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->bm_write is the number of the last batch successfully written.
  31  * conf->bm_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 bm_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/module.h>
  47 #include <linux/slab.h>
  48 #include <linux/highmem.h>
  49 #include <linux/bitops.h>
  50 #include <linux/kthread.h>
  51 #include <asm/atomic.h>
  52 #include "raid6.h"
  53
  54 #include <linux/raid/bitmap.h>
  55 #include <linux/async_tx.h>
  56
  57 /*
  58  * Stripe cache
  59  */
  60
  61 #define NR_STRIPES              256
  62 #define STRIPE_SIZE             PAGE_SIZE
  63 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
  64 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
  65 #define IO_THRESHOLD            1
  66 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
  67 #define HASH_MASK               (NR_HASH - 1)
  68
  69 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
  70
  71 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
  72  * order without overlap.  There may be several bio's per stripe+device, and
  73  * a bio could span several devices.
  74  * When walking this list for a particular stripe+device, we must never proceed
  75  * beyond a bio that extends past this device, as the next bio might no longer
  76  * be valid.
  77  * This macro is used to determine the 'next' bio in the list, given the sector
  78  * of the current stripe+device
  79  */
  80 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
  81 /*
  82  * The following can be used to debug the driver
  83  */
  84 #define RAID5_PARANOIA  1
  85 #if RAID5_PARANOIA && defined(CONFIG_SMP)
  86 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
  87 #else
  88 # define CHECK_DEVLOCK()
  89 #endif
  90
  91 #ifdef DEBUG
  92 #define inline
  93 #define __inline__
  94 #endif
  95
  96 #if !RAID6_USE_EMPTY_ZERO_PAGE
  97 /* In .bss so it's zeroed */
  98 const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
  99 #endif
 100
 101 static inline int raid6_next_disk(int disk, int raid_disks)
 102 {
 103         disk++;
 104         return (disk < raid_disks) ? disk : 0;
 105 }
 106
 107 static void return_io(struct bio *return_bi)
 108 {
 109         struct bio *bi = return_bi;
 110         while (bi) {
 111
 112                 return_bi = bi->bi_next;
 113                 bi->bi_next = NULL;
 114                 bi->bi_size = 0;
 115                 bi->bi_end_io(bi,
 116                               test_bit(BIO_UPTODATE, &bi->bi_flags)
 117                                 ? 0 : -EIO);
 118                 bi = return_bi;
 119         }
 120 }
 121
 122 static void print_raid5_conf (raid5_conf_t *conf);
 123
 124 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
 125 {
 126         if (atomic_dec_and_test(&sh->count)) {
 127                 BUG_ON(!list_empty(&sh->lru));
 128                 BUG_ON(atomic_read(&conf->active_stripes)==0);
 129                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
 130                         if (test_bit(STRIPE_DELAYED, &sh->state)) {
 131                                 list_add_tail(&sh->lru, &conf->delayed_list);
 132                                 blk_plug_device(conf->mddev->queue);
 133                         } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 134                                    sh->bm_seq - conf->seq_write > 0) {
 135                                 list_add_tail(&sh->lru, &conf->bitmap_list);
 136                                 blk_plug_device(conf->mddev->queue);
 137                         } else {
 138                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
 139                                 list_add_tail(&sh->lru, &conf->handle_list);
 140                         }
 141                         md_wakeup_thread(conf->mddev->thread);
 142                 } else {
 143                         BUG_ON(sh->ops.pending);
 144                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
 145                                 atomic_dec(&conf->preread_active_stripes);
 146                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
 147                                         md_wakeup_thread(conf->mddev->thread);
 148                         }
 149                         atomic_dec(&conf->active_stripes);
 150                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
 151                                 list_add_tail(&sh->lru, &conf->inactive_list);
 152                                 wake_up(&conf->wait_for_stripe);
 153                                 if (conf->retry_read_aligned)
 154                                         md_wakeup_thread(conf->mddev->thread);
 155                         }
 156                 }
 157         }
 158 }
 159 static void release_stripe(struct stripe_head *sh)
 160 {
 161         raid5_conf_t *conf = sh->raid_conf;
 162         unsigned long flags;
 163
 164         spin_lock_irqsave(&conf->device_lock, flags);
 165         __release_stripe(conf, sh);
 166         spin_unlock_irqrestore(&conf->device_lock, flags);
 167 }
 168
 169 static inline void remove_hash(struct stripe_head *sh)
 170 {
 171         pr_debug("remove_hash(), stripe %llu\n",
 172                 (unsigned long long)sh->sector);
 173
 174         hlist_del_init(&sh->hash);
 175 }
 176
 177 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
 178 {
 179         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 180
 181         pr_debug("insert_hash(), stripe %llu\n",
 182                 (unsigned long long)sh->sector);
 183
 184         CHECK_DEVLOCK();
 185         hlist_add_head(&sh->hash, hp);
 186 }
 187
 188
 189 /* find an idle stripe, make sure it is unhashed, and return it. */
 190 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
 191 {
 192         struct stripe_head *sh = NULL;
 193         struct list_head *first;
 194
 195         CHECK_DEVLOCK();
 196         if (list_empty(&conf->inactive_list))
 197                 goto out;
 198         first = conf->inactive_list.next;
 199         sh = list_entry(first, struct stripe_head, lru);
 200         list_del_init(first);
 201         remove_hash(sh);
 202         atomic_inc(&conf->active_stripes);
 203 out:
 204         return sh;
 205 }
 206
 207 static void shrink_buffers(struct stripe_head *sh, int num)
 208 {
 209         struct page *p;
 210         int i;
 211
 212         for (i=0; i<num ; i++) {
 213                 p = sh->dev[i].page;
 214                 if (!p)
 215                         continue;
 216                 sh->dev[i].page = NULL;
 217                 put_page(p);
 218         }
 219 }
 220
 221 static int grow_buffers(struct stripe_head *sh, int num)
 222 {
 223         int i;
 224
 225         for (i=0; i<num; i++) {
 226                 struct page *page;
 227
 228                 if (!(page = alloc_page(GFP_KERNEL))) {
 229                         return 1;
 230                 }
 231                 sh->dev[i].page = page;
 232         }
 233         return 0;
 234 }
 235
 236 static void raid5_build_block (struct stripe_head *sh, int i);
 237
 238 static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int disks)
 239 {
 240         raid5_conf_t *conf = sh->raid_conf;
 241         int i;
 242
 243         BUG_ON(atomic_read(&sh->count) != 0);
 244         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 245         BUG_ON(sh->ops.pending || sh->ops.ack || sh->ops.complete);
 246
 247         CHECK_DEVLOCK();
 248         pr_debug("init_stripe called, stripe %llu\n",
 249                 (unsigned long long)sh->sector);
 250
 251         remove_hash(sh);
 252
 253         sh->sector = sector;
 254         sh->pd_idx = pd_idx;
 255         sh->state = 0;
 256
 257         sh->disks = disks;
 258
 259         for (i = sh->disks; i--; ) {
 260                 struct r5dev *dev = &sh->dev[i];
 261
 262                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 263                     test_bit(R5_LOCKED, &dev->flags)) {
 264                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
 265                                (unsigned long long)sh->sector, i, dev->toread,
 266                                dev->read, dev->towrite, dev->written,
 267                                test_bit(R5_LOCKED, &dev->flags));
 268                         BUG();
 269                 }
 270                 dev->flags = 0;
 271                 raid5_build_block(sh, i);
 272         }
 273         insert_hash(conf, sh);
 274 }
 275
 276 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
 277 {
 278         struct stripe_head *sh;
 279         struct hlist_node *hn;
 280
 281         CHECK_DEVLOCK();
 282         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 283         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
 284                 if (sh->sector == sector && sh->disks == disks)
 285                         return sh;
 286         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 287         return NULL;
 288 }
 289
 290 static void unplug_slaves(mddev_t *mddev);
 291 static void raid5_unplug_device(struct request_queue *q);
 292
 293 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector, int disks,
 294                                              int pd_idx, int noblock)
 295 {
 296         struct stripe_head *sh;
 297
 298         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 299
 300         spin_lock_irq(&conf->device_lock);
 301
 302         do {
 303                 wait_event_lock_irq(conf->wait_for_stripe,
 304                                     conf->quiesce == 0,
 305                                     conf->device_lock, /* nothing */);
 306                 sh = __find_stripe(conf, sector, disks);
 307                 if (!sh) {
 308                         if (!conf->inactive_blocked)
 309                                 sh = get_free_stripe(conf);
 310                         if (noblock && sh == NULL)
 311                                 break;
 312                         if (!sh) {
 313                                 conf->inactive_blocked = 1;
 314                                 wait_event_lock_irq(conf->wait_for_stripe,
 315                                                     !list_empty(&conf->inactive_list) &&
 316                                                     (atomic_read(&conf->active_stripes)
 317                                                      < (conf->max_nr_stripes *3/4)
 318                                                      || !conf->inactive_blocked),
 319                                                     conf->device_lock,
 320                                                     raid5_unplug_device(conf->mddev->queue)
 321                                         );
 322                                 conf->inactive_blocked = 0;
 323                         } else
 324                                 init_stripe(sh, sector, pd_idx, disks);
 325                 } else {
 326                         if (atomic_read(&sh->count)) {
 327                           BUG_ON(!list_empty(&sh->lru));
 328                         } else {
 329                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 330                                         atomic_inc(&conf->active_stripes);
 331                                 if (list_empty(&sh->lru) &&
 332                                     !test_bit(STRIPE_EXPANDING, &sh->state))
 333                                         BUG();
 334                                 list_del_init(&sh->lru);
 335                         }
 336                 }
 337         } while (sh == NULL);
 338
 339         if (sh)
 340                 atomic_inc(&sh->count);
 341
 342         spin_unlock_irq(&conf->device_lock);
 343         return sh;
 344 }
 345
 346 /* test_and_ack_op() ensures that we only dequeue an operation once */
 347 #define test_and_ack_op(op, pend) \
 348 do {                                                    \
 349         if (test_bit(op, &sh->ops.pending) &&           \
 350                 !test_bit(op, &sh->ops.complete)) {     \
 351                 if (test_and_set_bit(op, &sh->ops.ack)) \
 352                         clear_bit(op, &pend);           \
 353                 else                                    \
 354                         ack++;                          \
 355         } else                                          \
 356                 clear_bit(op, &pend);                   \
 357 } while (0)
 358
 359 /* find new work to run, do not resubmit work that is already
 360  * in flight
 361  */
 362 static unsigned long get_stripe_work(struct stripe_head *sh)
 363 {
 364         unsigned long pending;
 365         int ack = 0;
 366
 367         pending = sh->ops.pending;
 368
 369         test_and_ack_op(STRIPE_OP_BIOFILL, pending);
 370         test_and_ack_op(STRIPE_OP_COMPUTE_BLK, pending);
 371         test_and_ack_op(STRIPE_OP_PREXOR, pending);
 372         test_and_ack_op(STRIPE_OP_BIODRAIN, pending);
 373         test_and_ack_op(STRIPE_OP_POSTXOR, pending);
 374         test_and_ack_op(STRIPE_OP_CHECK, pending);
 375         if (test_and_clear_bit(STRIPE_OP_IO, &sh->ops.pending))
 376                 ack++;
 377
 378         sh->ops.count -= ack;
 379         if (unlikely(sh->ops.count < 0)) {
 380                 printk(KERN_ERR "pending: %#lx ops.pending: %#lx ops.ack: %#lx "
 381                         "ops.complete: %#lx\n", pending, sh->ops.pending,
 382                         sh->ops.ack, sh->ops.complete);
 383                 BUG();
 384         }
 385
 386         return pending;
 387 }
 388
 389 static void
 390 raid5_end_read_request(struct bio *bi, int error);
 391 static void
 392 raid5_end_write_request(struct bio *bi, int error);
 393
 394 static void ops_run_io(struct stripe_head *sh)
 395 {
 396         raid5_conf_t *conf = sh->raid_conf;
 397         int i, disks = sh->disks;
 398
 399         might_sleep();
 400
 401         for (i = disks; i--; ) {
 402                 int rw;
 403                 struct bio *bi;
 404                 mdk_rdev_t *rdev;
 405                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
 406                         rw = WRITE;
 407                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
 408                         rw = READ;
 409                 else
 410                         continue;
 411
 412                 bi = &sh->dev[i].req;
 413
 414                 bi->bi_rw = rw;
 415                 if (rw == WRITE)
 416                         bi->bi_end_io = raid5_end_write_request;
 417                 else
 418                         bi->bi_end_io = raid5_end_read_request;
 419
 420                 rcu_read_lock();
 421                 rdev = rcu_dereference(conf->disks[i].rdev);
 422                 if (rdev && test_bit(Faulty, &rdev->flags))
 423                         rdev = NULL;
 424                 if (rdev)
 425                         atomic_inc(&rdev->nr_pending);
 426                 rcu_read_unlock();
 427
 428                 if (rdev) {
 429                         if (test_bit(STRIPE_SYNCING, &sh->state) ||
 430                                 test_bit(STRIPE_EXPAND_SOURCE, &sh->state) ||
 431                                 test_bit(STRIPE_EXPAND_READY, &sh->state))
 432                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
 433
 434                         bi->bi_bdev = rdev->bdev;
 435                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
 436                                 __FUNCTION__, (unsigned long long)sh->sector,
 437                                 bi->bi_rw, i);
 438                         atomic_inc(&sh->count);
 439                         bi->bi_sector = sh->sector + rdev->data_offset;
 440                         bi->bi_flags = 1 << BIO_UPTODATE;
 441                         bi->bi_vcnt = 1;
 442                         bi->bi_max_vecs = 1;
 443                         bi->bi_idx = 0;
 444                         bi->bi_io_vec = &sh->dev[i].vec;
 445                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
 446                         bi->bi_io_vec[0].bv_offset = 0;
 447                         bi->bi_size = STRIPE_SIZE;
 448                         bi->bi_next = NULL;
 449                         if (rw == WRITE &&
 450                             test_bit(R5_ReWrite, &sh->dev[i].flags))
 451                                 atomic_add(STRIPE_SECTORS,
 452                                         &rdev->corrected_errors);
 453                         generic_make_request(bi);
 454                 } else {
 455                         if (rw == WRITE)
 456                                 set_bit(STRIPE_DEGRADED, &sh->state);
 457                         pr_debug("skip op %ld on disc %d for sector %llu\n",
 458                                 bi->bi_rw, i, (unsigned long long)sh->sector);
 459                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 460                         set_bit(STRIPE_HANDLE, &sh->state);
 461                 }
 462         }
 463 }
 464
 465 static struct dma_async_tx_descriptor *
 466 async_copy_data(int frombio, struct bio *bio, struct page *page,
 467         sector_t sector, struct dma_async_tx_descriptor *tx)
 468 {
 469         struct bio_vec *bvl;
 470         struct page *bio_page;
 471         int i;
 472         int page_offset;
 473
 474         if (bio->bi_sector >= sector)
 475                 page_offset = (signed)(bio->bi_sector - sector) * 512;
 476         else
 477                 page_offset = (signed)(sector - bio->bi_sector) * -512;
 478         bio_for_each_segment(bvl, bio, i) {
 479                 int len = bio_iovec_idx(bio, i)->bv_len;
 480                 int clen;
 481                 int b_offset = 0;
 482
 483                 if (page_offset < 0) {
 484                         b_offset = -page_offset;
 485                         page_offset += b_offset;
 486                         len -= b_offset;
 487                 }
 488
 489                 if (len > 0 && page_offset + len > STRIPE_SIZE)
 490                         clen = STRIPE_SIZE - page_offset;
 491                 else
 492                         clen = len;
 493
 494                 if (clen > 0) {
 495                         b_offset += bio_iovec_idx(bio, i)->bv_offset;
 496                         bio_page = bio_iovec_idx(bio, i)->bv_page;
 497                         if (frombio)
 498                                 tx = async_memcpy(page, bio_page, page_offset,
 499                                         b_offset, clen,
 500                                         ASYNC_TX_DEP_ACK,
 501                                         tx, NULL, NULL);
 502                         else
 503                                 tx = async_memcpy(bio_page, page, b_offset,
 504                                         page_offset, clen,
 505                                         ASYNC_TX_DEP_ACK,
 506                                         tx, NULL, NULL);
 507                 }
 508                 if (clen < len) /* hit end of page */
 509                         break;
 510                 page_offset +=  len;
 511         }
 512
 513         return tx;
 514 }
 515
 516 static void ops_complete_biofill(void *stripe_head_ref)
 517 {
 518         struct stripe_head *sh = stripe_head_ref;
 519         struct bio *return_bi = NULL;
 520         raid5_conf_t *conf = sh->raid_conf;
 521         int i;
 522
 523         pr_debug("%s: stripe %llu\n", __FUNCTION__,
 524                 (unsigned long long)sh->sector);
 525
 526         /* clear completed biofills */
 527         for (i = sh->disks; i--; ) {
 528                 struct r5dev *dev = &sh->dev[i];
 529
 530                 /* acknowledge completion of a biofill operation */
 531                 /* and check if we need to reply to a read request,
 532                  * new R5_Wantfill requests are held off until
 533                  * !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending)
 534                  */
 535                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
 536                         struct bio *rbi, *rbi2;
 537
 538                         /* The access to dev->read is outside of the
 539                          * spin_lock_irq(&conf->device_lock), but is protected
 540                          * by the STRIPE_OP_BIOFILL pending bit
 541                          */
 542                         BUG_ON(!dev->read);
 543                         rbi = dev->read;
 544                         dev->read = NULL;
 545                         while (rbi && rbi->bi_sector <
 546                                 dev->sector + STRIPE_SECTORS) {
 547                                 rbi2 = r5_next_bio(rbi, dev->sector);
 548                                 spin_lock_irq(&conf->device_lock);
 549                                 if (--rbi->bi_phys_segments == 0) {
 550                                         rbi->bi_next = return_bi;
 551                                         return_bi = rbi;
 552                                 }
 553                                 spin_unlock_irq(&conf->device_lock);
 554                                 rbi = rbi2;
 555                         }
 556                 }
 557         }
 558         set_bit(STRIPE_OP_BIOFILL, &sh->ops.complete);
 559
 560         return_io(return_bi);
 561
 562         set_bit(STRIPE_HANDLE, &sh->state);
 563         release_stripe(sh);
 564 }
 565
 566 static void ops_run_biofill(struct stripe_head *sh)
 567 {
 568         struct dma_async_tx_descriptor *tx = NULL;
 569         raid5_conf_t *conf = sh->raid_conf;
 570         int i;
 571
 572         pr_debug("%s: stripe %llu\n", __FUNCTION__,
 573                 (unsigned long long)sh->sector);
 574
 575         for (i = sh->disks; i--; ) {
 576                 struct r5dev *dev = &sh->dev[i];
 577                 if (test_bit(R5_Wantfill, &dev->flags)) {
 578                         struct bio *rbi;
 579                         spin_lock_irq(&conf->device_lock);
 580                         dev->read = rbi = dev->toread;
 581                         dev->toread = NULL;
 582                         spin_unlock_irq(&conf->device_lock);
 583                         while (rbi && rbi->bi_sector <
 584                                 dev->sector + STRIPE_SECTORS) {
 585                                 tx = async_copy_data(0, rbi, dev->page,
 586                                         dev->sector, tx);
 587                                 rbi = r5_next_bio(rbi, dev->sector);
 588                         }
 589                 }
 590         }
 591
 592         atomic_inc(&sh->count);
 593         async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
 594                 ops_complete_biofill, sh);
 595 }
 596
 597 static void ops_complete_compute5(void *stripe_head_ref)
 598 {
 599         struct stripe_head *sh = stripe_head_ref;
 600         int target = sh->ops.target;
 601         struct r5dev *tgt = &sh->dev[target];
 602
 603         pr_debug("%s: stripe %llu\n", __FUNCTION__,
 604                 (unsigned long long)sh->sector);
 605
 606         set_bit(R5_UPTODATE, &tgt->flags);
 607         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
 608         clear_bit(R5_Wantcompute, &tgt->flags);
 609         set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
 610         set_bit(STRIPE_HANDLE, &sh->state);
 611         release_stripe(sh);
 612 }
 613
 614 static struct dma_async_tx_descriptor *
 615 ops_run_compute5(struct stripe_head *sh, unsigned long pending)
 616 {
 617         /* kernel stack size limits the total number of disks */
 618         int disks = sh->disks;
 619         struct page *xor_srcs[disks];
 620         int target = sh->ops.target;
 621         struct r5dev *tgt = &sh->dev[target];
 622         struct page *xor_dest = tgt->page;
 623         int count = 0;
 624         struct dma_async_tx_descriptor *tx;
 625         int i;
 626
 627         pr_debug("%s: stripe %llu block: %d\n",
 628                 __FUNCTION__, (unsigned long long)sh->sector, target);
 629         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
 630
 631         for (i = disks; i--; )
 632                 if (i != target)
 633                         xor_srcs[count++] = sh->dev[i].page;
 634
 635         atomic_inc(&sh->count);
 636
 637         if (unlikely(count == 1))
 638                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
 639                         0, NULL, ops_complete_compute5, sh);
 640         else
 641                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
 642                         ASYNC_TX_XOR_ZERO_DST, NULL,
 643                         ops_complete_compute5, sh);
 644
 645         /* ack now if postxor is not set to be run */
 646         if (tx && !test_bit(STRIPE_OP_POSTXOR, &pending))
 647                 async_tx_ack(tx);
 648
 649         return tx;
 650 }
 651
 652 static void ops_complete_prexor(void *stripe_head_ref)
 653 {
 654         struct stripe_head *sh = stripe_head_ref;
 655
 656         pr_debug("%s: stripe %llu\n", __FUNCTION__,
 657                 (unsigned long long)sh->sector);
 658
 659         set_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
 660 }
 661
 662 static struct dma_async_tx_descriptor *
 663 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
 664 {
 665         /* kernel stack size limits the total number of disks */
 666         int disks = sh->disks;
 667         struct page *xor_srcs[disks];
 668         int count = 0, pd_idx = sh->pd_idx, i;
 669
 670         /* existing parity data subtracted */
 671         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
 672
 673         pr_debug("%s: stripe %llu\n", __FUNCTION__,
 674                 (unsigned long long)sh->sector);
 675
 676         for (i = disks; i--; ) {
 677                 struct r5dev *dev = &sh->dev[i];
 678                 /* Only process blocks that are known to be uptodate */
 679                 if (dev->towrite && test_bit(R5_Wantprexor, &dev->flags))
 680                         xor_srcs[count++] = dev->page;
 681         }
 682
 683         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
 684                 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
 685                 ops_complete_prexor, sh);
 686
 687         return tx;
 688 }
 689
 690 static struct dma_async_tx_descriptor *
 691 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx,
 692                  unsigned long pending)
 693 {
 694         int disks = sh->disks;
 695         int pd_idx = sh->pd_idx, i;
 696
 697         /* check if prexor is active which means only process blocks
 698          * that are part of a read-modify-write (Wantprexor)
 699          */
 700         int prexor = test_bit(STRIPE_OP_PREXOR, &pending);
 701
 702         pr_debug("%s: stripe %llu\n", __FUNCTION__,
 703                 (unsigned long long)sh->sector);
 704
 705         for (i = disks; i--; ) {
 706                 struct r5dev *dev = &sh->dev[i];
 707                 struct bio *chosen;
 708                 int towrite;
 709
 710                 towrite = 0;
 711                 if (prexor) { /* rmw */
 712                         if (dev->towrite &&
 713                             test_bit(R5_Wantprexor, &dev->flags))
 714                                 towrite = 1;
 715                 } else { /* rcw */
 716                         if (i != pd_idx && dev->towrite &&
 717                                 test_bit(R5_LOCKED, &dev->flags))
 718                                 towrite = 1;
 719                 }
 720
 721                 if (towrite) {
 722                         struct bio *wbi;
 723
 724                         spin_lock(&sh->lock);
 725                         chosen = dev->towrite;
 726                         dev->towrite = NULL;
 727                         BUG_ON(dev->written);
 728                         wbi = dev->written = chosen;
 729                         spin_unlock(&sh->lock);
 730
 731                         while (wbi && wbi->bi_sector <
 732                                 dev->sector + STRIPE_SECTORS) {
 733                                 tx = async_copy_data(1, wbi, dev->page,
 734                                         dev->sector, tx);
 735                                 wbi = r5_next_bio(wbi, dev->sector);
 736                         }
 737                 }
 738         }
 739
 740         return tx;
 741 }
 742
 743 static void ops_complete_postxor(void *stripe_head_ref)
 744 {
 745         struct stripe_head *sh = stripe_head_ref;
 746
 747         pr_debug("%s: stripe %llu\n", __FUNCTION__,
 748                 (unsigned long long)sh->sector);
 749
 750         set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
 751         set_bit(STRIPE_HANDLE, &sh->state);
 752         release_stripe(sh);
 753 }
 754
 755 static void ops_complete_write(void *stripe_head_ref)
 756 {
 757         struct stripe_head *sh = stripe_head_ref;
 758         int disks = sh->disks, i, pd_idx = sh->pd_idx;
 759
 760         pr_debug("%s: stripe %llu\n", __FUNCTION__,
 761                 (unsigned long long)sh->sector);
 762
 763         for (i = disks; i--; ) {
 764                 struct r5dev *dev = &sh->dev[i];
 765                 if (dev->written || i == pd_idx)
 766                         set_bit(R5_UPTODATE, &dev->flags);
 767         }
 768
 769         set_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
 770         set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
 771
 772         set_bit(STRIPE_HANDLE, &sh->state);
 773         release_stripe(sh);
 774 }
 775
 776 static void
 777 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx,
 778                 unsigned long pending)
 779 {
 780         /* kernel stack size limits the total number of disks */
 781         int disks = sh->disks;
 782         struct page *xor_srcs[disks];
 783
 784         int count = 0, pd_idx = sh->pd_idx, i;
 785         struct page *xor_dest;
 786         int prexor = test_bit(STRIPE_OP_PREXOR, &pending);
 787         unsigned long flags;
 788         dma_async_tx_callback callback;
 789
 790         pr_debug("%s: stripe %llu\n", __FUNCTION__,
 791                 (unsigned long long)sh->sector);
 792
 793         /* check if prexor is active which means only process blocks
 794          * that are part of a read-modify-write (written)
 795          */
 796         if (prexor) {
 797                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
 798                 for (i = disks; i--; ) {
 799                         struct r5dev *dev = &sh->dev[i];
 800                         if (dev->written)
 801                                 xor_srcs[count++] = dev->page;
 802                 }
 803         } else {
 804                 xor_dest = sh->dev[pd_idx].page;
 805                 for (i = disks; i--; ) {
 806                         struct r5dev *dev = &sh->dev[i];
 807                         if (i != pd_idx)
 808                                 xor_srcs[count++] = dev->page;
 809                 }
 810         }
 811
 812         /* check whether this postxor is part of a write */
 813         callback = test_bit(STRIPE_OP_BIODRAIN, &pending) ?
 814                 ops_complete_write : ops_complete_postxor;
 815
 816         /* 1/ if we prexor'd then the dest is reused as a source
 817          * 2/ if we did not prexor then we are redoing the parity
 818          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
 819          * for the synchronous xor case
 820          */
 821         flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
 822                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
 823
 824         atomic_inc(&sh->count);
 825
 826         if (unlikely(count == 1)) {
 827                 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
 828                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
 829                         flags, tx, callback, sh);
 830         } else
 831                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
 832                         flags, tx, callback, sh);
 833 }
 834
 835 static void ops_complete_check(void *stripe_head_ref)
 836 {
 837         struct stripe_head *sh = stripe_head_ref;
 838         int pd_idx = sh->pd_idx;
 839
 840         pr_debug("%s: stripe %llu\n", __FUNCTION__,
 841                 (unsigned long long)sh->sector);
 842
 843         if (test_and_clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending) &&
 844                 sh->ops.zero_sum_result == 0)
 845                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
 846
 847         set_bit(STRIPE_OP_CHECK, &sh->ops.complete);
 848         set_bit(STRIPE_HANDLE, &sh->state);
 849         release_stripe(sh);
 850 }
 851
 852 static void ops_run_check(struct stripe_head *sh)
 853 {
 854         /* kernel stack size limits the total number of disks */
 855         int disks = sh->disks;
 856         struct page *xor_srcs[disks];
 857         struct dma_async_tx_descriptor *tx;
 858
 859         int count = 0, pd_idx = sh->pd_idx, i;
 860         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
 861
 862         pr_debug("%s: stripe %llu\n", __FUNCTION__,
 863                 (unsigned long long)sh->sector);
 864
 865         for (i = disks; i--; ) {
 866                 struct r5dev *dev = &sh->dev[i];
 867                 if (i != pd_idx)
 868                         xor_srcs[count++] = dev->page;
 869         }
 870
 871         tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
 872                 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
 873
 874         if (tx)
 875                 set_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
 876         else
 877                 clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
 878
 879         atomic_inc(&sh->count);
 880         tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
 881                 ops_complete_check, sh);
 882 }
 883
 884 static void raid5_run_ops(struct stripe_head *sh, unsigned long pending)
 885 {
 886         int overlap_clear = 0, i, disks = sh->disks;
 887         struct dma_async_tx_descriptor *tx = NULL;
 888
 889         if (test_bit(STRIPE_OP_BIOFILL, &pending)) {
 890                 ops_run_biofill(sh);
 891                 overlap_clear++;
 892         }
 893
 894         if (test_bit(STRIPE_OP_COMPUTE_BLK, &pending))
 895                 tx = ops_run_compute5(sh, pending);
 896
 897         if (test_bit(STRIPE_OP_PREXOR, &pending))
 898                 tx = ops_run_prexor(sh, tx);
 899
 900         if (test_bit(STRIPE_OP_BIODRAIN, &pending)) {
 901                 tx = ops_run_biodrain(sh, tx, pending);
 902                 overlap_clear++;
 903         }
 904
 905         if (test_bit(STRIPE_OP_POSTXOR, &pending))
 906                 ops_run_postxor(sh, tx, pending);
 907
 908         if (test_bit(STRIPE_OP_CHECK, &pending))
 909                 ops_run_check(sh);
 910
 911         if (test_bit(STRIPE_OP_IO, &pending))
 912                 ops_run_io(sh);
 913
 914         if (overlap_clear)
 915                 for (i = disks; i--; ) {
 916                         struct r5dev *dev = &sh->dev[i];
 917                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
 918                                 wake_up(&sh->raid_conf->wait_for_overlap);
 919                 }
 920 }
 921
 922 static int grow_one_stripe(raid5_conf_t *conf)
 923 {
 924         struct stripe_head *sh;
 925         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
 926         if (!sh)
 927                 return 0;
 928         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
 929         sh->raid_conf = conf;
 930         spin_lock_init(&sh->lock);
 931
 932         if (grow_buffers(sh, conf->raid_disks)) {
 933                 shrink_buffers(sh, conf->raid_disks);
 934                 kmem_cache_free(conf->slab_cache, sh);
 935                 return 0;
 936         }
 937         sh->disks = conf->raid_disks;
 938         /* we just created an active stripe so... */
 939         atomic_set(&sh->count, 1);
 940         atomic_inc(&conf->active_stripes);
 941         INIT_LIST_HEAD(&sh->lru);
 942         release_stripe(sh);
 943         return 1;
 944 }
 945
 946 static int grow_stripes(raid5_conf_t *conf, int num)
 947 {
 948         struct kmem_cache *sc;
 949         int devs = conf->raid_disks;
 950
 951         sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
 952         sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
 953         conf->active_name = 0;
 954         sc = kmem_cache_create(conf->cache_name[conf->active_name],
 955                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
 956                                0, 0, NULL);
 957         if (!sc)
 958                 return 1;
 959         conf->slab_cache = sc;
 960         conf->pool_size = devs;
 961         while (num--)
 962                 if (!grow_one_stripe(conf))
 963                         return 1;
 964         return 0;
 965 }
 966
 967 #ifdef CONFIG_MD_RAID5_RESHAPE
 968 static int resize_stripes(raid5_conf_t *conf, int newsize)
 969 {
 970         /* Make all the stripes able to hold 'newsize' devices.
 971          * New slots in each stripe get 'page' set to a new page.
 972          *
 973          * This happens in stages:
 974          * 1/ create a new kmem_cache and allocate the required number of
 975          *    stripe_heads.
 976          * 2/ gather all the old stripe_heads and tranfer the pages across
 977          *    to the new stripe_heads.  This will have the side effect of
 978          *    freezing the array as once all stripe_heads have been collected,
 979          *    no IO will be possible.  Old stripe heads are freed once their
 980          *    pages have been transferred over, and the old kmem_cache is
 981          *    freed when all stripes are done.
 982          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
 983          *    we simple return a failre status - no need to clean anything up.
 984          * 4/ allocate new pages for the new slots in the new stripe_heads.
 985          *    If this fails, we don't bother trying the shrink the
 986          *    stripe_heads down again, we just leave them as they are.
 987          *    As each stripe_head is processed the new one is released into
 988          *    active service.
 989          *
 990          * Once step2 is started, we cannot afford to wait for a write,
 991          * so we use GFP_NOIO allocations.
 992          */
 993         struct stripe_head *osh, *nsh;
 994         LIST_HEAD(newstripes);
 995         struct disk_info *ndisks;
 996         int err = 0;
 997         struct kmem_cache *sc;
 998         int i;
 999
1000         if (newsize <= conf->pool_size)
1001                 return 0; /* never bother to shrink */
1002
1003         md_allow_write(conf->mddev);
1004
1005         /* Step 1 */
1006         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1007                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1008                                0, 0, NULL);
1009         if (!sc)
1010                 return -ENOMEM;
1011
1012         for (i = conf->max_nr_stripes; i; i--) {
1013                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1014                 if (!nsh)
1015                         break;
1016
1017                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1018
1019                 nsh->raid_conf = conf;
1020                 spin_lock_init(&nsh->lock);
1021
1022                 list_add(&nsh->lru, &newstripes);
1023         }
1024         if (i) {
1025                 /* didn't get enough, give up */
1026                 while (!list_empty(&newstripes)) {
1027                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1028                         list_del(&nsh->lru);
1029                         kmem_cache_free(sc, nsh);
1030                 }
1031                 kmem_cache_destroy(sc);
1032                 return -ENOMEM;
1033         }
1034         /* Step 2 - Must use GFP_NOIO now.
1035          * OK, we have enough stripes, start collecting inactive
1036          * stripes and copying them over
1037          */
1038         list_for_each_entry(nsh, &newstripes, lru) {
1039                 spin_lock_irq(&conf->device_lock);
1040                 wait_event_lock_irq(conf->wait_for_stripe,
1041                                     !list_empty(&conf->inactive_list),
1042                                     conf->device_lock,
1043                                     unplug_slaves(conf->mddev)
1044                         );
1045                 osh = get_free_stripe(conf);
1046                 spin_unlock_irq(&conf->device_lock);
1047                 atomic_set(&nsh->count, 1);
1048                 for(i=0; i<conf->pool_size; i++)
1049                         nsh->dev[i].page = osh->dev[i].page;
1050                 for( ; i<newsize; i++)
1051                         nsh->dev[i].page = NULL;
1052                 kmem_cache_free(conf->slab_cache, osh);
1053         }
1054         kmem_cache_destroy(conf->slab_cache);
1055
1056         /* Step 3.
1057          * At this point, we are holding all the stripes so the array
1058          * is completely stalled, so now is a good time to resize
1059          * conf->disks.
1060          */
1061         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1062         if (ndisks) {
1063                 for (i=0; i<conf->raid_disks; i++)
1064                         ndisks[i] = conf->disks[i];
1065                 kfree(conf->disks);
1066                 conf->disks = ndisks;
1067         } else
1068                 err = -ENOMEM;
1069
1070         /* Step 4, return new stripes to service */
1071         while(!list_empty(&newstripes)) {
1072                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1073                 list_del_init(&nsh->lru);
1074                 for (i=conf->raid_disks; i < newsize; i++)
1075                         if (nsh->dev[i].page == NULL) {
1076                                 struct page *p = alloc_page(GFP_NOIO);
1077                                 nsh->dev[i].page = p;
1078                                 if (!p)
1079                                         err = -ENOMEM;
1080                         }
1081                 release_stripe(nsh);
1082         }
1083         /* critical section pass, GFP_NOIO no longer needed */
1084
1085         conf->slab_cache = sc;
1086         conf->active_name = 1-conf->active_name;
1087         conf->pool_size = newsize;
1088         return err;
1089 }
1090 #endif
1091
1092 static int drop_one_stripe(raid5_conf_t *conf)
1093 {
1094         struct stripe_head *sh;
1095
1096         spin_lock_irq(&conf->device_lock);
1097         sh = get_free_stripe(conf);
1098         spin_unlock_irq(&conf->device_lock);
1099         if (!sh)
1100                 return 0;
1101         BUG_ON(atomic_read(&sh->count));
1102         shrink_buffers(sh, conf->pool_size);
1103         kmem_cache_free(conf->slab_cache, sh);
1104         atomic_dec(&conf->active_stripes);
1105         return 1;
1106 }
1107
1108 static void shrink_stripes(raid5_conf_t *conf)
1109 {
1110         while (drop_one_stripe(conf))
1111                 ;
1112
1113         if (conf->slab_cache)
1114                 kmem_cache_destroy(conf->slab_cache);
1115         conf->slab_cache = NULL;
1116 }
1117
1118 static void raid5_end_read_request(struct bio * bi, int error)
1119 {
1120         struct stripe_head *sh = bi->bi_private;
1121         raid5_conf_t *conf = sh->raid_conf;
1122         int disks = sh->disks, i;
1123         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1124         char b[BDEVNAME_SIZE];
1125         mdk_rdev_t *rdev;
1126
1127
1128         for (i=0 ; i<disks; i++)
1129                 if (bi == &sh->dev[i].req)
1130                         break;
1131
1132         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1133                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1134                 uptodate);
1135         if (i == disks) {
1136                 BUG();
1137                 return;
1138         }
1139
1140         if (uptodate) {
1141                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1142                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1143                         rdev = conf->disks[i].rdev;
1144                         printk(KERN_INFO "raid5:%s: read error corrected (%lu sectors at %llu on %s)\n",
1145                                mdname(conf->mddev), STRIPE_SECTORS,
1146                                (unsigned long long)(sh->sector + rdev->data_offset),
1147                                bdevname(rdev->bdev, b));
1148                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1149                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1150                 }
1151                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1152                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1153         } else {
1154                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1155                 int retry = 0;
1156                 rdev = conf->disks[i].rdev;
1157
1158                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1159                 atomic_inc(&rdev->read_errors);
1160                 if (conf->mddev->degraded)
1161                         printk(KERN_WARNING "raid5:%s: read error not correctable (sector %llu on %s).\n",
1162                                mdname(conf->mddev),
1163                                (unsigned long long)(sh->sector + rdev->data_offset),
1164                                bdn);
1165                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1166                         /* Oh, no!!! */
1167                         printk(KERN_WARNING "raid5:%s: read error NOT corrected!! (sector %llu on %s).\n",
1168                                mdname(conf->mddev),
1169                                (unsigned long long)(sh->sector + rdev->data_offset),
1170                                bdn);
1171                 else if (atomic_read(&rdev->read_errors)
1172                          > conf->max_nr_stripes)
1173                         printk(KERN_WARNING
1174                                "raid5:%s: Too many read errors, failing device %s.\n",
1175                                mdname(conf->mddev), bdn);
1176                 else
1177                         retry = 1;
1178                 if (retry)
1179                         set_bit(R5_ReadError, &sh->dev[i].flags);
1180                 else {
1181                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1182                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1183                         md_error(conf->mddev, rdev);
1184                 }
1185         }
1186         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1187         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1188         set_bit(STRIPE_HANDLE, &sh->state);
1189         release_stripe(sh);
1190 }
1191
1192 static void raid5_end_write_request (struct bio *bi, int error)
1193 {
1194         struct stripe_head *sh = bi->bi_private;
1195         raid5_conf_t *conf = sh->raid_conf;
1196         int disks = sh->disks, i;
1197         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1198
1199         for (i=0 ; i<disks; i++)
1200                 if (bi == &sh->dev[i].req)
1201                         break;
1202
1203         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1204                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1205                 uptodate);
1206         if (i == disks) {
1207                 BUG();
1208                 return;
1209         }
1210
1211         if (!uptodate)
1212                 md_error(conf->mddev, conf->disks[i].rdev);
1213
1214         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1215
1216         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1217         set_bit(STRIPE_HANDLE, &sh->state);
1218         release_stripe(sh);
1219 }
1220
1221
1222 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1223
1224 static void raid5_build_block (struct stripe_head *sh, int i)
1225 {
1226         struct r5dev *dev = &sh->dev[i];
1227
1228         bio_init(&dev->req);
1229         dev->req.bi_io_vec = &dev->vec;
1230         dev->req.bi_vcnt++;
1231         dev->req.bi_max_vecs++;
1232         dev->vec.bv_page = dev->page;
1233         dev->vec.bv_len = STRIPE_SIZE;
1234         dev->vec.bv_offset = 0;
1235
1236         dev->req.bi_sector = sh->sector;
1237         dev->req.bi_private = sh;
1238
1239         dev->flags = 0;
1240         dev->sector = compute_blocknr(sh, i);
1241 }
1242
1243 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1244 {
1245         char b[BDEVNAME_SIZE];
1246         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1247         pr_debug("raid5: error called\n");
1248
1249         if (!test_bit(Faulty, &rdev->flags)) {
1250                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1251                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1252                         unsigned long flags;
1253                         spin_lock_irqsave(&conf->device_lock, flags);
1254                         mddev->degraded++;
1255                         spin_unlock_irqrestore(&conf->device_lock, flags);
1256                         /*
1257                          * if recovery was running, make sure it aborts.
1258                          */
1259                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
1260                 }
1261                 set_bit(Faulty, &rdev->flags);
1262                 printk (KERN_ALERT
1263                         "raid5: Disk failure on %s, disabling device."
1264                         " Operation continuing on %d devices\n",
1265                         bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1266         }
1267 }
1268
1269 /*
1270  * Input: a 'big' sector number,
1271  * Output: index of the data and parity disk, and the sector # in them.
1272  */
1273 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
1274                         unsigned int data_disks, unsigned int * dd_idx,
1275                         unsigned int * pd_idx, raid5_conf_t *conf)
1276 {
1277         long stripe;
1278         unsigned long chunk_number;
1279         unsigned int chunk_offset;
1280         sector_t new_sector;
1281         int sectors_per_chunk = conf->chunk_size >> 9;
1282
1283         /* First compute the information on this sector */
1284
1285         /*
1286          * Compute the chunk number and the sector offset inside the chunk
1287          */
1288         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1289         chunk_number = r_sector;
1290         BUG_ON(r_sector != chunk_number);
1291
1292         /*
1293          * Compute the stripe number
1294          */
1295         stripe = chunk_number / data_disks;
1296
1297         /*
1298          * Compute the data disk and parity disk indexes inside the stripe
1299          */
1300         *dd_idx = chunk_number % data_disks;
1301
1302         /*
1303          * Select the parity disk based on the user selected algorithm.
1304          */
1305         switch(conf->level) {
1306         case 4:
1307                 *pd_idx = data_disks;
1308                 break;
1309         case 5:
1310                 switch (conf->algorithm) {
1311                 case ALGORITHM_LEFT_ASYMMETRIC:
1312                         *pd_idx = data_disks - stripe % raid_disks;
1313                         if (*dd_idx >= *pd_idx)
1314                                 (*dd_idx)++;
1315                         break;
1316                 case ALGORITHM_RIGHT_ASYMMETRIC:
1317                         *pd_idx = stripe % raid_disks;
1318                         if (*dd_idx >= *pd_idx)
1319                                 (*dd_idx)++;
1320                         break;
1321                 case ALGORITHM_LEFT_SYMMETRIC:
1322                         *pd_idx = data_disks - stripe % raid_disks;
1323                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1324                         break;
1325                 case ALGORITHM_RIGHT_SYMMETRIC:
1326                         *pd_idx = stripe % raid_disks;
1327                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1328                         break;
1329                 default:
1330                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1331                                 conf->algorithm);
1332                 }
1333                 break;
1334         case 6:
1335
1336                 /**** FIX THIS ****/
1337                 switch (conf->algorithm) {
1338                 case ALGORITHM_LEFT_ASYMMETRIC:
1339                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1340                         if (*pd_idx == raid_disks-1)
1341                                 (*dd_idx)++;    /* Q D D D P */
1342                         else if (*dd_idx >= *pd_idx)
1343                                 (*dd_idx) += 2; /* D D P Q D */
1344                         break;
1345                 case ALGORITHM_RIGHT_ASYMMETRIC:
1346                         *pd_idx = stripe % raid_disks;
1347                         if (*pd_idx == raid_disks-1)
1348                                 (*dd_idx)++;    /* Q D D D P */
1349                         else if (*dd_idx >= *pd_idx)
1350                                 (*dd_idx) += 2; /* D D P Q D */
1351                         break;
1352                 case ALGORITHM_LEFT_SYMMETRIC:
1353                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1354                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1355                         break;
1356                 case ALGORITHM_RIGHT_SYMMETRIC:
1357                         *pd_idx = stripe % raid_disks;
1358                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1359                         break;
1360                 default:
1361                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1362                                 conf->algorithm);
1363                 }
1364                 break;
1365         }
1366
1367         /*
1368          * Finally, compute the new sector number
1369          */
1370         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1371         return new_sector;
1372 }
1373
1374
1375 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1376 {
1377         raid5_conf_t *conf = sh->raid_conf;
1378         int raid_disks = sh->disks;
1379         int data_disks = raid_disks - conf->max_degraded;
1380         sector_t new_sector = sh->sector, check;
1381         int sectors_per_chunk = conf->chunk_size >> 9;
1382         sector_t stripe;
1383         int chunk_offset;
1384         int chunk_number, dummy1, dummy2, dd_idx = i;
1385         sector_t r_sector;
1386
1387
1388         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1389         stripe = new_sector;
1390         BUG_ON(new_sector != stripe);
1391
1392         if (i == sh->pd_idx)
1393                 return 0;
1394         switch(conf->level) {
1395         case 4: break;
1396         case 5:
1397                 switch (conf->algorithm) {
1398                 case ALGORITHM_LEFT_ASYMMETRIC:
1399                 case ALGORITHM_RIGHT_ASYMMETRIC:
1400                         if (i > sh->pd_idx)
1401                                 i--;
1402                         break;
1403                 case ALGORITHM_LEFT_SYMMETRIC:
1404                 case ALGORITHM_RIGHT_SYMMETRIC:
1405                         if (i < sh->pd_idx)
1406                                 i += raid_disks;
1407                         i -= (sh->pd_idx + 1);
1408                         break;
1409                 default:
1410                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1411                                conf->algorithm);
1412                 }
1413                 break;
1414         case 6:
1415                 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
1416                         return 0; /* It is the Q disk */
1417                 switch (conf->algorithm) {
1418                 case ALGORITHM_LEFT_ASYMMETRIC:
1419                 case ALGORITHM_RIGHT_ASYMMETRIC:
1420                         if (sh->pd_idx == raid_disks-1)
1421                                 i--;    /* Q D D D P */
1422                         else if (i > sh->pd_idx)
1423                                 i -= 2; /* D D P Q D */
1424                         break;
1425                 case ALGORITHM_LEFT_SYMMETRIC:
1426                 case ALGORITHM_RIGHT_SYMMETRIC:
1427                         if (sh->pd_idx == raid_disks-1)
1428                                 i--; /* Q D D D P */
1429                         else {
1430                                 /* D D P Q D */
1431                                 if (i < sh->pd_idx)
1432                                         i += raid_disks;
1433                                 i -= (sh->pd_idx + 2);
1434                         }
1435                         break;
1436                 default:
1437                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1438                                 conf->algorithm);
1439                 }
1440                 break;
1441         }
1442
1443         chunk_number = stripe * data_disks + i;
1444         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1445
1446         check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
1447         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
1448                 printk(KERN_ERR "compute_blocknr: map not correct\n");
1449                 return 0;
1450         }
1451         return r_sector;
1452 }
1453
1454
1455
1456 /*
1457  * Copy data between a page in the stripe cache, and one or more bion
1458  * The page could align with the middle of the bio, or there could be
1459  * several bion, each with several bio_vecs, which cover part of the page
1460  * Multiple bion are linked together on bi_next.  There may be extras
1461  * at the end of this list.  We ignore them.
1462  */
1463 static void copy_data(int frombio, struct bio *bio,
1464                      struct page *page,
1465                      sector_t sector)
1466 {
1467         char *pa = page_address(page);
1468         struct bio_vec *bvl;
1469         int i;
1470         int page_offset;
1471
1472         if (bio->bi_sector >= sector)
1473                 page_offset = (signed)(bio->bi_sector - sector) * 512;
1474         else
1475                 page_offset = (signed)(sector - bio->bi_sector) * -512;
1476         bio_for_each_segment(bvl, bio, i) {
1477                 int len = bio_iovec_idx(bio,i)->bv_len;
1478                 int clen;
1479                 int b_offset = 0;
1480
1481                 if (page_offset < 0) {
1482                         b_offset = -page_offset;
1483                         page_offset += b_offset;
1484                         len -= b_offset;
1485                 }
1486
1487                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1488                         clen = STRIPE_SIZE - page_offset;
1489                 else clen = len;
1490
1491                 if (clen > 0) {
1492                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1493                         if (frombio)
1494                                 memcpy(pa+page_offset, ba+b_offset, clen);
1495                         else
1496                                 memcpy(ba+b_offset, pa+page_offset, clen);
1497                         __bio_kunmap_atomic(ba, KM_USER0);
1498                 }
1499                 if (clen < len) /* hit end of page */
1500                         break;
1501                 page_offset +=  len;
1502         }
1503 }
1504
1505 #define check_xor()     do {                                              \
1506                                 if (count == MAX_XOR_BLOCKS) {            \
1507                                 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1508                                 count = 0;                                \
1509                            }                                              \
1510                         } while(0)
1511
1512 static void compute_parity6(struct stripe_head *sh, int method)
1513 {
1514         raid6_conf_t *conf = sh->raid_conf;
1515         int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1516         struct bio *chosen;
1517         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1518         void *ptrs[disks];
1519
1520         qd_idx = raid6_next_disk(pd_idx, disks);
1521         d0_idx = raid6_next_disk(qd_idx, disks);
1522
1523         pr_debug("compute_parity, stripe %llu, method %d\n",
1524                 (unsigned long long)sh->sector, method);
1525
1526         switch(method) {
1527         case READ_MODIFY_WRITE:
1528                 BUG();          /* READ_MODIFY_WRITE N/A for RAID-6 */
1529         case RECONSTRUCT_WRITE:
1530                 for (i= disks; i-- ;)
1531                         if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1532                                 chosen = sh->dev[i].towrite;
1533                                 sh->dev[i].towrite = NULL;
1534
1535                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1536                                         wake_up(&conf->wait_for_overlap);
1537
1538                                 BUG_ON(sh->dev[i].written);
1539                                 sh->dev[i].written = chosen;
1540                         }
1541                 break;
1542         case CHECK_PARITY:
1543                 BUG();          /* Not implemented yet */
1544         }
1545
1546         for (i = disks; i--;)
1547                 if (sh->dev[i].written) {
1548                         sector_t sector = sh->dev[i].sector;
1549                         struct bio *wbi = sh->dev[i].written;
1550                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1551                                 copy_data(1, wbi, sh->dev[i].page, sector);
1552                                 wbi = r5_next_bio(wbi, sector);
1553                         }
1554
1555                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1556                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
1557                 }
1558
1559 //      switch(method) {
1560 //      case RECONSTRUCT_WRITE:
1561 //      case CHECK_PARITY:
1562 //      case UPDATE_PARITY:
1563                 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1564                 /* FIX: Is this ordering of drives even remotely optimal? */
1565                 count = 0;
1566                 i = d0_idx;
1567                 do {
1568                         ptrs[count++] = page_address(sh->dev[i].page);
1569                         if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1570                                 printk("block %d/%d not uptodate on parity calc\n", i,count);
1571                         i = raid6_next_disk(i, disks);
1572                 } while ( i != d0_idx );
1573 //              break;
1574 //      }
1575
1576         raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1577
1578         switch(method) {
1579         case RECONSTRUCT_WRITE:
1580                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1581                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1582                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
1583                 set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
1584                 break;
1585         case UPDATE_PARITY:
1586                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1587                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1588                 break;
1589         }
1590 }
1591
1592
1593 /* Compute one missing block */
1594 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1595 {
1596         int i, count, disks = sh->disks;
1597         void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1598         int pd_idx = sh->pd_idx;
1599         int qd_idx = raid6_next_disk(pd_idx, disks);
1600
1601         pr_debug("compute_block_1, stripe %llu, idx %d\n",
1602                 (unsigned long long)sh->sector, dd_idx);
1603
1604         if ( dd_idx == qd_idx ) {
1605                 /* We're actually computing the Q drive */
1606                 compute_parity6(sh, UPDATE_PARITY);
1607         } else {
1608                 dest = page_address(sh->dev[dd_idx].page);
1609                 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1610                 count = 0;
1611                 for (i = disks ; i--; ) {
1612                         if (i == dd_idx || i == qd_idx)
1613                                 continue;
1614                         p = page_address(sh->dev[i].page);
1615                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1616                                 ptr[count++] = p;
1617                         else
1618                                 printk("compute_block() %d, stripe %llu, %d"
1619                                        " not present\n", dd_idx,
1620                                        (unsigned long long)sh->sector, i);
1621
1622                         check_xor();
1623                 }
1624                 if (count)
1625                         xor_blocks(count, STRIPE_SIZE, dest, ptr);
1626                 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1627                 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1628         }
1629 }
1630
1631 /* Compute two missing blocks */
1632 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1633 {
1634         int i, count, disks = sh->disks;
1635         int pd_idx = sh->pd_idx;
1636         int qd_idx = raid6_next_disk(pd_idx, disks);
1637         int d0_idx = raid6_next_disk(qd_idx, disks);
1638         int faila, failb;
1639
1640         /* faila and failb are disk numbers relative to d0_idx */
1641         /* pd_idx become disks-2 and qd_idx become disks-1 */
1642         faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1643         failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1644
1645         BUG_ON(faila == failb);
1646         if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1647
1648         pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1649                (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1650
1651         if ( failb == disks-1 ) {
1652                 /* Q disk is one of the missing disks */
1653                 if ( faila == disks-2 ) {
1654                         /* Missing P+Q, just recompute */
1655                         compute_parity6(sh, UPDATE_PARITY);
1656                         return;
1657                 } else {
1658                         /* We're missing D+Q; recompute D from P */
1659                         compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1660                         compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1661                         return;
1662                 }
1663         }
1664
1665         /* We're missing D+P or D+D; build pointer table */
1666         {
1667                 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1668                 void *ptrs[disks];
1669
1670                 count = 0;
1671                 i = d0_idx;
1672                 do {
1673                         ptrs[count++] = page_address(sh->dev[i].page);
1674                         i = raid6_next_disk(i, disks);
1675                         if (i != dd_idx1 && i != dd_idx2 &&
1676                             !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1677                                 printk("compute_2 with missing block %d/%d\n", count, i);
1678                 } while ( i != d0_idx );
1679
1680                 if ( failb == disks-2 ) {
1681                         /* We're missing D+P. */
1682                         raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1683                 } else {
1684                         /* We're missing D+D. */
1685                         raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1686                 }
1687
1688                 /* Both the above update both missing blocks */
1689                 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1690                 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1691         }
1692 }
1693
1694 static int
1695 handle_write_operations5(struct stripe_head *sh, int rcw, int expand)
1696 {
1697         int i, pd_idx = sh->pd_idx, disks = sh->disks;
1698         int locked = 0;
1699
1700         if (rcw) {
1701                 /* if we are not expanding this is a proper write request, and
1702                  * there will be bios with new data to be drained into the
1703                  * stripe cache
1704                  */
1705                 if (!expand) {
1706                         set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1707                         sh->ops.count++;
1708                 }
1709
1710                 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1711                 sh->ops.count++;
1712
1713                 for (i = disks; i--; ) {
1714                         struct r5dev *dev = &sh->dev[i];
1715
1716                         if (dev->towrite) {
1717                                 set_bit(R5_LOCKED, &dev->flags);
1718                                 if (!expand)
1719                                         clear_bit(R5_UPTODATE, &dev->flags);
1720                                 locked++;
1721                         }
1722                 }
1723         } else {
1724                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1725                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1726
1727                 set_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
1728                 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1729                 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1730
1731                 sh->ops.count += 3;
1732
1733                 for (i = disks; i--; ) {
1734                         struct r5dev *dev = &sh->dev[i];
1735                         if (i == pd_idx)
1736                                 continue;
1737
1738                         /* For a read-modify write there may be blocks that are
1739                          * locked for reading while others are ready to be
1740                          * written so we distinguish these blocks by the
1741                          * R5_Wantprexor bit
1742                          */
1743                         if (dev->towrite &&
1744                             (test_bit(R5_UPTODATE, &dev->flags) ||
1745                             test_bit(R5_Wantcompute, &dev->flags))) {
1746                                 set_bit(R5_Wantprexor, &dev->flags);
1747                                 set_bit(R5_LOCKED, &dev->flags);
1748                                 clear_bit(R5_UPTODATE, &dev->flags);
1749                                 locked++;
1750                         }
1751                 }
1752         }
1753
1754         /* keep the parity disk locked while asynchronous operations
1755          * are in flight
1756          */
1757         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1758         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1759         locked++;
1760
1761         pr_debug("%s: stripe %llu locked: %d pending: %lx\n",
1762                 __FUNCTION__, (unsigned long long)sh->sector,
1763                 locked, sh->ops.pending);
1764
1765         return locked;
1766 }
1767
1768 /*
1769  * Each stripe/dev can have one or more bion attached.
1770  * toread/towrite point to the first in a chain.
1771  * The bi_next chain must be in order.
1772  */
1773 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1774 {
1775         struct bio **bip;
1776         raid5_conf_t *conf = sh->raid_conf;
1777         int firstwrite=0;
1778
1779         pr_debug("adding bh b#%llu to stripe s#%llu\n",
1780                 (unsigned long long)bi->bi_sector,
1781                 (unsigned long long)sh->sector);
1782
1783
1784         spin_lock(&sh->lock);
1785         spin_lock_irq(&conf->device_lock);
1786         if (forwrite) {
1787                 bip = &sh->dev[dd_idx].towrite;
1788                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1789                         firstwrite = 1;
1790         } else
1791                 bip = &sh->dev[dd_idx].toread;
1792         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1793                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1794                         goto overlap;
1795                 bip = & (*bip)->bi_next;
1796         }
1797         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1798                 goto overlap;
1799
1800         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1801         if (*bip)
1802                 bi->bi_next = *bip;
1803         *bip = bi;
1804         bi->bi_phys_segments ++;
1805         spin_unlock_irq(&conf->device_lock);
1806         spin_unlock(&sh->lock);
1807
1808         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1809                 (unsigned long long)bi->bi_sector,
1810                 (unsigned long long)sh->sector, dd_idx);
1811
1812         if (conf->mddev->bitmap && firstwrite) {
1813                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1814                                   STRIPE_SECTORS, 0);
1815                 sh->bm_seq = conf->seq_flush+1;
1816                 set_bit(STRIPE_BIT_DELAY, &sh->state);
1817         }
1818
1819         if (forwrite) {
1820                 /* check if page is covered */
1821                 sector_t sector = sh->dev[dd_idx].sector;
1822                 for (bi=sh->dev[dd_idx].towrite;
1823                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1824                              bi && bi->bi_sector <= sector;
1825                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1826                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1827                                 sector = bi->bi_sector + (bi->bi_size>>9);
1828                 }
1829                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1830                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1831         }
1832         return 1;
1833
1834  overlap:
1835         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1836         spin_unlock_irq(&conf->device_lock);
1837         spin_unlock(&sh->lock);
1838         return 0;
1839 }
1840
1841 static void end_reshape(raid5_conf_t *conf);
1842
1843 static int page_is_zero(struct page *p)
1844 {
1845         char *a = page_address(p);
1846         return ((*(u32*)a) == 0 &&
1847                 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1848 }
1849
1850 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1851 {
1852         int sectors_per_chunk = conf->chunk_size >> 9;
1853         int pd_idx, dd_idx;
1854         int chunk_offset = sector_div(stripe, sectors_per_chunk);
1855
1856         raid5_compute_sector(stripe * (disks - conf->max_degraded)
1857                              *sectors_per_chunk + chunk_offset,
1858                              disks, disks - conf->max_degraded,
1859                              &dd_idx, &pd_idx, conf);
1860         return pd_idx;
1861 }
1862
1863 static void
1864 handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh,
1865                                 struct stripe_head_state *s, int disks,
1866                                 struct bio **return_bi)
1867 {
1868         int i;
1869         for (i = disks; i--; ) {
1870                 struct bio *bi;
1871                 int bitmap_end = 0;
1872
1873                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1874                         mdk_rdev_t *rdev;
1875                         rcu_read_lock();
1876                         rdev = rcu_dereference(conf->disks[i].rdev);
1877                         if (rdev && test_bit(In_sync, &rdev->flags))
1878                                 /* multiple read failures in one stripe */
1879                                 md_error(conf->mddev, rdev);
1880                         rcu_read_unlock();
1881                 }
1882                 spin_lock_irq(&conf->device_lock);
1883                 /* fail all writes first */
1884                 bi = sh->dev[i].towrite;
1885                 sh->dev[i].towrite = NULL;
1886                 if (bi) {
1887                         s->to_write--;
1888                         bitmap_end = 1;
1889                 }
1890
1891                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1892                         wake_up(&conf->wait_for_overlap);
1893
1894                 while (bi && bi->bi_sector <
1895                         sh->dev[i].sector + STRIPE_SECTORS) {
1896                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1897                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1898                         if (--bi->bi_phys_segments == 0) {
1899                                 md_write_end(conf->mddev);
1900                                 bi->bi_next = *return_bi;
1901                                 *return_bi = bi;
1902                         }
1903                         bi = nextbi;
1904                 }
1905                 /* and fail all 'written' */
1906                 bi = sh->dev[i].written;
1907                 sh->dev[i].written = NULL;
1908                 if (bi) bitmap_end = 1;
1909                 while (bi && bi->bi_sector <
1910                        sh->dev[i].sector + STRIPE_SECTORS) {
1911                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1912                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1913                         if (--bi->bi_phys_segments == 0) {
1914                                 md_write_end(conf->mddev);
1915                                 bi->bi_next = *return_bi;
1916                                 *return_bi = bi;
1917                         }
1918                         bi = bi2;
1919                 }
1920
1921                 /* fail any reads if this device is non-operational and
1922                  * the data has not reached the cache yet.
1923                  */
1924                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
1925                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1926                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
1927                         bi = sh->dev[i].toread;
1928                         sh->dev[i].toread = NULL;
1929                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1930                                 wake_up(&conf->wait_for_overlap);
1931                         if (bi) s->to_read--;
1932                         while (bi && bi->bi_sector <
1933                                sh->dev[i].sector + STRIPE_SECTORS) {
1934                                 struct bio *nextbi =
1935                                         r5_next_bio(bi, sh->dev[i].sector);
1936                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1937                                 if (--bi->bi_phys_segments == 0) {
1938                                         bi->bi_next = *return_bi;
1939                                         *return_bi = bi;
1940                                 }
1941                                 bi = nextbi;
1942                         }
1943                 }
1944                 spin_unlock_irq(&conf->device_lock);
1945                 if (bitmap_end)
1946                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1947                                         STRIPE_SECTORS, 0, 0);
1948         }
1949
1950 }
1951
1952 /* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks
1953  * to process
1954  */
1955 static int __handle_issuing_new_read_requests5(struct stripe_head *sh,
1956                         struct stripe_head_state *s, int disk_idx, int disks)
1957 {
1958         struct r5dev *dev = &sh->dev[disk_idx];
1959         struct r5dev *failed_dev = &sh->dev[s->failed_num];
1960
1961         /* don't schedule compute operations or reads on the parity block while
1962          * a check is in flight
1963          */
1964         if ((disk_idx == sh->pd_idx) &&
1965              test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
1966                 return ~0;
1967
1968         /* is the data in this block needed, and can we get it? */
1969         if (!test_bit(R5_LOCKED, &dev->flags) &&
1970             !test_bit(R5_UPTODATE, &dev->flags) && (dev->toread ||
1971             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1972              s->syncing || s->expanding || (s->failed &&
1973              (failed_dev->toread || (failed_dev->towrite &&
1974              !test_bit(R5_OVERWRITE, &failed_dev->flags)
1975              ))))) {
1976                 /* 1/ We would like to get this block, possibly by computing it,
1977                  * but we might not be able to.
1978                  *
1979                  * 2/ Since parity check operations potentially make the parity
1980                  * block !uptodate it will need to be refreshed before any
1981                  * compute operations on data disks are scheduled.
1982                  *
1983                  * 3/ We hold off parity block re-reads until check operations
1984                  * have quiesced.
1985                  */
1986                 if ((s->uptodate == disks - 1) &&
1987                     !test_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
1988                         set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
1989                         set_bit(R5_Wantcompute, &dev->flags);
1990                         sh->ops.target = disk_idx;
1991                         s->req_compute = 1;
1992                         sh->ops.count++;
1993                         /* Careful: from this point on 'uptodate' is in the eye
1994                          * of raid5_run_ops which services 'compute' operations
1995                          * before writes. R5_Wantcompute flags a block that will
1996                          * be R5_UPTODATE by the time it is needed for a
1997                          * subsequent operation.
1998                          */
1999                         s->uptodate++;
2000                         return 0; /* uptodate + compute == disks */
2001                 } else if ((s->uptodate < disks - 1) &&
2002                         test_bit(R5_Insync, &dev->flags)) {
2003                         /* Note: we hold off compute operations while checks are
2004                          * in flight, but we still prefer 'compute' over 'read'
2005                          * hence we only read if (uptodate < * disks-1)
2006                          */
2007                         set_bit(R5_LOCKED, &dev->flags);
2008                         set_bit(R5_Wantread, &dev->flags);
2009                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2010                                 sh->ops.count++;
2011                         s->locked++;
2012                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2013                                 s->syncing);
2014                 }
2015         }
2016
2017         return ~0;
2018 }
2019
2020 static void handle_issuing_new_read_requests5(struct stripe_head *sh,
2021                         struct stripe_head_state *s, int disks)
2022 {
2023         int i;
2024
2025         /* Clear completed compute operations.  Parity recovery
2026          * (STRIPE_OP_MOD_REPAIR_PD) implies a write-back which is handled
2027          * later on in this routine
2028          */
2029         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2030                 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2031                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2032                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2033                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2034         }
2035
2036         /* look for blocks to read/compute, skip this if a compute
2037          * is already in flight, or if the stripe contents are in the
2038          * midst of changing due to a write
2039          */
2040         if (!test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2041                 !test_bit(STRIPE_OP_PREXOR, &sh->ops.pending) &&
2042                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2043                 for (i = disks; i--; )
2044                         if (__handle_issuing_new_read_requests5(
2045                                 sh, s, i, disks) == 0)
2046                                 break;
2047         }
2048         set_bit(STRIPE_HANDLE, &sh->state);
2049 }
2050
2051 static void handle_issuing_new_read_requests6(struct stripe_head *sh,
2052                         struct stripe_head_state *s, struct r6_state *r6s,
2053                         int disks)
2054 {
2055         int i;
2056         for (i = disks; i--; ) {
2057                 struct r5dev *dev = &sh->dev[i];
2058                 if (!test_bit(R5_LOCKED, &dev->flags) &&
2059                     !test_bit(R5_UPTODATE, &dev->flags) &&
2060                     (dev->toread || (dev->towrite &&
2061                      !test_bit(R5_OVERWRITE, &dev->flags)) ||
2062                      s->syncing || s->expanding ||
2063                      (s->failed >= 1 &&
2064                       (sh->dev[r6s->failed_num[0]].toread ||
2065                        s->to_write)) ||
2066                      (s->failed >= 2 &&
2067                       (sh->dev[r6s->failed_num[1]].toread ||
2068                        s->to_write)))) {
2069                         /* we would like to get this block, possibly
2070                          * by computing it, but we might not be able to
2071                          */
2072                         if (s->uptodate == disks-1) {
2073                                 pr_debug("Computing stripe %llu block %d\n",
2074                                        (unsigned long long)sh->sector, i);
2075                                 compute_block_1(sh, i, 0);
2076                                 s->uptodate++;
2077                         } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2078                                 /* Computing 2-failure is *very* expensive; only
2079                                  * do it if failed >= 2
2080                                  */
2081                                 int other;
2082                                 for (other = disks; other--; ) {
2083                                         if (other == i)
2084                                                 continue;
2085                                         if (!test_bit(R5_UPTODATE,
2086                                               &sh->dev[other].flags))
2087                                                 break;
2088                                 }
2089                                 BUG_ON(other < 0);
2090                                 pr_debug("Computing stripe %llu blocks %d,%d\n",
2091                                        (unsigned long long)sh->sector,
2092                                        i, other);
2093                                 compute_block_2(sh, i, other);
2094                                 s->uptodate += 2;
2095                         } else if (test_bit(R5_Insync, &dev->flags)) {
2096                                 set_bit(R5_LOCKED, &dev->flags);
2097                                 set_bit(R5_Wantread, &dev->flags);
2098                                 s->locked++;
2099                                 pr_debug("Reading block %d (sync=%d)\n",
2100                                         i, s->syncing);
2101                         }
2102                 }
2103         }
2104         set_bit(STRIPE_HANDLE, &sh->state);
2105 }
2106
2107
2108 /* handle_completed_write_requests
2109  * any written block on an uptodate or failed drive can be returned.
2110  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2111  * never LOCKED, so we don't need to test 'failed' directly.
2112  */
2113 static void handle_completed_write_requests(raid5_conf_t *conf,
2114         struct stripe_head *sh, int disks, struct bio **return_bi)
2115 {
2116         int i;
2117         struct r5dev *dev;
2118
2119         for (i = disks; i--; )
2120                 if (sh->dev[i].written) {
2121                         dev = &sh->dev[i];
2122                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2123                                 test_bit(R5_UPTODATE, &dev->flags)) {
2124                                 /* We can return any write requests */
2125                                 struct bio *wbi, *wbi2;
2126                                 int bitmap_end = 0;
2127                                 pr_debug("Return write for disc %d\n", i);
2128                                 spin_lock_irq(&conf->device_lock);
2129                                 wbi = dev->written;
2130                                 dev->written = NULL;
2131                                 while (wbi && wbi->bi_sector <
2132                                         dev->sector + STRIPE_SECTORS) {
2133                                         wbi2 = r5_next_bio(wbi, dev->sector);
2134                                         if (--wbi->bi_phys_segments == 0) {
2135                                                 md_write_end(conf->mddev);
2136                                                 wbi->bi_next = *return_bi;
2137                                                 *return_bi = wbi;
2138                                         }
2139                                         wbi = wbi2;
2140                                 }
2141                                 if (dev->towrite == NULL)
2142                                         bitmap_end = 1;
2143                                 spin_unlock_irq(&conf->device_lock);
2144                                 if (bitmap_end)
2145                                         bitmap_endwrite(conf->mddev->bitmap,
2146                                                         sh->sector,
2147                                                         STRIPE_SECTORS,
2148                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2149                                                         0);
2150                         }
2151                 }
2152 }
2153
2154 static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
2155                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2156 {
2157         int rmw = 0, rcw = 0, i;
2158         for (i = disks; i--; ) {
2159                 /* would I have to read this buffer for read_modify_write */
2160                 struct r5dev *dev = &sh->dev[i];
2161                 if ((dev->towrite || i == sh->pd_idx) &&
2162                     !test_bit(R5_LOCKED, &dev->flags) &&
2163                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2164                       test_bit(R5_Wantcompute, &dev->flags))) {
2165                         if (test_bit(R5_Insync, &dev->flags))
2166                                 rmw++;
2167                         else
2168                                 rmw += 2*disks;  /* cannot read it */
2169                 }
2170                 /* Would I have to read this buffer for reconstruct_write */
2171                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2172                     !test_bit(R5_LOCKED, &dev->flags) &&
2173                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2174                     test_bit(R5_Wantcompute, &dev->flags))) {
2175                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2176                         else
2177                                 rcw += 2*disks;
2178                 }
2179         }
2180         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2181                 (unsigned long long)sh->sector, rmw, rcw);
2182         set_bit(STRIPE_HANDLE, &sh->state);
2183         if (rmw < rcw && rmw > 0)
2184                 /* prefer read-modify-write, but need to get some data */
2185                 for (i = disks; i--; ) {
2186                         struct r5dev *dev = &sh->dev[i];
2187                         if ((dev->towrite || i == sh->pd_idx) &&
2188                             !test_bit(R5_LOCKED, &dev->flags) &&
2189                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2190                             test_bit(R5_Wantcompute, &dev->flags)) &&
2191                             test_bit(R5_Insync, &dev->flags)) {
2192                                 if (
2193                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2194                                         pr_debug("Read_old block "
2195                                                 "%d for r-m-w\n", i);
2196                                         set_bit(R5_LOCKED, &dev->flags);
2197                                         set_bit(R5_Wantread, &dev->flags);
2198                                         if (!test_and_set_bit(
2199                                                 STRIPE_OP_IO, &sh->ops.pending))
2200                                                 sh->ops.count++;
2201                                         s->locked++;
2202                                 } else {
2203                                         set_bit(STRIPE_DELAYED, &sh->state);
2204                                         set_bit(STRIPE_HANDLE, &sh->state);
2205                                 }
2206                         }
2207                 }
2208         if (rcw <= rmw && rcw > 0)
2209                 /* want reconstruct write, but need to get some data */
2210                 for (i = disks; i--; ) {
2211                         struct r5dev *dev = &sh->dev[i];
2212                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2213                             i != sh->pd_idx &&
2214                             !test_bit(R5_LOCKED, &dev->flags) &&
2215                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2216                             test_bit(R5_Wantcompute, &dev->flags)) &&
2217                             test_bit(R5_Insync, &dev->flags)) {
2218                                 if (
2219                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2220                                         pr_debug("Read_old block "
2221                                                 "%d for Reconstruct\n", i);
2222                                         set_bit(R5_LOCKED, &dev->flags);
2223                                         set_bit(R5_Wantread, &dev->flags);
2224                                         if (!test_and_set_bit(
2225                                                 STRIPE_OP_IO, &sh->ops.pending))
2226                                                 sh->ops.count++;
2227                                         s->locked++;
2228                                 } else {
2229                                         set_bit(STRIPE_DELAYED, &sh->state);
2230                                         set_bit(STRIPE_HANDLE, &sh->state);
2231                                 }
2232                         }
2233                 }
2234         /* now if nothing is locked, and if we have enough data,
2235          * we can start a write request
2236          */
2237         /* since handle_stripe can be called at any time we need to handle the
2238          * case where a compute block operation has been submitted and then a
2239          * subsequent call wants to start a write request.  raid5_run_ops only
2240          * handles the case where compute block and postxor are requested
2241          * simultaneously.  If this is not the case then new writes need to be
2242          * held off until the compute completes.
2243          */
2244         if ((s->req_compute ||
2245             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) &&
2246                 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2247                 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2248                 s->locked += handle_write_operations5(sh, rcw == 0, 0);
2249 }
2250
2251 static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
2252                 struct stripe_head *sh, struct stripe_head_state *s,
2253                 struct r6_state *r6s, int disks)
2254 {
2255         int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2256         int qd_idx = r6s->qd_idx;
2257         for (i = disks; i--; ) {
2258                 struct r5dev *dev = &sh->dev[i];
2259                 /* Would I have to read this buffer for reconstruct_write */
2260                 if (!test_bit(R5_OVERWRITE, &dev->flags)
2261                     && i != pd_idx && i != qd_idx
2262                     && (!test_bit(R5_LOCKED, &dev->flags)
2263                             ) &&
2264                     !test_bit(R5_UPTODATE, &dev->flags)) {
2265                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2266                         else {
2267                                 pr_debug("raid6: must_compute: "
2268                                         "disk %d flags=%#lx\n", i, dev->flags);
2269                                 must_compute++;
2270                         }
2271                 }
2272         }
2273         pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2274                (unsigned long long)sh->sector, rcw, must_compute);
2275         set_bit(STRIPE_HANDLE, &sh->state);
2276
2277         if (rcw > 0)
2278                 /* want reconstruct write, but need to get some data */
2279                 for (i = disks; i--; ) {
2280                         struct r5dev *dev = &sh->dev[i];
2281                         if (!test_bit(R5_OVERWRITE, &dev->flags)
2282                             && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2283                             && !test_bit(R5_LOCKED, &dev->flags) &&
2284                             !test_bit(R5_UPTODATE, &dev->flags) &&
2285                             test_bit(R5_Insync, &dev->flags)) {
2286                                 if (
2287                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2288                                         pr_debug("Read_old stripe %llu "
2289                                                 "block %d for Reconstruct\n",
2290                                              (unsigned long long)sh->sector, i);
2291                                         set_bit(R5_LOCKED, &dev->flags);
2292                                         set_bit(R5_Wantread, &dev->flags);
2293                                         s->locked++;
2294                                 } else {
2295                                         pr_debug("Request delayed stripe %llu "
2296                                                 "block %d for Reconstruct\n",
2297                                              (unsigned long long)sh->sector, i);
2298                                         set_bit(STRIPE_DELAYED, &sh->state);
2299                                         set_bit(STRIPE_HANDLE, &sh->state);
2300                                 }
2301                         }
2302                 }
2303         /* now if nothing is locked, and if we have enough data, we can start a
2304          * write request
2305          */
2306         if (s->locked == 0 && rcw == 0 &&
2307             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2308                 if (must_compute > 0) {
2309                         /* We have failed blocks and need to compute them */
2310                         switch (s->failed) {
2311                         case 0:
2312                                 BUG();
2313                         case 1:
2314                                 compute_block_1(sh, r6s->failed_num[0], 0);
2315                                 break;
2316                         case 2:
2317                                 compute_block_2(sh, r6s->failed_num[0],
2318                                                 r6s->failed_num[1]);
2319                                 break;
2320                         default: /* This request should have been failed? */
2321                                 BUG();
2322                         }
2323                 }
2324
2325                 pr_debug("Computing parity for stripe %llu\n",
2326                         (unsigned long long)sh->sector);
2327                 compute_parity6(sh, RECONSTRUCT_WRITE);
2328                 /* now every locked buffer is ready to be written */
2329                 for (i = disks; i--; )
2330                         if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2331                                 pr_debug("Writing stripe %llu block %d\n",
2332                                        (unsigned long long)sh->sector, i);
2333                                 s->locked++;
2334                                 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2335                         }
2336                 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2337                 set_bit(STRIPE_INSYNC, &sh->state);
2338
2339                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2340                         atomic_dec(&conf->preread_active_stripes);
2341                         if (atomic_read(&conf->preread_active_stripes) <
2342                             IO_THRESHOLD)
2343                                 md_wakeup_thread(conf->mddev->thread);
2344                 }
2345         }
2346 }
2347
2348 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2349                                 struct stripe_head_state *s, int disks)
2350 {
2351         set_bit(STRIPE_HANDLE, &sh->state);
2352         /* Take one of the following actions:
2353          * 1/ start a check parity operation if (uptodate == disks)
2354          * 2/ finish a check parity operation and act on the result
2355          * 3/ skip to the writeback section if we previously
2356          *    initiated a recovery operation
2357          */
2358         if (s->failed == 0 &&
2359             !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2360                 if (!test_and_set_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
2361                         BUG_ON(s->uptodate != disks);
2362                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2363                         sh->ops.count++;
2364                         s->uptodate--;
2365                 } else if (
2366                        test_and_clear_bit(STRIPE_OP_CHECK, &sh->ops.complete)) {
2367                         clear_bit(STRIPE_OP_CHECK, &sh->ops.ack);
2368                         clear_bit(STRIPE_OP_CHECK, &sh->ops.pending);
2369
2370                         if (sh->ops.zero_sum_result == 0)
2371                                 /* parity is correct (on disc,
2372                                  * not in buffer any more)
2373                                  */
2374                                 set_bit(STRIPE_INSYNC, &sh->state);
2375                         else {
2376                                 conf->mddev->resync_mismatches +=
2377                                         STRIPE_SECTORS;
2378                                 if (test_bit(
2379                                      MD_RECOVERY_CHECK, &conf->mddev->recovery))
2380                                         /* don't try to repair!! */
2381                                         set_bit(STRIPE_INSYNC, &sh->state);
2382                                 else {
2383                                         set_bit(STRIPE_OP_COMPUTE_BLK,
2384                                                 &sh->ops.pending);
2385                                         set_bit(STRIPE_OP_MOD_REPAIR_PD,
2386                                                 &sh->ops.pending);
2387                                         set_bit(R5_Wantcompute,
2388                                                 &sh->dev[sh->pd_idx].flags);
2389                                         sh->ops.target = sh->pd_idx;
2390                                         sh->ops.count++;
2391                                         s->uptodate++;
2392                                 }
2393                         }
2394                 }
2395         }
2396
2397         /* check if we can clear a parity disk reconstruct */
2398         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2399                 test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2400
2401                 clear_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending);
2402                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2403                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2404                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2405         }
2406
2407         /* Wait for check parity and compute block operations to complete
2408          * before write-back
2409          */
2410         if (!test_bit(STRIPE_INSYNC, &sh->state) &&
2411                 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending) &&
2412                 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) {
2413                 struct r5dev *dev;
2414                 /* either failed parity check, or recovery is happening */
2415                 if (s->failed == 0)
2416                         s->failed_num = sh->pd_idx;
2417                 dev = &sh->dev[s->failed_num];
2418                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2419                 BUG_ON(s->uptodate != disks);
2420
2421                 set_bit(R5_LOCKED, &dev->flags);
2422                 set_bit(R5_Wantwrite, &dev->flags);
2423                 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2424                         sh->ops.count++;
2425
2426                 clear_bit(STRIPE_DEGRADED, &sh->state);
2427                 s->locked++;
2428                 set_bit(STRIPE_INSYNC, &sh->state);
2429         }
2430 }
2431
2432
2433 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2434                                 struct stripe_head_state *s,
2435                                 struct r6_state *r6s, struct page *tmp_page,
2436                                 int disks)
2437 {
2438         int update_p = 0, update_q = 0;
2439         struct r5dev *dev;
2440         int pd_idx = sh->pd_idx;
2441         int qd_idx = r6s->qd_idx;
2442
2443         set_bit(STRIPE_HANDLE, &sh->state);
2444
2445         BUG_ON(s->failed > 2);
2446         BUG_ON(s->uptodate < disks);
2447         /* Want to check and possibly repair P and Q.
2448          * However there could be one 'failed' device, in which
2449          * case we can only check one of them, possibly using the
2450          * other to generate missing data
2451          */
2452
2453         /* If !tmp_page, we cannot do the calculations,
2454          * but as we have set STRIPE_HANDLE, we will soon be called
2455          * by stripe_handle with a tmp_page - just wait until then.
2456          */
2457         if (tmp_page) {
2458                 if (s->failed == r6s->q_failed) {
2459                         /* The only possible failed device holds 'Q', so it
2460                          * makes sense to check P (If anything else were failed,
2461                          * we would have used P to recreate it).
2462                          */
2463                         compute_block_1(sh, pd_idx, 1);
2464                         if (!page_is_zero(sh->dev[pd_idx].page)) {
2465                                 compute_block_1(sh, pd_idx, 0);
2466                                 update_p = 1;
2467                         }
2468                 }
2469                 if (!r6s->q_failed && s->failed < 2) {
2470                         /* q is not failed, and we didn't use it to generate
2471                          * anything, so it makes sense to check it
2472                          */
2473                         memcpy(page_address(tmp_page),
2474                                page_address(sh->dev[qd_idx].page),
2475                                STRIPE_SIZE);
2476                         compute_parity6(sh, UPDATE_PARITY);
2477                         if (memcmp(page_address(tmp_page),
2478                                    page_address(sh->dev[qd_idx].page),
2479                                    STRIPE_SIZE) != 0) {
2480                                 clear_bit(STRIPE_INSYNC, &sh->state);
2481                                 update_q = 1;
2482                         }
2483                 }
2484                 if (update_p || update_q) {
2485                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2486                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2487                                 /* don't try to repair!! */
2488                                 update_p = update_q = 0;
2489                 }
2490
2491                 /* now write out any block on a failed drive,
2492                  * or P or Q if they need it
2493                  */
2494
2495                 if (s->failed == 2) {
2496                         dev = &sh->dev[r6s->failed_num[1]];
2497                         s->locked++;
2498                         set_bit(R5_LOCKED, &dev->flags);
2499                         set_bit(R5_Wantwrite, &dev->flags);
2500                 }
2501                 if (s->failed >= 1) {
2502                         dev = &sh->dev[r6s->failed_num[0]];
2503                         s->locked++;
2504                         set_bit(R5_LOCKED, &dev->flags);
2505                         set_bit(R5_Wantwrite, &dev->flags);
2506                 }
2507
2508                 if (update_p) {
2509                         dev = &sh->dev[pd_idx];
2510                         s->locked++;
2511                         set_bit(R5_LOCKED, &dev->flags);
2512                         set_bit(R5_Wantwrite, &dev->flags);
2513                 }
2514                 if (update_q) {
2515                         dev = &sh->dev[qd_idx];
2516                         s->locked++;
2517                         set_bit(R5_LOCKED, &dev->flags);
2518                         set_bit(R5_Wantwrite, &dev->flags);
2519                 }
2520                 clear_bit(STRIPE_DEGRADED, &sh->state);
2521
2522                 set_bit(STRIPE_INSYNC, &sh->state);
2523         }
2524 }
2525
2526 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2527                                 struct r6_state *r6s)
2528 {
2529         int i;
2530
2531         /* We have read all the blocks in this stripe and now we need to
2532          * copy some of them into a target stripe for expand.
2533          */
2534         struct dma_async_tx_descriptor *tx = NULL;
2535         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2536         for (i = 0; i < sh->disks; i++)
2537                 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
2538                         int dd_idx, pd_idx, j;
2539                         struct stripe_head *sh2;
2540
2541                         sector_t bn = compute_blocknr(sh, i);
2542                         sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2543                                                 conf->raid_disks -
2544                                                 conf->max_degraded, &dd_idx,
2545                                                 &pd_idx, conf);
2546                         sh2 = get_active_stripe(conf, s, conf->raid_disks,
2547                                                 pd_idx, 1);
2548                         if (sh2 == NULL)
2549                                 /* so far only the early blocks of this stripe
2550                                  * have been requested.  When later blocks
2551                                  * get requested, we will try again
2552                                  */
2553                                 continue;
2554                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2555                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2556                                 /* must have already done this block */
2557                                 release_stripe(sh2);
2558                                 continue;
2559                         }
2560
2561                         /* place all the copies on one channel */
2562                         tx = async_memcpy(sh2->dev[dd_idx].page,
2563                                 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2564                                 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2565
2566                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2567                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2568                         for (j = 0; j < conf->raid_disks; j++)
2569                                 if (j != sh2->pd_idx &&
2570                                     (!r6s || j != raid6_next_disk(sh2->pd_idx,
2571                                                                  sh2->disks)) &&
2572                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2573                                         break;
2574                         if (j == conf->raid_disks) {
2575                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2576                                 set_bit(STRIPE_HANDLE, &sh2->state);
2577                         }
2578                         release_stripe(sh2);
2579
2580                 }
2581         /* done submitting copies, wait for them to complete */
2582         if (tx) {
2583                 async_tx_ack(tx);
2584                 dma_wait_for_async_tx(tx);
2585         }
2586 }
2587
2588 /*
2589  * handle_stripe - do things to a stripe.
2590  *
2591  * We lock the stripe and then examine the state of various bits
2592  * to see what needs to be done.
2593  * Possible results:
2594  *    return some read request which now have data
2595  *    return some write requests which are safely on disc
2596  *    schedule a read on some buffers
2597  *    schedule a write of some buffers
2598  *    return confirmation of parity correctness
2599  *
2600  * buffers are taken off read_list or write_list, and bh_cache buffers
2601  * get BH_Lock set before the stripe lock is released.
2602  *
2603  */
2604
2605 static void handle_stripe5(struct stripe_head *sh)
2606 {
2607         raid5_conf_t *conf = sh->raid_conf;
2608         int disks = sh->disks, i;
2609         struct bio *return_bi = NULL;
2610         struct stripe_head_state s;
2611         struct r5dev *dev;
2612         unsigned long pending = 0;
2613
2614         memset(&s, 0, sizeof(s));
2615         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2616                 "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
2617                 atomic_read(&sh->count), sh->pd_idx,
2618                 sh->ops.pending, sh->ops.ack, sh->ops.complete);
2619
2620         spin_lock(&sh->lock);
2621         clear_bit(STRIPE_HANDLE, &sh->state);
2622         clear_bit(STRIPE_DELAYED, &sh->state);
2623
2624         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2625         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2626         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2627         /* Now to look around and see what can be done */
2628
2629         /* clean-up completed biofill operations */
2630         if (test_bit(STRIPE_OP_BIOFILL, &sh->ops.complete)) {
2631                 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.pending);
2632                 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.ack);
2633                 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.complete);
2634         }
2635
2636         rcu_read_lock();
2637         for (i=disks; i--; ) {
2638                 mdk_rdev_t *rdev;
2639                 struct r5dev *dev = &sh->dev[i];
2640                 clear_bit(R5_Insync, &dev->flags);
2641
2642                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2643                         "written %p\n", i, dev->flags, dev->toread, dev->read,
2644                         dev->towrite, dev->written);
2645
2646                 /* maybe we can request a biofill operation
2647                  *
2648                  * new wantfill requests are only permitted while
2649                  * STRIPE_OP_BIOFILL is clear
2650                  */
2651                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2652                         !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2653                         set_bit(R5_Wantfill, &dev->flags);
2654
2655                 /* now count some things */
2656                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2657                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2658                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2659
2660                 if (test_bit(R5_Wantfill, &dev->flags))
2661                         s.to_fill++;
2662                 else if (dev->toread)
2663                         s.to_read++;
2664                 if (dev->towrite) {
2665                         s.to_write++;
2666                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2667                                 s.non_overwrite++;
2668                 }
2669                 if (dev->written)
2670                         s.written++;
2671                 rdev = rcu_dereference(conf->disks[i].rdev);
2672                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2673                         /* The ReadError flag will just be confusing now */
2674                         clear_bit(R5_ReadError, &dev->flags);
2675                         clear_bit(R5_ReWrite, &dev->flags);
2676                 }
2677                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2678                     || test_bit(R5_ReadError, &dev->flags)) {
2679                         s.failed++;
2680                         s.failed_num = i;
2681                 } else
2682                         set_bit(R5_Insync, &dev->flags);
2683         }
2684         rcu_read_unlock();
2685
2686         if (s.to_fill && !test_and_set_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2687                 sh->ops.count++;
2688
2689         pr_debug("locked=%d uptodate=%d to_read=%d"
2690                 " to_write=%d failed=%d failed_num=%d\n",
2691                 s.locked, s.uptodate, s.to_read, s.to_write,
2692                 s.failed, s.failed_num);
2693         /* check if the array has lost two devices and, if so, some requests might
2694          * need to be failed
2695          */
2696         if (s.failed > 1 && s.to_read+s.to_write+s.written)
2697                 handle_requests_to_failed_array(conf, sh, &s, disks,
2698                                                 &return_bi);
2699         if (s.failed > 1 && s.syncing) {
2700                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2701                 clear_bit(STRIPE_SYNCING, &sh->state);
2702                 s.syncing = 0;
2703         }
2704
2705         /* might be able to return some write requests if the parity block
2706          * is safe, or on a failed drive
2707          */
2708         dev = &sh->dev[sh->pd_idx];
2709         if ( s.written &&
2710              ((test_bit(R5_Insync, &dev->flags) &&
2711                !test_bit(R5_LOCKED, &dev->flags) &&
2712                test_bit(R5_UPTODATE, &dev->flags)) ||
2713                (s.failed == 1 && s.failed_num == sh->pd_idx)))
2714                 handle_completed_write_requests(conf, sh, disks, &return_bi);
2715
2716         /* Now we might consider reading some blocks, either to check/generate
2717          * parity, or to satisfy requests
2718          * or to load a block that is being partially written.
2719          */
2720         if (s.to_read || s.non_overwrite ||
2721             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding ||
2722             test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2723                 handle_issuing_new_read_requests5(sh, &s, disks);
2724
2725         /* Now we check to see if any write operations have recently
2726          * completed
2727          */
2728
2729         /* leave prexor set until postxor is done, allows us to distinguish
2730          * a rmw from a rcw during biodrain
2731          */
2732         if (test_bit(STRIPE_OP_PREXOR, &sh->ops.complete) &&
2733                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2734
2735                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
2736                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.ack);
2737                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
2738
2739                 for (i = disks; i--; )
2740                         clear_bit(R5_Wantprexor, &sh->dev[i].flags);
2741         }
2742
2743         /* if only POSTXOR is set then this is an 'expand' postxor */
2744         if (test_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete) &&
2745                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2746
2747                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
2748                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.ack);
2749                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
2750
2751                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2752                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2753                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2754
2755                 /* All the 'written' buffers and the parity block are ready to
2756                  * be written back to disk
2757                  */
2758                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2759                 for (i = disks; i--; ) {
2760                         dev = &sh->dev[i];
2761                         if (test_bit(R5_LOCKED, &dev->flags) &&
2762                                 (i == sh->pd_idx || dev->written)) {
2763                                 pr_debug("Writing block %d\n", i);
2764                                 set_bit(R5_Wantwrite, &dev->flags);
2765                                 if (!test_and_set_bit(
2766                                     STRIPE_OP_IO, &sh->ops.pending))
2767                                         sh->ops.count++;
2768                                 if (!test_bit(R5_Insync, &dev->flags) ||
2769                                     (i == sh->pd_idx && s.failed == 0))
2770                                         set_bit(STRIPE_INSYNC, &sh->state);
2771                         }
2772                 }
2773                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2774                         atomic_dec(&conf->preread_active_stripes);
2775                         if (atomic_read(&conf->preread_active_stripes) <
2776                                 IO_THRESHOLD)
2777                                 md_wakeup_thread(conf->mddev->thread);
2778                 }
2779         }
2780
2781         /* Now to consider new write requests and what else, if anything
2782          * should be read.  We do not handle new writes when:
2783          * 1/ A 'write' operation (copy+xor) is already in flight.
2784          * 2/ A 'check' operation is in flight, as it may clobber the parity
2785          *    block.
2786          */
2787         if (s.to_write && !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending) &&
2788                           !test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
2789                 handle_issuing_new_write_requests5(conf, sh, &s, disks);
2790
2791         /* maybe we need to check and possibly fix the parity for this stripe
2792          * Any reads will already have been scheduled, so we just see if enough
2793          * data is available.  The parity check is held off while parity
2794          * dependent operations are in flight.
2795          */
2796         if ((s.syncing && s.locked == 0 &&
2797              !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2798              !test_bit(STRIPE_INSYNC, &sh->state)) ||
2799               test_bit(STRIPE_OP_CHECK, &sh->ops.pending) ||
2800               test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending))
2801                 handle_parity_checks5(conf, sh, &s, disks);
2802
2803         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2804                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2805                 clear_bit(STRIPE_SYNCING, &sh->state);
2806         }
2807
2808         /* If the failed drive is just a ReadError, then we might need to progress
2809          * the repair/check process
2810          */
2811         if (s.failed == 1 && !conf->mddev->ro &&
2812             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2813             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2814             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2815                 ) {
2816                 dev = &sh->dev[s.failed_num];
2817                 if (!test_bit(R5_ReWrite, &dev->flags)) {
2818                         set_bit(R5_Wantwrite, &dev->flags);
2819                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2820                                 sh->ops.count++;
2821                         set_bit(R5_ReWrite, &dev->flags);
2822                         set_bit(R5_LOCKED, &dev->flags);
2823                         s.locked++;
2824                 } else {
2825                         /* let's read it back */
2826                         set_bit(R5_Wantread, &dev->flags);
2827                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2828                                 sh->ops.count++;
2829                         set_bit(R5_LOCKED, &dev->flags);
2830                         s.locked++;
2831                 }
2832         }
2833
2834         /* Finish postxor operations initiated by the expansion
2835          * process
2836          */
2837         if (test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete) &&
2838                 !test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending)) {
2839
2840                 clear_bit(STRIPE_EXPANDING, &sh->state);
2841
2842                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2843                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2844                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2845
2846                 for (i = conf->raid_disks; i--; ) {
2847                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
2848                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2849                                 sh->ops.count++;
2850                 }
2851         }
2852
2853         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2854                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2855                 /* Need to write out all blocks after computing parity */
2856                 sh->disks = conf->raid_disks;
2857                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2858                         conf->raid_disks);
2859                 s.locked += handle_write_operations5(sh, 1, 1);
2860         } else if (s.expanded &&
2861                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2862                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2863                 atomic_dec(&conf->reshape_stripes);
2864                 wake_up(&conf->wait_for_overlap);
2865                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2866         }
2867
2868         if (s.expanding && s.locked == 0 &&
2869             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2870                 handle_stripe_expansion(conf, sh, NULL);
2871
2872         if (sh->ops.count)
2873                 pending = get_stripe_work(sh);
2874
2875         spin_unlock(&sh->lock);
2876
2877         if (pending)
2878                 raid5_run_ops(sh, pending);
2879
2880         return_io(return_bi);
2881
2882 }
2883
2884 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2885 {
2886         raid6_conf_t *conf = sh->raid_conf;
2887         int disks = sh->disks;
2888         struct bio *return_bi = NULL;
2889         int i, pd_idx = sh->pd_idx;
2890         struct stripe_head_state s;
2891         struct r6_state r6s;
2892         struct r5dev *dev, *pdev, *qdev;
2893
2894         r6s.qd_idx = raid6_next_disk(pd_idx, disks);
2895         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2896                 "pd_idx=%d, qd_idx=%d\n",
2897                (unsigned long long)sh->sector, sh->state,
2898                atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2899         memset(&s, 0, sizeof(s));
2900
2901         spin_lock(&sh->lock);
2902         clear_bit(STRIPE_HANDLE, &sh->state);
2903         clear_bit(STRIPE_DELAYED, &sh->state);
2904
2905         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2906         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2907         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2908         /* Now to look around and see what can be done */
2909
2910         rcu_read_lock();
2911         for (i=disks; i--; ) {
2912                 mdk_rdev_t *rdev;
2913                 dev = &sh->dev[i];
2914                 clear_bit(R5_Insync, &dev->flags);
2915
2916                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2917                         i, dev->flags, dev->toread, dev->towrite, dev->written);
2918                 /* maybe we can reply to a read */
2919                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2920                         struct bio *rbi, *rbi2;
2921                         pr_debug("Return read for disc %d\n", i);
2922                         spin_lock_irq(&conf->device_lock);
2923                         rbi = dev->toread;
2924                         dev->toread = NULL;
2925                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2926                                 wake_up(&conf->wait_for_overlap);
2927                         spin_unlock_irq(&conf->device_lock);
2928                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2929                                 copy_data(0, rbi, dev->page, dev->sector);
2930                                 rbi2 = r5_next_bio(rbi, dev->sector);
2931                                 spin_lock_irq(&conf->device_lock);
2932                                 if (--rbi->bi_phys_segments == 0) {
2933                                         rbi->bi_next = return_bi;
2934                                         return_bi = rbi;
2935                                 }
2936                                 spin_unlock_irq(&conf->device_lock);
2937                                 rbi = rbi2;
2938                         }
2939                 }
2940
2941                 /* now count some things */
2942                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2943                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2944
2945
2946                 if (dev->toread)
2947                         s.to_read++;
2948                 if (dev->towrite) {
2949                         s.to_write++;
2950                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2951                                 s.non_overwrite++;
2952                 }
2953                 if (dev->written)
2954                         s.written++;
2955                 rdev = rcu_dereference(conf->disks[i].rdev);
2956                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2957                         /* The ReadError flag will just be confusing now */
2958                         clear_bit(R5_ReadError, &dev->flags);
2959                         clear_bit(R5_ReWrite, &dev->flags);
2960                 }
2961                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2962                     || test_bit(R5_ReadError, &dev->flags)) {
2963                         if (s.failed < 2)
2964                                 r6s.failed_num[s.failed] = i;
2965                         s.failed++;
2966                 } else
2967                         set_bit(R5_Insync, &dev->flags);
2968         }
2969         rcu_read_unlock();
2970         pr_debug("locked=%d uptodate=%d to_read=%d"
2971                " to_write=%d failed=%d failed_num=%d,%d\n",
2972                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
2973                r6s.failed_num[0], r6s.failed_num[1]);
2974         /* check if the array has lost >2 devices and, if so, some requests
2975          * might need to be failed