705
1706 static void shrink_stripes(struct r5conf *conf)
1707 {
1708         while (drop_one_stripe(conf))
1709                 ;
1710
1711         if (conf->slab_cache)
1712                 kmem_cache_destroy(conf->slab_cache);
1713         conf->slab_cache = NULL;
1714 }
1715
1716 static void raid5_end_read_request(struct bio * bi, int error)
1717 {
1718         struct stripe_head *sh = bi->bi_private;
1719         struct r5conf *conf = sh->raid_conf;
1720         int disks = sh->disks, i;
1721         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1722         char b[BDEVNAME_SIZE];
1723         struct md_rdev *rdev = NULL;
1724         sector_t s;
1725
1726         for (i=0 ; i<disks; i++)
1727                 if (bi == &sh->dev[i].req)
1728                         break;
1729
1730         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1731                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1732                 uptodate);
1733         if (i == disks) {
1734                 BUG();
1735                 return;
1736         }
1737         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1738                 /* If replacement finished while this request was outstanding,
1739                  * 'replacement' might be NULL already.
1740                  * In that case it moved down to 'rdev'.
1741                  * rdev is not removed until all requests are finished.
1742                  */
1743                 rdev = conf->disks[i].replacement;
1744         if (!rdev)
1745                 rdev = conf->disks[i].rdev;
1746
1747         if (use_new_offset(conf, sh))
1748                 s = sh->sector + rdev->new_data_offset;
1749         else
1750                 s = sh->sector + rdev->data_offset;
1751         if (uptodate) {
1752                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1753                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1754                         /* Note that this cannot happen on a
1755                          * replacement device.  We just fail those on
1756                          * any error
1757                          */
1758                         printk_ratelimited(
1759                                 KERN_INFO
1760                                 "md/raid:%s: read error corrected"
1761                                 " (%lu sectors at %llu on %s)\n",
1762                                 mdname(conf->mddev), STRIPE_SECTORS,
1763                                 (unsigned long long)s,
1764                                 bdevname(rdev->bdev, b));
1765                         atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1766                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1767                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1768                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
1769                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1770
1771                 if (atomic_read(&rdev->read_errors))
1772                         atomic_set(&rdev->read_errors, 0);
1773         } else {
1774                 const char *bdn = bdevname(rdev->bdev, b);
1775                 int retry = 0;
1776                 int set_bad = 0;
1777
1778                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1779                 atomic_inc(&rdev->read_errors);
1780                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1781                         printk_ratelimited(
1782                                 KERN_WARNING
1783                                 "md/raid:%s: read error on replacement device "
1784                                 "(sector %llu on %s).\n",
1785                                 mdname(conf->mddev),
1786                                 (unsigned long long)s,
1787                                 bdn);
1788                 else if (conf->mddev->degraded >= conf->max_degraded) {
1789                         set_bad = 1;
1790                         printk_ratelimited(
1791                                 KERN_WARNING
1792                                 "md/raid:%s: read error not correctable "
1793                                 "(sector %llu on %s).\n",
1794                                 mdname(conf->mddev),
1795                                 (unsigned long long)s,
1796                                 bdn);
1797                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
1798                         /* Oh, no!!! */
1799                         set_bad = 1;
1800                         printk_ratelimited(
1801                                 KERN_WARNING
1802                                 "md/raid:%s: read error NOT corrected!! "
1803                                 "(sector %llu on %s).\n",
1804                                 mdname(conf->mddev),
1805                                 (unsigned long long)s,
1806                                 bdn);
1807                 } else if (atomic_read(&rdev->read_errors)
1808                          > conf->max_nr_stripes)
1809                         printk(KERN_WARNING
1810                                "md/raid:%s: Too many read errors, failing device %s.\n",
1811                                mdname(conf->mddev), bdn);
1812                 else
1813                         retry = 1;
1814                 if (retry)
1815                         if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
1816                                 set_bit(R5_ReadError, &sh->dev[i].flags);
1817                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1818                         } else
1819                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1820                 else {
1821                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1822                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1823                         if (!(set_bad
1824                               && test_bit(In_sync, &rdev->flags)
1825                               && rdev_set_badblocks(
1826                                       rdev, sh->sector, STRIPE_SECTORS, 0)))
1827                                 md_error(conf->mddev, rdev);
1828                 }
1829         }
1830         rdev_dec_pending(rdev, conf->mddev);
1831         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1832         set_bit(STRIPE_HANDLE, &sh->state);
1833         release_stripe(sh);
1834 }
1835
1836 static void raid5_end_write_request(struct bio *bi, int error)
1837 {
1838         struct stripe_head *sh = bi->bi_private;
1839         struct r5conf *conf = sh->raid_conf;
1840         int disks = sh->disks, i;
1841         struct md_rdev *uninitialized_var(rdev);
1842         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1843         sector_t first_bad;
1844         int bad_sectors;
1845         int replacement = 0;
1846
1847         for (i = 0 ; i < disks; i++) {
1848                 if (bi == &sh->dev[i].req) {
1849                         rdev = conf->disks[i].rdev;
1850                         break;
1851                 }
1852                 if (bi == &sh->dev[i].rreq) {
1853                         rdev = conf->disks[i].replacement;
1854                         if (rdev)
1855                                 replacement = 1;
1856                         else
1857                                 /* rdev was removed and 'replacement'
1858                                  * replaced it.  rdev is not removed
1859                                  * until all requests are finished.
1860                                  */
1861                                 rdev = conf->disks[i].rdev;
1862                         break;
1863                 }
1864         }
1865         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1866                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1867                 uptodate);
1868         if (i == disks) {
1869                 BUG();
1870                 return;
1871         }
1872
1873         if (replacement) {
1874                 if (!uptodate)
1875                         md_error(conf->mddev, rdev);
1876                 else if (is_badblock(rdev, sh->sector,
1877                                      STRIPE_SECTORS,
1878                                      &first_bad, &bad_sectors))
1879                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
1880         } else {
1881                 if (!uptodate) {
1882                         set_bit(WriteErrorSeen, &rdev->flags);
1883                         set_bit(R5_WriteError, &sh->dev[i].flags);
1884                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
1885                                 set_bit(MD_RECOVERY_NEEDED,
1886                                         &rdev->mddev->recovery);
1887                 } else if (is_badblock(rdev, sh->sector,
1888                                        STRIPE_SECTORS,
1889                                        &first_bad, &bad_sectors)) {
1890                         set_bit(R5_MadeGood, &sh->dev[i].flags);
1891                         if (test_bit(R5_ReadError, &sh->dev[i].flags))
1892                                 /* That was a successful write so make
1893                                  * sure it looks like we already did
1894                                  * a re-write.
1895                                  */
1896                                 set_bit(R5_ReWrite, &sh->dev[i].flags);
1897                 }
1898         }
1899         rdev_dec_pending(rdev, conf->mddev);
1900
1901         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
1902                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1903         set_bit(STRIPE_HANDLE, &sh->state);
1904         release_stripe(sh);
1905 }
1906
1907 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1908
1909 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1910 {
1911         struct r5dev *dev = &sh->dev[i];
1912
1913         bio_init(&dev->req);
1914         dev->req.bi_io_vec = &dev->vec;
1915         dev->req.bi_vcnt++;
1916         dev->req.bi_max_vecs++;
1917         dev->req.bi_private = sh;
1918         dev->vec.bv_page = dev->page;
1919
1920         bio_init(&dev->rreq);
1921         dev->rreq.bi_io_vec = &dev->rvec;
1922         dev->rreq.bi_vcnt++;
1923         dev->rreq.bi_max_vecs++;
1924         dev->rreq.bi_private = sh;
1925         dev->rvec.bv_page = dev->page;
1926
1927         dev->flags = 0;
1928         dev->sector = compute_blocknr(sh, i, previous);
1929 }
1930
1931 static void error(struct mddev *mddev, struct md_rdev *rdev)
1932 {
1933         char b[BDEVNAME_SIZE];
1934         struct r5conf *conf = mddev->private;
1935         unsigned long flags;
1936         pr_debug("raid456: error called\n");
1937
1938         spin_lock_irqsave(&conf->device_lock, flags);
1939         clear_bit(In_sync, &rdev->flags);
1940         mddev->degraded = calc_degraded(conf);
1941         spin_unlock_irqrestore(&conf->device_lock, flags);
1942         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1943
1944         set_bit(Blocked, &rdev->flags);
1945         set_bit(Faulty, &rdev->flags);
1946         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1947         printk(KERN_ALERT
1948                "md/raid:%s: Disk failure on %s, disabling device.\n"
1949                "md/raid:%s: Operation continuing on %d devices.\n",
1950                mdname(mddev),
1951                bdevname(rdev->bdev, b),
1952                mdname(mddev),
1953                conf->raid_disks - mddev->degraded);
1954 }
1955
1956 /*
1957  * Input: a 'big' sector number,
1958  * Output: index of the data and parity disk, and the sector # in them.
1959  */
1960 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1961                                      int previous, int *dd_idx,
1962                                      struct stripe_head *sh)
1963 {
1964         sector_t stripe, stripe2;
1965         sector_t chunk_number;
1966         unsigned int chunk_offset;
1967         int pd_idx, qd_idx;
1968         int ddf_layout = 0;
1969         sector_t new_sector;
1970         int algorithm = previous ? conf->prev_algo
1971                                  : conf->algorithm;
1972         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1973                                          : conf->chunk_sectors;
1974         int raid_disks = previous ? conf->previous_raid_disks
1975                                   : conf->raid_disks;
1976         int data_disks = raid_disks - conf->max_degraded;
1977
1978         /* First compute the information on this sector */
1979
1980         /*
1981          * Compute the chunk number and the sector offset inside the chunk
1982          */
1983         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1984         chunk_number = r_sector;
1985
1986         /*
1987          * Compute the stripe number
1988          */
1989         stripe = chunk_number;
1990         *dd_idx = sector_div(stripe, data_disks);
1991         stripe2 = stripe;
1992         /*
1993          * Select the parity disk based on the user selected algorithm.
1994          */
1995         pd_idx = qd_idx = -1;
1996         switch(conf->level) {
1997         case 4:
1998                 pd_idx = data_disks;
1999                 break;
2000         case 5:
2001                 switch (algorithm) {
2002                 case ALGORITHM_LEFT_ASYMMETRIC:
2003                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2004                         if (*dd_idx >= pd_idx)
2005                                 (*dd_idx)++;
2006                         break;
2007                 case ALGORITHM_RIGHT_ASYMMETRIC:
2008                         pd_idx = sector_div(stripe2, raid_disks);
2009                         if (*dd_idx >= pd_idx)
2010                                 (*dd_idx)++;
2011                         break;
2012                 case ALGORITHM_LEFT_SYMMETRIC:
2013                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2014                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2015                         break;
2016                 case ALGORITHM_RIGHT_SYMMETRIC:
2017                         pd_idx = sector_div(stripe2, raid_disks);
2018                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2019                         break;
2020                 case ALGORITHM_PARITY_0:
2021                         pd_idx = 0;
2022                         (*dd_idx)++;
2023                         break;
2024                 case ALGORITHM_PARITY_N:
2025                         pd_idx = data_disks;
2026                         break;
2027                 default:
2028                         BUG();
2029                 }
2030                 break;
2031         case 6:
2032
2033                 switch (algorithm) {
2034                 case ALGORITHM_LEFT_ASYMMETRIC:
2035                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2036                         qd_idx = pd_idx + 1;
2037                         if (pd_idx == raid_disks-1) {
2038                                 (*dd_idx)++;    /* Q D D D P */
2039                                 qd_idx = 0;
2040                         } else if (*dd_idx >= pd_idx)
2041                                 (*dd_idx) += 2; /* D D P Q D */
2042                         break;
2043                 case ALGORITHM_RIGHT_ASYMMETRIC:
2044                         pd_idx = sector_div(stripe2, raid_disks);
2045                         qd_idx = pd_idx + 1;
2046                         if (pd_idx == raid_disks-1) {
2047                                 (*dd_idx)++;    /* Q D D D P */
2048                                 qd_idx = 0;
2049                         } else if (*dd_idx >= pd_idx)
2050                                 (*dd_idx) += 2; /* D D P Q D */
2051                         break;
2052                 case ALGORITHM_LEFT_SYMMETRIC:
2053                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2054                         qd_idx = (pd_idx + 1) % raid_disks;
2055                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2056                         break;
2057                 case ALGORITHM_RIGHT_SYMMETRIC:
2058                         pd_idx = sector_div(stripe2, raid_disks);
2059                         qd_idx = (pd_idx + 1) % raid_disks;
2060                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2061                         break;
2062
2063                 case ALGORITHM_PARITY_0:
2064                         pd_idx = 0;
2065                         qd_idx = 1;
2066                         (*dd_idx) += 2;
2067                         break;
2068                 case ALGORITHM_PARITY_N:
2069                         pd_idx = data_disks;
2070                         qd_idx = data_disks + 1;
2071                         break;
2072
2073                 case ALGORITHM_ROTATING_ZERO_RESTART:
2074                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
2075                          * of blocks for computing Q is different.
2076                          */
2077                         pd_idx = sector_div(stripe2, raid_disks);
2078                         qd_idx = pd_idx + 1;
2079                         if (pd_idx == raid_disks-1) {
2080                                 (*dd_idx)++;    /* Q D D D P */
2081                                 qd_idx = 0;
2082                         } else if (*dd_idx >= pd_idx)
2083                                 (*dd_idx) += 2; /* D D P Q D */
2084                         ddf_layout = 1;
2085                         break;
2086
2087                 case ALGORITHM_ROTATING_N_RESTART:
2088                         /* Same a left_asymmetric, by first stripe is
2089                          * D D D P Q  rather than
2090                          * Q D D D P
2091                          */
2092                         stripe2 += 1;
2093                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2094                         qd_idx = pd_idx + 1;
2095                         if (pd_idx == raid_disks-1) {
2096                                 (*dd_idx)++;    /* Q D D D P */
2097                                 qd_idx = 0;
2098                         } else if (*dd_idx >= pd_idx)
2099                                 (*dd_idx) += 2; /* D D P Q D */
2100                         ddf_layout = 1;
2101                         break;
2102
2103                 case ALGORITHM_ROTATING_N_CONTINUE:
2104                         /* Same as left_symmetric but Q is before P */
2105                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2106                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2107                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2108                         ddf_layout = 1;
2109                         break;
2110
2111                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2112                         /* RAID5 left_asymmetric, with Q on last device */
2113                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2114                         if (*dd_idx >= pd_idx)
2115                                 (*dd_idx)++;
2116                         qd_idx = raid_disks - 1;
2117                         break;
2118
2119                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2120                         pd_idx = sector_div(stripe2, raid_disks-1);
2121                         if (*dd_idx >= pd_idx)
2122                                 (*dd_idx)++;
2123                         qd_idx = raid_disks - 1;
2124                         break;
2125
2126                 case ALGORITHM_LEFT_SYMMETRIC_6:
2127                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2128                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2129                         qd_idx = raid_disks - 1;
2130                         break;
2131
2132                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2133                         pd_idx = sector_div(stripe2, raid_disks-1);
2134                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2135                         qd_idx = raid_disks - 1;
2136                         break;
2137
2138                 case ALGORITHM_PARITY_0_6:
2139                         pd_idx = 0;
2140                         (*dd_idx)++;
2141                         qd_idx = raid_disks - 1;
2142                         break;
2143
2144                 default:
2145                         BUG();
2146                 }
2147                 break;
2148         }
2149
2150         if (sh) {
2151                 sh->pd_idx = pd_idx;
2152                 sh->qd_idx = qd_idx;
2153                 sh->ddf_layout = ddf_layout;
2154         }
2155         /*
2156          * Finally, compute the new sector number
2157          */
2158         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2159         return new_sector;
2160 }
2161
2162
2163 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2164 {
2165         struct r5conf *conf = sh->raid_conf;
2166         int raid_disks = sh->disks;
2167         int data_disks = raid_disks - conf->max_degraded;
2168         sector_t new_sector = sh->sector, check;
2169         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2170                                          : conf->chunk_sectors;
2171         int algorithm = previous ? conf->prev_algo
2172                                  : conf->algorithm;
2173         sector_t stripe;
2174         int chunk_offset;
2175         sector_t chunk_number;
2176         int dummy1, dd_idx = i;
2177         sector_t r_sector;
2178         struct stripe_head sh2;
2179
2180
2181         chunk_offset = sector_div(new_sector, sectors_per_chunk);
2182         stripe = new_sector;
2183
2184         if (i == sh->pd_idx)
2185                 return 0;
2186         switch(conf->level) {
2187         case 4: break;
2188         case 5:
2189                 switch (algorithm) {
2190                 case ALGORITHM_LEFT_ASYMMETRIC:
2191                 case ALGORITHM_RIGHT_ASYMMETRIC:
2192                         if (i > sh->pd_idx)
2193                                 i--;
2194                         break;
2195                 case ALGORITHM_LEFT_SYMMETRIC:
2196                 case ALGORITHM_RIGHT_SYMMETRIC:
2197                         if (i < sh->pd_idx)
2198                                 i += raid_disks;
2199                         i -= (sh->pd_idx + 1);
2200                         break;
2201                 case ALGORITHM_PARITY_0:
2202                         i -= 1;
2203                         break;
2204                 case ALGORITHM_PARITY_N:
2205                         break;
2206                 default:
2207                         BUG();
2208                 }
2209                 break;
2210         case 6:
2211                 if (i == sh->qd_idx)
2212                         return 0; /* It is the Q disk */
2213                 switch (algorithm) {
2214                 case ALGORITHM_LEFT_ASYMMETRIC:
2215                 case ALGORITHM_RIGHT_ASYMMETRIC:
2216                 case ALGORITHM_ROTATING_ZERO_RESTART:
2217                 case ALGORITHM_ROTATING_N_RESTART:
2218                         if (sh->pd_idx == raid_disks-1)
2219                                 i--;    /* Q D D D P */
2220                         else if (i > sh->pd_idx)
2221                                 i -= 2; /* D D P Q D */
2222                         break;
2223                 case ALGORITHM_LEFT_SYMMETRIC:
2224                 case ALGORITHM_RIGHT_SYMMETRIC:
2225                         if (sh->pd_idx == raid_disks-1)
2226                                 i--; /* Q D D D P */
2227                         else {
2228                                 /* D D P Q D */
2229                                 if (i < sh->pd_idx)
2230                                         i += raid_disks;
2231                                 i -= (sh->pd_idx + 2);
2232                         }
2233                         break;
2234                 case ALGORITHM_PARITY_0:
2235                         i -= 2;
2236                         break;
2237                 case ALGORITHM_PARITY_N:
2238                         break;
2239                 case ALGORITHM_ROTATING_N_CONTINUE:
2240                         /* Like left_symmetric, but P is before Q */
2241                         if (sh->pd_idx == 0)
2242                                 i--;    /* P D D D Q */
2243                         else {
2244                                 /* D D Q P D */
2245                                 if (i < sh->pd_idx)
2246                                         i += raid_disks;
2247                                 i -= (sh->pd_idx + 1);
2248                         }
2249                         break;
2250                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2251                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2252                         if (i > sh->pd_idx)
2253                                 i--;
2254                         break;
2255                 case ALGORITHM_LEFT_SYMMETRIC_6:
2256                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2257                         if (i < sh->pd_idx)
2258                                 i += data_disks + 1;
2259                         i -= (sh->pd_idx + 1);
2260                         break;
2261                 case ALGORITHM_PARITY_0_6:
2262                         i -= 1;
2263                         break;
2264                 default:
2265                         BUG();
2266                 }
2267                 break;
2268         }
2269
2270         chunk_number = stripe * data_disks + i;
2271         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2272
2273         check = raid5_compute_sector(conf, r_sector,
2274                                      previous, &dummy1, &sh2);
2275         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2276                 || sh2.qd_idx != sh->qd_idx) {
2277                 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2278                        mdname(conf->mddev));
2279                 return 0;
2280         }
2281         return r_sector;
2282 }
2283
2284
2285 static void
2286 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2287                          int rcw, int expand)
2288 {
2289         int i, pd_idx = sh->pd_idx, disks = sh->disks;
2290         struct r5conf *conf = sh->raid_conf;
2291         int level = conf->level;
2292
2293         if (rcw) {
2294
2295                 for (i = disks; i--; ) {
2296                         struct r5dev *dev = &sh->dev[i];
2297
2298                         if (dev->towrite) {
2299                                 set_bit(R5_LOCKED, &dev->flags);
2300                                 set_bit(R5_Wantdrain, &dev->flags);
2301                                 if (!expand)
2302                                         clear_bit(R5_UPTODATE, &dev->flags);
2303                                 s->locked++;
2304                         }
2305                 }
2306                 /* if we are not expanding this is a proper write request, and
2307                  * there will be bios with new data to be drained into the
2308                  * stripe cache
2309                  */
2310                 if (!expand) {
2311                         if (!s->locked)
2312                                 /* False alarm, nothing to do */
2313                                 return;
2314                         sh->reconstruct_state = reconstruct_state_drain_run;
2315                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2316                 } else
2317                         sh->reconstruct_state = reconstruct_state_run;
2318
2319                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2320
2321                 if (s->locked + conf->max_degraded == disks)
2322                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2323                                 atomic_inc(&conf->pending_full_writes);
2324         } else {
2325                 BUG_ON(level == 6);
2326                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2327                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2328
2329                 for (i = disks; i--; ) {
2330                         struct r5dev *dev = &sh->dev[i];
2331                         if (i == pd_idx)
2332                                 continue;
2333
2334                         if (dev->towrite &&
2335                             (test_bit(R5_UPTODATE, &dev->flags) ||
2336                              test_bit(R5_Wantcompute, &dev->flags))) {
2337                                 set_bit(R5_Wantdrain, &dev->flags);
2338                                 set_bit(R5_LOCKED, &dev->flags);
2339                                 clear_bit(R5_UPTODATE, &dev->flags);
2340                                 s->locked++;
2341                         }
2342                 }
2343                 if (!s->locked)
2344                         /* False alarm - nothing to do */
2345                         return;
2346                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2347                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2348                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2349                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2350         }
2351
2352         /* keep the parity disk(s) locked while asynchronous operations
2353          * are in flight
2354          */
2355         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2356         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2357         s->locked++;
2358
2359         if (level == 6) {
2360                 int qd_idx = sh->qd_idx;
2361                 struct r5dev *dev = &sh->dev[qd_idx];
2362
2363                 set_bit(R5_LOCKED, &dev->flags);
2364                 clear_bit(R5_UPTODATE, &dev->flags);
2365                 s->locked++;
2366         }
2367
2368         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2369                 __func__, (unsigned long long)sh->sector,
2370                 s->locked, s->ops_request);
2371 }
2372
2373 /*
2374  * Each stripe/dev can have one or more bion attached.
2375  * toread/towrite point to the first in a chain.
2376  * The bi_next chain must be in order.
2377  */
2378 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2379 {
2380         struct bio **bip;
2381         struct r5conf *conf = sh->raid_conf;
2382         int firstwrite=0;
2383
2384         pr_debug("adding bi b#%llu to stripe s#%llu\n",
2385                 (unsigned long long)bi->bi_sector,
2386                 (unsigned long long)sh->sector);
2387
2388         /*
2389          * If several bio share a stripe. The bio bi_phys_segments acts as a
2390          * reference count to avoid race. The reference count should already be
2391          * increased before this function is called (for example, in
2392          * make_request()), so other bio sharing this stripe will not free the
2393          * stripe. If a stripe is owned by one stripe, the stripe lock will
2394          * protect it.
2395          */
2396         spin_lock_irq(&sh->stripe_lock);
2397         if (forwrite) {
2398                 bip = &sh->dev[dd_idx].towrite;
2399                 if (*bip == NULL)
2400                         firstwrite = 1;
2401         } else
2402                 bip = &sh->dev[dd_idx].toread;
2403         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2404                 if (bio_end_sector(*bip) > bi->bi_sector)
2405                         goto overlap;
2406                 bip = & (*bip)->bi_next;
2407         }
2408         if (*bip && (*bip)->bi_sector < bio_end_sector(bi))
2409                 goto overlap;
2410
2411         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2412         if (*bip)
2413                 bi->bi_next = *bip;
2414         *bip = bi;
2415         raid5_inc_bi_active_stripes(bi);
2416
2417         if (forwrite) {
2418                 /* check if page is covered */
2419                 sector_t sector = sh->dev[dd_idx].sector;
2420                 for (bi=sh->dev[dd_idx].towrite;
2421                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2422                              bi && bi->bi_sector <= sector;
2423                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2424                         if (bio_end_sector(bi) >= sector)
2425                                 sector = bio_end_sector(bi);
2426                 }
2427                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2428                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2429         }
2430
2431         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2432                 (unsigned long long)(*bip)->bi_sector,
2433                 (unsigned long long)sh->sector, dd_idx);
2434         spin_unlock_irq(&sh->stripe_lock);
2435
2436         if (conf->mddev->bitmap && firstwrite) {
2437                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2438                                   STRIPE_SECTORS, 0);
2439                 sh->bm_seq = conf->seq_flush+1;
2440                 set_bit(STRIPE_BIT_DELAY, &sh->state);
2441         }
2442         return 1;
2443
2444  overlap:
2445         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2446         spin_unlock_irq(&sh->stripe_lock);
2447         return 0;
2448 }
2449
2450 static void end_reshape(struct r5conf *conf);
2451
2452 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2453                             struct stripe_head *sh)
2454 {
2455         int sectors_per_chunk =
2456                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2457         int dd_idx;
2458         int chunk_offset = sector_div(stripe, sectors_per_chunk);
2459         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2460
2461         raid5_compute_sector(conf,
2462                              stripe * (disks - conf->max_degraded)
2463                              *sectors_per_chunk + chunk_offset,
2464                              previous,
2465                              &dd_idx, sh);
2466 }
2467
2468 static void
2469 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2470                                 struct stripe_head_state *s, int disks,
2471                                 struct bio **return_bi)
2472 {
2473         int i;
2474         for (i = disks; i--; ) {
2475                 struct bio *bi;
2476                 int bitmap_end = 0;
2477
2478                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2479                         struct md_rdev *rdev;
2480                         rcu_read_lock();
2481                         rdev = rcu_dereference(conf->disks[i].rdev);
2482                         if (rdev && test_bit(In_sync, &rdev->flags))
2483                                 atomic_inc(&rdev->nr_pending);
2484                         else
2485                                 rdev = NULL;
2486                         rcu_read_unlock();
2487                         if (rdev) {
2488                                 if (!rdev_set_badblocks(
2489                                             rdev,
2490                                             sh->sector,
2491                                             STRIPE_SECTORS, 0))
2492                                         md_error(conf->mddev, rdev);
2493                                 rdev_dec_pending(rdev, conf->mddev);
2494                         }
2495                 }
2496                 spin_lock_irq(&sh->stripe_lock);
2497                 /* fail all writes first */
2498                 bi = sh->dev[i].towrite;
2499                 sh->dev[i].towrite = NULL;
2500                 spin_unlock_irq(&sh->stripe_lock);
2501                 if (bi)
2502                         bitmap_end = 1;
2503
2504                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2505                         wake_up(&conf->wait_for_overlap);
2506
2507                 while (bi && bi->bi_sector <
2508                         sh->dev[i].sector + STRIPE_SECTORS) {
2509                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2510                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2511                         if (!raid5_dec_bi_active_stripes(bi)) {
2512                                 md_write_end(conf->mddev);
2513                                 bi->bi_next = *return_bi;
2514                                 *return_bi = bi;
2515                         }
2516                         bi = nextbi;
2517                 }
2518                 if (bitmap_end)
2519                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2520                                 STRIPE_SECTORS, 0, 0);
2521                 bitmap_end = 0;
2522                 /* and fail all 'written' */
2523                 bi = sh->dev[i].written;
2524                 sh->dev[i].written = NULL;
2525                 if (bi) bitmap_end = 1;
2526                 while (bi && bi->bi_sector <
2527                        sh->dev[i].sector + STRIPE_SECTORS) {
2528                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2529                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2530                         if (!raid5_dec_bi_active_stripes(bi)) {
2531                                 md_write_end(conf->mddev);
2532                                 bi->bi_next = *return_bi;
2533                                 *return_bi = bi;
2534                         }
2535                         bi = bi2;
2536                 }
2537
2538                 /* fail any reads if this device is non-operational and
2539                  * the data has not reached the cache yet.
2540                  */
2541                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2542                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2543                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
2544                         spin_lock_irq(&sh->stripe_lock);
2545                         bi = sh->dev[i].toread;
2546                         sh->dev[i].toread = NULL;
2547                         spin_unlock_irq(&sh->stripe_lock);
2548                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2549                                 wake_up(&conf->wait_for_overlap);
2550                         while (bi && bi->bi_sector <
2551                                sh->dev[i].sector + STRIPE_SECTORS) {
2552                                 struct bio *nextbi =
2553                                         r5_next_bio(bi, sh->dev[i].sector);
2554                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2555                                 if (!raid5_dec_bi_active_stripes(bi)) {
2556                                         bi->bi_next = *return_bi;
2557                                         *return_bi = bi;
2558                                 }
2559                                 bi = nextbi;
2560                         }
2561                 }
2562                 if (bitmap_end)
2563                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2564                                         STRIPE_SECTORS, 0, 0);
2565                 /* If we were in the middle of a write the parity block might
2566                  * still be locked - so just clear all R5_LOCKED flags
2567                  */
2568                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2569         }
2570
2571         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2572                 if (atomic_dec_and_test(&conf->pending_full_writes))
2573                         md_wakeup_thread(conf->mddev->thread);
2574 }
2575
2576 static void
2577 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2578                    struct stripe_head_state *s)
2579 {
2580         int abort = 0;
2581         int i;
2582
2583         clear_bit(STRIPE_SYNCING, &sh->state);
2584         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
2585                 wake_up(&conf->wait_for_overlap);
2586         s->syncing = 0;
2587         s->replacing = 0;
2588         /* There is nothing more to do for sync/check/repair.
2589          * Don't even need to abort as that is handled elsewhere
2590          * if needed, and not always wanted e.g. if there is a known
2591          * bad block here.
2592          * For recover/replace we need to record a bad block on all
2593          * non-sync devices, or abort the recovery
2594          */
2595         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2596                 /* During recovery devices cannot be removed, so
2597                  * locking and refcounting of rdevs is not needed
2598                  */
2599                 for (i = 0; i < conf->raid_disks; i++) {
2600                         struct md_rdev *rdev = conf->disks[i].rdev;
2601                         if (rdev
2602                             && !test_bit(Faulty, &rdev->flags)
2603                             && !test_bit(In_sync, &rdev->flags)
2604                             && !rdev_set_badblocks(rdev, sh->sector,
2605                                                    STRIPE_SECTORS, 0))
2606                                 abort = 1;
2607                         rdev = conf->disks[i].replacement;
2608                         if (rdev
2609                             && !test_bit(Faulty, &rdev->flags)
2610                             && !test_bit(In_sync, &rdev->flags)
2611                             && !rdev_set_badblocks(rdev, sh->sector,
2612                                                    STRIPE_SECTORS, 0))
2613                                 abort = 1;
2614                 }
2615                 if (abort)
2616                         conf->recovery_disabled =
2617                                 conf->mddev->recovery_disabled;
2618         }
2619         md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
2620 }
2621
2622 static int want_replace(struct stripe_head *sh, int disk_idx)
2623 {
2624         struct md_rdev *rdev;
2625         int rv = 0;
2626         /* Doing recovery so rcu locking not required */
2627         rdev = sh->raid_conf->disks[disk_idx].replacement;
2628         if (rdev
2629             && !test_bit(Faulty, &rdev->flags)
2630             && !test_bit(In_sync, &rdev->flags)
2631             && (rdev->recovery_offset <= sh->sector
2632                 || rdev->mddev->recovery_cp <= sh->sector))
2633                 rv = 1;
2634
2635         return rv;
2636 }
2637
2638 /* fetch_block - checks the given member device to see if its data needs
2639  * to be read or computed to satisfy a request.
2640  *
2641  * Returns 1 when no more member devices need to be checked, otherwise returns
2642  * 0 to tell the loop in handle_stripe_fill to continue
2643  */
2644 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2645                        int disk_idx, int disks)
2646 {
2647         struct r5dev *dev = &sh->dev[disk_idx];
2648         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2649                                   &sh->dev[s->failed_num[1]] };
2650
2651         /* is the data in this block needed, and can we get it? */
2652         if (!test_bit(R5_LOCKED, &dev->flags) &&
2653             !test_bit(R5_UPTODATE, &dev->flags) &&
2654             (dev->toread ||
2655              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2656              s->syncing || s->expanding ||
2657              (s->replacing && want_replace(sh, disk_idx)) ||
2658              (s->failed >= 1 && fdev[0]->toread) ||
2659              (s->failed >= 2 && fdev[1]->toread) ||
2660              (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2661               !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2662              (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2663                 /* we would like to get this block, possibly by computing it,
2664                  * otherwise read it if the backing disk is insync
2665                  */
2666                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2667                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2668                 if ((s->uptodate == disks - 1) &&
2669                     (s->failed && (disk_idx == s->failed_num[0] ||
2670                                    disk_idx == s->failed_num[1]))) {
2671                         /* have disk failed, and we're requested to fetch it;
2672                          * do compute it
2673                          */
2674                         pr_debug("Computing stripe %llu block %d\n",
2675                                (unsigned long long)sh->sector, disk_idx);
2676                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2677                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2678                         set_bit(R5_Wantcompute, &dev->flags);
2679                         sh->ops.target = disk_idx;
2680                         sh->ops.target2 = -1; /* no 2nd target */
2681                         s->req_compute = 1;
2682                         /* Careful: from this point on 'uptodate' is in the eye
2683                          * of raid_run_ops which services 'compute' operations
2684                          * before writes. R5_Wantcompute flags a block that will
2685                          * be R5_UPTODATE by the time it is needed for a
2686                          * subsequent operation.
2687                          */
2688                         s->uptodate++;
2689                         return 1;
2690                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2691                         /* Computing 2-failure is *very* expensive; only
2692                          * do it if failed >= 2
2693                          */
2694                         int other;
2695                         for (other = disks; other--; ) {
2696                                 if (other == disk_idx)
2697                                         continue;
2698                                 if (!test_bit(R5_UPTODATE,
2699                                       &sh->dev[other].flags))
2700                                         break;
2701                         }
2702                         BUG_ON(other < 0);
2703                         pr_debug("Computing stripe %llu blocks %d,%d\n",
2704                                (unsigned long long)sh->sector,
2705                                disk_idx, other);
2706                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2707                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2708                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2709                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
2710                         sh->ops.target = disk_idx;
2711                         sh->ops.target2 = other;
2712                         s->uptodate += 2;
2713                         s->req_compute = 1;
2714                         return 1;
2715                 } else if (test_bit(R5_Insync, &dev->flags)) {
2716                         set_bit(R5_LOCKED, &dev->flags);
2717                         set_bit(R5_Wantread, &dev->flags);
2718                         s->locked++;
2719                         pr_debug("Reading block %d (sync=%d)\n",
2720                                 disk_idx, s->syncing);
2721                 }
2722         }
2723
2724         return 0;
2725 }
2726
2727 /**
2728  * handle_stripe_fill - read or compute data to satisfy pending requests.
2729  */
2730 static void handle_stripe_fill(struct stripe_head *sh,
2731                                struct stripe_head_state *s,
2732                                int disks)
2733 {
2734         int i;
2735
2736         /* look for blocks to read/compute, skip this if a compute
2737          * is already in flight, or if the stripe contents are in the
2738          * midst of changing due to a write
2739          */
2740         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2741             !sh->reconstruct_state)
2742                 for (i = disks; i--; )
2743                         if (fetch_block(sh, s, i, disks))
2744                                 break;
2745         set_bit(STRIPE_HANDLE, &sh->state);
2746 }
2747
2748
2749 /* handle_stripe_clean_event
2750  * any written block on an uptodate or failed drive can be returned.
2751  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2752  * never LOCKED, so we don't need to test 'failed' directly.
2753  */
2754 static void handle_stripe_clean_event(struct r5conf *conf,
2755         struct stripe_head *sh, int disks, struct bio **return_bi)
2756 {
2757         int i;
2758         struct r5dev *dev;
2759         int discard_pending = 0;
2760
2761         for (i = disks; i--; )
2762                 if (sh->dev[i].written) {
2763                         dev = &sh->dev[i];
2764                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2765                             (test_bit(R5_UPTODATE, &dev->flags) ||
2766                              test_bit(R5_Discard, &dev->flags))) {
2767                                 /* We can return any write requests */
2768                                 struct bio *wbi, *wbi2;
2769                                 pr_debug("Return write for disc %d\n", i);
2770                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
2771                                         clear_bit(R5_UPTODATE, &dev->flags);
2772                                 wbi = dev->written;
2773                                 dev->written = NULL;
2774                                 while (wbi && wbi->bi_sector <
2775                                         dev->sector + STRIPE_SECTORS) {
2776                                         wbi2 = r5_next_bio(wbi, dev->sector);
2777                                         if (!raid5_dec_bi_active_stripes(wbi)) {
2778                                                 md_write_end(conf->mddev);
2779                                                 wbi->bi_next = *return_bi;
2780                                                 *return_bi = wbi;
2781                                         }
2782                                         wbi = wbi2;
2783                                 }
2784                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2785                                                 STRIPE_SECTORS,
2786                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2787                                                 0);
2788                         } else if (test_bit(R5_Discard, &dev->flags))
2789                                 discard_pending = 1;
2790                 }
2791         if (!discard_pending &&
2792             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
2793                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
2794                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2795                 if (sh->qd_idx >= 0) {
2796                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
2797                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
2798                 }
2799                 /* now that discard is done we can proceed with any sync */
2800                 clear_bit(STRIPE_DISCARD, &sh->state);
2801                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2802                         set_bit(STRIPE_HANDLE, &sh->state);
2803
2804         }
2805
2806         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2807                 if (atomic_dec_and_test(&conf->pending_full_writes))
2808                         md_wakeup_thread(conf->mddev->thread);
2809 }
2810
2811 static void handle_stripe_dirtying(struct r5conf *conf,
2812                                    struct stripe_head *sh,
2813                                    struct stripe_head_state *s,
2814                                    int disks)
2815 {
2816         int rmw = 0, rcw = 0, i;
2817         sector_t recovery_cp = conf->mddev->recovery_cp;
2818
2819         /* RAID6 requires 'rcw' in current implementation.
2820          * Otherwise, check whether resync is now happening or should start.
2821          * If yes, then the array is dirty (after unclean shutdown or
2822          * initial creation), so parity in some stripes might be inconsistent.
2823          * In this case, we need to always do reconstruct-write, to ensure
2824          * that in case of drive failure or read-error correction, we
2825          * generate correct data from the parity.
2826          */
2827         if (conf->max_degraded == 2 ||
2828             (recovery_cp < MaxSector && sh->sector >= recovery_cp)) {
2829                 /* Calculate the real rcw later - for now make it
2830                  * look like rcw is cheaper
2831                  */
2832                 rcw = 1; rmw = 2;
2833                 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
2834                          conf->max_degraded, (unsigned long long)recovery_cp,
2835                          (unsigned long long)sh->sector);
2836         } else for (i = disks; i--; ) {
2837                 /* would I have to read this buffer for read_modify_write */
2838                 struct r5dev *dev = &sh->dev[i];
2839                 if ((dev->towrite || i == sh->pd_idx) &&
2840                     !test_bit(R5_LOCKED, &dev->flags) &&
2841                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2842                       test_bit(R5_Wantcompute, &dev->flags))) {
2843                         if (test_bit(R5_Insync, &dev->flags))
2844                                 rmw++;
2845                         else
2846                                 rmw += 2*disks;  /* cannot read it */
2847                 }
2848                 /* Would I have to read this buffer for reconstruct_write */
2849                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2850                     !test_bit(R5_LOCKED, &dev->flags) &&
2851                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2852                     test_bit(R5_Wantcompute, &dev->flags))) {
2853                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2854                         else
2855                                 rcw += 2*disks;
2856                 }
2857         }
2858         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2859                 (unsigned long long)sh->sector, rmw, rcw);
2860         set_bit(STRIPE_HANDLE, &sh->state);
2861         if (rmw < rcw && rmw > 0) {
2862                 /* prefer read-modify-write, but need to get some data */
2863                 if (conf->mddev->queue)
2864                         blk_add_trace_msg(conf->mddev->queue,
2865                                           "raid5 rmw %llu %d",
2866                                           (unsigned long long)sh->sector, rmw);
2867                 for (i = disks; i--; ) {
2868                         struct r5dev *dev = &sh->dev[i];
2869                         if ((dev->towrite || i == sh->pd_idx) &&
2870                             !test_bit(R5_LOCKED, &dev->flags) &&
2871                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2872                             test_bit(R5_Wantcompute, &dev->flags)) &&
2873                             test_bit(R5_Insync, &dev->flags)) {
2874                                 if (
2875                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2876                                         pr_debug("Read_old block "
2877                                                  "%d for r-m-w\n", i);
2878                                         set_bit(R5_LOCKED, &dev->flags);
2879                                         set_bit(R5_Wantread, &dev->flags);
2880                                         s->locked++;
2881                                 } else {
2882                                         set_bit(STRIPE_DELAYED, &sh->state);
2883                                         set_bit(STRIPE_HANDLE, &sh->state);
2884                                 }
2885                         }
2886                 }
2887         }
2888         if (rcw <= rmw && rcw > 0) {
2889                 /* want reconstruct write, but need to get some data */
2890                 int qread =0;
2891                 rcw = 0;
2892                 for (i = disks; i--; ) {
2893                         struct r5dev *dev = &sh->dev[i];
2894                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2895                             i != sh->pd_idx && i != sh->qd_idx &&
2896                             !test_bit(R5_LOCKED, &dev->flags) &&
2897                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2898                               test_bit(R5_Wantcompute, &dev->flags))) {
2899                                 rcw++;
2900                                 if (!test_bit(R5_Insync, &dev->flags))
2901                                         continue; /* it's a failed drive */
2902                                 if (
2903                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2904                                         pr_debug("Read_old block "
2905                                                 "%d for Reconstruct\n", i);
2906                                         set_bit(R5_LOCKED, &dev->flags);
2907                                         set_bit(R5_Wantread, &dev->flags);
2908                                         s->locked++;
2909                                         qread++;
2910                                 } else {
2911                                         set_bit(STRIPE_DELAYED, &sh->state);
2912                                         set_bit(STRIPE_HANDLE, &sh->state);
2913                                 }
2914                         }
2915                 }
2916                 if (rcw && conf->mddev->queue)
2917                         blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
2918                                           (unsigned long long)sh->sector,
2919                                           rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
2920         }
2921         /* now if nothing is locked, and if we have enough data,
2922          * we can start a write request
2923          */
2924         /* since handle_stripe can be called at any time we need to handle the
2925          * case where a compute block operation has been submitted and then a
2926          * subsequent call wants to start a write request.  raid_run_ops only
2927          * handles the case where compute block and reconstruct are requested
2928          * simultaneously.  If this is not the case then new writes need to be
2929          * held off until the compute completes.
2930          */
2931         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2932             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2933             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2934                 schedule_reconstruction(sh, s, rcw == 0, 0);
2935 }
2936
2937 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2938                                 struct stripe_head_state *s, int disks)
2939 {
2940         struct r5dev *dev = NULL;
2941
2942         set_bit(STRIPE_HANDLE, &sh->state);
2943
2944         switch (sh->check_state) {
2945         case check_state_idle:
2946                 /* start a new check operation if there are no failures */
2947                 if (s->failed == 0) {
2948                         BUG_ON(s->uptodate != disks);
2949                         sh->check_state = check_state_run;
2950                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2951                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2952                         s->uptodate--;
2953                         break;
2954                 }
2955                 dev = &sh->dev[s->failed_num[0]];
2956                 /* fall through */
2957         case check_state_compute_result:
2958                 sh->check_state = check_state_idle;
2959                 if (!dev)
2960                         dev = &sh->dev[sh->pd_idx];
2961
2962                 /* check that a write has not made the stripe insync */
2963                 if (test_bit(STRIPE_INSYNC, &sh->state))
2964                         break;
2965
2966                 /* either failed parity check, or recovery is happening */
2967                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2968                 BUG_ON(s->uptodate != disks);
2969
2970                 set_bit(R5_LOCKED, &dev->flags);
2971                 s->locked++;
2972                 set_bit(R5_Wantwrite, &dev->flags);
2973
2974                 clear_bit(STRIPE_DEGRADED, &sh->state);
2975                 set_bit(STRIPE_INSYNC, &sh->state);
2976                 break;
2977         case check_state_run:
2978                 break; /* we will be called again upon completion */
2979         case check_state_check_result:
2980                 sh->check_state = check_state_idle;
2981
2982                 /* if a failure occurred during the check operation, leave
2983                  * STRIPE_INSYNC not set and let the stripe be handled again
2984                  */
2985                 if (s->failed)
2986                         break;
2987
2988                 /* handle a successful check operation, if parity is correct
2989                  * we are done.  Otherwise update the mismatch count and repair
2990                  * parity if !MD_RECOVERY_CHECK
2991                  */
2992                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2993                         /* parity is correct (on disc,
2994                          * not in buffer any more)
2995                          */
2996                         set_bit(STRIPE_INSYNC, &sh->state);
2997                 else {
2998                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
2999                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3000                                 /* don't try to repair!! */
3001                                 set_bit(STRIPE_INSYNC, &sh->state);
3002                         else {
3003                                 sh->check_state = check_state_compute_run;
3004                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3005                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3006                                 set_bit(R5_Wantcompute,
3007                                         &sh->dev[sh->pd_idx].flags);
3008                                 sh->ops.target = sh->pd_idx;
3009                                 sh->ops.target2 = -1;
3010                                 s->uptodate++;
3011                         }
3012                 }
3013                 break;
3014         case check_state_compute_run:
3015                 break;
3016         default:
3017                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3018                        __func__, sh->check_state,
3019                        (unsigned long long) sh->sector);
3020                 BUG();
3021         }
3022 }
3023
3024
3025 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3026                                   struct stripe_head_state *s,
3027                                   int disks)
3028 {
3029         int pd_idx = sh->pd_idx;
3030         int qd_idx = sh->qd_idx;
3031         struct r5dev *dev;
3032
3033         set_bit(STRIPE_HANDLE, &sh->state);
3034
3035         BUG_ON(s->failed > 2);
3036
3037         /* Want to check and possibly repair P and Q.
3038          * However there could be one 'failed' device, in which
3039          * case we can only check one of them, possibly using the
3040          * other to generate missing data
3041          */
3042
3043         switch (sh->check_state) {
3044         case check_state_idle:
3045                 /* start a new check operation if there are < 2 failures */
3046                 if (s->failed == s->q_failed) {
3047                         /* The only possible failed device holds Q, so it
3048                          * makes sense to check P (If anything else were failed,
3049                          * we would have used P to recreate it).
3050                          */
3051                         sh->check_state = check_state_run;
3052                 }
3053                 if (!s->q_failed && s->failed < 2) {
3054                         /* Q is not failed, and we didn't use it to generate
3055                          * anything, so it makes sense to check it
3056                          */
3057                         if (sh->check_state == check_state_run)
3058                                 sh->check_state = check_state_run_pq;
3059                         else
3060                                 sh->check_state = check_state_run_q;
3061                 }
3062
3063                 /* discard potentially stale zero_sum_result */
3064                 sh->ops.zero_sum_result = 0;
3065
3066                 if (sh->check_state == check_state_run) {
3067                         /* async_xor_zero_sum destroys the contents of P */
3068                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3069                         s->uptodate--;
3070                 }
3071                 if (sh->check_state >= check_state_run &&
3072                     sh->check_state <= check_state_run_pq) {
3073                         /* async_syndrome_zero_sum preserves P and Q, so
3074                          * no need to mark them !uptodate here
3075                          */
3076                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
3077                         break;
3078                 }
3079
3080                 /* we have 2-disk failure */
3081                 BUG_ON(s->failed != 2);
3082                 /* fall through */
3083         case check_state_compute_result:
3084                 sh->check_state = check_state_idle;
3085
3086                 /* check that a write has not made the stripe insync */
3087                 if (test_bit(STRIPE_INSYNC, &sh->state))
3088                         break;
3089
3090                 /* now write out any block on a failed drive,
3091                  * or P or Q if they were recomputed
3092                  */
3093                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3094                 if (s->failed == 2) {
3095                         dev = &sh->dev[s->failed_num[1]];
3096                         s->locked++;
3097                         set_bit(R5_LOCKED, &dev->flags);
3098                         set_bit(R5_Wantwrite, &dev->flags);
3099                 }
3100                 if (s->failed >= 1) {
3101                         dev = &sh->dev[s->failed_num[0]];
3102                         s->locked++;
3103                         set_bit(R5_LOCKED, &dev->flags);
3104                         set_bit(R5_Wantwrite, &dev->flags);
3105                 }
3106                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3107                         dev = &sh->dev[pd_idx];
3108                         s->locked++;
3109                         set_bit(R5_LOCKED, &dev->flags);
3110                         set_bit(R5_Wantwrite, &dev->flags);
3111                 }
3112                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3113                         dev = &sh->dev[qd_idx];
3114                         s->locked++;
3115                         set_bit(R5_LOCKED, &dev->flags);
3116                         set_bit(R5_Wantwrite, &dev->flags);
3117                 }
3118                 clear_bit(STRIPE_DEGRADED, &sh->state);
3119
3120                 set_bit(STRIPE_INSYNC, &sh->state);
3121                 break;
3122         case check_state_run:
3123         case check_state_run_q:
3124         case check_state_run_pq:
3125                 break; /* we will be called again upon completion */
3126         case check_state_check_result:
3127                 sh->check_state = check_state_idle;
3128
3129                 /* handle a successful check operation, if parity is correct
3130                  * we are done.  Otherwise update the mismatch count and repair
3131                  * parity if !MD_RECOVERY_CHECK
3132                  */
3133                 if (sh->ops.zero_sum_result == 0) {
3134                         /* both parities are correct */
3135                         if (!s->failed)
3136                                 set_bit(STRIPE_INSYNC, &sh->state);
3137                         else {
3138                                 /* in contrast to the raid5 case we can validate
3139                                  * parity, but still have a failure to write
3140                                  * back
3141                                  */
3142                                 sh->check_state = check_state_compute_result;
3143                                 /* Returning at this point means that we may go
3144                                  * off and bring p and/or q uptodate again so
3145                                  * we make sure to check zero_sum_result again
3146                                  * to verify if p or q need writeback
3147                                  */
3148                         }
3149                 } else {
3150                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3151                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3152                                 /* don't try to repair!! */
3153                                 set_bit(STRIPE_INSYNC, &sh->state);
3154                         else {
3155                                 int *target = &sh->ops.target;
3156
3157                                 sh->ops.target = -1;
3158                                 sh->ops.target2 = -1;
3159                                 sh->check_state = check_state_compute_run;
3160                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3161                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3162                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3163                                         set_bit(R5_Wantcompute,
3164                                                 &sh->dev[pd_idx].flags);
3165                                         *target = pd_idx;
3166                                         target = &sh->ops.target2;
3167                                         s->uptodate++;
3168                                 }
3169                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3170                                         set_bit(R5_Wantcompute,
3171                                                 &sh->dev[qd_idx].flags);
3172                                         *target = qd_idx;
3173                                         s->uptodate++;
3174                                 }
3175                         }
3176                 }
3177                 break;
3178         case check_state_compute_run:
3179                 break;
3180         default:
3181                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3182                        __func__, sh->check_state,
3183                        (unsigned long long) sh->sector);
3184                 BUG();
3185         }
3186 }
3187
3188 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3189 {
3190         int i;
3191
3192         /* We have read all the blocks in this stripe and now we need to
3193          * copy some of them into a target stripe for expand.
3194          */
3195         struct dma_async_tx_descriptor *tx = NULL;
3196         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3197         for (i = 0; i < sh->disks; i++)
3198                 if (i != sh->pd_idx && i != sh->qd_idx) {
3199                         int dd_idx, j;
3200                         struct stripe_head *sh2;
3201                         struct async_submit_ctl submit;
3202
3203                         sector_t bn = compute_blocknr(sh, i, 1);
3204                         sector_t s = raid5_compute_sector(conf, bn, 0,
3205                                                           &dd_idx, NULL);
3206                         sh2 = get_active_stripe(conf, s, 0, 1, 1);
3207                         if (sh2 == NULL)
3208                                 /* so far only the early blocks of this stripe
3209                                  * have been requested.  When later blocks
3210                                  * get requested, we will try again
3211                                  */
3212                                 continue;
3213                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3214                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3215                                 /* must have already done this block */
3216                                 release_stripe(sh2);
3217                                 continue;
3218                         }
3219
3220                         /* place all the copies on one channel */
3221                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3222                         tx = async_memcpy(sh2->dev[dd_idx].page,
3223                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
3224                                           &submit);
3225
3226                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3227                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3228                         for (j = 0; j < conf->raid_disks; j++)
3229                                 if (j != sh2->pd_idx &&
3230                                     j != sh2->qd_idx &&
3231                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
3232                                         break;
3233                         if (j == conf->raid_disks) {
3234                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3235                                 set_bit(STRIPE_HANDLE, &sh2->state);
3236                         }
3237                         release_stripe(sh2);
3238
3239                 }
3240         /* done submitting copies, wait for them to complete */
3241         async_tx_quiesce(&tx);
3242 }
3243
3244 /*
3245  * handle_stripe - do things to a stripe.
3246  *
3247  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3248  * state of various bits to see what needs to be done.
3249  * Possible results:
3250  *    return some read requests which now have data
3251  *    return some write requests which are safely on storage
3252  *    schedule a read on some buffers
3253  *    schedule a write of some buffers
3254  *    return confirmation of parity correctness
3255  *
3256  */
3257
3258 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3259 {
3260         struct r5conf *conf = sh->raid_conf;
3261         int disks = sh->disks;
3262         struct r5dev *dev;
3263         int i;
3264         int do_recovery = 0;
3265
3266         memset(s, 0, sizeof(*s));
3267
3268         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3269         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3270         s->failed_num[0] = -1;
3271         s->failed_num[1] = -1;
3272
3273         /* Now to look around and see what can be done */
3274         rcu_read_lock();
3275         for (i=disks; i--; ) {
3276                 struct md_rdev *rdev;
3277                 sector_t first_bad;
3278                 int bad_sectors;
3279                 int is_bad = 0;
3280
3281                 dev = &sh->dev[i];
3282
3283                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3284                          i, dev->flags,
3285                          dev->toread, dev->towrite, dev->written);
3286                 /* maybe we can reply to a read
3287                  *
3288                  * new wantfill requests are only permitted while
3289                  * ops_complete_biofill is guaranteed to be inactive
3290                  */
3291                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3292                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3293                         set_bit(R5_Wantfill, &dev->flags);
3294
3295                 /* now count some things */
3296                 if (test_bit(R5_LOCKED, &dev->flags))
3297                         s->locked++;
3298                 if (test_bit(R5_UPTODATE, &dev->flags))
3299                         s->uptodate++;
3300                 if (test_bit(R5_Wantcompute, &dev->flags)) {
3301                         s->compute++;
3302                         BUG_ON(s->compute > 2);
3303                 }
3304
3305                 if (test_bit(R5_Wantfill, &dev->flags))
3306                         s->to_fill++;
3307                 else if (dev->toread)
3308                         s->to_read++;
3309                 if (dev->towrite) {
3310                         s->to_write++;
3311                         if (!test_bit(R5_OVERWRITE, &dev->flags))
3312                                 s->non_overwrite++;
3313                 }
3314                 if (dev->written)
3315                         s->written++;
3316                 /* Prefer to use the replacement for reads, but only
3317                  * if it is recovered enough and has no bad blocks.
3318                  */
3319                 rdev = rcu_dereference(conf->disks[i].replacement);
3320                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3321                     rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3322                     !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3323                                  &first_bad, &bad_sectors))
3324                         set_bit(R5_ReadRepl, &dev->flags);
3325                 else {
3326                         if (rdev)
3327                                 set_bit(R5_NeedReplace, &dev->flags);
3328                         rdev = rcu_dereference(conf->disks[i].rdev);
3329                         clear_bit(R5_ReadRepl, &dev->flags);
3330                 }
3331                 if (rdev && test_bit(Faulty, &rdev->flags))
3332                         rdev = NULL;
3333                 if (rdev) {
3334                         is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3335                                              &first_bad, &bad_sectors);
3336                         if (s->blocked_rdev == NULL
3337                             && (test_bit(Blocked, &rdev->flags)
3338                                 || is_bad < 0)) {
3339                                 if (is_bad < 0)
3340                                         set_bit(BlockedBadBlocks,
3341                                                 &rdev->flags);
3342                                 s->blocked_rdev = rdev;
3343                                 atomic_inc(&rdev->nr_pending);
3344                         }
3345                 }
3346                 clear_bit(R5_Insync, &dev->flags);
3347                 if (!rdev)
3348                         /* Not in-sync */;
3349                 else if (is_bad) {
3350                         /* also not in-sync */
3351                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3352                             test_bit(R5_UPTODATE, &dev->flags)) {
3353                                 /* treat as in-sync, but with a read error
3354                                  * which we can now try to correct
3355                                  */
3356                                 set_bit(R5_Insync, &dev->flags);
3357                                 set_bit(R5_ReadError, &dev->flags);
3358                         }
3359                 } else if (test_bit(In_sync, &rdev->flags))
3360                         set_bit(R5_Insync, &dev->flags);
3361                 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3362                         /* in sync if before recovery_offset */
3363                         set_bit(R5_Insync, &dev->flags);
3364                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3365                          test_bit(R5_Expanded, &dev->flags))
3366                         /* If we've reshaped into here, we assume it is Insync.
3367                          * We will shortly update recovery_offset to make
3368                          * it official.
3369                          */
3370                         set_bit(R5_Insync, &dev->flags);
3371
3372                 if (rdev && test_bit(R5_WriteError, &dev->flags)) {
3373                         /* This flag does not apply to '.replacement'
3374                          * only to .rdev, so make sure to check that*/
3375                         struct md_rdev *rdev2 = rcu_dereference(
3376                                 conf->disks[i].rdev);
3377                         if (rdev2 == rdev)
3378                                 clear_bit(R5_Insync, &dev->flags);
3379                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3380                                 s->handle_bad_blocks = 1;
3381                                 atomic_inc(&rdev2->nr_pending);
3382                         } else
3383                                 clear_bit(R5_WriteError, &dev->flags);
3384                 }
3385                 if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
3386                         /* This flag does not apply to '.replacement'
3387                          * only to .rdev, so make sure to check that*/
3388                         struct md_rdev *rdev2 = rcu_dereference(
3389                                 conf->disks[i].rdev);
3390                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3391                                 s->handle_bad_blocks = 1;
3392                                 atomic_inc(&rdev2->nr_pending);
3393                         } else
3394                                 clear_bit(R5_MadeGood, &dev->flags);
3395                 }
3396                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3397                         struct md_rdev *rdev2 = rcu_dereference(
3398                                 conf->disks[i].replacement);
3399                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3400                                 s->handle_bad_blocks = 1;
3401                                 atomic_inc(&rdev2->nr_pending);
3402                         } else
3403                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
3404                 }
3405                 if (!test_bit(R5_Insync, &dev->flags)) {
3406                         /* The ReadError flag will just be confusing now */
3407                         clear_bit(R5_ReadError, &dev->flags);
3408                         clear_bit(R5_ReWrite, &dev->flags);
3409                 }
3410                 if (test_bit(R5_ReadError, &dev->flags))
3411                         clear_bit(R5_Insync, &dev->flags);
3412                 if (!test_bit(R5_Insync, &dev->flags)) {
3413                         if (s->failed < 2)
3414                                 s->failed_num[s->failed] = i;
3415                         s->failed++;
3416                         if (rdev && !test_bit(Faulty, &rdev->flags))
3417                                 do_recovery = 1;
3418                 }
3419         }
3420         if (test_bit(STRIPE_SYNCING, &sh->state)) {
3421                 /* If there is a failed device being replaced,
3422                  *     we must be recovering.
3423                  * else if we are after recovery_cp, we must be syncing
3424                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3425                  * else we can only be replacing
3426                  * sync and recovery both need to read all devices, and so
3427                  * use the same flag.
3428                  */
3429                 if (do_recovery ||
3430                     sh->sector >= conf->mddev->recovery_cp ||
3431                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3432                         s->syncing = 1;
3433                 else
3434                         s->replacing = 1;
3435         }
3436         rcu_read_unlock();
3437 }
3438
3439 static void handle_stripe(struct stripe_head *sh)
3440 {
3441         struct stripe_head_state s;
3442         struct r5conf *conf = sh->raid_conf;
3443         int i;
3444         int prexor;
3445         int disks = sh->disks;
3446         struct r5dev *pdev, *qdev;
3447
3448         clear_bit(STRIPE_HANDLE, &sh->state);
3449         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3450                 /* already being handled, ensure it gets handled
3451                  * again when current action finishes */
3452                 set_bit(STRIPE_HANDLE, &sh->state);
3453                 return;
3454         }
3455
3456         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3457                 spin_lock(&sh->stripe_lock);
3458                 /* Cannot process 'sync' concurrently with 'discard' */
3459                 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
3460                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3461                         set_bit(STRIPE_SYNCING, &sh->state);
3462                         clear_bit(STRIPE_INSYNC, &sh->state);
3463                 }
3464                 spin_unlock(&sh->stripe_lock);
3465         }
3466         clear_bit(STRIPE_DELAYED, &sh->state);
3467
3468         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3469                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3470                (unsigned long long)sh->sector, sh->state,
3471                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3472                sh->check_state, sh->reconstruct_state);
3473
3474         analyse_stripe(sh, &s);
3475
3476         if (s.handle_bad_blocks) {
3477                 set_bit(STRIPE_HANDLE, &sh->state);
3478                 goto finish;
3479         }
3480
3481         if (unlikely(s.blocked_rdev)) {
3482                 if (s.syncing || s.expanding || s.expanded ||
3483                     s.replacing || s.to_write || s.written) {
3484                         set_bit(STRIPE_HANDLE, &sh->state);
3485                         goto finish;
3486                 }
3487                 /* There is nothing for the blocked_rdev to block */
3488                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3489                 s.blocked_rdev = NULL;
3490         }
3491
3492         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3493                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3494                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3495         }
3496
3497         pr_debug("locked=%d uptodate=%d to_read=%d"
3498                " to_write=%d failed=%d failed_num=%d,%d\n",
3499                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3500                s.failed_num[0], s.failed_num[1]);
3501         /* check if the array has lost more than max_degraded devices and,
3502          * if so, some requests might need to be failed.
3503          */
3504         if (s.failed > conf->max_degraded) {
3505                 sh->check_state = 0;
3506                 sh->reconstruct_state = 0;
3507                 if (s.to_read+s.to_write+s.written)
3508                         handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3509                 if (s.syncing + s.replacing)
3510                         handle_failed_sync(conf, sh, &s);
3511         }
3512
3513         /* Now we check to see if any write operations have recently
3514          * completed
3515          */
3516         prexor = 0;
3517         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3518                 prexor = 1;
3519         if (sh->reconstruct_state == reconstruct_state_drain_result ||
3520             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3521                 sh->reconstruct_state = reconstruct_state_idle;
3522
3523                 /* All the 'written' buffers and the parity block are ready to
3524                  * be written back to disk
3525                  */
3526                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
3527                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
3528                 BUG_ON(sh->qd_idx >= 0 &&
3529                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
3530                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
3531                 for (i = disks; i--; ) {
3532                         struct r5dev *dev = &sh->dev[i];
3533                         if (test_bit(R5_LOCKED, &dev->flags) &&
3534                                 (i == sh->pd_idx || i == sh->qd_idx ||
3535                                  dev->written)) {
3536                                 pr_debug("Writing block %d\n", i);
3537                                 set_bit(R5_Wantwrite, &dev->flags);
3538                                 if (prexor)
3539                                         continue;
3540                                 if (!test_bit(R5_Insync, &dev->flags) ||
3541                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
3542                                      s.failed == 0))
3543                                         set_bit(STRIPE_INSYNC, &sh->state);
3544                         }
3545                 }
3546                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3547                         s.dec_preread_active = 1;
3548         }
3549
3550         /*
3551          * might be able to return some write requests if the parity blocks
3552          * are safe, or on a failed drive
3553          */
3554         pdev = &sh->dev[sh->pd_idx];
3555         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3556                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3557         qdev = &sh->dev[sh->qd_idx];
3558         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3559                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3560                 || conf->level < 6;
3561
3562         if (s.written &&
3563             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3564                              && !test_bit(R5_LOCKED, &pdev->flags)
3565                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
3566                                  test_bit(R5_Discard, &pdev->flags))))) &&
3567             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3568                              && !test_bit(R5_LOCKED, &qdev->flags)
3569                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
3570                                  test_bit(R5_Discard, &qdev->flags))))))
3571                 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3572
3573         /* Now we might consider reading some blocks, either to check/generate
3574          * parity, or to satisfy requests
3575          * or to load a block that is being partially written.
3576          */
3577         if (s.to_read || s.non_overwrite
3578             || (conf->level == 6 && s.to_write && s.failed)
3579             || (s.syncing && (s.uptodate + s.compute < disks))
3580             || s.replacing
3581             || s.expanding)
3582                 handle_stripe_fill(sh, &s, disks);
3583
3584         /* Now to consider new write requests and what else, if anything
3585          * should be read.  We do not handle new writes when:
3586          * 1/ A 'write' operation (copy+xor) is already in flight.
3587          * 2/ A 'check' operation is in flight, as it may clobber the parity
3588          *    block.
3589          */
3590         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3591                 handle_stripe_dirtying(conf, sh, &s, disks);
3592
3593         /* maybe we need to check and possibly fix the parity for this stripe
3594          * Any reads will already have been scheduled, so we just see if enough
3595          * data is available.  The parity check is held off while parity
3596          * dependent operations are in flight.
3597          */
3598         if (sh->check_state ||
3599             (s.syncing && s.locked == 0 &&
3600              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3601              !test_bit(STRIPE_INSYNC, &sh->state))) {
3602                 if (conf->level == 6)
3603                         handle_parity_checks6(conf, sh, &s, disks);
3604                 else
3605                         handle_parity_checks5(conf, sh, &s, disks);
3606         }
3607
3608         if (s.replacing && s.locked == 0
3609             && !test_bit(STRIPE_INSYNC, &sh->state)) {
3610                 /* Write out to replacement devices where possible */
3611                 for (i = 0; i < conf->raid_disks; i++)
3612                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags) &&
3613                             test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3614                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
3615                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
3616                                 s.locked++;
3617                         }
3618                 set_bit(STRIPE_INSYNC, &sh->state);
3619         }
3620         if ((s.syncing || s.replacing) && s.locked == 0 &&
3621             test_bit(STRIPE_INSYNC, &sh->state)) {
3622                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3623                 clear_bit(STRIPE_SYNCING, &sh->state);
3624                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3625                         wake_up(&conf->wait_for_overlap);
3626         }
3627
3628         /* If the failed drives are just a ReadError, then we might need
3629          * to progress the repair/check process
3630          */
3631         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3632                 for (i = 0; i < s.failed; i++) {
3633                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
3634                         if (test_bit(R5_ReadError, &dev->flags)
3635                             && !test_bit(R5_LOCKED, &dev->flags)
3636                             && test_bit(R5_UPTODATE, &dev->flags)
3637                                 ) {
3638                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3639                                         set_bit(R5_Wantwrite, &dev->flags);
3640                                         set_bit(R5_ReWrite, &dev->flags);
3641                                         set_bit(R5_LOCKED, &dev->flags);
3642                                         s.locked++;
3643                                 } else {
3644                                         /* let's read it back */
3645                                         set_bit(R5_Wantread, &dev->flags);
3646                                         set_bit(R5_LOCKED, &dev->flags);
3647                                         s.locked++;
3648                                 }
3649                         }
3650                 }
3651
3652
3653         /* Finish reconstruct operations initiated by the expansion process */
3654         if (sh->reconstruct_state == reconstruct_state_result) {
3655                 struct stripe_head *sh_src
3656                         = get_active_stripe(conf, sh->sector, 1, 1, 1);
3657                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3658                         /* sh cannot be written until sh_src has been read.
3659                          * so arrange for sh to be delayed a little
3660                          */
3661                         set_bit(STRIPE_DELAYED, &sh->state);
3662                         set_bit(STRIPE_HANDLE, &sh->state);
3663                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3664                                               &sh_src->state))
3665                                 atomic_inc(&conf->preread_active_stripes);
3666                         release_stripe(sh_src);
3667                         goto finish;
3668                 }
3669                 if (sh_src)
3670                         release_stripe(sh_src);
3671
3672                 sh->reconstruct_state = reconstruct_state_idle;
3673                 clear_bit(STRIPE_EXPANDING, &sh->state);
3674                 for (i = conf->raid_disks; i--; ) {
3675                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3676                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3677                         s.locked++;
3678                 }
3679         }
3680
3681         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3682             !sh->reconstruct_state) {
3683                 /* Need to write out all blocks after computing parity */
3684                 sh->disks = conf->raid_disks;
3685                 stripe_set_idx(sh->sector, conf, 0, sh);
3686                 schedule_reconstruction(sh, &s, 1, 1);
3687         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3688                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3689                 atomic_dec(&conf->reshape_stripes);
3690                 wake_up(&conf->wait_for_overlap);
3691                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3692         }
3693
3694         if (s.expanding && s.locked == 0 &&
3695             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3696                 handle_stripe_expansion(conf, sh);
3697
3698 finish:
3699         /* wait for this device to become unblocked */
3700         if (unlikely(s.blocked_rdev)) {
3701                 if (conf->mddev->external)
3702                         md_wait_for_blocked_rdev(s.blocked_rdev,
3703                                                  conf->mddev);
3704                 else
3705                         /* Internal metadata will immediately
3706                          * be written by raid5d, so we don't
3707                          * need to wait here.
3708                          */
3709                         rdev_dec_pending(s.blocked_rdev,
3710                                          conf->mddev);
3711         }
3712
3713         if (s.handle_bad_blocks)
3714                 for (i = disks; i--; ) {
3715                         struct md_rdev *rdev;
3716                         struct r5dev *dev = &sh->dev[i];
3717                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3718                                 /* We own a safe reference to the rdev */
3719                                 rdev = conf->disks[i].rdev;
3720                                 if (!rdev_set_badblocks(rdev, sh->sector,
3721                                                         STRIPE_SECTORS, 0))
3722                                         md_error(conf->mddev, rdev);
3723                                 rdev_dec_pending(rdev, conf->mddev);
3724                         }
3725                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3726                                 rdev = conf->disks[i].rdev;
3727                                 rdev_clear_badblocks(rdev, sh->sector,
3728                                                      STRIPE_SECTORS, 0);
3729                                 rdev_dec_pending(rdev, conf->mddev);
3730                         }
3731                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
3732                                 rdev = conf->disks[i].replacement;
3733                                 if (!rdev)
3734                                         /* rdev have been moved down */
3735                                         rdev = conf->disks[i].rdev;
3736                                 rdev_clear_badblocks(rdev, sh->sector,
3737                                                      STRIPE_SECTORS, 0);
3738                                 rdev_dec_pending(rdev, conf->mddev);
3739                         }
3740                 }
3741
3742         if (s.ops_request)
3743                 raid_run_ops(sh, s.ops_request);
3744
3745         ops_run_io(sh, &s);
3746
3747         if (s.dec_preread_active) {
3748                 /* We delay this until after ops_run_io so that if make_request
3749                  * is waiting on a flush, it won't continue until the writes
3750                  * have actually been submitted.
3751                  */
3752                 atomic_dec(&conf->preread_active_stripes);
3753                 if (atomic_read(&conf->preread_active_stripes) <
3754                     IO_THRESHOLD)
3755                         md_wakeup_thread(conf->mddev->thread);
3756         }
3757
3758         return_io(s.return_bi);
3759
3760         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3761 }
3762
3763 static void raid5_activate_delayed(struct r5conf *conf)
3764 {
3765         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3766                 while (!list_empty(&conf->delayed_list)) {
3767                         struct list_head *l = conf->delayed_list.next;
3768                         struct stripe_head *sh;
3769                         sh = list_entry(l, struct stripe_head, lru);
3770                         list_del_init(l);
3771                         clear_bit(STRIPE_DELAYED, &sh->state);
3772                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3773                                 atomic_inc(&conf->preread_active_stripes);
3774                         list_add_tail(&sh->lru, &conf->hold_list);
3775                 }
3776         }
3777 }
3778
3779 static void activate_bit_delay(struct r5conf *conf)
3780 {
3781         /* device_lock is held */
3782         struct list_head head;
3783         list_add(&head, &conf->bitmap_list);
3784         list_del_init(&conf->bitmap_list);
3785         while (!list_empty(&head)) {
3786                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3787                 list_del_init(&sh->lru);
3788                 atomic_inc(&sh->count);
3789                 __release_stripe(conf, sh);
3790         }
3791 }
3792
3793 int md_raid5_congested(struct mddev *mddev, int bits)
3794 {
3795         struct r5conf *conf = mddev->private;
3796
3797         /* No difference between reads and writes.  Just check
3798          * how busy the stripe_cache is
3799          */
3800
3801         if (conf->inactive_blocked)
3802                 return 1;
3803         if (conf->quiesce)
3804                 return 1;
3805         if (list_empty_careful(&conf->inactive_list))
3806                 return 1;
3807
3808         return 0;
3809 }
3810 EXPORT_SYMBOL_GPL(md_raid5_congested);
3811
3812 static int raid5_congested(void *data, int bits)
3813 {
3814         struct mddev *mddev = data;
3815
3816         return mddev_congested(mddev, bits) ||
3817                 md_raid5_congested(mddev, bits);
3818 }
3819
3820 /* We want read requests to align with chunks where possible,
3821  * but write requests don't need to.
3822  */
3823 static int raid5_mergeable_bvec(struct request_queue *q,
3824                                 struct bvec_merge_data *bvm,
3825                                 struct bio_vec *biovec)
3826 {
3827         struct mddev *mddev = q->queuedata;
3828         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3829         int max;
3830         unsigned int chunk_sectors = mddev->chunk_sectors;
3831         unsigned int bio_sectors = bvm->bi_size >> 9;
3832
3833         if ((bvm->bi_rw & 1) == WRITE)
3834                 return biovec->bv_len; /* always allow writes to be mergeable */
3835
3836         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3837                 chunk_sectors = mddev->new_chunk_sectors;
3838         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3839         if (max < 0) max = 0;
3840         if (max <= biovec->bv_len && bio_sectors == 0)
3841                 return biovec->bv_len;
3842         else
3843                 return max;
3844 }
3845
3846
3847 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3848 {
3849         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3850         unsigned int chunk_sectors = mddev->chunk_sectors;
3851         unsigned int bio_sectors = bio_sectors(bio);
3852
3853         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3854                 chunk_sectors = mddev->new_chunk_sectors;
3855         return  chunk_sectors >=
3856                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3857 }
3858
3859 /*
3860  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3861  *  later sampled by raid5d.
3862  */
3863 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3864 {
3865         unsigned long flags;
3866
3867         spin_lock_irqsave(&conf->device_lock, flags);
3868
3869         bi->bi_next = conf->retry_read_aligned_list;
3870         conf->retry_read_aligned_list = bi;
3871
3872         spin_unlock_irqrestore(&conf->device_lock, flags);
3873         md_wakeup_thread(conf->mddev->thread);
3874 }
3875
3876
3877 static struct bio *remove_bio_from_retry(struct r5conf *conf)
3878 {
3879         struct bio *bi;
3880
3881         bi = conf->retry_read_aligned;
3882         if (bi) {
3883                 conf->retry_read_aligned = NULL;
3884                 return bi;
3885         }
3886         bi = conf->retry_read_aligned_list;
3887         if(bi) {
3888                 conf->retry_read_aligned_list = bi->bi_next;
3889                 bi->bi_next = NULL;
3890                 /*
3891                  * this sets the active strip count to 1 and the processed
3892                  * strip count to zero (upper 8 bits)
3893                  */
3894                 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
3895         }
3896
3897         return bi;
3898 }
3899
3900
3901 /*
3902  *  The "raid5_align_endio" should check if the read succeeded and if it
3903  *  did, call bio_endio on the original bio (having bio_put the new bio
3904  *  first).
3905  *  If the read failed..
3906  */
3907 static void raid5_align_endio(struct bio *bi, int error)
3908 {
3909         struct bio* raid_bi  = bi->bi_private;
3910         struct mddev *mddev;
3911         struct r5conf *conf;
3912         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3913         struct md_rdev *rdev;
3914
3915         bio_put(bi);
3916
3917         rdev = (void*)raid_bi->bi_next;
3918         raid_bi->bi_next = NULL;
3919         mddev = rdev->mddev;
3920         conf = mddev->private;
3921
3922         rdev_dec_pending(rdev, conf->mddev);
3923
3924         if (!error && uptodate) {
3925                 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
3926                                          raid_bi, 0);
3927                 bio_endio(raid_bi, 0);
3928                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3929                         wake_up(&conf->wait_for_stripe);
3930                 return;
3931         }
3932
3933
3934         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3935
3936         add_bio_to_retry(raid_bi, conf);
3937 }
3938
3939 static int bio_fits_rdev(struct bio *bi)
3940 {
3941         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3942
3943         if (bio_sectors(bi) > queue_max_sectors(q))
3944                 return 0;
3945         blk_recount_segments(q, bi);
3946         if (bi->bi_phys_segments > queue_max_segments(q))
3947                 return 0;
3948
3949         if (q->merge_bvec_fn)
3950                 /* it's too hard to apply the merge_bvec_fn at this stage,
3951                  * just just give up
3952                  */
3953                 return 0;
3954
3955         return 1;
3956 }
3957
3958
3959 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3960 {
3961         struct r5conf *conf = mddev->private;
3962         int dd_idx;
3963         struct bio* align_bi;
3964         struct md_rdev *rdev;
3965         sector_t end_sector;
3966
3967         if (!in_chunk_boundary(mddev, raid_bio)) {
3968                 pr_debug("chunk_aligned_read : non aligned\n");
3969                 return 0;
3970         }
3971         /*
3972          * use bio_clone_mddev to make a copy of the bio
3973          */
3974         align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3975         if (!align_bi)
3976                 return 0;
3977         /*
3978          *   set bi_end_io to a new function, and set bi_private to the
3979          *     original bio.
3980          */
3981         align_bi->bi_end_io  = raid5_align_endio;
3982         align_bi->bi_private = raid_bio;
3983         /*
3984          *      compute position
3985          */
3986         align_bi->bi_sector =  raid5_compute_sector(conf, raid_bio->bi_sector,
3987                                                     0,
3988                                                     &dd_idx, NULL);
3989
3990         end_sector = bio_end_sector(align_bi);
3991         rcu_read_lock();
3992         rdev = rcu_dereference(conf->disks[dd_idx].replacement);
3993         if (!rdev || test_bit(Faulty, &rdev->flags) ||
3994             rdev->recovery_offset < end_sector) {
3995                 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3996                 if (rdev &&
3997                     (test_bit(Faulty, &rdev->flags) ||
3998                     !(test_bit(In_sync, &rdev->flags) ||
3999                       rdev->recovery_offset >= end_sector)))
4000                         rdev = NULL;
4001         }
4002         if (rdev) {
4003                 sector_t first_bad;
4004                 int bad_sectors;
4005
4006                 atomic_inc(&rdev->nr_pending);
4007                 rcu_read_unlock();
4008                 raid_bio->bi_next = (void*)rdev;
4009                 align_bi->bi_bdev =  rdev->bdev;
4010                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
4011
4012                 if (!bio_fits_rdev(align_bi) ||
4013                     is_badblock(rdev, align_bi->bi_sector, bio_sectors(align_bi),
4014                                 &first_bad, &bad_sectors)) {
4015                         /* too big in some way, or has a known bad block */
4016                         bio_put(align_bi);
4017                         rdev_dec_pending(rdev, mddev);
4018                         return 0;
4019                 }
4020
4021                 /* No reshape active, so we can trust rdev->data_offset */
4022                 align_bi->bi_sector += rdev->data_offset;
4023
4024                 spin_lock_irq(&conf->device_lock);
4025                 wait_event_lock_irq(conf->wait_for_stripe,
4026                                     conf->quiesce == 0,
4027                                     conf->device_lock);
4028                 atomic_inc(&conf->active_aligned_reads);
4029                 spin_unlock_irq(&conf->device_lock);
4030
4031                 if (mddev->gendisk)
4032                         trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4033                                               align_bi, disk_devt(mddev->gendisk),
4034                                               raid_bio->bi_sector);
4035                 generic_make_request(align_bi);
4036                 return 1;
4037         } else {
4038                 rcu_read_unlock();
4039                 bio_put(align_bi);
4040                 return 0;
4041         }
4042 }
4043
4044 /* __get_priority_stripe - get the next stripe to process
4045  *
4046  * Full stripe writes are allowed to pass preread active stripes up until
4047  * the bypass_threshold is exceeded.  In general the bypass_count
4048  * increments when the handle_list is handled before the hold_list; however, it
4049  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4050  * stripe with in flight i/o.  The bypass_count will be reset when the
4051  * head of the hold_list has changed, i.e. the head was promoted to the
4052  * handle_list.
4053  */
4054 static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
4055 {
4056         struct stripe_head *sh;
4057
4058         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4059                   __func__,
4060                   list_empty(&conf->handle_list) ? "empty" : "busy",
4061                   list_empty(&conf->hold_list) ? "empty" : "busy",
4062                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
4063
4064         if (!list_empty(&conf->handle_list)) {
4065                 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
4066
4067                 if (list_empty(&conf->hold_list))
4068                         conf->bypass_count = 0;
4069                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4070                         if (conf->hold_list.next == conf->last_hold)
4071                                 conf->bypass_count++;
4072                         else {
4073                                 conf->last_hold = conf->hold_list.next;
4074                                 conf->bypass_count -= conf->bypass_threshold;
4075                                 if (conf->bypass_count < 0)
4076                                         conf->bypass_count = 0;
4077                         }
4078                 }
4079         } else if (!list_empty(&conf->hold_list) &&
4080                    ((conf->bypass_threshold &&
4081                      conf->bypass_count > conf->bypass_threshold) ||
4082                     atomic_read(&conf->pending_full_writes) == 0)) {
4083                 sh = list_entry(conf->hold_list.next,
4084                                 typeof(*sh), lru);
4085                 conf->bypass_count -= conf->bypass_threshold;
4086                 if (conf->bypass_count < 0)
4087                         conf->bypass_count = 0;
4088         } else
4089                 return NULL;
4090
4091         list_del_init(&sh->lru);
4092         atomic_inc(&sh->count);
4093         BUG_ON(atomic_read(&sh->count) != 1);
4094         return sh;
4095 }
4096
4097 struct raid5_plug_cb {
4098         struct blk_plug_cb      cb;
4099         struct list_head        list;
4100 };
4101
4102 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4103 {
4104         struct raid5_plug_cb *cb = container_of(
4105                 blk_cb, struct raid5_plug_cb, cb);
4106         struct stripe_head *sh;
4107         struct mddev *mddev = cb->cb.data;
4108         struct r5conf *conf = mddev->private;
4109         int cnt = 0;
4110
4111         if (cb->list.next && !list_empty(&cb->list)) {
4112                 spin_lock_irq(&conf->device_lock);
4113                 while (!list_empty(&cb->list)) {
4114                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
4115                         list_del_init(&sh->lru);
4116                         /*
4117                          * avoid race release_stripe_plug() sees
4118                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
4119                          * is still in our list
4120                          */
4121                         smp_mb__before_clear_bit();
4122                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4123                         __release_stripe(conf, sh);
4124                         cnt++;
4125                 }
4126                 spin_unlock_irq(&conf->device_lock);
4127         }
4128         if (mddev->queue)
4129                 trace_block_unplug(mddev->queue, cnt, !from_schedule);
4130         kfree(cb);
4131 }
4132
4133 static void release_stripe_plug(struct mddev *mddev,
4134                                 struct stripe_head *sh)
4135 {
4136         struct blk_plug_cb *blk_cb = blk_check_plugged(
4137                 raid5_unplug, mddev,
4138                 sizeof(struct raid5_plug_cb));
4139         struct raid5_plug_cb *cb;
4140
4141         if (!blk_cb) {
4142                 release_stripe(sh);
4143                 return;
4144         }
4145
4146         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4147
4148         if (cb->list.next == NULL)
4149                 INIT_LIST_HEAD(&cb->list);
4150
4151         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4152                 list_add_tail(&sh->lru, &cb->list);
4153         else
4154                 release_stripe(sh);
4155 }
4156
4157 static void make_discard_request(struct mddev *mddev, struct bio *bi)
4158 {
4159         struct r5conf *conf = mddev->private;
4160         sector_t logical_sector, last_sector;
4161         struct stripe_head *sh;
4162         int remaining;
4163         int stripe_sectors;
4164
4165         if (mddev->reshape_position != MaxSector)
4166                 /* Skip discard while reshape is happening */
4167                 return;
4168
4169         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4170         last_sector = bi->bi_sector + (bi->bi_size>>9);
4171
4172         bi->bi_next = NULL;
4173         bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4174
4175         stripe_sectors = conf->chunk_sectors *
4176                 (conf->raid_disks - conf->max_degraded);
4177         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4178                                                stripe_sectors);
4179         sector_div(last_sector, stripe_sectors);
4180
4181         logical_sector *= conf->chunk_sectors;
4182         last_sector *= conf->chunk_sectors;
4183
4184         for (; logical_sector < last_sector;
4185              logical_sector += STRIPE_SECTORS) {
4186                 DEFINE_WAIT(w);
4187                 int d;
4188         again:
4189                 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4190                 prepare_to_wait(&conf->wait_for_overlap, &w,
4191                                 TASK_UNINTERRUPTIBLE);
4192                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4193                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4194                         release_stripe(sh);
4195                         schedule();
4196                         goto again;
4197                 }
4198                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4199                 spin_lock_irq(&sh->stripe_lock);
4200                 for (d = 0; d < conf->raid_disks; d++) {
4201                         if (d == sh->pd_idx || d == sh->qd_idx)
4202                                 continue;
4203                         if (sh->dev[d].towrite || sh->dev[d].toread) {
4204                                 set_bit(R5_Overlap, &sh->dev[d].flags);
4205                                 spin_unlock_irq(&sh->stripe_lock);
4206                                 release_stripe(sh);
4207                                 schedule();
4208                                 goto again;
4209                         }
4210                 }
4211                 set_bit(STRIPE_DISCARD, &sh->state);
4212                 finish_wait(&conf->wait_for_overlap, &w);
4213                 for (d = 0; d < conf->raid_disks; d++) {
4214                         if (d == sh->pd_idx || d == sh->qd_idx)
4215                                 continue;
4216                         sh->dev[d].towrite = bi;
4217                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4218                         raid5_inc_bi_active_stripes(bi);
4219                 }
4220                 spin_unlock_irq(&sh->stripe_lock);
4221                 if (conf->mddev->bitmap) {
4222                         for (d = 0;
4223                              d < conf->raid_disks - conf->max_degraded;
4224                              d++)
4225                                 bitmap_startwrite(mddev->bitmap,
4226                                                   sh->sector,
4227                                                   STRIPE_SECTORS,
4228                                                   0);
4229                         sh->bm_seq = conf->seq_flush + 1;
4230                         set_bit(STRIPE_BIT_DELAY, &sh->state);
4231                 }
4232
4233                 set_bit(STRIPE_HANDLE, &sh->state);
4234                 clear_bit(STRIPE_DELAYED, &sh->state);
4235                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4236                         atomic_inc(&conf->preread_active_stripes);
4237                 release_stripe_plug(mddev, sh);
4238         }
4239
4240         remaining = raid5_dec_bi_active_stripes(bi);
4241         if (remaining == 0) {
4242                 md_write_end(mddev);
4243                 bio_endio(bi, 0);
4244         }
4245 }
4246
4247 static void make_request(struct mddev *mddev, struct bio * bi)
4248 {
4249         struct r5conf *conf = mddev->private;
4250         int dd_idx;
4251         sector_t new_sector;
4252         sector_t logical_sector, last_sector;
4253         struct stripe_head *sh;
4254         const int rw = bio_data_dir(bi);
4255         int remaining;
4256
4257         if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4258                 md_flush_request(mddev, bi);
4259                 return;
4260         }
4261
4262         md_write_start(mddev, bi);
4263
4264         if (rw == READ &&
4265              mddev->reshape_position == MaxSector &&
4266              chunk_aligned_read(mddev,bi))
4267                 return;
4268
4269         if (unlikely(bi->bi_rw & REQ_DISCARD)) {
4270                 make_discard_request(mddev, bi);
4271                 return;
4272         }
4273
4274         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4275         last_sector = bio_end_sector(bi);
4276         bi->bi_next = NULL;
4277         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
4278
4279         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4280                 DEFINE_WAIT(w);
4281                 int previous;
4282
4283         retry:
4284                 previous = 0;
4285                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4286                 if (unlikely(conf->reshape_progress != MaxSector)) {
4287                         /* spinlock is needed as reshape_progress may be
4288                          * 64bit on a 32bit platform, and so it might be
4289                          * possible to see a half-updated value
4290                          * Of course reshape_progress could change after
4291                          * the lock is dropped, so once we get a reference
4292                          * to the stripe that we think it is, we will have
4293                          * to check again.
4294                          */
4295                         spin_lock_irq(&conf->device_lock);
4296                         if (mddev->reshape_backwards
4297                             ? logical_sector < conf->reshape_progress
4298                             : logical_sector >= conf->reshape_progress) {
4299                                 previous = 1;
4300                         } else {
4301                                 if (mddev->reshape_backwards
4302                                     ? logical_sector < conf->reshape_safe
4303                                     : logical_sector >= conf->reshape_safe) {
4304                                         spin_unlock_irq(&conf->device_lock);
4305                                         schedule();
4306                                         goto retry;
4307                                 }
4308                         }
4309                         spin_unlock_irq(&conf->device_lock);
4310                 }
4311
4312                 new_sector = raid5_compute_sector(conf, logical_sector,
4313                                                   previous,
4314                                                   &dd_idx, NULL);
4315                 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4316                         (unsigned long long)new_sector,
4317                         (unsigned long long)logical_sector);
4318
4319                 sh = get_active_stripe(conf, new_sector, previous,
4320                                        (bi->bi_rw&RWA_MASK), 0);
4321                 if (sh) {
4322                         if (unlikely(previous)) {
4323                                 /* expansion might have moved on while waiting for a
4324                                  * stripe, so we must do the range check again.
4325                                  * Expansion could still move past after this
4326                                  * test, but as we are holding a reference to
4327                                  * 'sh', we know that if that happens,
4328                                  *  STRIPE_EXPANDING will get set and the expansion
4329                                  * won't proceed until we finish with the stripe.
4330                                  */
4331                                 int must_retry = 0;
4332                                 spin_lock_irq(&conf->device_lock);
4333                                 if (mddev->reshape_backwards
4334                                     ? logical_sector >= conf->reshape_progress
4335                                     : logical_sector < conf->reshape_progress)
4336                                         /* mismatch, need to try again */
4337                                         must_retry = 1;
4338                                 spin_unlock_irq(&conf->device_lock);
4339                                 if (must_retry) {
4340                                         release_stripe(sh);
4341                                         schedule();
4342                                         goto retry;
4343                                 }
4344                         }
4345
4346                         if (rw == WRITE &&
4347                             logical_sector >= mddev->suspend_lo &&
4348                             logical_sector < mddev->suspend_hi) {
4349                                 release_stripe(sh);
4350                                 /* As the suspend_* range is controlled by
4351                                  * userspace, we want an interruptible
4352                                  * wait.
4353                                  */
4354                                 flush_signals(current);
4355                                 prepare_to_wait(&conf->wait_for_overlap,
4356                                                 &w, TASK_INTERRUPTIBLE);
4357                                 if (logical_sector >= mddev->suspend_lo &&
4358                                     logical_sector < mddev->suspend_hi)
4359                                         schedule();
4360                                 goto retry;
4361                         }
4362
4363                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4364                             !add_stripe_bio(sh, bi, dd_idx, rw)) {
4365                                 /* Stripe is busy expanding or
4366                                  * add failed due to overlap.  Flush everything
4367                                  * and wait a while
4368                                  */
4369                                 md_wakeup_thread(mddev->thread);
4370                                 release_stripe(sh);
4371                                 schedule();
4372                                 goto retry;
4373                         }
4374                         finish_wait(&conf->wait_for_overlap, &w);
4375                         set_bit(STRIPE_HANDLE, &sh->state);
4376                         clear_bit(STRIPE_DELAYED, &sh->state);
4377                         if ((bi->bi_rw & REQ_SYNC) &&
4378                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4379                                 atomic_inc(&conf->preread_active_stripes);
4380                         release_stripe_plug(mddev, sh);
4381                 } else {
4382                         /* cannot get stripe for read-ahead, just give-up */
4383                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
4384                         finish_wait(&conf->wait_for_overlap, &w);
4385                         break;
4386                 }
4387         }
4388
4389         remaining = raid5_dec_bi_active_stripes(bi);
4390         if (remaining == 0) {
4391
4392                 if ( rw == WRITE )
4393                         md_write_end(mddev);
4394
4395                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
4396                                          bi, 0);
4397                 bio_endio(bi, 0);
4398         }
4399 }
4400
4401 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4402
4403 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4404 {
4405         /* reshaping is quite different to recovery/resync so it is
4406          * handled quite separately ... here.
4407          *
4408          * On each call to sync_request, we gather one chunk worth of
4409          * destination stripes and flag them as expanding.
4410          * Then we find all the source stripes and request reads.
4411          * As the reads complete, handle_stripe will copy the data
4412          * into the destination stripe and release that stripe.
4413          */
4414         struct r5conf *conf = mddev->private;
4415         struct stripe_head *sh;
4416         sector_t first_sector, last_sector;
4417         int raid_disks = conf->previous_raid_disks;
4418         int data_disks = raid_disks - conf->max_degraded;
4419         int new_data_disks = conf->raid_disks - conf->max_degraded;
4420         int i;
4421         int dd_idx;
4422         sector_t writepos, readpos, safepos;
4423         sector_t stripe_addr;
4424         int reshape_sectors;
4425         struct list_head stripes;
4426
4427         if (sector_nr == 0) {
4428                 /* If restarting in the middle, skip the initial sectors */
4429                 if (mddev->reshape_backwards &&
4430                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4431                         sector_nr = raid5_size(mddev, 0, 0)
4432                                 - conf->reshape_progress;
4433                 } else if (!mddev->reshape_backwards &&
4434                            conf->reshape_progress > 0)
4435                         sector_nr = conf->reshape_progress;
4436                 sector_div(sector_nr, new_data_disks);
4437                 if (sector_nr) {
4438                         mddev->curr_resync_completed = sector_nr;
4439                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4440                         *skipped = 1;
4441                         return sector_nr;
4442                 }
4443         }
4444
4445         /* We need to process a full chunk at a time.
4446          * If old and new chunk sizes differ, we need to process the
4447          * largest of these
4448          */
4449         if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4450                 reshape_sectors = mddev->new_chunk_sectors;
4451         else
4452                 reshape_sectors = mddev->chunk_sectors;
4453
4454         /* We update the metadata at least every 10 seconds, or when
4455          * the data about to be copied would over-write the source of
4456          * the data at the front of the range.  i.e. one new_stripe
4457          * along from reshape_progress new_maps to after where
4458          * reshape_safe old_maps to
4459          */
4460         writepos = conf->reshape_progress;
4461         sector_div(writepos, new_data_disks);
4462         readpos = conf->reshape_progress;
4463         sector_div(readpos, data_disks);
4464         safepos = conf->reshape_safe;
4465         sector_div(safepos, data_disks);
4466         if (mddev->reshape_backwards) {
4467                 writepos -= min_t(sector_t, reshape_sectors, writepos);
4468                 readpos += reshape_sectors;
4469                 safepos += reshape_sectors;
4470         } else {
4471                 writepos += reshape_sectors;
4472                 readpos -= min_t(sector_t, reshape_sectors, readpos);
4473                 safepos -= min_t(sector_t, reshape_sectors, safepos);
4474         }
4475
4476         /* Having calculated the 'writepos' possibly use it
4477          * to set 'stripe_addr' which is where we will write to.
4478          */
4479         if (mddev->reshape_backwards) {
4480                 BUG_ON(conf->reshape_progress == 0);
4481                 stripe_addr = writepos;
4482                 BUG_ON((mddev->dev_sectors &
4483                         ~((sector_t)reshape_sectors - 1))
4484                        - reshape_sectors - stripe_addr
4485                        != sector_nr);
4486         } else {
4487                 BUG_ON(writepos != sector_nr + reshape_sectors);
4488                 stripe_addr = sector_nr;
4489         }
4490
4491         /* 'writepos' is the most advanced device address we might write.
4492          * 'readpos' is the least advanced device address we might read.
4493          * 'safepos' is the least address recorded in the metadata as having
4494          *     been reshaped.
4495          * If there is a min_offset_diff, these are adjusted either by
4496          * increasing the safepos/readpos if diff is negative, or
4497          * increasing writepos if diff is positive.
4498          * If 'readpos' is then behind 'writepos', there is no way that we can
4499          * ensure safety in the face of a crash - that must be done by userspace
4500          * making a backup of the data.  So in that case there is no particular
4501          * rush to update metadata.
4502          * Otherwise if 'safepos' is behind 'writepos', then we really need to
4503          * update the metadata to advance 'safepos' to match 'readpos' so that
4504          * we can be safe in the event of a crash.
4505          * So we insist on updating metadata if safepos is behind writepos and
4506          * readpos is beyond writepos.
4507          * In any case, update the metadata every 10 seconds.
4508          * Maybe that number should be configurable, but I'm not sure it is
4509          * worth it.... maybe it could be a multiple of safemode_delay???
4510          */
4511         if (conf->min_offset_diff < 0) {
4512                 safepos += -conf->min_offset_diff;
4513                 readpos += -conf->min_offset_diff;
4514         } else
4515                 writepos += conf->min_offset_diff;
4516
4517         if ((mddev->reshape_backwards
4518              ? (safepos > writepos && readpos < writepos)
4519              : (safepos < writepos && readpos > writepos)) ||
4520             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4521                 /* Cannot proceed until we've updated the superblock... */
4522                 wait_event(conf->wait_for_overlap,
4523                            atomic_read(&conf->reshape_stripes)==0);
4524                 mddev->reshape_position = conf->reshape_progress;
4525                 mddev->curr_resync_completed = sector_nr;
4526                 conf->reshape_checkpoint = jiffies;
4527                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4528                 md_wakeup_thread(mddev->thread);
4529                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4530                            kthread_should_stop());
4531                 spin_lock_irq(&conf->device_lock);
4532                 conf->reshape_safe = mddev->reshape_position;
4533                 spin_unlock_irq(&conf->device_lock);
4534                 wake_up(&conf->wait_for_overlap);
4535                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4536         }
4537
4538         INIT_LIST_HEAD(&stripes);
4539         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4540                 int j;
4541                 int skipped_disk = 0;
4542                 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4543                 set_bit(STRIPE_EXPANDING, &sh->state);
4544                 atomic_inc(&conf->reshape_stripes);
4545                 /* If any of this stripe is beyond the end of the old
4546                  * array, then we need to zero those blocks
4547                  */
4548                 for (j=sh->disks; j--;) {
4549                         sector_t s;
4550                         if (j == sh->pd_idx)
4551                                 continue;
4552                         if (conf->level == 6 &&
4553                             j == sh->qd_idx)
4554                                 continue;
4555                         s = compute_blocknr(sh, j, 0);
4556                         if (s < raid5_size(mddev, 0, 0)) {
4557                                 skipped_disk = 1;
4558                                 continue;
4559                         }
4560                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4561                         set_bit(R5_Expanded, &sh->dev[j].flags);
4562                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
4563                 }
4564                 if (!skipped_disk) {
4565                         set_bit(STRIPE_EXPAND_READY, &sh->state);
4566                         set_bit(STRIPE_HANDLE, &sh->state);
4567                 }
4568                 list_add(&sh->lru, &stripes);
4569         }
4570         spin_lock_irq(&conf->device_lock);
4571         if (mddev->reshape_backwards)
4572                 conf->reshape_progress -= reshape_sectors * new_data_disks;
4573         else
4574                 conf->reshape_progress += reshape_sectors * new_data_disks;
4575         spin_unlock_irq(&conf->device_lock);
4576         /* Ok, those stripe are ready. We can start scheduling
4577          * reads on the source stripes.
4578          * The source stripes are determined by mapping the first and last
4579          * block on the destination stripes.
4580          */
4581         first_sector =
4582                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4583                                      1, &dd_idx, NULL);
4584         last_sector =
4585                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4586                                             * new_data_disks - 1),
4587                                      1, &dd_idx, NULL);
4588         if (last_sector >= mddev->dev_sectors)
4589                 last_sector = mddev->dev_sectors - 1;
4590         while (first_sector <= last_sector) {
4591                 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4592                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4593                 set_bit(STRIPE_HANDLE, &sh->state);
4594                 release_stripe(sh);
4595                 first_sector += STRIPE_SECTORS;
4596         }
4597         /* Now that the sources are clearly marked, we can release
4598          * the destination stripes
4599          */
4600         while (!list_empty(&stripes)) {
4601                 sh = list_entry(stripes.next, struct stripe_head, lru);
4602                 list_del_init(&sh->lru);
4603                 release_stripe(sh);
4604         }
4605         /* If this takes us to the resync_max point where we have to pause,
4606          * then we need to write out the superblock.
4607          */
4608         sector_nr += reshape_sectors;
4609         if ((sector_nr - mddev->curr_resync_completed) * 2
4610             >= mddev->resync_max - mddev->curr_resync_completed) {
4611                 /* Cannot proceed until we've updated the superblock... */
4612                 wait_event(conf->wait_for_overlap,
4613                            atomic_read(&conf->reshape_stripes) == 0);
4614                 mddev->reshape_position = conf->reshape_progress;
4615                 mddev->curr_resync_completed = sector_nr;
4616                 conf->reshape_checkpoint = jiffies;
4617                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4618                 md_wakeup_thread(mddev->thread);
4619                 wait_event(mddev->sb_wait,
4620                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4621                            || kthread_should_stop());
4622                 spin_lock_irq(&conf->device_lock);
4623                 conf->reshape_safe = mddev->reshape_position;
4624                 spin_unlock_irq(&conf->device_lock);
4625                 wake_up(&conf->wait_for_overlap);
4626                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4627         }
4628         return reshape_sectors;
4629 }
4630
4631 /* FIXME go_faster isn't used */
4632 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4633 {
4634         struct r5conf *conf = mddev->private;
4635         struct stripe_head *sh;
4636         sector_t max_sector = mddev->dev_sectors;
4637         sector_t sync_blocks;
4638         int still_degraded = 0;
4639         int i;
4640
4641         if (sector_nr >= max_sector) {
4642                 /* just being told to finish up .. nothing much to do */
4643
4644                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4645                         end_reshape(conf);
4646                         return 0;
4647                 }
4648
4649                 if (mddev->curr_resync < max_sector) /* aborted */
4650                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4651                                         &sync_blocks, 1);
4652                 else /* completed sync */
4653                         conf->fullsync = 0;
4654                 bitmap_close_sync(mddev->bitmap);
4655
4656                 return 0;
4657         }
4658
4659         /* Allow raid5_quiesce to complete */
4660         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4661
4662         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4663                 return reshape_request(mddev, sector_nr, skipped);
4664
4665         /* No need to check resync_max as we never do more than one
4666          * stripe, and as resync_max will always be on a chunk boundary,
4667          * if the check in md_do_sync didn't fire, there is no chance
4668          * of overstepping resync_max here
4669          */
4670
4671         /* if there is too many failed drives and we are trying
4672          * to resync, then assert that we are finished, because there is
4673          * nothing we can do.
4674          */
4675         if (mddev->degraded >= conf->max_degraded &&
4676             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4677                 sector_t rv = mddev->dev_sectors - sector_nr;
4678                 *skipped = 1;
4679                 return rv;
4680         }
4681         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4682             !conf->fullsync &&
4683             !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4684             sync_blocks >= STRIPE_SECTORS) {
4685                 /* we can skip this block, and probably more */
4686                 sync_blocks /= STRIPE_SECTORS;
4687                 *skipped = 1;
4688                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4689         }
4690
4691         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4692
4693         sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4694         if (sh == NULL) {
4695                 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4696                 /* make sure we don't swamp the stripe cache if someone else
4697                  * is trying to get access
4698                  */
4699                 schedule_timeout_uninterruptible(1);
4700         }
4701         /* Need to check if array will still be degraded after recovery/resync
4702          * We don't need to check the 'failed' flag as when that gets set,
4703          * recovery aborts.
4704          */
4705         for (i = 0; i < conf->raid_disks; i++)
4706                 if (conf->disks[i].rdev == NULL)
4707                         still_degraded = 1;
4708
4709         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4710
4711         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4712
4713         handle_stripe(sh);
4714         release_stripe(sh);
4715
4716         return STRIPE_SECTORS;
4717 }
4718
4719 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4720 {
4721         /* We may not be able to submit a whole bio at once as there
4722          * may not be enough stripe_heads available.
4723          * We cannot pre-allocate enough stripe_heads as we may need
4724          * more than exist in the cache (if we allow ever large chunks).
4725          * So we do one stripe head at a time and record in
4726          * ->bi_hw_segments how many have been done.
4727          *
4728          * We *know* that this entire raid_bio is in one chunk, so
4729          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4730          */
4731         struct stripe_head *sh;
4732         int dd_idx;
4733         sector_t sector, logical_sector, last_sector;
4734         int scnt = 0;
4735         int remaining;
4736         int handled = 0;
4737
4738         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4739         sector = raid5_compute_sector(conf, logical_sector,
4740                                       0, &dd_idx, NULL);
4741         last_sector = bio_end_sector(raid_bio);
4742
4743         for (; logical_sector < last_sector;
4744              logical_sector += STRIPE_SECTORS,
4745                      sector += STRIPE_SECTORS,
4746                      scnt++) {
4747
4748                 if (scnt < raid5_bi_processed_stripes(raid_bio))
4749                         /* already done this stripe */
4750                         continue;
4751
4752                 sh = get_active_stripe(conf, sector, 0, 1, 0);
4753
4754                 if (!sh) {
4755                         /* failed to get a stripe - must wait */
4756                         raid5_set_bi_processed_stripes(raid_bio, scnt);
4757                         conf->retry_read_aligned = raid_bio;
4758                         return handled;
4759                 }
4760
4761                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4762                         release_stripe(sh);
4763                         raid5_set_bi_processed_stripes(raid_bio, scnt);
4764                         conf->retry_read_aligned = raid_bio;
4765                         return handled;
4766                 }
4767
4768                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
4769                 handle_stripe(sh);
4770                 release_stripe(sh);
4771                 handled++;
4772         }
4773         remaining = raid5_dec_bi_active_stripes(raid_bio);
4774         if (remaining == 0) {
4775                 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
4776                                          raid_bio, 0);
4777                 bio_endio(raid_bio, 0);
4778         }
4779         if (atomic_dec_and_test(&conf->active_aligned_reads))
4780                 wake_up(&conf->wait_for_stripe);
4781         return handled;
4782 }
4783
4784 #define MAX_STRIPE_BATCH 8
4785 static int handle_active_stripes(struct r5conf *conf)
4786 {
4787         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
4788         int i, batch_size = 0;
4789
4790         while (batch_size < MAX_STRIPE_BATCH &&
4791                         (sh = __get_priority_stripe(conf)) != NULL)
4792                 batch[batch_size++] = sh;
4793
4794         if (batch_size == 0)
4795                 return batch_size;
4796         spin_unlock_irq(&conf->device_lock);
4797
4798         for (i = 0; i < batch_size; i++)
4799                 handle_stripe(batch[i]);
4800
4801         cond_resched();
4802
4803         spin_lock_irq(&conf->device_lock);
4804         for (i = 0; i < batch_size; i++)
4805                 __release_stripe(conf, batch[i]);
4806         return batch_size;
4807 }
4808
4809 /*
4810  * This is our raid5 kernel thread.
4811  *
4812  * We scan the hash table for stripes which can be handled now.
4813  * During the scan, completed stripes are saved for us by the interrupt
4814  * handler, so that they will not have to wait for our next wakeup.
4815  */
4816 static void raid5d(struct md_thread *thread)
4817 {
4818         struct mddev *mddev = thread->mddev;
4819         struct r5conf *conf = mddev->private;
4820         int handled;
4821         struct blk_plug plug;
4822
4823         pr_debug("+++ raid5d active\n");
4824
4825         md_check_recovery(mddev);
4826
4827         blk_start_plug(&plug);
4828         handled = 0;
4829         spin_lock_irq(&conf->device_lock);
4830         while (1) {
4831                 struct bio *bio;
4832                 int batch_size;
4833
4834                 if (
4835                     !list_empty(&conf->bitmap_list)) {
4836                         /* Now is a good time to flush some bitmap updates */
4837                         conf->seq_flush++;
4838                         spin_unlock_irq(&conf->device_lock);
4839                         bitmap_unplug(mddev->bitmap);
4840                         spin_lock_irq(&conf->device_lock);
4841                         conf->seq_write = conf->seq_flush;
4842                         activate_bit_delay(conf);
4843                 }
4844                 raid5_activate_delayed(conf);
4845
4846                 while ((bio = remove_bio_from_retry(conf))) {
4847                         int ok;
4848                         spin_unlock_irq(&conf->device_lock);
4849                         ok = retry_aligned_read(conf, bio);
4850                         spin_lock_irq(&conf->device_lock);
4851                         if (!ok)
4852                                 break;
4853                         handled++;
4854                 }
4855
4856                 batch_size = handle_active_stripes(conf);
4857                 if (!batch_size)
4858                         break;
4859                 handled += batch_size;
4860
4861                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
4862                         spin_unlock_irq(&conf->device_lock);
4863                         md_check_recovery(mddev);
4864                         spin_lock_irq(&conf->device_lock);
4865                 }
4866         }
4867         pr_debug("%d stripes handled\n", handled);
4868
4869         spin_unlock_irq(&conf->device_lock);
4870
4871         async_tx_issue_pending_all();
4872         blk_finish_plug(&plug);
4873
4874         pr_debug("--- raid5d inactive\n");
4875 }
4876
4877 static ssize_t
4878 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4879 {
4880         struct r5conf *conf = mddev->private;
4881         if (conf)
4882                 return sprintf(page, "%d\n", conf->max_nr_stripes);
4883         else
4884                 return 0;
4885 }
4886
4887 int
4888 raid5_set_cache_size(struct mddev *mddev, int size)
4889 {
4890         struct r5conf *conf = mddev->private;
4891         int err;
4892
4893         if (size <= 16 || size > 32768)
4894                 return -EINVAL;
4895         while (size < conf->max_nr_stripes) {
4896                 if (drop_one_stripe(conf))
4897                         conf->max_nr_stripes--;
4898                 else
4899                         break;
4900         }
4901         err = md_allow_write(mddev);
4902         if (err)
4903                 return err;
4904         while (size > conf->max_nr_stripes) {
4905                 if (grow_one_stripe(conf))
4906                         conf->max_nr_stripes++;
4907                 else break;
4908         }
4909         return 0;
4910 }
4911 EXPORT_SYMBOL(raid5_set_cache_size);
4912
4913 static ssize_t
4914 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4915 {
4916         struct r5conf *conf = mddev->private;
4917         unsigned long new;
4918         int err;
4919
4920         if (len >= PAGE_SIZE)
4921                 return -EINVAL;
4922         if (!conf)
4923                 return -ENODEV;
4924
4925         if (strict_strtoul(page, 10, &new))
4926                 return -EINVAL;
4927         err = raid5_set_cache_size(mddev, new);
4928         if (err)
4929                 return err;
4930         return len;
4931 }
4932
4933 static struct md_sysfs_entry
4934 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4935                                 raid5_show_stripe_cache_size,
4936                                 raid5_store_stripe_cache_size);
4937
4938 static ssize_t
4939 raid5_show_preread_threshold(struct mddev *mddev, char *page)
4940 {
4941         struct r5conf *conf = mddev->private;
4942         if (conf)
4943                 return sprintf(page, "%d\n", conf->bypass_threshold);
4944         else
4945                 return 0;
4946 }
4947
4948 static ssize_t
4949 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4950 {
4951         struct r5conf *conf = mddev->private;
4952         unsigned long new;
4953         if (len >= PAGE_SIZE)
4954                 return -EINVAL;
4955         if (!conf)
4956                 return -ENODEV;
4957
4958         if (strict_strtoul(page, 10, &new))
4959                 return -EINVAL;
4960         if (new > conf->max_nr_stripes)
4961                 return -EINVAL;
4962         conf->bypass_threshold = new;
4963         return len;
4964 }
4965
4966 static struct md_sysfs_entry
4967 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4968                                         S_IRUGO | S_IWUSR,
4969                                         raid5_show_preread_threshold,
4970                                         raid5_store_preread_threshold);
4971
4972 static ssize_t
4973 stripe_cache_active_show(struct mddev *mddev, char *page)
4974 {
4975         struct r5conf *conf = mddev->private;
4976         if (conf)
4977                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4978         else
4979                 return 0;
4980 }
4981
4982 static struct md_sysfs_entry
4983 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4984
4985 static struct attribute *raid5_attrs[] =  {
4986         &raid5_stripecache_size.attr,
4987         &raid5_stripecache_active.attr,
4988         &raid5_preread_bypass_threshold.attr,
4989         NULL,
4990 };
4991 static struct attribute_group raid5_attrs_group = {
4992         .name = NULL,
4993         .attrs = raid5_attrs,
4994 };
4995
4996 static sector_t
4997 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
4998 {
4999         struct r5conf *conf = mddev->private;
5000
5001         if (!sectors)
5002                 sectors = mddev->dev_sectors;
5003         if (!raid_disks)
5004                 /* size is defined by the smallest of previous and new size */
5005                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
5006
5007         sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5008         sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
5009         return sectors * (raid_disks - conf->max_degraded);
5010 }
5011
5012 static void raid5_free_percpu(struct r5conf *conf)
5013 {
5014         struct raid5_percpu *percpu;
5015         unsigned long cpu;
5016
5017         if (!conf->percpu)
5018                 return;
5019
5020         get_online_cpus();
5021         for_each_possible_cpu(cpu) {
5022                 percpu = per_cpu_ptr(conf->percpu, cpu);
5023                 safe_put_page(percpu->spare_page);
5024                 kfree(percpu->scribble);
5025         }
5026 #ifdef CONFIG_HOTPLUG_CPU
5027         unregister_cpu_notifier(&conf->cpu_notify);
5028 #endif
5029         put_online_cpus();
5030
5031         free_percpu(conf->percpu);
5032 }
5033
5034 static void free_conf(struct r5conf *conf)
5035 {
5036         shrink_stripes(conf);
5037         raid5_free_percpu(conf);
5038         kfree(conf->disks);
5039         kfree(conf->stripe_hashtbl);
5040         kfree(conf);
5041 }
5042
5043 #ifdef CONFIG_HOTPLUG_CPU
5044 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
5045                               void *hcpu)
5046 {
5047         struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
5048         long cpu = (long)hcpu;
5049         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
5050
5051         switch (action) {
5052         case CPU_UP_PREPARE:
5053         case CPU_UP_PREPARE_FROZEN:
5054                 if (conf->level == 6 && !percpu->spare_page)
5055                         percpu->spare_page = alloc_page(GFP_KERNEL);
5056                 if (!percpu->scribble)
5057                         percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5058
5059                 if (!percpu->scribble ||
5060                     (conf->level == 6 && !percpu->spare_page)) {
5061                         safe_put_page(percpu->spare_page);
5062                         kfree(percpu->scribble);
5063                         pr_err("%s: failed memory allocation for cpu%ld\n",
5064                                __func__, cpu);
5065                         return notifier_from_errno(-ENOMEM);
5066                 }
5067                 break;
5068         case CPU_DEAD:
5069         case CPU_DEAD_FROZEN:
5070                 safe_put_page(percpu->spare_page);
5071                 kfree(percpu->scribble);
5072                 percpu->spare_page = NULL;
5073                 percpu->scribble = NULL;
5074                 break;
5075         default:
5076                 break;
5077         }
5078         return NOTIFY_OK;
5079 }
5080 #endif
5081
5082 static int raid5_alloc_percpu(struct r5conf *conf)
5083 {
5084         unsigned long cpu;
5085         struct page *spare_page;
5086         struct raid5_percpu __percpu *allcpus;
5087         void *scribble;
5088         int err;
5089
5090         allcpus = alloc_percpu(struct raid5_percpu);
5091         if (!allcpus)
5092                 return -ENOMEM;
5093         conf->percpu = allcpus;
5094
5095         get_online_cpus();
5096         err = 0;
5097         for_each_present_cpu(cpu) {
5098                 if (conf->level == 6) {
5099                         spare_page = alloc_page(GFP_KERNEL);
5100                         if (!spare_page) {
5101                                 err = -ENOMEM;
5102                                 break;
5103                         }
5104                         per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
5105                 }
5106                 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5107                 if (!scribble) {
5108                         err = -ENOMEM;
5109                         break;
5110                 }
5111                 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
5112         }
5113 #ifdef CONFIG_HOTPLUG_CPU
5114         conf->cpu_notify.notifier_call = raid456_cpu_notify;
5115         conf->cpu_notify.priority = 0;
5116         if (err == 0)
5117                 err = register_cpu_notifier(&conf->cpu_notify);
5118 #endif
5119         put_online_cpus();
5120
5121         return err;
5122 }
5123
5124 static struct r5conf *setup_conf(struct mddev *mddev)
5125 {
5126         struct r5conf *conf;
5127         int raid_disk, memory, max_disks;
5128         struct md_rdev *rdev;
5129         struct disk_info *disk;
5130         char pers_name[6];
5131
5132         if (mddev->new_level != 5
5133             && mddev->new_level != 4
5134             && mddev->new_level != 6) {
5135                 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5136                        mdname(mddev), mddev->new_level);
5137                 return ERR_PTR(-EIO);
5138         }
5139         if ((mddev->new_level == 5
5140              && !algorithm_valid_raid5(mddev->new_layout)) ||
5141             (mddev->new_level == 6
5142              && !algorithm_valid_raid6(mddev->new_layout))) {
5143                 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
5144                        mdname(mddev), mddev->new_layout);
5145                 return ERR_PTR(-EIO);
5146         }
5147         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
5148                 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5149                        mdname(mddev), mddev->raid_disks);
5150                 return ERR_PTR(-EINVAL);
5151         }
5152
5153         if (!mddev->new_chunk_sectors ||
5154             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
5155             !is_power_of_2(mddev->new_chunk_sectors)) {
5156                 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
5157                        mdname(mddev), mddev->new_chunk_sectors << 9);
5158                 return ERR_PTR(-EINVAL);
5159         }
5160
5161         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
5162         if (conf == NULL)
5163                 goto abort;
5164         spin_lock_init(&conf->device_lock);
5165         init_waitqueue_head(&conf->wait_for_stripe);
5166         init_waitqueue_head(&conf->wait_for_overlap);
5167         INIT_LIST_HEAD(&conf->handle_list);
5168         INIT_LIST_HEAD(&conf->hold_list);
5169         INIT_LIST_HEAD(&conf->delayed_list);
5170         INIT_LIST_HEAD(&conf->bitmap_list);
5171         INIT_LIST_HEAD(&conf->inactive_list);
5172         atomic_set(&conf->active_stripes, 0);
5173         atomic_set(&conf->preread_active_stripes, 0);
5174         atomic_set(&conf->active_aligned_reads, 0);
5175         conf->bypass_threshold = BYPASS_THRESHOLD;
5176         conf->recovery_disabled = mddev->recovery_disabled - 1;
5177
5178         conf->raid_disks = mddev->raid_disks;
5179         if (mddev->reshape_position == MaxSector)
5180                 conf->previous_raid_disks = mddev->raid_disks;
5181         else
5182                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
5183         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
5184         conf->scribble_len = scribble_len(max_disks);
5185
5186         conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
5187                               GFP_KERNEL);
5188         if (!conf->disks)
5189                 goto abort;
5190
5191         conf->mddev = mddev;
5192
5193         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
5194                 goto abort;
5195
5196         conf->level = mddev->new_level;
5197         if (raid5_alloc_percpu(conf) != 0)
5198                 goto abort;
5199
5200         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
5201
5202         rdev_for_each(rdev, mddev) {
5203                 raid_disk = rdev->raid_disk;
5204                 if (raid_disk >= max_disks
5205                     || raid_disk < 0)
5206                         continue;
5207                 disk = conf->disks + raid_disk;
5208
5209                 if (test_bit(Replacement, &rdev->flags)) {
5210                         if (disk->replacement)
5211                                 goto abort;
5212                         disk->replacement = rdev;
5213                 } else {
5214                         if (disk->rdev)
5215                                 goto abort;
5216                         disk->rdev = rdev;
5217                 }
5218
5219                 if (test_bit(In_sync, &rdev->flags)) {
5220                         char b[BDEVNAME_SIZE];
5221                         printk(KERN_INFO "md/raid:%s: device %s operational as raid"
5222                                " disk %d\n",
5223                                mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
5224                 } else if (rdev->saved_raid_disk != raid_disk)
5225                         /* Cannot rely on bitmap to complete recovery */
5226                         conf->fullsync = 1;
5227         }
5228
5229         conf->chunk_sectors = mddev->new_chunk_sectors;
5230         conf->level = mddev->new_level;
5231         if (conf->level == 6)
5232                 conf->max_degraded = 2;
5233         else
5234                 conf->max_degraded = 1;
5235         conf->algorithm = mddev->new_layout;
5236         conf->max_nr_stripes = NR_STRIPES;
5237         conf->reshape_progress = mddev->reshape_position;
5238         if (conf->reshape_progress != MaxSector) {
5239                 conf->prev_chunk_sectors = mddev->chunk_sectors;
5240                 conf->prev_algo = mddev->layout;
5241         }
5242
5243         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
5244                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
5245         if (grow_stripes(conf, conf->max_nr_stripes)) {
5246                 printk(KERN_ERR
5247                        "md/raid:%s: couldn't allocate %dkB for buffers\n",
5248                        mdname(mddev), memory);
5249                 goto abort;
5250         } else
5251                 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
5252                        mdname(mddev), memory);
5253
5254         sprintf(pers_name, "raid%d", mddev->new_level);
5255         conf->thread = md_register_thread(raid5d, mddev, pers_name);
5256         if (!conf->thread) {
5257                 printk(KERN_ERR
5258                        "md/raid:%s: couldn't allocate thread.\n",
5259                        mdname(mddev));
5260                 goto abort;
5261         }
5262
5263         return conf;
5264
5265  abort:
5266         if (conf) {
5267                 free_conf(conf);
5268                 return ERR_PTR(-EIO);
5269         } else
5270                 return ERR_PTR(-ENOMEM);
5271 }
5272
5273
5274 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5275 {
5276         switch (algo) {
5277         case ALGORITHM_PARITY_0:
5278                 if (raid_disk < max_degraded)
5279                         return 1;
5280                 break;
5281         case ALGORITHM_PARITY_N:
5282                 if (raid_disk >= raid_disks - max_degraded)
5283                         return 1;
5284                 break;
5285         case ALGORITHM_PARITY_0_6:
5286                 if (raid_disk == 0 ||
5287                     raid_disk == raid_disks - 1)
5288                         return 1;
5289                 break;
5290         case ALGORITHM_LEFT_ASYMMETRIC_6:
5291         case ALGORITHM_RIGHT_ASYMMETRIC_6:
5292         case ALGORITHM_LEFT_SYMMETRIC_6:
5293         case ALGORITHM_RIGHT_SYMMETRIC_6:
5294                 if (raid_disk == raid_disks - 1)
5295                         return 1;
5296         }
5297         return 0;
5298 }
5299
5300 static int run(struct mddev *mddev)
5301 {
5302         struct r5conf *conf;
5303         int working_disks = 0;
5304         int dirty_parity_disks = 0;
5305         struct md_rdev *rdev;
5306         sector_t reshape_offset = 0;
5307         int i;
5308         long long min_offset_diff = 0;
5309         int first = 1;
5310
5311         if (mddev->recovery_cp != MaxSector)
5312                 printk(KERN_NOTICE "md/raid:%s: not clean"
5313                        " -- starting background reconstruction\n",
5314                        mdname(mddev));
5315
5316         rdev_for_each(rdev, mddev) {
5317                 long long diff;
5318                 if (rdev->raid_disk < 0)
5319                         continue;
5320                 diff = (rdev->new_data_offset - rdev->data_offset);
5321                 if (first) {
5322                         min_offset_diff = diff;
5323                         first = 0;
5324                 } else if (mddev->reshape_backwards &&
5325                          diff < min_offset_diff)
5326                         min_offset_diff = diff;
5327                 else if (!mddev->reshape_backwards &&
5328                          diff > min_offset_diff)
5329                         min_offset_diff = diff;
5330         }
5331
5332         if (mddev->reshape_position != MaxSector) {
5333                 /* Check that we can continue the reshape.
5334                  * Difficulties arise if the stripe we would write to
5335                  * next is at or after the stripe we would read from next.
5336                  * For a reshape that changes the number of devices, this
5337                  * is only possible for a very short time, and mdadm makes
5338                  * sure that time appears to have past before assembling
5339                  * the array.  So we fail if that time hasn't passed.
5340                  * For a reshape that keeps the number of devices the same
5341                  * mdadm must be monitoring the reshape can keeping the
5342                  * critical areas read-only and backed up.  It will start
5343                  * the array in read-only mode, so we check for that.
5344                  */
5345                 sector_t here_new, here_old;
5346                 int old_disks;
5347                 int max_degraded = (mddev->level == 6 ? 2 : 1);
5348
5349                 if (mddev->new_level != mddev->level) {
5350                         printk(KERN_ERR "md/raid:%s: unsupported reshape "
5351                                "required - aborting.\n",
5352                                mdname(mddev));
5353                         return -EINVAL;
5354                 }
5355                 old_disks = mddev->raid_disks - mddev->delta_disks;
5356                 /* reshape_position must be on a new-stripe boundary, and one
5357                  * further up in new geometry must map after here in old
5358                  * geometry.
5359                  */
5360                 here_new = mddev->reshape_position;
5361                 if (sector_div(here_new, mddev->new_chunk_sectors *
5362                                (mddev->raid_disks - max_degraded))) {
5363                         printk(KERN_ERR "md/raid:%s: reshape_position not "
5364                                "on a stripe boundary\n", mdname(mddev));
5365                         return -EINVAL;
5366                 }
5367                 reshape_offset = here_new * mddev->new_chunk_sectors;
5368                 /* here_new is the stripe we will write to */
5369                 here_old = mddev->reshape_position;
5370                 sector_div(here_old, mddev->chunk_sectors *
5371                            (old_disks-max_degraded));
5372                 /* here_old is the first stripe that we might need to read
5373                  * from */
5374                 if (mddev->delta_disks == 0) {
5375                         if ((here_new * mddev->new_chunk_sectors !=
5376                              here_old * mddev->chunk_sectors)) {
5377                                 printk(KERN_ERR "md/raid:%s: reshape position is"
5378                                        " confused - aborting\n", mdname(mddev));
5379                                 return -EINVAL;
5380                         }
5381                         /* We cannot be sure it is safe to start an in-place
5382                          * reshape.  It is only safe if user-space is monitoring
5383                          * and taking constant backups.
5384                          * mdadm always starts a situation like this in
5385                          * readonly mode so it can take control before
5386                          * allowing any writes.  So just check for that.
5387                          */
5388                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
5389                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
5390                                 /* not really in-place - so OK */;
5391                         else if (mddev->ro == 0) {
5392                                 printk(KERN_ERR "md/raid:%s: in-place reshape "
5393                                        "must be started in read-only mode "
5394                                        "- aborting\n",
5395                                        mdname(mddev));
5396                                 return -EINVAL;
5397                         }
5398                 } else if (mddev->reshape_backwards
5399                     ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
5400                        here_old * mddev->chunk_sectors)
5401                     : (here_new * mddev->new_chunk_sectors >=
5402                        here_old * mddev->chunk_sectors + (-min_offset_diff))) {
5403                         /* Reading from the same stripe as writing to - bad */
5404                         printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5405                                "auto-recovery - aborting.\n",
5406                                mdname(mddev));
5407                         return -EINVAL;
5408                 }
5409                 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5410                        mdname(mddev));
5411                 /* OK, we should be able to continue; */
5412         } else {
5413                 BUG_ON(mddev->level != mddev->new_level);
5414                 BUG_ON(mddev->layout != mddev->new_layout);
5415                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
5416                 BUG_ON(mddev->delta_disks != 0);
5417         }
5418
5419         if (mddev->private == NULL)
5420                 conf = setup_conf(mddev);
5421         else
5422                 conf = mddev->private;
5423
5424         if (IS_ERR(conf))
5425                 return PTR_ERR(conf);
5426
5427         conf->min_offset_diff = min_offset_diff;
5428         mddev->thread = conf->thread;
5429         conf->thread = NULL;
5430         mddev->private = conf;
5431
5432         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
5433              i++) {
5434                 rdev = conf->disks[i].rdev;
5435                 if (!rdev && conf->disks[i].replacement) {
5436                         /* The replacement is all we have yet */
5437                         rdev = conf->disks[i].replacement;
5438                         conf->disks[i].replacement = NULL;
5439                         clear_bit(Replacement, &rdev->flags);
5440                         conf->disks[i].rdev = rdev;
5441                 }
5442                 if (!rdev)
5443                         continue;
5444                 if (conf->disks[i].replacement &&
5445                     conf->reshape_progress != MaxSector) {
5446                         /* replacements and reshape simply do not mix. */
5447                         printk(KERN_ERR "md: cannot handle concurrent "
5448                                "replacement and reshape.\n");
5449                         goto abort;
5450                 }
5451                 if (test_bit(In_sync, &rdev->flags)) {
5452                         working_disks++;
5453                         continue;
5454                 }
5455                 /* This disc is not fully in-sync.  However if it
5456                  * just stored parity (beyond the recovery_offset),
5457                  * when we don't need to be concerned about the
5458                  * array being dirty.
5459                  * When reshape goes 'backwards', we never have
5460                  * partially completed devices, so we only need
5461                  * to worry about reshape going forwards.
5462                  */
5463                 /* Hack because v0.91 doesn't store recovery_offset properly. */
5464                 if (mddev->major_version == 0 &&
5465                     mddev->minor_version > 90)
5466                         rdev->recovery_offset = reshape_offset;
5467
5468                 if (rdev->recovery_offset < reshape_offset) {
5469                         /* We need to check old and new layout */
5470                         if (!only_parity(rdev->raid_disk,
5471                                          conf->algorithm,
5472                                          conf->raid_disks,
5473                                          conf->max_degraded))
5474                                 continue;
5475                 }
5476                 if (!only_parity(rdev->raid_disk,
5477                                  conf->prev_algo,
5478                                  conf->previous_raid_disks,
5479                                  conf->max_degraded))
5480                         continue;
5481                 dirty_parity_disks++;
5482         }
5483
5484         /*
5485          * 0 for a fully functional array, 1 or 2 for a degraded array.
5486          */
5487         mddev->degraded = calc_degraded(conf);
5488
5489         if (has_failed(conf)) {
5490                 printk(KERN_ERR "md/raid:%s: not enough operational devices"
5491                         " (%d/%d failed)\n",
5492                         mdname(mddev), mddev->degraded, conf->raid_disks);
5493                 goto abort;
5494         }
5495
5496         /* device size must be a multiple of chunk size */
5497         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5498         mddev->resync_max_sectors = mddev->dev_sectors;
5499
5500         if (mddev->degraded > dirty_parity_disks &&
5501             mddev->recovery_cp != MaxSector) {
5502                 if (mddev->ok_start_degraded)
5503                         printk(KERN_WARNING
5504                                "md/raid:%s: starting dirty degraded array"
5505                                " - data corruption possible.\n",
5506                                mdname(mddev));
5507                 else {
5508                         printk(KERN_ERR
5509                                "md/raid:%s: cannot start dirty degraded array.\n",
5510                                mdname(mddev));
5511                         goto abort;
5512                 }
5513         }
5514
5515         if (mddev->degraded == 0)
5516                 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5517                        " devices, algorithm %d\n", mdname(mddev), conf->level,
5518                        mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5519                        mddev->new_layout);
5520         else
5521                 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5522                        " out of %d devices, algorithm %d\n",
5523                        mdname(mddev), conf->level,
5524                        mddev->raid_disks - mddev->degraded,
5525                        mddev->raid_disks, mddev->new_layout);
5526
5527         print_raid5_conf(conf);
5528
5529         if (conf->reshape_progress != MaxSector) {
5530                 conf->reshape_safe = conf->reshape_progress;
5531                 atomic_set(&conf->reshape_stripes, 0);
5532                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5533                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5534                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5535                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5536                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5537                                                         "reshape");
5538         }
5539
5540
5541         /* Ok, everything is just fine now */
5542         if (mddev->to_remove == &raid5_attrs_group)
5543                 mddev->to_remove = NULL;
5544         else if (mddev->kobj.sd &&
5545             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5546                 printk(KERN_WARNING
5547                        "raid5: failed to create sysfs attributes for %s\n",
5548                        mdname(mddev));
5549         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5550
5551         if (mddev->queue) {
5552                 int chunk_size;
5553                 bool discard_supported = true;
5554                 /* read-ahead size must cover two whole stripes, which
5555                  * is 2 * (datadisks) * chunksize where 'n' is the
5556                  * number of raid devices
5557                  */
5558                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5559                 int stripe = data_disks *
5560                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5561                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5562                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5563
5564                 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5565
5566                 mddev->queue->backing_dev_info.congested_data = mddev;
5567                 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5568
5569                 chunk_size = mddev->chunk_sectors << 9;
5570                 blk_queue_io_min(mddev->queue, chunk_size);
5571                 blk_queue_io_opt(mddev->queue, chunk_size *
5572                                  (conf->raid_disks - conf->max_degraded));
5573                 /*
5574                  * We can only discard a whole stripe. It doesn't make sense to
5575                  * discard data disk but write parity disk
5576                  */
5577                 stripe = stripe * PAGE_SIZE;
5578                 /* Round up to power of 2, as discard handling
5579                  * currently assumes that */
5580                 while ((stripe-1) & stripe)
5581                         stripe = (stripe | (stripe-1)) + 1;
5582                 mddev->queue->limits.discard_alignment = stripe;
5583                 mddev->queue->limits.discard_granularity = stripe;
5584                 /*
5585                  * unaligned part of discard request will be ignored, so can't
5586                  * guarantee discard_zerors_data
5587                  */
5588                 mddev->queue->limits.discard_zeroes_data = 0;
5589
5590                 rdev_for_each(rdev, mddev) {
5591                         disk_stack_limits(mddev->gendisk, rdev->bdev,
5592                                           rdev->data_offset << 9);
5593                         disk_stack_limits(mddev->gendisk, rdev->bdev,
5594                                           rdev->new_data_offset << 9);
5595                         /*
5596                          * discard_zeroes_data is required, otherwise data
5597                          * could be lost. Consider a scenario: discard a stripe
5598                          * (the stripe could be inconsistent if
5599                          * discard_zeroes_data is 0); write one disk of the
5600                          * stripe (the stripe could be inconsistent again
5601                          * depending on which disks are used to calculate
5602                          * parity); the disk is broken; The stripe data of this
5603                          * disk is lost.
5604                          */
5605                         if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
5606                             !bdev_get_queue(rdev->bdev)->
5607                                                 limits.discard_zeroes_data)
5608                                 discard_supported = false;
5609                 }
5610
5611                 if (discard_supported &&
5612                    mddev->queue->limits.max_discard_sectors >= stripe &&
5613                    mddev->queue->limits.discard_granularity >= stripe)
5614                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
5615                                                 mddev->queue);
5616                 else
5617                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
5618                                                 mddev->queue);
5619         }
5620
5621         return 0;
5622 abort:
5623         md_unregister_thread(&mddev->thread);
5624         print_raid5_conf(conf);
5625         free_conf(conf);
5626         mddev->private = NULL;
5627         printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5628         return -EIO;
5629 }
5630
5631 static int stop(struct mddev *mddev)
5632 {
5633         struct r5conf *conf = mddev->private;
5634
5635         md_unregister_thread(&mddev->thread);
5636         if (mddev->queue)
5637                 mddev->queue->backing_dev_info.congested_fn = NULL;
5638         free_conf(conf);
5639         mddev->private = NULL;
5640         mddev->to_remove = &raid5_attrs_group;
5641         return 0;
5642 }
5643
5644 static void status(struct seq_file *seq, struct mddev *mddev)
5645 {
5646         struct r5conf *conf = mddev->private;
5647         int i;
5648
5649         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5650                 mddev->chunk_sectors / 2, mddev->layout);
5651         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5652         for (i = 0; i < conf->raid_disks; i++)
5653                 seq_printf (seq, "%s",
5654                                conf->disks[i].rdev &&
5655                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5656         seq_printf (seq, "]");
5657 }
5658
5659 static void print_raid5_conf (struct r5conf *conf)
5660 {
5661         int i;
5662         struct disk_info *tmp;
5663
5664         printk(KERN_DEBUG "RAID conf printout:\n");
5665         if (!conf) {
5666                 printk("(conf==NULL)\n");
5667                 return;
5668         }
5669         printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5670                conf->raid_disks,
5671                conf->raid_disks - conf->mddev->degraded);
5672
5673         for (i = 0; i < conf->raid_disks; i++) {
5674                 char b[BDEVNAME_SIZE];
5675                 tmp = conf->disks + i;
5676                 if (tmp->rdev)
5677                         printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5678                                i, !test_bit(Faulty, &tmp->rdev->flags),
5679                                bdevname(tmp->rdev->bdev, b));
5680         }
5681 }
5682
5683 static int raid5_spare_active(struct mddev *mddev)
5684 {
5685         int i;
5686         struct r5conf *conf = mddev->private;
5687         struct disk_info *tmp;
5688         int count = 0;
5689         unsigned long flags;
5690
5691         for (i = 0; i < conf->raid_disks; i++) {
5692                 tmp = conf->disks + i;
5693                 if (tmp->replacement
5694                     && tmp->replacement->recovery_offset == MaxSector
5695                     && !test_bit(Faulty, &tmp->replacement->flags)
5696                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
5697                         /* Replacement has just become active. */
5698                         if (!tmp->rdev
5699                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
5700                                 count++;
5701                         if (tmp->rdev) {
5702                                 /* Replaced device not technically faulty,
5703                                  * but we need to be sure it gets removed
5704                                  * and never re-added.
5705                                  */
5706                                 set_bit(Faulty, &tmp->rdev->flags);
5707                                 sysfs_notify_dirent_safe(
5708                                         tmp->rdev->sysfs_state);
5709                         }
5710                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
5711                 } else if (tmp->rdev
5712                     && tmp->rdev->recovery_offset == MaxSector
5713                     && !test_bit(Faulty, &tmp->rdev->flags)
5714                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5715                         count++;
5716                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5717                 }
5718         }
5719         spin_lock_irqsave(&conf->device_lock, flags);
5720         mddev->degraded = calc_degraded(conf);
5721         spin_unlock_irqrestore(&conf->device_lock, flags);
5722         print_raid5_conf(conf);
5723         return count;
5724 }
5725
5726 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
5727 {
5728         struct r5conf *conf = mddev->private;
5729         int err = 0;
5730         int number = rdev->raid_disk;
5731         struct md_rdev **rdevp;
5732         struct disk_info *p = conf->disks + number;
5733
5734         print_raid5_conf(conf);
5735         if (rdev == p->rdev)
5736                 rdevp = &p->rdev;
5737         else if (rdev == p->replacement)
5738                 rdevp = &p->replacement;
5739         else
5740                 return 0;
5741
5742         if (number >= conf->raid_disks &&
5743             conf->reshape_progress == MaxSector)
5744                 clear_bit(In_sync, &rdev->flags);
5745
5746         if (test_bit(In_sync, &rdev->flags) ||
5747             atomic_read(&rdev->nr_pending)) {
5748                 err = -EBUSY;
5749                 goto abort;
5750         }
5751         /* Only remove non-faulty devices if recovery
5752          * isn't possible.
5753          */
5754         if (!test_bit(Faulty, &rdev->flags) &&
5755             mddev->recovery_disabled != conf->recovery_disabled &&
5756             !has_failed(conf) &&
5757             (!p->replacement || p->replacement == rdev) &&
5758             number < conf->raid_disks) {
5759                 err = -EBUSY;
5760                 goto abort;
5761         }
5762         *rdevp = NULL;
5763         synchronize_rcu();
5764         if (atomic_read(&rdev->nr_pending)) {
5765                 /* lost the race, try later */
5766                 err = -EBUSY;
5767                 *rdevp = rdev;
5768         } else if (p->replacement) {
5769                 /* We must have just cleared 'rdev' */
5770                 p->rdev = p->replacement;
5771                 clear_bit(Replacement, &p->replacement->flags);
5772                 smp_mb(); /* Make sure other CPUs may see both as identical
5773                            * but will never see neither - if they are careful
5774                            */
5775                 p->replacement = NULL;
5776                 clear_bit(WantReplacement, &rdev->flags);
5777         } else
5778                 /* We might have just removed the Replacement as faulty-
5779                  * clear the bit just in case
5780                  */
5781                 clear_bit(WantReplacement, &rdev->flags);
5782 abort:
5783
5784         print_raid5_conf(conf);
5785         return err;
5786 }
5787
5788 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5789 {
5790         struct r5conf *conf = mddev->private;
5791         int err = -EEXIST;
5792         int disk;
5793         struct disk_info *p;
5794         int first = 0;
5795         int last = conf->raid_disks - 1;
5796
5797         if (mddev->recovery_disabled == conf->recovery_disabled)
5798                 return -EBUSY;
5799
5800         if (rdev->saved_raid_disk < 0 && has_failed(conf))
5801                 /* no point adding a device */
5802                 return -EINVAL;
5803
5804         if (rdev->raid_disk >= 0)
5805                 first = last = rdev->raid_disk;
5806
5807         /*
5808          * find the disk ... but prefer rdev->saved_raid_disk
5809          * if possible.
5810          */
5811         if (rdev->saved_raid_disk >= 0 &&
5812             rdev->saved_raid_disk >= first &&
5813             conf->disks[rdev->saved_raid_disk].rdev == NULL)
5814                 first = rdev->saved_raid_disk;
5815
5816         for (disk = first; disk <= last; disk++) {
5817                 p = conf->disks + disk;
5818                 if (p->rdev == NULL) {
5819                         clear_bit(In_sync, &rdev->flags);
5820                         rdev->raid_disk = disk;
5821                         err = 0;
5822                         if (rdev->saved_raid_disk != disk)
5823                                 conf->fullsync = 1;
5824                         rcu_assign_pointer(p->rdev, rdev);
5825                         goto out;
5826                 }
5827         }
5828         for (disk = first; disk <= last; disk++) {
5829                 p = conf->disks + disk;
5830                 if (test_bit(WantReplacement, &p->rdev->flags) &&
5831                     p->replacement == NULL) {
5832                         clear_bit(In_sync, &rdev->flags);
5833                         set_bit(Replacement, &rdev->flags);
5834                         rdev->raid_disk = disk;
5835                         err = 0;
5836                         conf->fullsync = 1;
5837                         rcu_assign_pointer(p->replacement, rdev);
5838                         break;
5839                 }
5840         }
5841 out:
5842         print_raid5_conf(conf);
5843         return err;
5844 }
5845
5846 static int raid5_resize(struct mddev *mddev, sector_t sectors)
5847 {
5848         /* no resync is happening, and there is enough space
5849          * on all devices, so we can resize.
5850          * We need to make sure resync covers any new space.
5851          * If the array is shrinking we should possibly wait until
5852          * any io in the removed space completes, but it hardly seems
5853          * worth it.
5854          */
5855         sector_t newsize;
5856         sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5857         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
5858         if (mddev->external_size &&
5859             mddev->array_sectors > newsize)
5860                 return -EINVAL;
5861         if (mddev->bitmap) {
5862                 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
5863                 if (ret)
5864                         return ret;
5865         }
5866         md_set_array_sectors(mddev, newsize);
5867         set_capacity(mddev->gendisk, mddev->array_sectors);
5868         revalidate_disk(mddev->gendisk);
5869         if (sectors > mddev->dev_sectors &&
5870             mddev->recovery_cp > mddev->dev_sectors) {
5871                 mddev->recovery_cp = mddev->dev_sectors;
5872                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5873         }
5874         mddev->dev_sectors = sectors;
5875         mddev->resync_max_sectors = sectors;
5876         return 0;
5877 }
5878
5879 static int check_stripe_cache(struct mddev *mddev)
5880 {
5881         /* Can only proceed if there are plenty of stripe_heads.
5882          * We need a minimum of one full stripe,, and for sensible progress
5883          * it is best to have about 4 times that.
5884          * If we require 4 times, then the default 256 4K stripe_heads will
5885          * allow for chunk sizes up to 256K, which is probably OK.
5886          * If the chunk size is greater, user-space should request more
5887          * stripe_heads first.
5888          */
5889         struct r5conf *conf = mddev->private;
5890         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5891             > conf->max_nr_stripes ||
5892             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5893             > conf->max_nr_stripes) {
5894                 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
5895                        mdname(mddev),
5896                        ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5897                         / STRIPE_SIZE)*4);
5898                 return 0;
5899         }
5900         return 1;
5901 }
5902
5903 static int check_reshape(struct mddev *mddev)
5904 {
5905         struct r5conf *conf = mddev->private;
5906
5907         if (mddev->delta_disks == 0 &&
5908             mddev->new_layout == mddev->layout &&
5909             mddev->new_chunk_sectors == mddev->chunk_sectors)
5910                 return 0; /* nothing to do */
5911         if (has_failed(conf))
5912                 return -EINVAL;
5913         if (mddev->delta_disks < 0) {
5914                 /* We might be able to shrink, but the devices must
5915                  * be made bigger first.
5916                  * For raid6, 4 is the minimum size.
5917                  * Otherwise 2 is the minimum
5918                  */
5919                 int min = 2;
5920                 if (mddev->level == 6)
5921                         min = 4;
5922                 if (mddev->raid_disks + mddev->delta_disks < min)
5923                         return -EINVAL;
5924         }
5925
5926         if (!check_stripe_cache(mddev))
5927                 return -ENOSPC;
5928
5929         return resize_stripes(conf, (conf->previous_raid_disks
5930                                      + mddev->delta_disks));
5931 }
5932
5933 static int raid5_start_reshape(struct mddev *mddev)
5934 {
5935         struct r5conf *conf = mddev->private;
5936         struct md_rdev *rdev;
5937         int spares = 0;
5938         unsigned long flags;
5939
5940         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5941                 return -EBUSY;
5942
5943         if (!check_stripe_cache(mddev))
5944                 return -ENOSPC;
5945
5946         if (has_failed(conf))
5947                 return -EINVAL;
5948
5949         rdev_for_each(rdev, mddev) {
5950                 if (!test_bit(In_sync, &rdev->flags)
5951                     && !test_bit(Faulty, &rdev->flags))
5952                         spares++;
5953         }
5954
5955         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5956                 /* Not enough devices even to make a degraded array
5957                  * of that size
5958                  */
5959                 return -EINVAL;
5960
5961         /* Refuse to reduce size of the array.  Any reductions in
5962          * array size must be through explicit setting of array_size
5963          * attribute.
5964          */
5965         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5966             < mddev->array_sectors) {
5967                 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5968                        "before number of disks\n", mdname(mddev));
5969                 return -EINVAL;
5970         }
5971
5972         atomic_set(&conf->reshape_stripes, 0);
5973         spin_lock_irq(&conf->device_lock);
5974         conf->previous_raid_disks = conf->raid_disks;
5975         conf->raid_disks += mddev->delta_disks;
5976         conf->prev_chunk_sectors = conf->chunk_sectors;
5977         conf->chunk_sectors = mddev->new_chunk_sectors;
5978         conf->prev_algo = conf->algorithm;
5979         conf->algorithm = mddev->new_layout;
5980         conf->generation++;
5981         /* Code that selects data_offset needs to see the generation update
5982          * if reshape_progress has been set - so a memory barrier needed.
5983          */
5984         smp_mb();
5985         if (mddev->reshape_backwards)
5986                 conf->reshape_progress = raid5_size(mddev, 0, 0);
5987         else
5988                 conf->reshape_progress = 0;
5989         conf->reshape_safe = conf->reshape_progress;
5990         spin_unlock_irq(&conf->device_lock);
5991
5992         /* Add some new drives, as many as will fit.
5993          * We know there are enough to make the newly sized array work.
5994          * Don't add devices if we are reducing the number of
5995          * devices in the array.  This is because it is not possible
5996          * to correctly record the "partially reconstructed" state of
5997          * such devices during the reshape and confusion could result.
5998          */
5999         if (mddev->delta_disks >= 0) {
6000                 rdev_for_each(rdev, mddev)
6001                         if (rdev->raid_disk < 0 &&
6002                             !test_bit(Faulty, &rdev->flags)) {
6003                                 if (raid5_add_disk(mddev, rdev) == 0) {
6004                                         if (rdev->raid_disk
6005                                             >= conf->previous_raid_disks)
6006                                                 set_bit(In_sync, &rdev->flags);
6007                                         else
6008                                                 rdev->recovery_offset = 0;
6009
6010                                         if (sysfs_link_rdev(mddev, rdev))
6011                                                 /* Failure here is OK */;
6012                                 }
6013                         } else if (rdev->raid_disk >= conf->previous_raid_disks
6014                                    && !test_bit(Faulty, &rdev->flags)) {
6015                                 /* This is a spare that was manually added */
6016                                 set_bit(In_sync, &rdev->flags);
6017                         }
6018
6019                 /* When a reshape changes the number of devices,
6020                  * ->degraded is measured against the larger of the
6021                  * pre and post number of devices.
6022                  */
6023                 spin_lock_irqsave(&conf->device_lock, flags);
6024                 mddev->degraded = calc_degraded(conf);
6025                 spin_unlock_irqrestore(&conf->device_lock, flags);
6026         }
6027         mddev->raid_disks = conf->raid_disks;
6028         mddev->reshape_position = conf->reshape_progress;
6029         set_bit(MD_CHANGE_DEVS, &mddev->flags);
6030
6031         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6032         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6033         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6034         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6035         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6036                                                 "reshape");
6037         if (!mddev->sync_thread) {
6038                 mddev->recovery = 0;
6039                 spin_lock_irq(&conf->device_lock);
6040                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
6041                 rdev_for_each(rdev, mddev)
6042                         rdev->new_data_offset = rdev->data_offset;
6043                 smp_wmb();
6044                 conf->reshape_progress = MaxSector;
6045                 mddev->reshape_position = MaxSector;
6046                 spin_unlock_irq(&conf->device_lock);
6047                 return -EAGAIN;
6048         }
6049         conf->reshape_checkpoint = jiffies;
6050         md_wakeup_thread(mddev->sync_thread);
6051         md_new_event(mddev);
6052         return 0;
6053 }
6054
6055 /* This is called from the reshape thread and should make any
6056  * changes needed in 'conf'
6057  */
6058 static void end_reshape(struct r5conf *conf)
6059 {
6060
6061         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
6062                 struct md_rdev *rdev;
6063
6064                 spin_lock_irq(&conf->device_lock);
6065                 conf->previous_raid_disks = conf->raid_disks;
6066                 rdev_for_each(rdev, conf->mddev)
6067                         rdev->data_offset = rdev->new_data_offset;
6068                 smp_wmb();
6069                 conf->reshape_progress = MaxSector;
6070                 spin_unlock_irq(&conf->device_lock);
6071                 wake_up(&conf->wait_for_overlap);
6072
6073                 /* read-ahead size must cover two whole stripes, which is
6074                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6075                  */
6076                 if (conf->mddev->queue) {
6077                         int data_disks = conf->raid_disks - conf->max_degraded;
6078                         int stripe = data_disks * ((conf->chunk_sectors << 9)
6079                                                    / PAGE_SIZE);
6080                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6081                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6082                 }
6083         }
6084 }
6085
6086 /* This is called from the raid5d thread with mddev_lock held.
6087  * It makes config changes to the device.
6088  */
6089 static void raid5_finish_reshape(struct mddev *mddev)
6090 {
6091         struct r5conf *conf = mddev->private;
6092
6093         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
6094
6095                 if (mddev->delta_disks > 0) {
6096                         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6097                         set_capacity(mddev->gendisk, mddev->array_sectors);
6098                         revalidate_disk(mddev->gendisk);
6099                 } else {
6100                         int d;
6101                         spin_lock_irq(&conf->device_lock);
6102                         mddev->degraded = calc_degraded(conf);
6103                         spin_unlock_irq(&conf->device_lock);
6104                         for (d = conf->raid_disks ;
6105                              d < conf->raid_disks - mddev->delta_disks;
6106                              d++) {
6107                                 struct md_rdev *rdev = conf->disks[d].rdev;
6108                                 if (rdev)
6109                                         clear_bit(In_sync, &rdev->flags);
6110                                 rdev = conf->disks[d].replacement;
6111                                 if (rdev)
6112                                         clear_bit(In_sync, &rdev->flags);
6113                         }
6114                 }
6115                 mddev->layout = conf->algorithm;
6116                 mddev->chunk_sectors = conf->chunk_sectors;
6117                 mddev->reshape_position = MaxSector;
6118                 mddev->delta_disks = 0;
6119                 mddev->reshape_backwards = 0;
6120         }
6121 }
6122
6123 static void raid5_quiesce(struct mddev *mddev, int state)
6124 {
6125         struct r5conf *conf = mddev->private;
6126
6127         switch(state) {
6128         case 2: /* resume for a suspend */
6129                 wake_up(&conf->wait_for_overlap);
6130                 break;
6131
6132         case 1: /* stop all writes */
6133                 spin_lock_irq(&conf->device_lock);
6134                 /* '2' tells resync/reshape to pause so that all
6135                  * active stripes can drain
6136                  */
6137                 conf->quiesce = 2;
6138                 wait_event_lock_irq(conf->wait_for_stripe,
6139                                     atomic_read(&conf->active_stripes) == 0 &&
6140                                     atomic_read(&conf->active_aligned_reads) == 0,
6141                                     conf->device_lock);
6142                 conf->quiesce = 1;
6143                 spin_unlock_irq(&conf->device_lock);
6144                 /* allow reshape to continue */
6145                 wake_up(&conf->wait_for_overlap);
6146                 break;
6147
6148         case 0: /* re-enable writes */
6149                 spin_lock_irq(&conf->device_lock);
6150                 conf->quiesce = 0;
6151                 wake_up(&conf->wait_for_stripe);
6152                 wake_up(&conf->wait_for_overlap);
6153                 spin_unlock_irq(&conf->device_lock);
6154                 break;
6155         }
6156 }
6157
6158
6159 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
6160 {
6161         struct r0conf *raid0_conf = mddev->private;
6162         sector_t sectors;
6163
6164         /* for raid0 takeover only one zone is supported */
6165         if (raid0_conf->nr_strip_zones > 1) {
6166                 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6167                        mdname(mddev));
6168                 return ERR_PTR(-EINVAL);
6169         }
6170
6171         sectors = raid0_conf->strip_zone[0].zone_end;
6172         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
6173         mddev->dev_sectors = sectors;
6174         mddev->new_level = level;
6175         mddev->new_layout = ALGORITHM_PARITY_N;
6176         mddev->new_chunk_sectors = mddev->chunk_sectors;
6177         mddev->raid_disks += 1;
6178         mddev->delta_disks = 1;
6179         /* make sure it will be not marked as dirty */
6180         mddev->recovery_cp = MaxSector;
6181
6182         return setup_conf(mddev);
6183 }
6184
6185
6186 static void *raid5_takeover_raid1(struct mddev *mddev)
6187 {
6188         int chunksect;
6189
6190         if (mddev->raid_disks != 2 ||
6191             mddev->degraded > 1)
6192                 return ERR_PTR(-EINVAL);
6193
6194         /* Should check if there are write-behind devices? */
6195
6196         chunksect = 64*2; /* 64K by default */
6197
6198         /* The array must be an exact multiple of chunksize */
6199         while (chunksect && (mddev->array_sectors & (chunksect-1)))
6200                 chunksect >>= 1;
6201
6202         if ((chunksect<<9) < STRIPE_SIZE)
6203                 /* array size does not allow a suitable chunk size */
6204                 return ERR_PTR(-EINVAL);
6205
6206         mddev->new_level = 5;
6207         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
6208         mddev->new_chunk_sectors = chunksect;
6209
6210         return setup_conf(mddev);
6211 }
6212
6213 static void *raid5_takeover_raid6(struct mddev *mddev)
6214 {
6215         int new_layout;
6216
6217         switch (mddev->layout) {
6218         case ALGORITHM_LEFT_ASYMMETRIC_6:
6219                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
6220                 break;
6221         case ALGORITHM_RIGHT_ASYMMETRIC_6:
6222                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
6223                 break;
6224         case ALGORITHM_LEFT_SYMMETRIC_6:
6225                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
6226                 break;
6227         case ALGORITHM_RIGHT_SYMMETRIC_6:
6228                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
6229                 break;
6230         case ALGORITHM_PARITY_0_6:
6231                 new_layout = ALGORITHM_PARITY_0;
6232                 break;
6233         case ALGORITHM_PARITY_N:
6234                 new_layout = ALGORITHM_PARITY_N;
6235                 break;
6236         default:
6237                 return ERR_PTR(-EINVAL);
6238         }
6239         mddev->new_level = 5;
6240         mddev->new_layout = new_layout;
6241         mddev->delta_disks = -1;
6242         mddev->raid_disks -= 1;
6243         return setup_conf(mddev);
6244 }
6245
6246
6247 static int raid5_check_reshape(struct mddev *mddev)
6248 {
6249         /* For a 2-drive array, the layout and chunk size can be changed
6250          * immediately as not restriping is needed.
6251          * For larger arrays we record the new value - after validation
6252          * to be used by a reshape pass.
6253          */
6254         struct r5conf *conf = mddev->private;
6255         int new_chunk = mddev->new_chunk_sectors;
6256
6257         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
6258                 return -EINVAL;
6259         if (new_chunk > 0) {
6260                 if (!is_power_of_2(new_chunk))
6261                         return -EINVAL;
6262                 if (new_chunk < (PAGE_SIZE>>9))
6263                         return -EINVAL;
6264                 if (mddev->array_sectors & (new_chunk-1))
6265                         /* not factor of array size */
6266                         return -EINVAL;
6267         }
6268
6269         /* They look valid */
6270
6271         if (mddev->raid_disks == 2) {
6272                 /* can make the change immediately */
6273                 if (mddev->new_layout >= 0) {
6274                         conf->algorithm = mddev->new_layout;
6275                         mddev->layout = mddev->new_layout;
6276                 }
6277                 if (new_chunk > 0) {
6278                         conf->chunk_sectors = new_chunk ;
6279                         mddev->chunk_sectors = new_chunk;
6280                 }
6281                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6282                 md_wakeup_thread(mddev->thread);
6283         }
6284         return check_reshape(mddev);
6285 }
6286
6287 static int raid6_check_reshape(struct mddev *mddev)
6288 {
6289         int new_chunk = mddev->new_chunk_sectors;
6290
6291         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
6292                 return -EINVAL;
6293         if (new_chunk > 0) {
6294                 if (!is_power_of_2(new_chunk))
6295                         return -EINVAL;
6296                 if (new_chunk < (PAGE_SIZE >> 9))
6297                         return -EINVAL;
6298                 if (mddev->array_sectors & (new_chunk-1))
6299                         /* not factor of array size */
6300                         return -EINVAL;
6301         }
6302
6303         /* They look valid */
6304         return check_reshape(mddev);
6305 }
6306
6307 static void *raid5_takeover(struct mddev *mddev)
6308 {
6309         /* raid5 can take over:
6310          *  raid0 - if there is only one strip zone - make it a raid4 layout
6311          *  raid1 - if there are two drives.  We need to know the chunk size
6312          *  raid4 - trivial - just use a raid4 layout.
6313          *  raid6 - Providing it is a *_6 layout
6314          */
6315         if (mddev->level == 0)
6316                 return raid45_takeover_raid0(mddev, 5);
6317         if (mddev->level == 1)
6318                 return raid5_takeover_raid1(mddev);
6319         if (mddev->level == 4) {
6320                 mddev->new_layout = ALGORITHM_PARITY_N;
6321                 mddev->new_level = 5;
6322                 return setup_conf(mddev);
6323         }
6324         if (mddev->level == 6)
6325                 return raid5_takeover_raid6(mddev);
6326
6327         return ERR_PTR(-EINVAL);
6328 }
6329
6330 static void *raid4_takeover(struct mddev *mddev)
6331 {
6332         /* raid4 can take over:
6333          *  raid0 - if there is only one strip zone
6334          *  raid5 - if layout is right
6335          */
6336         if (mddev->level == 0)
6337                 return raid45_takeover_raid0(mddev, 4);
6338         if (mddev->level == 5 &&
6339             mddev->layout == ALGORITHM_PARITY_N) {
6340                 mddev->new_layout = 0;
6341                 mddev->new_level = 4;
6342                 return setup_conf(mddev);
6343         }
6344         return ERR_PTR(-EINVAL);
6345 }
6346
6347 static struct md_personality raid5_personality;
6348
6349 static void *raid6_takeover(struct mddev *mddev)
6350 {
6351         /* Currently can only take over a raid5.  We map the
6352          * personality to an equivalent raid6 personality
6353          * with the Q block at the end.
6354          */
6355         int new_layout;
6356
6357         if (mddev->pers != &raid5_personality)
6358                 return ERR_PTR(-EINVAL);
6359         if (mddev->degraded > 1)
6360                 return ERR_PTR(-EINVAL);
6361         if (mddev->raid_disks > 253)
6362                 return ERR_PTR(-EINVAL);
6363         if (mddev->raid_disks < 3)
6364                 return ERR_PTR(-EINVAL);
6365
6366         switch (mddev->layout) {
6367         case ALGORITHM_LEFT_ASYMMETRIC:
6368                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
6369                 break;
6370         case ALGORITHM_RIGHT_ASYMMETRIC:
6371                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
6372                 break;
6373         case ALGORITHM_LEFT_SYMMETRIC:
6374                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
6375                 break;
6376         case ALGORITHM_RIGHT_SYMMETRIC:
6377                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
6378                 break;
6379         case ALGORITHM_PARITY_0:
6380                 new_layout = ALGORITHM_PARITY_0_6;
6381                 break;
6382         case ALGORITHM_PARITY_N:
6383                 new_layout = ALGORITHM_PARITY_N;
6384                 break;
6385         default:
6386                 return ERR_PTR(-EINVAL);
6387         }
6388         mddev->new_level = 6;
6389         mddev->new_layout = new_layout;
6390         mddev->delta_disks = 1;
6391         mddev->raid_disks += 1;
6392         return setup_conf(mddev);
6393 }
6394
6395
6396 static struct md_personality raid6_personality =
6397 {
6398         .name           = "raid6",
6399         .level          = 6,
6400         .owner          = THIS_MODULE,
6401         .make_request   = make_request,
6402         .run            = run,
6403         .stop           = stop,
6404         .status         = status,
6405         .error_handler  = error,
6406         .hot_add_disk   = raid5_add_disk,
6407         .hot_remove_disk= raid5_remove_disk,
6408         .spare_active   = raid5_spare_active,
6409         .sync_request   = sync_request,
6410         .resize         = raid5_resize,
6411         .size           = raid5_size,
6412         .check_reshape  = raid6_check_reshape,
6413         .start_reshape  = raid5_start_reshape,
6414         .finish_reshape = raid5_finish_reshape,
6415         .quiesce        = raid5_quiesce,
6416         .takeover       = raid6_takeover,
6417 };
6418 static struct md_personality raid5_personality =
6419 {
6420         .name           = "raid5",
6421         .level          = 5,
6422         .owner          = THIS_MODULE,
6423         .make_request   = make_request,
6424         .run            = run,
6425         .stop           = stop,
6426         .status         = status,
6427         .error_handler  = error,
6428         .hot_add_disk   = raid5_add_disk,
6429         .hot_remove_disk= raid5_remove_disk,
6430         .spare_active   = raid5_spare_active,
6431         .sync_request   = sync_request,
6432         .resize         = raid5_resize,
6433         .size           = raid5_size,
6434         .check_reshape  = raid5_check_reshape,
6435         .start_reshape  = raid5_start_reshape,
6436         .finish_reshape = raid5_finish_reshape,
6437         .quiesce        = raid5_quiesce,
6438         .takeover       = raid5_takeover,
6439 };
6440
6441 static struct md_personality raid4_personality =
6442 {
6443         .name           = "raid4",
6444         .level          = 4,
6445         .owner          = THIS_MODULE,
6446         .make_request   = make_request,
6447         .run            = run,
6448         .stop           = stop,
6449         .status         = status,
6450         .error_handler  = error,
6451         .hot_add_disk   = raid5_add_disk,
6452         .hot_remove_disk= raid5_remove_disk,
6453         .spare_active   = raid5_spare_active,
6454         .sync_request   = sync_request,
6455         .resize         = raid5_resize,
6456         .size           = raid5_size,
6457         .check_reshape  = raid5_check_reshape,
6458         .start_reshape  = raid5_start_reshape,
6459         .finish_reshape = raid5_finish_reshape,
6460         .quiesce        = raid5_quiesce,
6461         .takeover       = raid4_takeover,
6462 };
6463
6464 static int __init raid5_init(void)
6465 {
6466         register_md_personality(&raid6_personality);
6467         register_md_personality(&raid5_personality);
6468         register_md_personality(&raid4_personality);
6469         return 0;
6470 }
6471
6472 static void raid5_exit(void)
6473 {
6474         unregister_md_personality(&raid6_personality);
6475         unregister_md_personality(&raid5_personality);
6476         unregister_md_personality(&raid4_personality);
6477 }
6478
6479 module_init(raid5_init);
6480 module_exit(raid5_exit);
6481 MODULE_LICENSE("GPL");
6482 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6483 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6484 MODULE_ALIAS("md-raid5");
6485 MODULE_ALIAS("md-raid4");
6486 MODULE_ALIAS("md-level-5");
6487 MODULE_ALIAS("md-level-4");
6488 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6489 MODULE_ALIAS("md-raid6");
6490 MODULE_ALIAS("md-level-6");
6491
6492 /* This used to be two separate modules, they were: */
6493 MODULE_ALIAS("raid5");
6494 MODULE_ALIAS("raid6");