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 {
 693         int disks = sh->disks;
 694         int pd_idx = sh->pd_idx, i;
 695
 696         /* check if prexor is active which means only process blocks
 697          * that are part of a read-modify-write (Wantprexor)
 698          */
 699         int prexor = test_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
 700
 701         pr_debug("%s: stripe %llu\n", __FUNCTION__,
 702                 (unsigned long long)sh->sector);
 703
 704         for (i = disks; i--; ) {
 705                 struct r5dev *dev = &sh->dev[i];
 706                 struct bio *chosen;
 707                 int towrite;
 708
 709                 towrite = 0;
 710                 if (prexor) { /* rmw */
 711                         if (dev->towrite &&
 712                             test_bit(R5_Wantprexor, &dev->flags))
 713                                 towrite = 1;
 714                 } else { /* rcw */
 715                         if (i != pd_idx && dev->towrite &&
 716                                 test_bit(R5_LOCKED, &dev->flags))
 717                                 towrite = 1;
 718                 }
 719
 720                 if (towrite) {
 721                         struct bio *wbi;
 722
 723                         spin_lock(&sh->lock);
 724                         chosen = dev->towrite;
 725                         dev->towrite = NULL;
 726                         BUG_ON(dev->written);
 727                         wbi = dev->written = chosen;
 728                         spin_unlock(&sh->lock);
 729
 730                         while (wbi && wbi->bi_sector <
 731                                 dev->sector + STRIPE_SECTORS) {
 732                                 tx = async_copy_data(1, wbi, dev->page,
 733                                         dev->sector, tx);
 734                                 wbi = r5_next_bio(wbi, dev->sector);
 735                         }
 736                 }
 737         }
 738
 739         return tx;
 740 }
 741
 742 static void ops_complete_postxor(void *stripe_head_ref)
 743 {
 744         struct stripe_head *sh = stripe_head_ref;
 745
 746         pr_debug("%s: stripe %llu\n", __FUNCTION__,
 747                 (unsigned long long)sh->sector);
 748
 749         set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
 750         set_bit(STRIPE_HANDLE, &sh->state);
 751         release_stripe(sh);
 752 }
 753
 754 static void ops_complete_write(void *stripe_head_ref)
 755 {
 756         struct stripe_head *sh = stripe_head_ref;
 757         int disks = sh->disks, i, pd_idx = sh->pd_idx;
 758
 759         pr_debug("%s: stripe %llu\n", __FUNCTION__,
 760                 (unsigned long long)sh->sector);
 761
 762         for (i = disks; i--; ) {
 763                 struct r5dev *dev = &sh->dev[i];
 764                 if (dev->written || i == pd_idx)
 765                         set_bit(R5_UPTODATE, &dev->flags);
 766         }
 767
 768         set_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
 769         set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
 770
 771         set_bit(STRIPE_HANDLE, &sh->state);
 772         release_stripe(sh);
 773 }
 774
 775 static void
 776 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
 777 {
 778         /* kernel stack size limits the total number of disks */
 779         int disks = sh->disks;
 780         struct page *xor_srcs[disks];
 781
 782         int count = 0, pd_idx = sh->pd_idx, i;
 783         struct page *xor_dest;
 784         int prexor = test_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
 785         unsigned long flags;
 786         dma_async_tx_callback callback;
 787
 788         pr_debug("%s: stripe %llu\n", __FUNCTION__,
 789                 (unsigned long long)sh->sector);
 790
 791         /* check if prexor is active which means only process blocks
 792          * that are part of a read-modify-write (written)
 793          */
 794         if (prexor) {
 795                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
 796                 for (i = disks; i--; ) {
 797                         struct r5dev *dev = &sh->dev[i];
 798                         if (dev->written)
 799                                 xor_srcs[count++] = dev->page;
 800                 }
 801         } else {
 802                 xor_dest = sh->dev[pd_idx].page;
 803                 for (i = disks; i--; ) {
 804                         struct r5dev *dev = &sh->dev[i];
 805                         if (i != pd_idx)
 806                                 xor_srcs[count++] = dev->page;
 807                 }
 808         }
 809
 810         /* check whether this postxor is part of a write */
 811         callback = test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending) ?
 812                 ops_complete_write : ops_complete_postxor;
 813
 814         /* 1/ if we prexor'd then the dest is reused as a source
 815          * 2/ if we did not prexor then we are redoing the parity
 816          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
 817          * for the synchronous xor case
 818          */
 819         flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
 820                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
 821
 822         atomic_inc(&sh->count);
 823
 824         if (unlikely(count == 1)) {
 825                 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
 826                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
 827                         flags, tx, callback, sh);
 828         } else
 829                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
 830                         flags, tx, callback, sh);
 831 }
 832
 833 static void ops_complete_check(void *stripe_head_ref)
 834 {
 835         struct stripe_head *sh = stripe_head_ref;
 836         int pd_idx = sh->pd_idx;
 837
 838         pr_debug("%s: stripe %llu\n", __FUNCTION__,
 839                 (unsigned long long)sh->sector);
 840
 841         if (test_and_clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending) &&
 842                 sh->ops.zero_sum_result == 0)
 843                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
 844
 845         set_bit(STRIPE_OP_CHECK, &sh->ops.complete);
 846         set_bit(STRIPE_HANDLE, &sh->state);
 847         release_stripe(sh);
 848 }
 849
 850 static void ops_run_check(struct stripe_head *sh)
 851 {
 852         /* kernel stack size limits the total number of disks */
 853         int disks = sh->disks;
 854         struct page *xor_srcs[disks];
 855         struct dma_async_tx_descriptor *tx;
 856
 857         int count = 0, pd_idx = sh->pd_idx, i;
 858         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
 859
 860         pr_debug("%s: stripe %llu\n", __FUNCTION__,
 861                 (unsigned long long)sh->sector);
 862
 863         for (i = disks; i--; ) {
 864                 struct r5dev *dev = &sh->dev[i];
 865                 if (i != pd_idx)
 866                         xor_srcs[count++] = dev->page;
 867         }
 868
 869         tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
 870                 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
 871
 872         if (tx)
 873                 set_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
 874         else
 875                 clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
 876
 877         atomic_inc(&sh->count);
 878         tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
 879                 ops_complete_check, sh);
 880 }
 881
 882 static void raid5_run_ops(struct stripe_head *sh, unsigned long pending)
 883 {
 884         int overlap_clear = 0, i, disks = sh->disks;
 885         struct dma_async_tx_descriptor *tx = NULL;
 886
 887         if (test_bit(STRIPE_OP_BIOFILL, &pending)) {
 888                 ops_run_biofill(sh);
 889                 overlap_clear++;
 890         }
 891
 892         if (test_bit(STRIPE_OP_COMPUTE_BLK, &pending))
 893                 tx = ops_run_compute5(sh, pending);
 894
 895         if (test_bit(STRIPE_OP_PREXOR, &pending))
 896                 tx = ops_run_prexor(sh, tx);
 897
 898         if (test_bit(STRIPE_OP_BIODRAIN, &pending)) {
 899                 tx = ops_run_biodrain(sh, tx);
 900                 overlap_clear++;
 901         }
 902
 903         if (test_bit(STRIPE_OP_POSTXOR, &pending))
 904                 ops_run_postxor(sh, tx);
 905
 906         if (test_bit(STRIPE_OP_CHECK, &pending))
 907                 ops_run_check(sh);
 908
 909         if (test_bit(STRIPE_OP_IO, &pending))
 910                 ops_run_io(sh);
 911
 912         if (overlap_clear)
 913                 for (i = disks; i--; ) {
 914                         struct r5dev *dev = &sh->dev[i];
 915                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
 916                                 wake_up(&sh->raid_conf->wait_for_overlap);
 917                 }
 918 }
 919
 920 static int grow_one_stripe(raid5_conf_t *conf)
 921 {
 922         struct stripe_head *sh;
 923         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
 924         if (!sh)
 925                 return 0;
 926         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
 927         sh->raid_conf = conf;
 928         spin_lock_init(&sh->lock);
 929
 930         if (grow_buffers(sh, conf->raid_disks)) {
 931                 shrink_buffers(sh, conf->raid_disks);
 932                 kmem_cache_free(conf->slab_cache, sh);
 933                 return 0;
 934         }
 935         sh->disks = conf->raid_disks;
 936         /* we just created an active stripe so... */
 937         atomic_set(&sh->count, 1);
 938         atomic_inc(&conf->active_stripes);
 939         INIT_LIST_HEAD(&sh->lru);
 940         release_stripe(sh);
 941         return 1;
 942 }
 943
 944 static int grow_stripes(raid5_conf_t *conf, int num)
 945 {
 946         struct kmem_cache *sc;
 947         int devs = conf->raid_disks;
 948
 949         sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
 950         sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
 951         conf->active_name = 0;
 952         sc = kmem_cache_create(conf->cache_name[conf->active_name],
 953                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
 954                                0, 0, NULL);
 955         if (!sc)
 956                 return 1;
 957         conf->slab_cache = sc;
 958         conf->pool_size = devs;
 959         while (num--)
 960                 if (!grow_one_stripe(conf))
 961                         return 1;
 962         return 0;
 963 }
 964
 965 #ifdef CONFIG_MD_RAID5_RESHAPE
 966 static int resize_stripes(raid5_conf_t *conf, int newsize)
 967 {
 968         /* Make all the stripes able to hold 'newsize' devices.
 969          * New slots in each stripe get 'page' set to a new page.
 970          *
 971          * This happens in stages:
 972          * 1/ create a new kmem_cache and allocate the required number of
 973          *    stripe_heads.
 974          * 2/ gather all the old stripe_heads and tranfer the pages across
 975          *    to the new stripe_heads.  This will have the side effect of
 976          *    freezing the array as once all stripe_heads have been collected,
 977          *    no IO will be possible.  Old stripe heads are freed once their
 978          *    pages have been transferred over, and the old kmem_cache is
 979          *    freed when all stripes are done.
 980          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
 981          *    we simple return a failre status - no need to clean anything up.
 982          * 4/ allocate new pages for the new slots in the new stripe_heads.
 983          *    If this fails, we don't bother trying the shrink the
 984          *    stripe_heads down again, we just leave them as they are.
 985          *    As each stripe_head is processed the new one is released into
 986          *    active service.
 987          *
 988          * Once step2 is started, we cannot afford to wait for a write,
 989          * so we use GFP_NOIO allocations.
 990          */
 991         struct stripe_head *osh, *nsh;
 992         LIST_HEAD(newstripes);
 993         struct disk_info *ndisks;
 994         int err = 0;
 995         struct kmem_cache *sc;
 996         int i;
 997
 998         if (newsize <= conf->pool_size)
 999                 return 0; /* never bother to shrink */
1000
1001         md_allow_write(conf->mddev);
1002
1003         /* Step 1 */
1004         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1005                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1006                                0, 0, NULL);
1007         if (!sc)
1008                 return -ENOMEM;
1009
1010         for (i = conf->max_nr_stripes; i; i--) {
1011                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1012                 if (!nsh)
1013                         break;
1014
1015                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1016
1017                 nsh->raid_conf = conf;
1018                 spin_lock_init(&nsh->lock);
1019
1020                 list_add(&nsh->lru, &newstripes);
1021         }
1022         if (i) {
1023                 /* didn't get enough, give up */
1024                 while (!list_empty(&newstripes)) {
1025                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1026                         list_del(&nsh->lru);
1027                         kmem_cache_free(sc, nsh);
1028                 }
1029                 kmem_cache_destroy(sc);
1030                 return -ENOMEM;
1031         }
1032         /* Step 2 - Must use GFP_NOIO now.
1033          * OK, we have enough stripes, start collecting inactive
1034          * stripes and copying them over
1035          */
1036         list_for_each_entry(nsh, &newstripes, lru) {
1037                 spin_lock_irq(&conf->device_lock);
1038                 wait_event_lock_irq(conf->wait_for_stripe,
1039                                     !list_empty(&conf->inactive_list),
1040                                     conf->device_lock,
1041                                     unplug_slaves(conf->mddev)
1042                         );
1043                 osh = get_free_stripe(conf);
1044                 spin_unlock_irq(&conf->device_lock);
1045                 atomic_set(&nsh->count, 1);
1046                 for(i=0; i<conf->pool_size; i++)
1047                         nsh->dev[i].page = osh->dev[i].page;
1048                 for( ; i<newsize; i++)
1049                         nsh->dev[i].page = NULL;
1050                 kmem_cache_free(conf->slab_cache, osh);
1051         }
1052         kmem_cache_destroy(conf->slab_cache);
1053
1054         /* Step 3.
1055          * At this point, we are holding all the stripes so the array
1056          * is completely stalled, so now is a good time to resize
1057          * conf->disks.
1058          */
1059         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1060         if (ndisks) {
1061                 for (i=0; i<conf->raid_disks; i++)
1062                         ndisks[i] = conf->disks[i];
1063                 kfree(conf->disks);
1064                 conf->disks = ndisks;
1065         } else
1066                 err = -ENOMEM;
1067
1068         /* Step 4, return new stripes to service */
1069         while(!list_empty(&newstripes)) {
1070                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1071                 list_del_init(&nsh->lru);
1072                 for (i=conf->raid_disks; i < newsize; i++)
1073                         if (nsh->dev[i].page == NULL) {
1074                                 struct page *p = alloc_page(GFP_NOIO);
1075                                 nsh->dev[i].page = p;
1076                                 if (!p)
1077                                         err = -ENOMEM;
1078                         }
1079                 release_stripe(nsh);
1080         }
1081         /* critical section pass, GFP_NOIO no longer needed */
1082
1083         conf->slab_cache = sc;
1084         conf->active_name = 1-conf->active_name;
1085         conf->pool_size = newsize;
1086         return err;
1087 }
1088 #endif
1089
1090 static int drop_one_stripe(raid5_conf_t *conf)
1091 {
1092         struct stripe_head *sh;
1093
1094         spin_lock_irq(&conf->device_lock);
1095         sh = get_free_stripe(conf);
1096         spin_unlock_irq(&conf->device_lock);
1097         if (!sh)
1098                 return 0;
1099         BUG_ON(atomic_read(&sh->count));
1100         shrink_buffers(sh, conf->pool_size);
1101         kmem_cache_free(conf->slab_cache, sh);
1102         atomic_dec(&conf->active_stripes);
1103         return 1;
1104 }
1105
1106 static void shrink_stripes(raid5_conf_t *conf)
1107 {
1108         while (drop_one_stripe(conf))
1109                 ;
1110
1111         if (conf->slab_cache)
1112                 kmem_cache_destroy(conf->slab_cache);
1113         conf->slab_cache = NULL;
1114 }
1115
1116 static void raid5_end_read_request(struct bio * bi, int error)
1117 {
1118         struct stripe_head *sh = bi->bi_private;
1119         raid5_conf_t *conf = sh->raid_conf;
1120         int disks = sh->disks, i;
1121         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1122         char b[BDEVNAME_SIZE];
1123         mdk_rdev_t *rdev;
1124
1125
1126         for (i=0 ; i<disks; i++)
1127                 if (bi == &sh->dev[i].req)
1128                         break;
1129
1130         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1131                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1132                 uptodate);
1133         if (i == disks) {
1134                 BUG();
1135                 return;
1136         }
1137
1138         if (uptodate) {
1139                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1140                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1141                         rdev = conf->disks[i].rdev;
1142                         printk(KERN_INFO "raid5:%s: read error corrected (%lu sectors at %llu on %s)\n",
1143                                mdname(conf->mddev), STRIPE_SECTORS,
1144                                (unsigned long long)sh->sector + rdev->data_offset,
1145                                bdevname(rdev->bdev, b));
1146                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1147                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1148                 }
1149                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1150                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1151         } else {
1152                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1153                 int retry = 0;
1154                 rdev = conf->disks[i].rdev;
1155
1156                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1157                 atomic_inc(&rdev->read_errors);
1158                 if (conf->mddev->degraded)
1159                         printk(KERN_WARNING "raid5:%s: read error not correctable (sector %llu on %s).\n",
1160                                mdname(conf->mddev),
1161                                (unsigned long long)sh->sector + rdev->data_offset,
1162                                bdn);
1163                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1164                         /* Oh, no!!! */
1165                         printk(KERN_WARNING "raid5:%s: read error NOT corrected!! (sector %llu on %s).\n",
1166                                mdname(conf->mddev),
1167                                (unsigned long long)sh->sector + rdev->data_offset,
1168                                bdn);
1169                 else if (atomic_read(&rdev->read_errors)
1170                          > conf->max_nr_stripes)
1171                         printk(KERN_WARNING
1172                                "raid5:%s: Too many read errors, failing device %s.\n",
1173                                mdname(conf->mddev), bdn);
1174                 else
1175                         retry = 1;
1176                 if (retry)
1177                         set_bit(R5_ReadError, &sh->dev[i].flags);
1178                 else {
1179                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1180                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1181                         md_error(conf->mddev, rdev);
1182                 }
1183         }
1184         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1185         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1186         set_bit(STRIPE_HANDLE, &sh->state);
1187         release_stripe(sh);
1188 }
1189
1190 static void raid5_end_write_request (struct bio *bi, int error)
1191 {
1192         struct stripe_head *sh = bi->bi_private;
1193         raid5_conf_t *conf = sh->raid_conf;
1194         int disks = sh->disks, i;
1195         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1196
1197         for (i=0 ; i<disks; i++)
1198                 if (bi == &sh->dev[i].req)
1199                         break;
1200
1201         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1202                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1203                 uptodate);
1204         if (i == disks) {
1205                 BUG();
1206                 return;
1207         }
1208
1209         if (!uptodate)
1210                 md_error(conf->mddev, conf->disks[i].rdev);
1211
1212         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1213
1214         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1215         set_bit(STRIPE_HANDLE, &sh->state);
1216         release_stripe(sh);
1217 }
1218
1219
1220 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1221
1222 static void raid5_build_block (struct stripe_head *sh, int i)
1223 {
1224         struct r5dev *dev = &sh->dev[i];
1225
1226         bio_init(&dev->req);
1227         dev->req.bi_io_vec = &dev->vec;
1228         dev->req.bi_vcnt++;
1229         dev->req.bi_max_vecs++;
1230         dev->vec.bv_page = dev->page;
1231         dev->vec.bv_len = STRIPE_SIZE;
1232         dev->vec.bv_offset = 0;
1233
1234         dev->req.bi_sector = sh->sector;
1235         dev->req.bi_private = sh;
1236
1237         dev->flags = 0;
1238         dev->sector = compute_blocknr(sh, i);
1239 }
1240
1241 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1242 {
1243         char b[BDEVNAME_SIZE];
1244         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1245         pr_debug("raid5: error called\n");
1246
1247         if (!test_bit(Faulty, &rdev->flags)) {
1248                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1249                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1250                         unsigned long flags;
1251                         spin_lock_irqsave(&conf->device_lock, flags);
1252                         mddev->degraded++;
1253                         spin_unlock_irqrestore(&conf->device_lock, flags);
1254                         /*
1255                          * if recovery was running, make sure it aborts.
1256                          */
1257                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
1258                 }
1259                 set_bit(Faulty, &rdev->flags);
1260                 printk (KERN_ALERT
1261                         "raid5: Disk failure on %s, disabling device."
1262                         " Operation continuing on %d devices\n",
1263                         bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1264         }
1265 }
1266
1267 /*
1268  * Input: a 'big' sector number,
1269  * Output: index of the data and parity disk, and the sector # in them.
1270  */
1271 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
1272                         unsigned int data_disks, unsigned int * dd_idx,
1273                         unsigned int * pd_idx, raid5_conf_t *conf)
1274 {
1275         long stripe;
1276         unsigned long chunk_number;
1277         unsigned int chunk_offset;
1278         sector_t new_sector;
1279         int sectors_per_chunk = conf->chunk_size >> 9;
1280
1281         /* First compute the information on this sector */
1282
1283         /*
1284          * Compute the chunk number and the sector offset inside the chunk
1285          */
1286         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1287         chunk_number = r_sector;
1288         BUG_ON(r_sector != chunk_number);
1289
1290         /*
1291          * Compute the stripe number
1292          */
1293         stripe = chunk_number / data_disks;
1294
1295         /*
1296          * Compute the data disk and parity disk indexes inside the stripe
1297          */
1298         *dd_idx = chunk_number % data_disks;
1299
1300         /*
1301          * Select the parity disk based on the user selected algorithm.
1302          */
1303         switch(conf->level) {
1304         case 4:
1305                 *pd_idx = data_disks;
1306                 break;
1307         case 5:
1308                 switch (conf->algorithm) {
1309                 case ALGORITHM_LEFT_ASYMMETRIC:
1310                         *pd_idx = data_disks - stripe % raid_disks;
1311                         if (*dd_idx >= *pd_idx)
1312                                 (*dd_idx)++;
1313                         break;
1314                 case ALGORITHM_RIGHT_ASYMMETRIC:
1315                         *pd_idx = stripe % raid_disks;
1316                         if (*dd_idx >= *pd_idx)
1317                                 (*dd_idx)++;
1318                         break;
1319                 case ALGORITHM_LEFT_SYMMETRIC:
1320                         *pd_idx = data_disks - stripe % raid_disks;
1321                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1322                         break;
1323                 case ALGORITHM_RIGHT_SYMMETRIC:
1324                         *pd_idx = stripe % raid_disks;
1325                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1326                         break;
1327                 default:
1328                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1329                                 conf->algorithm);
1330                 }
1331                 break;
1332         case 6:
1333
1334                 /**** FIX THIS ****/
1335                 switch (conf->algorithm) {
1336                 case ALGORITHM_LEFT_ASYMMETRIC:
1337                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1338                         if (*pd_idx == raid_disks-1)
1339                                 (*dd_idx)++;    /* Q D D D P */
1340                         else if (*dd_idx >= *pd_idx)
1341                                 (*dd_idx) += 2; /* D D P Q D */
1342                         break;
1343                 case ALGORITHM_RIGHT_ASYMMETRIC:
1344                         *pd_idx = stripe % raid_disks;
1345                         if (*pd_idx == raid_disks-1)
1346                                 (*dd_idx)++;    /* Q D D D P */
1347                         else if (*dd_idx >= *pd_idx)
1348                                 (*dd_idx) += 2; /* D D P Q D */
1349                         break;
1350                 case ALGORITHM_LEFT_SYMMETRIC:
1351                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1352                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1353                         break;
1354                 case ALGORITHM_RIGHT_SYMMETRIC:
1355                         *pd_idx = stripe % raid_disks;
1356                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1357                         break;
1358                 default:
1359                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1360                                 conf->algorithm);
1361                 }
1362                 break;
1363         }
1364
1365         /*
1366          * Finally, compute the new sector number
1367          */
1368         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1369         return new_sector;
1370 }
1371
1372
1373 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1374 {
1375         raid5_conf_t *conf = sh->raid_conf;
1376         int raid_disks = sh->disks;
1377         int data_disks = raid_disks - conf->max_degraded;
1378         sector_t new_sector = sh->sector, check;
1379         int sectors_per_chunk = conf->chunk_size >> 9;
1380         sector_t stripe;
1381         int chunk_offset;
1382         int chunk_number, dummy1, dummy2, dd_idx = i;
1383         sector_t r_sector;
1384
1385
1386         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1387         stripe = new_sector;
1388         BUG_ON(new_sector != stripe);
1389
1390         if (i == sh->pd_idx)
1391                 return 0;
1392         switch(conf->level) {
1393         case 4: break;
1394         case 5:
1395                 switch (conf->algorithm) {
1396                 case ALGORITHM_LEFT_ASYMMETRIC:
1397                 case ALGORITHM_RIGHT_ASYMMETRIC:
1398                         if (i > sh->pd_idx)
1399                                 i--;
1400                         break;
1401                 case ALGORITHM_LEFT_SYMMETRIC:
1402                 case ALGORITHM_RIGHT_SYMMETRIC:
1403                         if (i < sh->pd_idx)
1404                                 i += raid_disks;
1405                         i -= (sh->pd_idx + 1);
1406                         break;
1407                 default:
1408                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1409                                conf->algorithm);
1410                 }
1411                 break;
1412         case 6:
1413                 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
1414                         return 0; /* It is the Q disk */
1415                 switch (conf->algorithm) {
1416                 case ALGORITHM_LEFT_ASYMMETRIC:
1417                 case ALGORITHM_RIGHT_ASYMMETRIC:
1418                         if (sh->pd_idx == raid_disks-1)
1419                                 i--;    /* Q D D D P */
1420                         else if (i > sh->pd_idx)
1421                                 i -= 2; /* D D P Q D */
1422                         break;
1423                 case ALGORITHM_LEFT_SYMMETRIC:
1424                 case ALGORITHM_RIGHT_SYMMETRIC:
1425                         if (sh->pd_idx == raid_disks-1)
1426                                 i--; /* Q D D D P */
1427                         else {
1428                                 /* D D P Q D */
1429                                 if (i < sh->pd_idx)
1430                                         i += raid_disks;
1431                                 i -= (sh->pd_idx + 2);
1432                         }
1433                         break;
1434                 default:
1435                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1436                                 conf->algorithm);
1437                 }
1438                 break;
1439         }
1440
1441         chunk_number = stripe * data_disks + i;
1442         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1443
1444         check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
1445         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
1446                 printk(KERN_ERR "compute_blocknr: map not correct\n");
1447                 return 0;
1448         }
1449         return r_sector;
1450 }
1451
1452
1453
1454 /*
1455  * Copy data between a page in the stripe cache, and one or more bion
1456  * The page could align with the middle of the bio, or there could be
1457  * several bion, each with several bio_vecs, which cover part of the page
1458  * Multiple bion are linked together on bi_next.  There may be extras
1459  * at the end of this list.  We ignore them.
1460  */
1461 static void copy_data(int frombio, struct bio *bio,
1462                      struct page *page,
1463                      sector_t sector)
1464 {
1465         char *pa = page_address(page);
1466         struct bio_vec *bvl;
1467         int i;
1468         int page_offset;
1469
1470         if (bio->bi_sector >= sector)
1471                 page_offset = (signed)(bio->bi_sector - sector) * 512;
1472         else
1473                 page_offset = (signed)(sector - bio->bi_sector) * -512;
1474         bio_for_each_segment(bvl, bio, i) {
1475                 int len = bio_iovec_idx(bio,i)->bv_len;
1476                 int clen;
1477                 int b_offset = 0;
1478
1479                 if (page_offset < 0) {
1480                         b_offset = -page_offset;
1481                         page_offset += b_offset;
1482                         len -= b_offset;
1483                 }
1484
1485                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1486                         clen = STRIPE_SIZE - page_offset;
1487                 else clen = len;
1488
1489                 if (clen > 0) {
1490                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1491                         if (frombio)
1492                                 memcpy(pa+page_offset, ba+b_offset, clen);
1493                         else
1494                                 memcpy(ba+b_offset, pa+page_offset, clen);
1495                         __bio_kunmap_atomic(ba, KM_USER0);
1496                 }
1497                 if (clen < len) /* hit end of page */
1498                         break;
1499                 page_offset +=  len;
1500         }
1501 }
1502
1503 #define check_xor()     do {                                              \
1504                                 if (count == MAX_XOR_BLOCKS) {            \
1505                                 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1506                                 count = 0;                                \
1507                            }                                              \
1508                         } while(0)
1509
1510 static void compute_parity6(struct stripe_head *sh, int method)
1511 {
1512         raid6_conf_t *conf = sh->raid_conf;
1513         int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1514         struct bio *chosen;
1515         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1516         void *ptrs[disks];
1517
1518         qd_idx = raid6_next_disk(pd_idx, disks);
1519         d0_idx = raid6_next_disk(qd_idx, disks);
1520
1521         pr_debug("compute_parity, stripe %llu, method %d\n",
1522                 (unsigned long long)sh->sector, method);
1523
1524         switch(method) {
1525         case READ_MODIFY_WRITE:
1526                 BUG();          /* READ_MODIFY_WRITE N/A for RAID-6 */
1527         case RECONSTRUCT_WRITE:
1528                 for (i= disks; i-- ;)
1529                         if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1530                                 chosen = sh->dev[i].towrite;
1531                                 sh->dev[i].towrite = NULL;
1532
1533                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1534                                         wake_up(&conf->wait_for_overlap);
1535
1536                                 BUG_ON(sh->dev[i].written);
1537                                 sh->dev[i].written = chosen;
1538                         }
1539                 break;
1540         case CHECK_PARITY:
1541                 BUG();          /* Not implemented yet */
1542         }
1543
1544         for (i = disks; i--;)
1545                 if (sh->dev[i].written) {
1546                         sector_t sector = sh->dev[i].sector;
1547                         struct bio *wbi = sh->dev[i].written;
1548                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1549                                 copy_data(1, wbi, sh->dev[i].page, sector);
1550                                 wbi = r5_next_bio(wbi, sector);
1551                         }
1552
1553                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1554                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
1555                 }
1556
1557 //      switch(method) {
1558 //      case RECONSTRUCT_WRITE:
1559 //      case CHECK_PARITY:
1560 //      case UPDATE_PARITY:
1561                 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1562                 /* FIX: Is this ordering of drives even remotely optimal? */
1563                 count = 0;
1564                 i = d0_idx;
1565                 do {
1566                         ptrs[count++] = page_address(sh->dev[i].page);
1567                         if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1568                                 printk("block %d/%d not uptodate on parity calc\n", i,count);
1569                         i = raid6_next_disk(i, disks);
1570                 } while ( i != d0_idx );
1571 //              break;
1572 //      }
1573
1574         raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1575
1576         switch(method) {
1577         case RECONSTRUCT_WRITE:
1578                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1579                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1580                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
1581                 set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
1582                 break;
1583         case UPDATE_PARITY:
1584                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1585                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1586                 break;
1587         }
1588 }
1589
1590
1591 /* Compute one missing block */
1592 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1593 {
1594         int i, count, disks = sh->disks;
1595         void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1596         int pd_idx = sh->pd_idx;
1597         int qd_idx = raid6_next_disk(pd_idx, disks);
1598
1599         pr_debug("compute_block_1, stripe %llu, idx %d\n",
1600                 (unsigned long long)sh->sector, dd_idx);
1601
1602         if ( dd_idx == qd_idx ) {
1603                 /* We're actually computing the Q drive */
1604                 compute_parity6(sh, UPDATE_PARITY);
1605         } else {
1606                 dest = page_address(sh->dev[dd_idx].page);
1607                 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1608                 count = 0;
1609                 for (i = disks ; i--; ) {
1610                         if (i == dd_idx || i == qd_idx)
1611                                 continue;
1612                         p = page_address(sh->dev[i].page);
1613                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1614                                 ptr[count++] = p;
1615                         else
1616                                 printk("compute_block() %d, stripe %llu, %d"
1617                                        " not present\n", dd_idx,
1618                                        (unsigned long long)sh->sector, i);
1619
1620                         check_xor();
1621                 }
1622                 if (count)
1623                         xor_blocks(count, STRIPE_SIZE, dest, ptr);
1624                 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1625                 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1626         }
1627 }
1628
1629 /* Compute two missing blocks */
1630 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1631 {
1632         int i, count, disks = sh->disks;
1633         int pd_idx = sh->pd_idx;
1634         int qd_idx = raid6_next_disk(pd_idx, disks);
1635         int d0_idx = raid6_next_disk(qd_idx, disks);
1636         int faila, failb;
1637
1638         /* faila and failb are disk numbers relative to d0_idx */
1639         /* pd_idx become disks-2 and qd_idx become disks-1 */
1640         faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1641         failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1642
1643         BUG_ON(faila == failb);
1644         if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1645
1646         pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1647                (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1648
1649         if ( failb == disks-1 ) {
1650                 /* Q disk is one of the missing disks */
1651                 if ( faila == disks-2 ) {
1652                         /* Missing P+Q, just recompute */
1653                         compute_parity6(sh, UPDATE_PARITY);
1654                         return;
1655                 } else {
1656                         /* We're missing D+Q; recompute D from P */
1657                         compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1658                         compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1659                         return;
1660                 }
1661         }
1662
1663         /* We're missing D+P or D+D; build pointer table */
1664         {
1665                 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1666                 void *ptrs[disks];
1667
1668                 count = 0;
1669                 i = d0_idx;
1670                 do {
1671                         ptrs[count++] = page_address(sh->dev[i].page);
1672                         i = raid6_next_disk(i, disks);
1673                         if (i != dd_idx1 && i != dd_idx2 &&
1674                             !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1675                                 printk("compute_2 with missing block %d/%d\n", count, i);
1676                 } while ( i != d0_idx );
1677
1678                 if ( failb == disks-2 ) {
1679                         /* We're missing D+P. */
1680                         raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1681                 } else {
1682                         /* We're missing D+D. */
1683                         raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1684                 }
1685
1686                 /* Both the above update both missing blocks */
1687                 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1688                 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1689         }
1690 }
1691
1692 static int
1693 handle_write_operations5(struct stripe_head *sh, int rcw, int expand)
1694 {
1695         int i, pd_idx = sh->pd_idx, disks = sh->disks;
1696         int locked = 0;
1697
1698         if (rcw) {
1699                 /* if we are not expanding this is a proper write request, and
1700                  * there will be bios with new data to be drained into the
1701                  * stripe cache
1702                  */
1703                 if (!expand) {
1704                         set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1705                         sh->ops.count++;
1706                 }
1707
1708                 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1709                 sh->ops.count++;
1710
1711                 for (i = disks; i--; ) {
1712                         struct r5dev *dev = &sh->dev[i];
1713
1714                         if (dev->towrite) {
1715                                 set_bit(R5_LOCKED, &dev->flags);
1716                                 if (!expand)
1717                                         clear_bit(R5_UPTODATE, &dev->flags);
1718                                 locked++;
1719                         }
1720                 }
1721         } else {
1722                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1723                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1724
1725                 set_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
1726                 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1727                 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1728
1729                 sh->ops.count += 3;
1730
1731                 for (i = disks; i--; ) {
1732                         struct r5dev *dev = &sh->dev[i];
1733                         if (i == pd_idx)
1734                                 continue;
1735
1736                         /* For a read-modify write there may be blocks that are
1737                          * locked for reading while others are ready to be
1738                          * written so we distinguish these blocks by the
1739                          * R5_Wantprexor bit
1740                          */
1741                         if (dev->towrite &&
1742                             (test_bit(R5_UPTODATE, &dev->flags) ||
1743                             test_bit(R5_Wantcompute, &dev->flags))) {
1744                                 set_bit(R5_Wantprexor, &dev->flags);
1745                                 set_bit(R5_LOCKED, &dev->flags);
1746                                 clear_bit(R5_UPTODATE, &dev->flags);
1747                                 locked++;
1748                         }
1749                 }
1750         }
1751
1752         /* keep the parity disk locked while asynchronous operations
1753          * are in flight
1754          */
1755         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1756         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1757         locked++;
1758
1759         pr_debug("%s: stripe %llu locked: %d pending: %lx\n",
1760                 __FUNCTION__, (unsigned long long)sh->sector,
1761                 locked, sh->ops.pending);
1762
1763         return locked;
1764 }
1765
1766 /*
1767  * Each stripe/dev can have one or more bion attached.
1768  * toread/towrite point to the first in a chain.
1769  * The bi_next chain must be in order.
1770  */
1771 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1772 {
1773         struct bio **bip;
1774         raid5_conf_t *conf = sh->raid_conf;
1775         int firstwrite=0;
1776
1777         pr_debug("adding bh b#%llu to stripe s#%llu\n",
1778                 (unsigned long long)bi->bi_sector,
1779                 (unsigned long long)sh->sector);
1780
1781
1782         spin_lock(&sh->lock);
1783         spin_lock_irq(&conf->device_lock);
1784         if (forwrite) {
1785                 bip = &sh->dev[dd_idx].towrite;
1786                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1787                         firstwrite = 1;
1788         } else
1789                 bip = &sh->dev[dd_idx].toread;
1790         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1791                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1792                         goto overlap;
1793                 bip = & (*bip)->bi_next;
1794         }
1795         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1796                 goto overlap;
1797
1798         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1799         if (*bip)
1800                 bi->bi_next = *bip;
1801         *bip = bi;
1802         bi->bi_phys_segments ++;
1803         spin_unlock_irq(&conf->device_lock);
1804         spin_unlock(&sh->lock);
1805
1806         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1807                 (unsigned long long)bi->bi_sector,
1808                 (unsigned long long)sh->sector, dd_idx);
1809
1810         if (conf->mddev->bitmap && firstwrite) {
1811                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1812                                   STRIPE_SECTORS, 0);
1813                 sh->bm_seq = conf->seq_flush+1;
1814                 set_bit(STRIPE_BIT_DELAY, &sh->state);
1815         }
1816
1817         if (forwrite) {
1818                 /* check if page is covered */
1819                 sector_t sector = sh->dev[dd_idx].sector;
1820                 for (bi=sh->dev[dd_idx].towrite;
1821                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1822                              bi && bi->bi_sector <= sector;
1823                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1824                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1825                                 sector = bi->bi_sector + (bi->bi_size>>9);
1826                 }
1827                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1828                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1829         }
1830         return 1;
1831
1832  overlap:
1833         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1834         spin_unlock_irq(&conf->device_lock);
1835         spin_unlock(&sh->lock);
1836         return 0;
1837 }
1838
1839 static void end_reshape(raid5_conf_t *conf);
1840
1841 static int page_is_zero(struct page *p)
1842 {
1843         char *a = page_address(p);
1844         return ((*(u32*)a) == 0 &&
1845                 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1846 }
1847
1848 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1849 {
1850         int sectors_per_chunk = conf->chunk_size >> 9;
1851         int pd_idx, dd_idx;
1852         int chunk_offset = sector_div(stripe, sectors_per_chunk);
1853
1854         raid5_compute_sector(stripe * (disks - conf->max_degraded)
1855                              *sectors_per_chunk + chunk_offset,
1856                              disks, disks - conf->max_degraded,
1857                              &dd_idx, &pd_idx, conf);
1858         return pd_idx;
1859 }
1860
1861 static void
1862 handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh,
1863                                 struct stripe_head_state *s, int disks,
1864                                 struct bio **return_bi)
1865 {
1866         int i;
1867         for (i = disks; i--; ) {
1868                 struct bio *bi;
1869                 int bitmap_end = 0;
1870
1871                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1872                         mdk_rdev_t *rdev;
1873                         rcu_read_lock();
1874                         rdev = rcu_dereference(conf->disks[i].rdev);
1875                         if (rdev && test_bit(In_sync, &rdev->flags))
1876                                 /* multiple read failures in one stripe */
1877                                 md_error(conf->mddev, rdev);
1878                         rcu_read_unlock();
1879                 }
1880                 spin_lock_irq(&conf->device_lock);
1881                 /* fail all writes first */
1882                 bi = sh->dev[i].towrite;
1883                 sh->dev[i].towrite = NULL;
1884                 if (bi) {
1885                         s->to_write--;
1886                         bitmap_end = 1;
1887                 }
1888
1889                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1890                         wake_up(&conf->wait_for_overlap);
1891
1892                 while (bi && bi->bi_sector <
1893                         sh->dev[i].sector + STRIPE_SECTORS) {
1894                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1895                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1896                         if (--bi->bi_phys_segments == 0) {
1897                                 md_write_end(conf->mddev);
1898                                 bi->bi_next = *return_bi;
1899                                 *return_bi = bi;
1900                         }
1901                         bi = nextbi;
1902                 }
1903                 /* and fail all 'written' */
1904                 bi = sh->dev[i].written;
1905                 sh->dev[i].written = NULL;
1906                 if (bi) bitmap_end = 1;
1907                 while (bi && bi->bi_sector <
1908                        sh->dev[i].sector + STRIPE_SECTORS) {
1909                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1910                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1911                         if (--bi->bi_phys_segments == 0) {
1912                                 md_write_end(conf->mddev);
1913                                 bi->bi_next = *return_bi;
1914                                 *return_bi = bi;
1915                         }
1916                         bi = bi2;
1917                 }
1918
1919                 /* fail any reads if this device is non-operational and
1920                  * the data has not reached the cache yet.
1921                  */
1922                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
1923                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1924                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
1925                         bi = sh->dev[i].toread;
1926                         sh->dev[i].toread = NULL;
1927                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1928                                 wake_up(&conf->wait_for_overlap);
1929                         if (bi) s->to_read--;
1930                         while (bi && bi->bi_sector <
1931                                sh->dev[i].sector + STRIPE_SECTORS) {
1932                                 struct bio *nextbi =
1933                                         r5_next_bio(bi, sh->dev[i].sector);
1934                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1935                                 if (--bi->bi_phys_segments == 0) {
1936                                         bi->bi_next = *return_bi;
1937                                         *return_bi = bi;
1938                                 }
1939                                 bi = nextbi;
1940                         }
1941                 }
1942                 spin_unlock_irq(&conf->device_lock);
1943                 if (bitmap_end)
1944                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1945                                         STRIPE_SECTORS, 0, 0);
1946         }
1947
1948 }
1949
1950 /* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks
1951  * to process
1952  */
1953 static int __handle_issuing_new_read_requests5(struct stripe_head *sh,
1954                         struct stripe_head_state *s, int disk_idx, int disks)
1955 {
1956         struct r5dev *dev = &sh->dev[disk_idx];
1957         struct r5dev *failed_dev = &sh->dev[s->failed_num];
1958
1959         /* don't schedule compute operations or reads on the parity block while
1960          * a check is in flight
1961          */
1962         if ((disk_idx == sh->pd_idx) &&
1963              test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
1964                 return ~0;
1965
1966         /* is the data in this block needed, and can we get it? */
1967         if (!test_bit(R5_LOCKED, &dev->flags) &&
1968             !test_bit(R5_UPTODATE, &dev->flags) && (dev->toread ||
1969             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1970              s->syncing || s->expanding || (s->failed &&
1971              (failed_dev->toread || (failed_dev->towrite &&
1972              !test_bit(R5_OVERWRITE, &failed_dev->flags)
1973              ))))) {
1974                 /* 1/ We would like to get this block, possibly by computing it,
1975                  * but we might not be able to.
1976                  *
1977                  * 2/ Since parity check operations potentially make the parity
1978                  * block !uptodate it will need to be refreshed before any
1979                  * compute operations on data disks are scheduled.
1980                  *
1981                  * 3/ We hold off parity block re-reads until check operations
1982                  * have quiesced.
1983                  */
1984                 if ((s->uptodate == disks - 1) &&
1985                     !test_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
1986                         set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
1987                         set_bit(R5_Wantcompute, &dev->flags);
1988                         sh->ops.target = disk_idx;
1989                         s->req_compute = 1;
1990                         sh->ops.count++;
1991                         /* Careful: from this point on 'uptodate' is in the eye
1992                          * of raid5_run_ops which services 'compute' operations
1993                          * before writes. R5_Wantcompute flags a block that will
1994                          * be R5_UPTODATE by the time it is needed for a
1995                          * subsequent operation.
1996                          */
1997                         s->uptodate++;
1998                         return 0; /* uptodate + compute == disks */
1999                 } else if ((s->uptodate < disks - 1) &&
2000                         test_bit(R5_Insync, &dev->flags)) {
2001                         /* Note: we hold off compute operations while checks are
2002                          * in flight, but we still prefer 'compute' over 'read'
2003                          * hence we only read if (uptodate < * disks-1)
2004                          */
2005                         set_bit(R5_LOCKED, &dev->flags);
2006                         set_bit(R5_Wantread, &dev->flags);
2007                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2008                                 sh->ops.count++;
2009                         s->locked++;
2010                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2011                                 s->syncing);
2012                 }
2013         }
2014
2015         return ~0;
2016 }
2017
2018 static void handle_issuing_new_read_requests5(struct stripe_head *sh,
2019                         struct stripe_head_state *s, int disks)
2020 {
2021         int i;
2022
2023         /* Clear completed compute operations.  Parity recovery
2024          * (STRIPE_OP_MOD_REPAIR_PD) implies a write-back which is handled
2025          * later on in this routine
2026          */
2027         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2028                 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2029                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2030                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2031                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2032         }
2033
2034         /* look for blocks to read/compute, skip this if a compute
2035          * is already in flight, or if the stripe contents are in the
2036          * midst of changing due to a write
2037          */
2038         if (!test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2039                 !test_bit(STRIPE_OP_PREXOR, &sh->ops.pending) &&
2040                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2041                 for (i = disks; i--; )
2042                         if (__handle_issuing_new_read_requests5(
2043                                 sh, s, i, disks) == 0)
2044                                 break;
2045         }
2046         set_bit(STRIPE_HANDLE, &sh->state);
2047 }
2048
2049 static void handle_issuing_new_read_requests6(struct stripe_head *sh,
2050                         struct stripe_head_state *s, struct r6_state *r6s,
2051                         int disks)
2052 {
2053         int i;
2054         for (i = disks; i--; ) {
2055                 struct r5dev *dev = &sh->dev[i];
2056                 if (!test_bit(R5_LOCKED, &dev->flags) &&
2057                     !test_bit(R5_UPTODATE, &dev->flags) &&
2058                     (dev->toread || (dev->towrite &&
2059                      !test_bit(R5_OVERWRITE, &dev->flags)) ||
2060                      s->syncing || s->expanding ||
2061                      (s->failed >= 1 &&
2062                       (sh->dev[r6s->failed_num[0]].toread ||
2063                        s->to_write)) ||
2064                      (s->failed >= 2 &&
2065                       (sh->dev[r6s->failed_num[1]].toread ||
2066                        s->to_write)))) {
2067                         /* we would like to get this block, possibly
2068                          * by computing it, but we might not be able to
2069                          */
2070                         if (s->uptodate == disks-1) {
2071                                 pr_debug("Computing stripe %llu block %d\n",
2072                                        (unsigned long long)sh->sector, i);
2073                                 compute_block_1(sh, i, 0);
2074                                 s->uptodate++;
2075                         } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2076                                 /* Computing 2-failure is *very* expensive; only
2077                                  * do it if failed >= 2
2078                                  */
2079                                 int other;
2080                                 for (other = disks; other--; ) {
2081                                         if (other == i)
2082                                                 continue;
2083                                         if (!test_bit(R5_UPTODATE,
2084                                               &sh->dev[other].flags))
2085                                                 break;
2086                                 }
2087                                 BUG_ON(other < 0);
2088                                 pr_debug("Computing stripe %llu blocks %d,%d\n",
2089                                        (unsigned long long)sh->sector,
2090                                        i, other);
2091                                 compute_block_2(sh, i, other);
2092                                 s->uptodate += 2;
2093                         } else if (test_bit(R5_Insync, &dev->flags)) {
2094                                 set_bit(R5_LOCKED, &dev->flags);
2095                                 set_bit(R5_Wantread, &dev->flags);
2096                                 s->locked++;
2097                                 pr_debug("Reading block %d (sync=%d)\n",
2098                                         i, s->syncing);
2099                         }
2100                 }
2101         }
2102         set_bit(STRIPE_HANDLE, &sh->state);
2103 }
2104
2105
2106 /* handle_completed_write_requests
2107  * any written block on an uptodate or failed drive can be returned.
2108  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2109  * never LOCKED, so we don't need to test 'failed' directly.
2110  */
2111 static void handle_completed_write_requests(raid5_conf_t *conf,
2112         struct stripe_head *sh, int disks, struct bio **return_bi)
2113 {
2114         int i;
2115         struct r5dev *dev;
2116
2117         for (i = disks; i--; )
2118                 if (sh->dev[i].written) {
2119                         dev = &sh->dev[i];
2120                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2121                                 test_bit(R5_UPTODATE, &dev->flags)) {
2122                                 /* We can return any write requests */
2123                                 struct bio *wbi, *wbi2;
2124                                 int bitmap_end = 0;
2125                                 pr_debug("Return write for disc %d\n", i);
2126                                 spin_lock_irq(&conf->device_lock);
2127                                 wbi = dev->written;
2128                                 dev->written = NULL;
2129                                 while (wbi && wbi->bi_sector <
2130                                         dev->sector + STRIPE_SECTORS) {
2131                                         wbi2 = r5_next_bio(wbi, dev->sector);
2132                                         if (--wbi->bi_phys_segments == 0) {
2133                                                 md_write_end(conf->mddev);
2134                                                 wbi->bi_next = *return_bi;
2135                                                 *return_bi = wbi;
2136                                         }
2137                                         wbi = wbi2;
2138                                 }
2139                                 if (dev->towrite == NULL)
2140                                         bitmap_end = 1;
2141                                 spin_unlock_irq(&conf->device_lock);
2142                                 if (bitmap_end)
2143                                         bitmap_endwrite(conf->mddev->bitmap,
2144                                                         sh->sector,
2145                                                         STRIPE_SECTORS,
2146                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2147                                                         0);
2148                         }
2149                 }
2150 }
2151
2152 static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
2153                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2154 {
2155         int rmw = 0, rcw = 0, i;
2156         for (i = disks; i--; ) {
2157                 /* would I have to read this buffer for read_modify_write */
2158                 struct r5dev *dev = &sh->dev[i];
2159                 if ((dev->towrite || i == sh->pd_idx) &&
2160                     !test_bit(R5_LOCKED, &dev->flags) &&
2161                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2162                       test_bit(R5_Wantcompute, &dev->flags))) {
2163                         if (test_bit(R5_Insync, &dev->flags))
2164                                 rmw++;
2165                         else
2166                                 rmw += 2*disks;  /* cannot read it */
2167                 }
2168                 /* Would I have to read this buffer for reconstruct_write */
2169                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2170                     !test_bit(R5_LOCKED, &dev->flags) &&
2171                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2172                     test_bit(R5_Wantcompute, &dev->flags))) {
2173                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2174                         else
2175                                 rcw += 2*disks;
2176                 }
2177         }
2178         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2179                 (unsigned long long)sh->sector, rmw, rcw);
2180         set_bit(STRIPE_HANDLE, &sh->state);
2181         if (rmw < rcw && rmw > 0)
2182                 /* prefer read-modify-write, but need to get some data */
2183                 for (i = disks; i--; ) {
2184                         struct r5dev *dev = &sh->dev[i];
2185                         if ((dev->towrite || i == sh->pd_idx) &&
2186                             !test_bit(R5_LOCKED, &dev->flags) &&
2187                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2188                             test_bit(R5_Wantcompute, &dev->flags)) &&
2189                             test_bit(R5_Insync, &dev->flags)) {
2190                                 if (
2191                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2192                                         pr_debug("Read_old block "
2193                                                 "%d for r-m-w\n", i);
2194                                         set_bit(R5_LOCKED, &dev->flags);
2195                                         set_bit(R5_Wantread, &dev->flags);
2196                                         if (!test_and_set_bit(
2197                                                 STRIPE_OP_IO, &sh->ops.pending))
2198                                                 sh->ops.count++;
2199                                         s->locked++;
2200                                 } else {
2201                                         set_bit(STRIPE_DELAYED, &sh->state);
2202                                         set_bit(STRIPE_HANDLE, &sh->state);
2203                                 }
2204                         }
2205                 }
2206         if (rcw <= rmw && rcw > 0)
2207                 /* want reconstruct write, but need to get some data */
2208                 for (i = disks; i--; ) {
2209                         struct r5dev *dev = &sh->dev[i];
2210                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2211                             i != sh->pd_idx &&
2212                             !test_bit(R5_LOCKED, &dev->flags) &&
2213                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2214                             test_bit(R5_Wantcompute, &dev->flags)) &&
2215                             test_bit(R5_Insync, &dev->flags)) {
2216                                 if (
2217                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2218                                         pr_debug("Read_old block "
2219                                                 "%d for Reconstruct\n", i);
2220                                         set_bit(R5_LOCKED, &dev->flags);
2221                                         set_bit(R5_Wantread, &dev->flags);
2222                                         if (!test_and_set_bit(
2223                                                 STRIPE_OP_IO, &sh->ops.pending))
2224                                                 sh->ops.count++;
2225                                         s->locked++;
2226                                 } else {
2227                                         set_bit(STRIPE_DELAYED, &sh->state);
2228                                         set_bit(STRIPE_HANDLE, &sh->state);
2229                                 }
2230                         }
2231                 }
2232         /* now if nothing is locked, and if we have enough data,
2233          * we can start a write request
2234          */
2235         /* since handle_stripe can be called at any time we need to handle the
2236          * case where a compute block operation has been submitted and then a
2237          * subsequent call wants to start a write request.  raid5_run_ops only
2238          * handles the case where compute block and postxor are requested
2239          * simultaneously.  If this is not the case then new writes need to be
2240          * held off until the compute completes.
2241          */
2242         if ((s->req_compute ||
2243             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) &&
2244                 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2245                 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2246                 s->locked += handle_write_operations5(sh, rcw == 0, 0);
2247 }
2248
2249 static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
2250                 struct stripe_head *sh, struct stripe_head_state *s,
2251                 struct r6_state *r6s, int disks)
2252 {
2253         int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2254         int qd_idx = r6s->qd_idx;
2255         for (i = disks; i--; ) {
2256                 struct r5dev *dev = &sh->dev[i];
2257                 /* Would I have to read this buffer for reconstruct_write */
2258                 if (!test_bit(R5_OVERWRITE, &dev->flags)
2259                     && i != pd_idx && i != qd_idx
2260                     && (!test_bit(R5_LOCKED, &dev->flags)
2261                             ) &&
2262                     !test_bit(R5_UPTODATE, &dev->flags)) {
2263                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2264                         else {
2265                                 pr_debug("raid6: must_compute: "
2266                                         "disk %d flags=%#lx\n", i, dev->flags);
2267                                 must_compute++;
2268                         }
2269                 }
2270         }
2271         pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2272                (unsigned long long)sh->sector, rcw, must_compute);
2273         set_bit(STRIPE_HANDLE, &sh->state);
2274
2275         if (rcw > 0)
2276                 /* want reconstruct write, but need to get some data */
2277                 for (i = disks; i--; ) {
2278                         struct r5dev *dev = &sh->dev[i];
2279                         if (!test_bit(R5_OVERWRITE, &dev->flags)
2280                             && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2281                             && !test_bit(R5_LOCKED, &dev->flags) &&
2282                             !test_bit(R5_UPTODATE, &dev->flags) &&
2283                             test_bit(R5_Insync, &dev->flags)) {
2284                                 if (
2285                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2286                                         pr_debug("Read_old stripe %llu "
2287                                                 "block %d for Reconstruct\n",
2288                                              (unsigned long long)sh->sector, i);
2289                                         set_bit(R5_LOCKED, &dev->flags);
2290                                         set_bit(R5_Wantread, &dev->flags);
2291                                         s->locked++;
2292                                 } else {
2293                                         pr_debug("Request delayed stripe %llu "
2294                                                 "block %d for Reconstruct\n",
2295                                              (unsigned long long)sh->sector, i);
2296                                         set_bit(STRIPE_DELAYED, &sh->state);
2297                                         set_bit(STRIPE_HANDLE, &sh->state);
2298                                 }
2299                         }
2300                 }
2301         /* now if nothing is locked, and if we have enough data, we can start a
2302          * write request
2303          */
2304         if (s->locked == 0 && rcw == 0 &&
2305             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2306                 if (must_compute > 0) {
2307                         /* We have failed blocks and need to compute them */
2308                         switch (s->failed) {
2309                         case 0:
2310                                 BUG();
2311                         case 1:
2312                                 compute_block_1(sh, r6s->failed_num[0], 0);
2313                                 break;
2314                         case 2:
2315                                 compute_block_2(sh, r6s->failed_num[0],
2316                                                 r6s->failed_num[1]);
2317                                 break;
2318                         default: /* This request should have been failed? */
2319                                 BUG();
2320                         }
2321                 }
2322
2323                 pr_debug("Computing parity for stripe %llu\n",
2324                         (unsigned long long)sh->sector);
2325                 compute_parity6(sh, RECONSTRUCT_WRITE);
2326                 /* now every locked buffer is ready to be written */
2327                 for (i = disks; i--; )
2328                         if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2329                                 pr_debug("Writing stripe %llu block %d\n",
2330                                        (unsigned long long)sh->sector, i);
2331                                 s->locked++;
2332                                 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2333                         }
2334                 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2335                 set_bit(STRIPE_INSYNC, &sh->state);
2336
2337                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2338                         atomic_dec(&conf->preread_active_stripes);
2339                         if (atomic_read(&conf->preread_active_stripes) <
2340                             IO_THRESHOLD)
2341                                 md_wakeup_thread(conf->mddev->thread);
2342                 }
2343         }
2344 }
2345
2346 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2347                                 struct stripe_head_state *s, int disks)
2348 {
2349         set_bit(STRIPE_HANDLE, &sh->state);
2350         /* Take one of the following actions:
2351          * 1/ start a check parity operation if (uptodate == disks)
2352          * 2/ finish a check parity operation and act on the result
2353          * 3/ skip to the writeback section if we previously
2354          *    initiated a recovery operation
2355          */
2356         if (s->failed == 0 &&
2357             !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2358                 if (!test_and_set_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
2359                         BUG_ON(s->uptodate != disks);
2360                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2361                         sh->ops.count++;
2362                         s->uptodate--;
2363                 } else if (
2364                        test_and_clear_bit(STRIPE_OP_CHECK, &sh->ops.complete)) {
2365                         clear_bit(STRIPE_OP_CHECK, &sh->ops.ack);
2366                         clear_bit(STRIPE_OP_CHECK, &sh->ops.pending);
2367
2368                         if (sh->ops.zero_sum_result == 0)
2369                                 /* parity is correct (on disc,
2370                                  * not in buffer any more)
2371                                  */
2372                                 set_bit(STRIPE_INSYNC, &sh->state);
2373                         else {
2374                                 conf->mddev->resync_mismatches +=
2375                                         STRIPE_SECTORS;
2376                                 if (test_bit(
2377                                      MD_RECOVERY_CHECK, &conf->mddev->recovery))
2378                                         /* don't try to repair!! */
2379                                         set_bit(STRIPE_INSYNC, &sh->state);
2380                                 else {
2381                                         set_bit(STRIPE_OP_COMPUTE_BLK,
2382                                                 &sh->ops.pending);
2383                                         set_bit(STRIPE_OP_MOD_REPAIR_PD,
2384                                                 &sh->ops.pending);
2385                                         set_bit(R5_Wantcompute,
2386                                                 &sh->dev[sh->pd_idx].flags);
2387                                         sh->ops.target = sh->pd_idx;
2388                                         sh->ops.count++;
2389                                         s->uptodate++;
2390                                 }
2391                         }
2392                 }
2393         }
2394
2395         /* check if we can clear a parity disk reconstruct */
2396         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2397                 test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2398
2399                 clear_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending);
2400                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2401                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2402                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2403         }
2404
2405         /* Wait for check parity and compute block operations to complete
2406          * before write-back
2407          */
2408         if (!test_bit(STRIPE_INSYNC, &sh->state) &&
2409                 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending) &&
2410                 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) {
2411                 struct r5dev *dev;
2412                 /* either failed parity check, or recovery is happening */
2413                 if (s->failed == 0)
2414                         s->failed_num = sh->pd_idx;
2415                 dev = &sh->dev[s->failed_num];
2416                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2417                 BUG_ON(s->uptodate != disks);
2418
2419                 set_bit(R5_LOCKED, &dev->flags);
2420                 set_bit(R5_Wantwrite, &dev->flags);
2421                 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2422                         sh->ops.count++;
2423
2424                 clear_bit(STRIPE_DEGRADED, &sh->state);
2425                 s->locked++;
2426                 set_bit(STRIPE_INSYNC, &sh->state);
2427         }
2428 }
2429
2430
2431 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2432                                 struct stripe_head_state *s,
2433                                 struct r6_state *r6s, struct page *tmp_page,
2434                                 int disks)
2435 {
2436         int update_p = 0, update_q = 0;
2437         struct r5dev *dev;
2438         int pd_idx = sh->pd_idx;
2439         int qd_idx = r6s->qd_idx;
2440
2441         set_bit(STRIPE_HANDLE, &sh->state);
2442
2443         BUG_ON(s->failed > 2);
2444         BUG_ON(s->uptodate < disks);
2445         /* Want to check and possibly repair P and Q.
2446          * However there could be one 'failed' device, in which
2447          * case we can only check one of them, possibly using the
2448          * other to generate missing data
2449          */
2450
2451         /* If !tmp_page, we cannot do the calculations,
2452          * but as we have set STRIPE_HANDLE, we will soon be called
2453          * by stripe_handle with a tmp_page - just wait until then.
2454          */
2455         if (tmp_page) {
2456                 if (s->failed == r6s->q_failed) {
2457                         /* The only possible failed device holds 'Q', so it
2458                          * makes sense to check P (If anything else were failed,
2459                          * we would have used P to recreate it).
2460                          */
2461                         compute_block_1(sh, pd_idx, 1);
2462                         if (!page_is_zero(sh->dev[pd_idx].page)) {
2463                                 compute_block_1(sh, pd_idx, 0);
2464                                 update_p = 1;
2465                         }
2466                 }
2467                 if (!r6s->q_failed && s->failed < 2) {
2468                         /* q is not failed, and we didn't use it to generate
2469                          * anything, so it makes sense to check it
2470                          */
2471                         memcpy(page_address(tmp_page),
2472                                page_address(sh->dev[qd_idx].page),
2473                                STRIPE_SIZE);
2474                         compute_parity6(sh, UPDATE_PARITY);
2475                         if (memcmp(page_address(tmp_page),
2476                                    page_address(sh->dev[qd_idx].page),
2477                                    STRIPE_SIZE) != 0) {
2478                                 clear_bit(STRIPE_INSYNC, &sh->state);
2479                                 update_q = 1;
2480                         }
2481                 }
2482                 if (update_p || update_q) {
2483                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2484                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2485                                 /* don't try to repair!! */
2486                                 update_p = update_q = 0;
2487                 }
2488
2489                 /* now write out any block on a failed drive,
2490                  * or P or Q if they need it
2491                  */
2492
2493                 if (s->failed == 2) {
2494                         dev = &sh->dev[r6s->failed_num[1]];
2495                         s->locked++;
2496                         set_bit(R5_LOCKED, &dev->flags);
2497                         set_bit(R5_Wantwrite, &dev->flags);
2498                 }
2499                 if (s->failed >= 1) {
2500                         dev = &sh->dev[r6s->failed_num[0]];
2501                         s->locked++;
2502                         set_bit(R5_LOCKED, &dev->flags);
2503                         set_bit(R5_Wantwrite, &dev->flags);
2504                 }
2505
2506                 if (update_p) {
2507                         dev = &sh->dev[pd_idx];
2508                         s->locked++;
2509                         set_bit(R5_LOCKED, &dev->flags);
2510                         set_bit(R5_Wantwrite, &dev->flags);
2511                 }
2512                 if (update_q) {
2513                         dev = &sh->dev[qd_idx];
2514                         s->locked++;
2515                         set_bit(R5_LOCKED, &dev->flags);
2516                         set_bit(R5_Wantwrite, &dev->flags);
2517                 }
2518                 clear_bit(STRIPE_DEGRADED, &sh->state);
2519
2520                 set_bit(STRIPE_INSYNC, &sh->state);
2521         }
2522 }
2523
2524 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2525                                 struct r6_state *r6s)
2526 {
2527         int i;
2528
2529         /* We have read all the blocks in this stripe and now we need to
2530          * copy some of them into a target stripe for expand.
2531          */
2532         struct dma_async_tx_descriptor *tx = NULL;
2533         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2534         for (i = 0; i < sh->disks; i++)
2535                 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
2536                         int dd_idx, pd_idx, j;
2537                         struct stripe_head *sh2;
2538
2539                         sector_t bn = compute_blocknr(sh, i);
2540                         sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2541                                                 conf->raid_disks -
2542                                                 conf->max_degraded, &dd_idx,
2543                                                 &pd_idx, conf);
2544                         sh2 = get_active_stripe(conf, s, conf->raid_disks,
2545                                                 pd_idx, 1);
2546                         if (sh2 == NULL)
2547                                 /* so far only the early blocks of this stripe
2548                                  * have been requested.  When later blocks
2549                                  * get requested, we will try again
2550                                  */
2551                                 continue;
2552                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2553                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2554                                 /* must have already done this block */
2555                                 release_stripe(sh2);
2556                                 continue;
2557                         }
2558
2559                         /* place all the copies on one channel */
2560                         tx = async_memcpy(sh2->dev[dd_idx].page,
2561                                 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2562                                 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2563
2564                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2565                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2566                         for (j = 0; j < conf->raid_disks; j++)
2567                                 if (j != sh2->pd_idx &&
2568                                     (!r6s || j != raid6_next_disk(sh2->pd_idx,
2569                                                                  sh2->disks)) &&
2570                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2571                                         break;
2572                         if (j == conf->raid_disks) {
2573                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2574                                 set_bit(STRIPE_HANDLE, &sh2->state);
2575                         }
2576                         release_stripe(sh2);
2577
2578                 }
2579         /* done submitting copies, wait for them to complete */
2580         if (tx) {
2581                 async_tx_ack(tx);
2582                 dma_wait_for_async_tx(tx);
2583         }
2584 }
2585
2586 /*
2587  * handle_stripe - do things to a stripe.
2588  *
2589  * We lock the stripe and then examine the state of various bits
2590  * to see what needs to be done.
2591  * Possible results:
2592  *    return some read request which now have data
2593  *    return some write requests which are safely on disc
2594  *    schedule a read on some buffers
2595  *    schedule a write of some buffers
2596  *    return confirmation of parity correctness
2597  *
2598  * buffers are taken off read_list or write_list, and bh_cache buffers
2599  * get BH_Lock set before the stripe lock is released.
2600  *
2601  */
2602
2603 static void handle_stripe5(struct stripe_head *sh)
2604 {
2605         raid5_conf_t *conf = sh->raid_conf;
2606         int disks = sh->disks, i;
2607         struct bio *return_bi = NULL;
2608         struct stripe_head_state s;
2609         struct r5dev *dev;
2610         unsigned long pending = 0;
2611
2612         memset(&s, 0, sizeof(s));
2613         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2614                 "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
2615                 atomic_read(&sh->count), sh->pd_idx,
2616                 sh->ops.pending, sh->ops.ack, sh->ops.complete);
2617
2618         spin_lock(&sh->lock);
2619         clear_bit(STRIPE_HANDLE, &sh->state);
2620         clear_bit(STRIPE_DELAYED, &sh->state);
2621
2622         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2623         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2624         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2625         /* Now to look around and see what can be done */
2626
2627         /* clean-up completed biofill operations */
2628         if (test_bit(STRIPE_OP_BIOFILL, &sh->ops.complete)) {
2629                 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.pending);
2630                 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.ack);
2631                 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.complete);
2632         }
2633
2634         rcu_read_lock();
2635         for (i=disks; i--; ) {
2636                 mdk_rdev_t *rdev;
2637                 struct r5dev *dev = &sh->dev[i];
2638                 clear_bit(R5_Insync, &dev->flags);
2639
2640                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2641                         "written %p\n", i, dev->flags, dev->toread, dev->read,
2642                         dev->towrite, dev->written);
2643
2644                 /* maybe we can request a biofill operation
2645                  *
2646                  * new wantfill requests are only permitted while
2647                  * STRIPE_OP_BIOFILL is clear
2648                  */
2649                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2650                         !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2651                         set_bit(R5_Wantfill, &dev->flags);
2652
2653                 /* now count some things */
2654                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2655                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2656                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2657
2658                 if (test_bit(R5_Wantfill, &dev->flags))
2659                         s.to_fill++;
2660                 else if (dev->toread)
2661                         s.to_read++;
2662                 if (dev->towrite) {
2663                         s.to_write++;
2664                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2665                                 s.non_overwrite++;
2666                 }
2667                 if (dev->written)
2668                         s.written++;
2669                 rdev = rcu_dereference(conf->disks[i].rdev);
2670                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2671                         /* The ReadError flag will just be confusing now */
2672                         clear_bit(R5_ReadError, &dev->flags);
2673                         clear_bit(R5_ReWrite, &dev->flags);
2674                 }
2675                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2676                     || test_bit(R5_ReadError, &dev->flags)) {
2677                         s.failed++;
2678                         s.failed_num = i;
2679                 } else
2680                         set_bit(R5_Insync, &dev->flags);
2681         }
2682         rcu_read_unlock();
2683
2684         if (s.to_fill && !test_and_set_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2685                 sh->ops.count++;
2686
2687         pr_debug("locked=%d uptodate=%d to_read=%d"
2688                 " to_write=%d failed=%d failed_num=%d\n",
2689                 s.locked, s.uptodate, s.to_read, s.to_write,
2690                 s.failed, s.failed_num);
2691         /* check if the array has lost two devices and, if so, some requests might
2692          * need to be failed
2693          */
2694         if (s.failed > 1 && s.to_read+s.to_write+s.written)
2695                 handle_requests_to_failed_array(conf, sh, &s, disks,
2696                                                 &return_bi);
2697         if (s.failed > 1 && s.syncing) {
2698                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2699                 clear_bit(STRIPE_SYNCING, &sh->state);
2700                 s.syncing = 0;
2701         }
2702
2703         /* might be able to return some write requests if the parity block
2704          * is safe, or on a failed drive
2705          */
2706         dev = &sh->dev[sh->pd_idx];
2707         if ( s.written &&
2708              ((test_bit(R5_Insync, &dev->flags) &&
2709                !test_bit(R5_LOCKED, &dev->flags) &&
2710                test_bit(R5_UPTODATE, &dev->flags)) ||
2711                (s.failed == 1 && s.failed_num == sh->pd_idx)))
2712                 handle_completed_write_requests(conf, sh, disks, &return_bi);
2713
2714         /* Now we might consider reading some blocks, either to check/generate
2715          * parity, or to satisfy requests
2716          * or to load a block that is being partially written.
2717          */
2718         if (s.to_read || s.non_overwrite ||
2719             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding ||
2720             test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2721                 handle_issuing_new_read_requests5(sh, &s, disks);
2722
2723         /* Now we check to see if any write operations have recently
2724          * completed
2725          */
2726
2727         /* leave prexor set until postxor is done, allows us to distinguish
2728          * a rmw from a rcw during biodrain
2729          */
2730         if (test_bit(STRIPE_OP_PREXOR, &sh->ops.complete) &&
2731                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2732
2733                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
2734                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.ack);
2735                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
2736
2737                 for (i = disks; i--; )
2738                         clear_bit(R5_Wantprexor, &sh->dev[i].flags);
2739         }
2740
2741         /* if only POSTXOR is set then this is an 'expand' postxor */
2742         if (test_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete) &&
2743                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2744
2745                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
2746                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.ack);
2747                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
2748
2749                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2750                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2751                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2752
2753                 /* All the 'written' buffers and the parity block are ready to
2754                  * be written back to disk
2755                  */
2756                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2757                 for (i = disks; i--; ) {
2758                         dev = &sh->dev[i];
2759                         if (test_bit(R5_LOCKED, &dev->flags) &&
2760                                 (i == sh->pd_idx || dev->written)) {
2761                                 pr_debug("Writing block %d\n", i);
2762                                 set_bit(R5_Wantwrite, &dev->flags);
2763                                 if (!test_and_set_bit(
2764                                     STRIPE_OP_IO, &sh->ops.pending))
2765                                         sh->ops.count++;
2766                                 if (!test_bit(R5_Insync, &dev->flags) ||
2767                                     (i == sh->pd_idx && s.failed == 0))
2768                                         set_bit(STRIPE_INSYNC, &sh->state);
2769                         }
2770                 }
2771                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2772                         atomic_dec(&conf->preread_active_stripes);
2773                         if (atomic_read(&conf->preread_active_stripes) <
2774                                 IO_THRESHOLD)
2775                                 md_wakeup_thread(conf->mddev->thread);
2776                 }
2777         }
2778
2779         /* Now to consider new write requests and what else, if anything
2780          * should be read.  We do not handle new writes when:
2781          * 1/ A 'write' operation (copy+xor) is already in flight.
2782          * 2/ A 'check' operation is in flight, as it may clobber the parity
2783          *    block.
2784          */
2785         if (s.to_write && !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending) &&
2786                           !test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
2787                 handle_issuing_new_write_requests5(conf, sh, &s, disks);
2788
2789         /* maybe we need to check and possibly fix the parity for this stripe
2790          * Any reads will already have been scheduled, so we just see if enough
2791          * data is available.  The parity check is held off while parity
2792          * dependent operations are in flight.
2793          */
2794         if ((s.syncing && s.locked == 0 &&
2795              !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2796              !test_bit(STRIPE_INSYNC, &sh->state)) ||
2797               test_bit(STRIPE_OP_CHECK, &sh->ops.pending) ||
2798               test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending))
2799                 handle_parity_checks5(conf, sh, &s, disks);
2800
2801         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2802                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2803                 clear_bit(STRIPE_SYNCING, &sh->state);
2804         }
2805
2806         /* If the failed drive is just a ReadError, then we might need to progress
2807          * the repair/check process
2808          */
2809         if (s.failed == 1 && !conf->mddev->ro &&
2810             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2811             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2812             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2813                 ) {
2814                 dev = &sh->dev[s.failed_num];
2815                 if (!test_bit(R5_ReWrite, &dev->flags)) {
2816                         set_bit(R5_Wantwrite, &dev->flags);
2817                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2818                                 sh->ops.count++;
2819                         set_bit(R5_ReWrite, &dev->flags);
2820                         set_bit(R5_LOCKED, &dev->flags);
2821                         s.locked++;
2822                 } else {
2823                         /* let's read it back */
2824                         set_bit(R5_Wantread, &dev->flags);
2825                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2826                                 sh->ops.count++;
2827                         set_bit(R5_LOCKED, &dev->flags);
2828                         s.locked++;
2829                 }
2830         }
2831
2832         /* Finish postxor operations initiated by the expansion
2833          * process
2834          */
2835         if (test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete) &&
2836                 !test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending)) {
2837
2838                 clear_bit(STRIPE_EXPANDING, &sh->state);
2839
2840                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2841                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2842                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2843
2844                 for (i = conf->raid_disks; i--; ) {
2845                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
2846                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2847                                 sh->ops.count++;
2848                 }
2849         }
2850
2851         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2852                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2853                 /* Need to write out all blocks after computing parity */
2854                 sh->disks = conf->raid_disks;
2855                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2856                         conf->raid_disks);
2857                 s.locked += handle_write_operations5(sh, 1, 1);
2858         } else if (s.expanded &&
2859                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2860                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2861                 atomic_dec(&conf->reshape_stripes);
2862                 wake_up(&conf->wait_for_overlap);
2863                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2864         }
2865
2866         if (s.expanding && s.locked == 0)
2867                 handle_stripe_expansion(conf, sh, NULL);
2868
2869         if (sh->ops.count)
2870                 pending = get_stripe_work(sh);
2871
2872         spin_unlock(&sh->lock);
2873
2874         if (pending)
2875                 raid5_run_ops(sh, pending);
2876
2877         return_io(return_bi);
2878
2879 }
2880
2881 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2882 {
2883         raid6_conf_t *conf = sh->raid_conf;
2884         int disks = sh->disks;
2885         struct bio *return_bi = NULL;
2886         int i, pd_idx = sh->pd_idx;
2887         struct stripe_head_state s;
2888         struct r6_state r6s;
2889         struct r5dev *dev, *pdev, *qdev;
2890
2891         r6s.qd_idx = raid6_next_disk(pd_idx, disks);
2892         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2893                 "pd_idx=%d, qd_idx=%d\n",
2894                (unsigned long long)sh->sector, sh->state,
2895                atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2896         memset(&s, 0, sizeof(s));
2897
2898         spin_lock(&sh->lock);
2899         clear_bit(STRIPE_HANDLE, &sh->state);
2900         clear_bit(STRIPE_DELAYED, &sh->state);
2901
2902         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2903         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2904         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2905         /* Now to look around and see what can be done */
2906
2907         rcu_read_lock();
2908         for (i=disks; i--; ) {
2909                 mdk_rdev_t *rdev;
2910                 dev = &sh->dev[i];
2911                 clear_bit(R5_Insync, &dev->flags);
2912
2913                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2914                         i, dev->flags, dev->toread, dev->towrite, dev->written);
2915                 /* maybe we can reply to a read */
2916                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2917                         struct bio *rbi, *rbi2;
2918                         pr_debug("Return read for disc %d\n", i);
2919                         spin_lock_irq(&conf->device_lock);
2920                         rbi = dev->toread;
2921                         dev->toread = NULL;
2922                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2923                                 wake_up(&conf->wait_for_overlap);
2924                         spin_unlock_irq(&conf->device_lock);
2925                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2926                                 copy_data(0, rbi, dev->page, dev->sector);
2927                                 rbi2 = r5_next_bio(rbi, dev->sector);
2928                                 spin_lock_irq(&conf->device_lock);
2929                                 if (--rbi->bi_phys_segments == 0) {
2930                                         rbi->bi_next = return_bi;
2931                                         return_bi = rbi;
2932                                 }
2933                                 spin_unlock_irq(&conf->device_lock);
2934                                 rbi = rbi2;
2935                         }
2936                 }
2937
2938                 /* now count some things */
2939                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2940                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2941
2942
2943                 if (dev->toread)
2944                         s.to_read++;
2945                 if (dev->towrite) {
2946                         s.to_write++;
2947                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2948                                 s.non_overwrite++;
2949                 }
2950                 if (dev->written)
2951                         s.written++;
2952                 rdev = rcu_dereference(conf->disks[i].rdev);
2953                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2954                         /* The ReadError flag will just be confusing now */
2955                         clear_bit(R5_ReadError, &dev->flags);
2956                         clear_bit(R5_ReWrite, &dev->flags);
2957                 }
2958                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2959                     || test_bit(R5_ReadError, &dev->flags)) {
2960                         if (s.failed < 2)
2961                                 r6s.failed_num[s.failed] = i;
2962                         s.failed++;
2963                 } else
2964                         set_bit(R5_Insync, &dev->flags);
2965         }
2966         rcu_read_unlock();
2967         pr_debug("locked=%d uptodate=%d to_read=%d"
2968                " to_write=%d failed=%d failed_num=%d,%d\n",
2969                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
2970                r6s.failed_num[0], r6s.failed_num[1]);
2971         /* check if the array has lost >2 devices and, if so, some requests
2972          * might need to be failed
2973          */
2974         if (s.failed > 2 && s.to_read+s.to_write+s.written)
2975                 handle_requests_to_failed_array(conf, sh, &s, disks,
2976                                                 &return_bi);
2977         if (s.failed > 2 && s.syncing) {
2978                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2979                 clear_bit(STRIPE_SYNCING, &sh->state);
2980                 s.syncing = 0;
2981         }
2982
2983         /*
2984          * might be able to return some write requests if the parity blocks
2985          * are safe, or on a failed drive
2986          */
2987         pdev = &sh->dev[pd_idx];
2988         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
2989                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
2990         qdev = &sh->dev[r6s.qd_idx];
2991         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
2992                 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
2993
2994         if ( s.written &&
2995              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
2996                              && !test_bit(R5_LOCKED, &pdev->flags)
2997                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
2998              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
2999                              && !test_bit(R5_LOCKED, &qdev->flags)
3000                              && test_bit(R5_UPTODATE, &qdev->flags)))))
3001                 handle_completed_write_requests(conf, sh, disks, &return_bi);
3002
3003         /* Now we might consider reading some blocks, either to check/generate
3004          * parity, or to satisfy requests
3005          * or to load a block that is being partially written.
3006          */
3007         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3008             (s.syncing && (s.uptodate < disks)) || s.expanding)
3009                 handle_issuing_new_read_requests6(sh, &s, &r6s, disks);
3010
3011         /* now to consider writing and what else, if anything should be read */
3012         if (s.to_write)
3013                 handle_issuing_new_write_requests6(conf, sh, &s, &r6s, disks);
3014
3015         /* maybe we need to check and possibly fix the parity for this stripe
3016          * Any reads will already have been scheduled, so we just see if enough
3017          * data is available
3018          */
3019         if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3020                 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3021
3022         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3023                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3024                 clear_bit(STRIPE_SYNCING, &sh->state);
3025         }
3026
3027         /* If the failed drives are just a ReadError, then we might need
3028          * to progress the repair/check process
3029          */
3030         if (s.failed <= 2 && !conf->mddev->ro)
3031                 for (i = 0; i < s.failed; i++) {
3032                         dev = &sh->dev[r6s.failed_num[i]];
3033                         if (test_bit(R5_ReadError, &dev->flags)
3034                             && !test_bit(R5_LOCKED, &dev->flags)
3035                             && test_bit(R5_UPTODATE, &dev->flags)
3036                                 ) {
3037                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3038                                         set_bit(R5_Wantwrite, &dev->flags);
3039                                         set_bit(R5_ReWrite, &dev->flags);
3040                                         set_bit(R5_LOCKED, &dev->flags);
3041                                 } else {
3042                                         /* let's read it back */
3043                                         set_bit(R5_Wantread, &dev->flags);
3044                                         set_bit(R5_LOCKED, &dev->flags);
3045                                 }
3046                         }
3047                 }
3048
3049         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3050                 /* Need to write out all blocks after computing P&Q */
3051                 sh->disks = conf->raid_disks;
3052                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
3053                                              conf->raid_disks);
3054                 compute_parity6(sh, RECONSTRUCT_WRITE);
3055                 for (i = conf->raid_disks ; i-- ;  ) {
3056                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3057                         s.locked++;
3058                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3059                 }
3060                 clear_bit(STRIPE_EXPANDING, &sh->state);
3061         } else if (s.expanded) {
3062                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3063                 atomic_dec(&conf->reshape_stripes);
3064                 wake_up(&conf->wait_for_overlap);
3065                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3066         }
3067
3068         if (s.expanding && s.locked == 0)
3069                 handle_stripe_expansion(conf, sh, &r6s);
3070
3071         spin_unlock(&sh->lock);
3072
3073         return_io(return_bi);
3074
3075         for (i=disks; i-- ;) {
3076                 int rw;
3077                 struct bio *bi;
3078                 mdk_rdev_t *rdev;
3079                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
3080                         rw = WRITE;
3081                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
3082                         rw = READ;
3083                 else
3084                         continue;
3085
3086                 bi = &sh->dev[i].req;
3087
3088                 bi->bi_rw = rw;
3089                 if (rw == WRITE)
3090                         bi->bi_end_io = raid5_end_write_request;
3091                 else
3092                         bi->bi_end_io = raid5_end_read_request;
3093
3094                 rcu_read_lock();
3095                 rdev = rcu_dereference(conf->disks[i].rdev);
3096                 if (rdev && test_bit(Faulty, &rdev->flags))
3097                         rdev = NULL;
3098                 if (rdev)
3099                         atomic_inc(&rdev->nr_pending);
3100                 rcu_read_unlock();
3101
3102                 if (rdev) {
3103                         if (s.syncing || s.expanding || s.expanded)
3104                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
3105
3106                         bi->bi_bdev = rdev->bdev;
3107                         pr_debug("for %llu schedule op %ld on disc %d\n",
3108                                 (unsigned long long)sh->sector, bi->bi_rw, i);
3109                         atomic_inc(&sh->count);
3110                         bi->bi_sector = sh->sector + rdev->data_offset;
3111                         bi->bi_flags = 1 << BIO_UPTODATE;
3112                         bi->bi_vcnt = 1;
3113                         bi->bi_max_vecs = 1;
3114                         bi->bi_idx = 0;
3115                         bi->bi_io_vec = &sh->dev[i].vec;
3116                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
3117                         bi->bi_io_vec[0].bv_offset = 0;
3118                         bi->bi_size = STRIPE_SIZE;
3119                         bi->bi_next = NULL;
3120                         if (rw == WRITE &&
3121                             test_bit(R5_ReWrite, &sh->dev[i].flags))
3122                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
3123                         generic_make_request(bi);
3124                 } else {
3125                         if (rw == WRITE)
3126                                 set_bit(STRIPE_DEGRADED, &sh->state);
3127                         pr_debug("skip op %ld on disc %d for sector %llu\n",
3128                                 bi->bi_rw, i, (unsigned long long)sh->sector);
3129                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
3130                         set_bit(STRIPE_HANDLE, &sh->state);
3131                 }
3132         }
3133 }
3134
3135 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3136 {
3137         if (sh->raid_conf->level == 6)
3138                 handle_stripe6(sh, tmp_page);
3139         else
3140                 handle_stripe5(sh);
3141 }
3142
3143
3144
3145 static void raid5_activate_delayed(raid5_conf_t *conf)
3146 {
3147         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3148                 while (!list_empty(&conf->delayed_list)) {
3149                         struct list_head *l = conf->delayed_list.next;
3150                         struct stripe_head *sh;
3151                         sh = list_entry(l, struct stripe_head, lru);
3152                         list_del_init(l);
3153                         clear_bit(STRIPE_DELAYED, &sh->state);
3154                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3155                                 atomic_inc(&conf->preread_active_stripes);
3156                         list_add_tail(&sh->lru, &conf->handle_list);
3157                 }
3158         }
3159 }
3160
3161 static void activate_bit_delay(raid5_conf_t *conf)
3162 {
3163         /* device_lock is held */
3164         struct list_head head;
3165         list_add(&head, &conf->bitmap_list);
3166         list_del_init(&conf->bitmap_list);
3167         while (!list_empty(&head)) {
3168                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3169                 list_del_init(&sh->lru);
3170                 atomic_inc(&sh->count);
3171                 __release_stripe(conf, sh);
3172         }
3173 }
3174
3175 static void unplug_slaves(mddev_t *mddev)
3176 {
3177         raid5_conf_t *conf = mddev_to_conf(mddev);
3178         int i;
3179
3180         rcu_read_lock();
3181         for (i=0; i<mddev->raid_disks; i++) {
3182                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3183                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3184                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3185
3186                         atomic_inc(&rdev->nr_pending);
3187                         rcu_read_unlock();
3188
3189                         blk_unplug(r_queue);
3190
3191                         rdev_dec_pending(rdev, mddev);
3192                         rcu_read_lock();
3193                 }
3194         }
3195         rcu_read_unlock();
3196 }
3197
3198 static void raid5_unplug_device(struct request_queue *q)
3199 {
3200         mddev_t *mddev = q->queuedata;
3201         raid5_conf_t *conf = mddev_to_conf(mddev);
3202         unsigned long flags;
3203
3204         spin_lock_irqsave(&conf->device_lock, flags);
3205
3206         if (blk_remove_plug(q)) {
3207                 conf->seq_flush++;
3208                 raid5_activate_delayed(conf);
3209         }
3210         md_wakeup_thread(mddev->thread);
3211
3212         spin_unlock_irqrestore(&conf->device_lock, flags);
3213
3214         unplug_slaves(mddev);
3215 }
3216
3217 static int raid5_congested(void *data, int bits)
3218 {
3219         mddev_t *mddev = data;
3220         raid5_conf_t *conf = mddev_to_conf(mddev);
3221
3222         /* No difference between reads and writes.  Just check
3223          * how busy the stripe_cache is
3224          */
3225         if (conf->inactive_blocked)
3226                 return 1;
3227         if (conf->quiesce)
3228                 return 1;
3229         if (list_empty_careful(&conf->inactive_list))
3230                 return 1;
3231
3232         return 0;
3233 }
3234
3235 /* We want read requests to align with chunks where possible,
3236  * but write requests don't need to.
3237  */
3238 static int raid5_mergeable_bvec(struct request_queue *q, struct bio *bio, struct bio_vec *biovec)
3239 {
3240         mddev_t *mddev = q->queuedata;
3241         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3242         int max;
3243         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3244         unsigned int bio_sectors = bio->bi_size >> 9;
3245
3246         if (bio_data_dir(bio) == WRITE)
3247                 return biovec->bv_len; /* always allow writes to be mergeable */
3248
3249         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3250         if (max < 0) max = 0;
3251         if (max <= biovec->bv_len && bio_sectors == 0)
3252                 return biovec->bv_len;
3253         else
3254                 return max;
3255 }
3256
3257
3258 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3259 {
3260         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3261         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3262         unsigned int bio_sectors = bio->bi_size >> 9;
3263
3264         return  chunk_sectors >=
3265                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3266 }
3267
3268 /*
3269  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3270  *  later sampled by raid5d.
3271  */
3272 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3273 {
3274         unsigned long flags;
3275
3276         spin_lock_irqsave(&conf->device_lock, flags);
3277
3278         bi->bi_next = conf->retry_read_aligned_list;
3279         conf->retry_read_aligned_list = bi;
3280
3281         spin_unlock_irqrestore(&conf->device_lock, flags);
3282         md_wakeup_thread(conf->mddev->thread);
3283 }
3284
3285
3286 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3287 {
3288         struct bio *bi;
3289
3290         bi = conf->retry_read_aligned;
3291         if (bi) {
3292                 conf->retry_read_aligned = NULL;
3293                 return bi;
3294         }
3295         bi = conf->retry_read_aligned_list;
3296         if(bi) {
3297                 conf->retry_read_aligned_list = bi->bi_next;
3298                 bi->bi_next = NULL;
3299                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3300                 bi->bi_hw_segments = 0; /* count of processed stripes */
3301         }
3302
3303         return bi;
3304 }
3305
3306
3307 /*
3308  *  The "raid5_align_endio" should check if the read succeeded and if it
3309  *  did, call bio_endio on the original bio (having bio_put the new bio
3310  *  first).
3311  *  If the read failed..
3312  */
3313 static void raid5_align_endio(struct bio *bi, int error)
3314 {
3315         struct bio* raid_bi  = bi->bi_private;
3316         mddev_t *mddev;
3317         raid5_conf_t *conf;
3318         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3319         mdk_rdev_t *rdev;
3320
3321         bio_put(bi);
3322
3323         mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3324         conf = mddev_to_conf(mddev);
3325         rdev = (void*)raid_bi->bi_next;
3326         raid_bi->bi_next = NULL;
3327
3328         rdev_dec_pending(rdev, conf->mddev);
3329
3330         if (!error && uptodate) {
3331                 bio_endio(raid_bi, 0);
3332                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3333                         wake_up(&conf->wait_for_stripe);
3334                 return;
3335         }
3336
3337
3338         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3339
3340         add_bio_to_retry(raid_bi, conf);
3341 }
3342
3343 static int bio_fits_rdev(struct bio *bi)
3344 {
3345         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3346
3347         if ((bi->bi_size>>9) > q->max_sectors)
3348                 return 0;
3349         blk_recount_segments(q, bi);
3350         if (bi->bi_phys_segments > q->max_phys_segments ||
3351             bi->bi_hw_segments > q->max_hw_segments)
3352                 return 0;
3353
3354         if (q->merge_bvec_fn)
3355                 /* it's too hard to apply the merge_bvec_fn at this stage,
3356                  * just just give up
3357                  */
3358                 return 0;
3359
3360         return 1;
3361 }
3362
3363
3364 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3365 {
3366         mddev_t *mddev = q->queuedata;
3367         raid5_conf_t *conf = mddev_to_conf(mddev);
3368         const unsigned int raid_disks = conf->raid_disks;
3369         const unsigned int data_disks = raid_disks - conf->max_degraded;
3370         unsigned int dd_idx, pd_idx;
3371         struct bio* align_bi;
3372         mdk_rdev_t *rdev;
3373
3374         if (!in_chunk_boundary(mddev, raid_bio)) {
3375                 pr_debug("chunk_aligned_read : non aligned\n");
3376                 return 0;
3377         }
3378         /*
3379          * use bio_clone to make a copy of the bio
3380          */
3381         align_bi = bio_clone(raid_bio, GFP_NOIO);
3382         if (!align_bi)
3383                 return 0;
3384         /*
3385          *   set bi_end_io to a new function, and set bi_private to the
3386          *     original bio.
3387          */
3388         align_bi->bi_end_io  = raid5_align_endio;
3389         align_bi->bi_private = raid_bio;
3390         /*
3391          *      compute position
3392          */
3393         align_bi->bi_sector =  raid5_compute_sector(raid_bio->bi_sector,
3394                                         raid_disks,
3395                                         data_disks,
3396                                         &dd_idx,
3397                                         &pd_idx,
3398                                         conf);
3399
3400         rcu_read_lock();
3401         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3402         if (rdev && test_bit(In_sync, &rdev->flags)) {
3403                 atomic_inc(&rdev->nr_pending);
3404                 rcu_read_unlock();
3405                 raid_bio->bi_next = (void*)rdev;
3406                 align_bi->bi_bdev =  rdev->bdev;
3407                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3408                 align_bi->bi_sector += rdev->data_offset;
3409
3410                 if (!bio_fits_rdev(align_bi)) {
3411                         /* too big in some way */
3412                         bio_put(align_bi);
3413                         rdev_dec_pending(rdev, mddev);
3414                         return 0;
3415                 }
3416
3417                 spin_lock_irq(&conf->device_lock);
3418                 wait_event_lock_irq(conf->wait_for_stripe,
3419                                     conf->quiesce == 0,
3420                                     conf->device_lock, /* nothing */);
3421                 atomic_inc(&conf->active_aligned_reads);
3422                 spin_unlock_irq(&conf->device_lock);
3423
3424                 generic_make_request(align_bi);
3425                 return 1;
3426         } else {
3427                 rcu_read_unlock();
3428                 bio_put(align_bi);
3429                 return 0;
3430         }
3431 }
3432
3433
3434 static int make_request(struct request_queue *q, struct bio * bi)
3435 {
3436         mddev_t *mddev = q->queuedata;
3437         raid5_conf_t *conf = mddev_to_conf(mddev);
3438         unsigned int dd_idx, pd_idx;
3439         sector_t new_sector;
3440         sector_t logical_sector, last_sector;
3441         struct stripe_head *sh;
3442         const int rw = bio_data_dir(bi);
3443         int remaining;
3444
3445         if (unlikely(bio_barrier(bi))) {
3446                 bio_endio(bi, -EOPNOTSUPP);
3447                 return 0;
3448         }
3449
3450         md_write_start(mddev, bi);
3451
3452         disk_stat_inc(mddev->gendisk, ios[rw]);
3453         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
3454
3455         if (rw == READ &&
3456              mddev->reshape_position == MaxSector &&
3457              chunk_aligned_read(q,bi))
3458                 return 0;
3459
3460         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3461         last_sector = bi->bi_sector + (bi->bi_size>>9);
3462         bi->bi_next = NULL;
3463         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
3464
3465         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3466                 DEFINE_WAIT(w);
3467                 int disks, data_disks;
3468
3469         retry:
3470                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3471                 if (likely(conf->expand_progress == MaxSector))
3472                         disks = conf->raid_disks;
3473                 else {
3474                         /* spinlock is needed as expand_progress may be
3475                          * 64bit on a 32bit platform, and so it might be
3476                          * possible to see a half-updated value
3477                          * Ofcourse expand_progress could change after
3478                          * the lock is dropped, so once we get a reference
3479                          * to the stripe that we think it is, we will have
3480                          * to check again.
3481                          */
3482                         spin_lock_irq(&conf->device_lock);
3483                         disks = conf->raid_disks;
3484                         if (logical_sector >= conf->expand_progress)
3485                                 disks = conf->previous_raid_disks;
3486                         else {
3487                                 if (logical_sector >= conf->expand_lo) {
3488                                         spin_unlock_irq(&conf->device_lock);
3489                                         schedule();
3490                                         goto retry;
3491                                 }
3492                         }
3493                         spin_unlock_irq(&conf->device_lock);
3494                 }
3495                 data_disks = disks - conf->max_degraded;
3496
3497                 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
3498                                                   &dd_idx, &pd_idx, conf);
3499                 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3500                         (unsigned long long)new_sector,
3501                         (unsigned long long)logical_sector);
3502
3503                 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
3504                 if (sh) {
3505                         if (unlikely(conf->expand_progress != MaxSector)) {
3506                                 /* expansion might have moved on while waiting for a
3507                                  * stripe, so we must do the range check again.
3508                                  * Expansion could still move past after this
3509                                  * test, but as we are holding a reference to
3510                                  * 'sh', we know that if that happens,
3511                                  *  STRIPE_EXPANDING will get set and the expansion
3512                                  * won't proceed until we finish with the stripe.
3513                                  */
3514                                 int must_retry = 0;
3515                                 spin_lock_irq(&conf->device_lock);
3516                                 if (logical_sector <  conf->expand_progress &&
3517                                     disks == conf->previous_raid_disks)
3518                                         /* mismatch, need to try again */
3519                                         must_retry = 1;
3520                                 spin_unlock_irq(&conf->device_lock);
3521                                 if (must_retry) {
3522                                         release_stripe(sh);
3523                                         goto retry;
3524                                 }
3525                         }
3526                         /* FIXME what if we get a false positive because these
3527                          * are being updated.
3528                          */
3529                         if (logical_sector >= mddev->suspend_lo &&
3530                             logical_sector < mddev->suspend_hi) {
3531                                 release_stripe(sh);
3532                                 schedule();
3533                                 goto retry;
3534                         }
3535
3536                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3537                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3538                                 /* Stripe is busy expanding or
3539                                  * add failed due to overlap.  Flush everything
3540                                  * and wait a while
3541                                  */
3542                                 raid5_unplug_device(mddev->queue);
3543                                 release_stripe(sh);
3544                                 schedule();
3545                                 goto retry;
3546                         }
3547                         finish_wait(&conf->wait_for_overlap, &w);
3548                         handle_stripe(sh, NULL);
3549                         release_stripe(sh);
3550                 } else {
3551                         /* cannot get stripe for read-ahead, just give-up */
3552                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
3553                         finish_wait(&conf->wait_for_overlap, &w);
3554                         break;
3555                 }
3556
3557         }
3558         spin_lock_irq(&conf->device_lock);
3559         remaining = --bi->bi_phys_segments;
3560         spin_unlock_irq(&conf->device_lock);
3561         if (remaining == 0) {
3562
3563                 if ( rw == WRITE )
3564                         md_write_end(mddev);
3565
3566                 bi->bi_end_io(bi,
3567                               test_bit(BIO_UPTODATE, &bi->bi_flags)
3568                                 ? 0 : -EIO);
3569         }
3570         return 0;
3571 }
3572
3573 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3574 {
3575         /* reshaping is quite different to recovery/resync so it is
3576          * handled quite separately ... here.
3577          *
3578          * On each call to sync_request, we gather one chunk worth of
3579          * destination stripes and flag them as expanding.
3580          * Then we find all the source stripes and request reads.
3581          * As the reads complete, handle_stripe will copy the data
3582          * into the destination stripe and release that stripe.
3583          */
3584         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3585         struct stripe_head *sh;
3586         int pd_idx;
3587         sector_t first_sector, last_sector;
3588         int raid_disks = conf->previous_raid_disks;
3589         int data_disks = raid_disks - conf->max_degraded;
3590         int new_data_disks = conf->raid_disks - conf->max_degraded;
3591         int i;
3592         int dd_idx;
3593         sector_t writepos, safepos, gap;
3594
3595         if (sector_nr == 0 &&
3596             conf->expand_progress != 0) {
3597                 /* restarting in the middle, skip the initial sectors */
3598                 sector_nr = conf->expand_progress;
3599                 sector_div(sector_nr, new_data_disks);
3600                 *skipped = 1;
3601                 return sector_nr;
3602         }
3603
3604         /* we update the metadata when there is more than 3Meg
3605          * in the block range (that is rather arbitrary, should
3606          * probably be time based) or when the data about to be
3607          * copied would over-write the source of the data at
3608          * the front of the range.
3609          * i.e. one new_stripe forward from expand_progress new_maps
3610          * to after where expand_lo old_maps to
3611          */
3612         writepos = conf->expand_progress +
3613                 conf->chunk_size/512*(new_data_disks);
3614         sector_div(writepos, new_data_disks);
3615         safepos = conf->expand_lo;
3616         sector_div(safepos, data_disks);
3617         gap = conf->expand_progress - conf->expand_lo;
3618
3619         if (writepos >= safepos ||
3620             gap > (new_data_disks)*3000*2 /*3Meg*/) {
3621                 /* Cannot proceed until we've updated the superblock... */
3622                 wait_event(conf->wait_for_overlap,
3623                            atomic_read(&conf->reshape_stripes)==0);
3624                 mddev->reshape_position = conf->expand_progress;
3625                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3626                 md_wakeup_thread(mddev->thread);
3627                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3628                            kthread_should_stop());
3629                 spin_lock_irq(&conf->device_lock);
3630                 conf->expand_lo = mddev->reshape_position;
3631                 spin_unlock_irq(&conf->device_lock);
3632                 wake_up(&conf->wait_for_overlap);
3633         }
3634
3635         for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3636                 int j;
3637                 int skipped = 0;
3638                 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3639                 sh = get_active_stripe(conf, sector_nr+i,
3640                                        conf->raid_disks, pd_idx, 0);
3641                 set_bit(STRIPE_EXPANDING, &sh->state);
3642                 atomic_inc(&conf->reshape_stripes);
3643                 /* If any of this stripe is beyond the end of the old
3644                  * array, then we need to zero those blocks
3645                  */
3646                 for (j=sh->disks; j--;) {
3647                         sector_t s;
3648                         if (j == sh->pd_idx)
3649                                 continue;
3650                         if (conf->level == 6 &&
3651                             j == raid6_next_disk(sh->pd_idx, sh->disks))
3652                                 continue;
3653                         s = compute_blocknr(sh, j);
3654                         if (s < (mddev->array_size<<1)) {
3655                                 skipped = 1;
3656                                 continue;
3657                         }
3658                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3659                         set_bit(R5_Expanded, &sh->dev[j].flags);
3660                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
3661                 }
3662                 if (!skipped) {
3663                         set_bit(STRIPE_EXPAND_READY, &sh->state);
3664                         set_bit(STRIPE_HANDLE, &sh->state);
3665                 }
3666                 release_stripe(sh);
3667         }
3668         spin_lock_irq(&conf->device_lock);
3669         conf->expand_progress = (sector_nr + i) * new_data_disks;
3670         spin_unlock_irq(&conf->device_lock);
3671         /* Ok, those stripe are ready. We can start scheduling
3672          * reads on the source stripes.
3673          * The source stripes are determined by mapping the first and last
3674          * block on the destination stripes.
3675          */
3676         first_sector =
3677                 raid5_compute_sector(sector_nr*(new_data_disks),
3678                                      raid_disks, data_disks,
3679                                      &dd_idx, &pd_idx, conf);
3680         last_sector =
3681                 raid5_compute_sector((sector_nr+conf->chunk_size/512)
3682                                      *(new_data_disks) -1,
3683                                      raid_disks, data_disks,
3684                                      &dd_idx, &pd_idx, conf);
3685         if (last_sector >= (mddev->size<<1))
3686                 last_sector = (mddev->size<<1)-1;
3687         while (first_sector <= last_sector) {
3688                 pd_idx = stripe_to_pdidx(first_sector, conf,
3689                                          conf->previous_raid_disks);
3690                 sh = get_active_stripe(conf, first_sector,
3691                                        conf->previous_raid_disks, pd_idx, 0);
3692                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3693                 set_bit(STRIPE_HANDLE, &sh->state);
3694                 release_stripe(sh);
3695                 first_sector += STRIPE_SECTORS;
3696         }
3697         return conf->chunk_size>>9;
3698 }
3699
3700 /* FIXME go_faster isn't used */
3701 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3702 {
3703         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3704         struct stripe_head *sh;
3705         int pd_idx;
3706         int raid_disks = conf->raid_disks;
3707         sector_t max_sector = mddev->size << 1;
3708         int sync_blocks;
3709         int still_degraded = 0;
3710         int i;
3711
3712         if (sector_nr >= max_sector) {
3713                 /* just being told to finish up .. nothing much to do */
3714                 unplug_slaves(mddev);
3715                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3716                         end_reshape(conf);
3717                         return 0;
3718                 }
3719
3720                 if (mddev->curr_resync < max_sector) /* aborted */
3721                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3722                                         &sync_blocks, 1);
3723                 else /* completed sync */
3724                         conf->fullsync = 0;
3725                 bitmap_close_sync(mddev->bitmap);
3726
3727                 return 0;
3728         }
3729
3730         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3731                 return reshape_request(mddev, sector_nr, skipped);
3732
3733         /* if there is too many failed drives and we are trying
3734          * to resync, then assert that we are finished, because there is
3735          * nothing we can do.
3736          */
3737         if (mddev->degraded >= conf->max_degraded &&
3738             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3739                 sector_t rv = (mddev->size << 1) - sector_nr;
3740                 *skipped = 1;
3741                 return rv;
3742         }
3743         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3744             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3745             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3746                 /* we can skip this block, and probably more */
3747                 sync_blocks /= STRIPE_SECTORS;
3748                 *skipped = 1;
3749                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3750         }
3751
3752         pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3753         sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3754         if (sh == NULL) {
3755                 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3756                 /* make sure we don't swamp the stripe cache if someone else
3757                  * is trying to get access
3758                  */
3759                 schedule_timeout_uninterruptible(1);
3760         }
3761         /* Need to check if array will still be degraded after recovery/resync
3762          * We don't need to check the 'failed' flag as when that gets set,
3763          * recovery aborts.
3764          */
3765         for (i=0; i<mddev->raid_disks; i++)
3766                 if (conf->disks[i].rdev == NULL)
3767                         still_degraded = 1;
3768
3769         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3770
3771         spin_lock(&sh->lock);
3772         set_bit(STRIPE_SYNCING, &sh->state);
3773         clear_bit(STRIPE_INSYNC, &sh->state);
3774         spin_unlock(&sh->lock);
3775
3776         handle_stripe(sh, NULL);
3777         release_stripe(sh);
3778
3779         return STRIPE_SECTORS;
3780 }
3781
3782 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3783 {
3784         /* We may not be able to submit a whole bio at once as there
3785          * may not be enough stripe_heads available.
3786          * We cannot pre-allocate enough stripe_heads as we may need
3787          * more than exist in the cache (if we allow ever large chunks).
3788          * So we do one stripe head at a time and record in
3789          * ->bi_hw_segments how many have been done.
3790          *
3791          * We *know* that this entire raid_bio is in one chunk, so
3792          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3793          */
3794         struct stripe_head *sh;
3795         int dd_idx, pd_idx;
3796         sector_t sector, logical_sector, last_sector;
3797         int scnt = 0;
3798         int remaining;
3799         int handled = 0;
3800
3801         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3802         sector = raid5_compute_sector(  logical_sector,
3803                                         conf->raid_disks,
3804                                         conf->raid_disks - conf->max_degraded,
3805                                         &dd_idx,
3806                                         &pd_idx,
3807                                         conf);
3808         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3809
3810         for (; logical_sector < last_sector;
3811              logical_sector += STRIPE_SECTORS,
3812                      sector += STRIPE_SECTORS,
3813                      scnt++) {
3814
3815                 if (scnt < raid_bio->bi_hw_segments)
3816                         /* already done this stripe */
3817                         continue;
3818
3819                 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3820
3821                 if (!sh) {
3822                         /* failed to get a stripe - must wait */
3823                         raid_bio->bi_hw_segments = scnt;
3824                         conf->retry_read_aligned = raid_bio;
3825                         return handled;
3826                 }
3827
3828                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3829                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3830                         release_stripe(sh);
3831                         raid_bio->bi_hw_segments = scnt;
3832                         conf->retry_read_aligned = raid_bio;
3833                         return handled;
3834                 }
3835
3836                 handle_stripe(sh, NULL);
3837                 release_stripe(sh);
3838                 handled++;
3839         }
3840         spin_lock_irq(&conf->device_lock);
3841         remaining = --raid_bio->bi_phys_segments;
3842         spin_unlock_irq(&conf->device_lock);
3843         if (remaining == 0) {
3844
3845                 raid_bio->bi_end_io(raid_bio,
3846                               test_bit(BIO_UPTODATE, &raid_bio->bi_flags)
3847                                 ? 0 : -EIO);
3848         }
3849         if (atomic_dec_and_test(&conf->active_aligned_reads))
3850                 wake_up(&conf->wait_for_stripe);
3851         return handled;
3852 }
3853
3854
3855
3856 /*
3857  * This is our raid5 kernel thread.
3858  *
3859  * We scan the hash table for stripes which can be handled now.
3860  * During the scan, completed stripes are saved for us by the interrupt
3861  * handler, so that they will not have to wait for our next wakeup.
3862  */
3863 static void raid5d (mddev_t *mddev)
3864 {
3865         struct stripe_head *sh;
3866         raid5_conf_t *conf = mddev_to_conf(mddev);
3867         int handled;
3868
3869         pr_debug("+++ raid5d active\n");
3870
3871         md_check_recovery(mddev);
3872
3873         handled = 0;
3874         spin_lock_irq(&conf->device_lock);
3875         while (1) {
3876                 struct list_head *first;
3877                 struct bio *bio;
3878
3879                 if (conf->seq_flush != conf->seq_write) {
3880                         int seq = conf->seq_flush;
3881                         spin_unlock_irq(&conf->device_lock);
3882                         bitmap_unplug(mddev->bitmap);
3883                         spin_lock_irq(&conf->device_lock);
3884                         conf->seq_write = seq;
3885                         activate_bit_delay(conf);
3886                 }
3887
3888                 if (list_empty(&conf->handle_list) &&
3889                     atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
3890                     !blk_queue_plugged(mddev->queue) &&
3891                     !list_empty(&conf->delayed_list))
3892                         raid5_activate_delayed(conf);
3893
3894                 while ((bio = remove_bio_from_retry(conf))) {
3895                         int ok;
3896                         spin_unlock_irq(&conf->device_lock);
3897                         ok = retry_aligned_read(conf, bio);
3898                         spin_lock_irq(&conf->device_lock);
3899                         if (!ok)
3900                                 break;
3901                         handled++;
3902                 }
3903
3904                 if (list_empty(&conf->handle_list)) {
3905                         async_tx_issue_pending_all();
3906                         break;
3907                 }
3908
3909                 first = conf->handle_list.next;
3910                 sh = list_entry(first, struct stripe_head, lru);
3911
3912                 list_del_init(first);
3913                 atomic_inc(&sh->count);
3914                 BUG_ON(atomic_read(&sh->count)!= 1);
3915                 spin_unlock_irq(&conf->device_lock);
3916
3917                 handled++;
3918                 handle_stripe(sh, conf->spare_page);
3919                 release_stripe(sh);
3920
3921                 spin_lock_irq(&conf->device_lock);
3922         }
3923         pr_debug("%d stripes handled\n", handled);
3924
3925         spin_unlock_irq(&conf->device_lock);
3926
3927         unplug_slaves(mddev);
3928
3929         pr_debug("--- raid5d inactive\n");
3930 }
3931
3932 static ssize_t
3933 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
3934 {
3935         raid5_conf_t *conf = mddev_to_conf(mddev);
3936         if (conf)
3937                 return sprintf(page, "%d\n", conf->max_nr_stripes);
3938         else
3939                 return 0;
3940 }
3941
3942 static ssize_t
3943 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
3944 {
3945         raid5_conf_t *conf = mddev_to_conf(mddev);
3946         char *end;
3947         int new;
3948         if (len >= PAGE_SIZE)
3949                 return -EINVAL;
3950         if (!conf)
3951                 return -ENODEV;
3952
3953         new = simple_strtoul(page, &end, 10);
3954         if (!*page || (*end && *end != '\n') )
3955                 return -EINVAL;
3956         if (new <= 16 || new > 32768)
3957                 return -EINVAL;
3958         while (new < conf->max_nr_stripes) {
3959                 if (drop_one_stripe(conf))
3960                         conf->max_nr_stripes--;
3961                 else
3962                         break;
3963         }
3964         md_allow_write(mddev);
3965         while (new > conf->max_nr_stripes) {
3966                 if (grow_one_stripe(conf))
3967                         conf->max_nr_stripes++;
3968                 else break;
3969         }
3970         return len;
3971 }
3972
3973 static struct md_sysfs_entry
3974 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
3975                                 raid5_show_stripe_cache_size,
3976                                 raid5_store_stripe_cache_size);
3977
3978 static ssize_t
3979 stripe_cache_active_show(mddev_t *mddev, char *page)
3980 {
3981         raid5_conf_t *conf = mddev_to_conf(mddev);
3982         if (conf)
3983                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
3984         else
3985                 return 0;
3986 }
3987
3988 static struct md_sysfs_entry
3989 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
3990
3991 static struct attribute *raid5_attrs[] =  {
3992         &raid5_stripecache_size.attr,
3993         &raid5_stripecache_active.attr,
3994         NULL,
3995 };
3996 static struct attribute_group raid5_attrs_group = {
3997         .name = NULL,
3998         .attrs = raid5_attrs,
3999 };
4000
4001 static int run(mddev_t *mddev)
4002 {
4003         raid5_conf_t *conf;
4004         int raid_disk, memory;
4005         mdk_rdev_t *rdev;
4006         struct disk_info *disk;
4007         struct list_head *tmp;
4008         int working_disks = 0;
4009
4010         if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
4011                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4012                        mdname(mddev), mddev->level);
4013                 return -EIO;
4014         }
4015
4016         if (mddev->reshape_position != MaxSector) {
4017                 /* Check that we can continue the reshape.
4018                  * Currently only disks can change, it must
4019                  * increase, and we must be past the point where
4020                  * a stripe over-writes itself
4021                  */
4022                 sector_t here_new, here_old;
4023                 int old_disks;
4024                 int max_degraded = (mddev->level == 5 ? 1 : 2);
4025
4026                 if (mddev->new_level != mddev->level ||
4027                     mddev->new_layout != mddev->layout ||
4028                     mddev->new_chunk != mddev->chunk_size) {
4029                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4030                                "required - aborting.\n",
4031                                mdname(mddev));
4032                         return -EINVAL;
4033                 }
4034                 if (mddev->delta_disks <= 0) {
4035                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4036                                "(reduce disks) required - aborting.\n",
4037                                mdname(mddev));
4038                         return -EINVAL;
4039                 }
4040                 old_disks = mddev->raid_disks - mddev->delta_disks;
4041                 /* reshape_position must be on a new-stripe boundary, and one
4042                  * further up in new geometry must map after here in old
4043                  * geometry.
4044                  */
4045                 here_new = mddev->reshape_position;
4046                 if (sector_div(here_new, (mddev->chunk_size>>9)*
4047                                (mddev->raid_disks - max_degraded))) {
4048                         printk(KERN_ERR "raid5: reshape_position not "
4049                                "on a stripe boundary\n");
4050                         return -EINVAL;
4051                 }
4052                 /* here_new is the stripe we will write to */
4053                 here_old = mddev->reshape_position;
4054                 sector_div(here_old, (mddev->chunk_size>>9)*
4055                            (old_disks-max_degraded));
4056                 /* here_old is the first stripe that we might need to read
4057                  * from */
4058                 if (here_new >= here_old) {
4059                         /* Reading from the same stripe as writing to - bad */
4060                         printk(KERN_ERR "raid5: reshape_position too early for "
4061                                "auto-recovery - aborting.\n");
4062                         return -EINVAL;
4063                 }
4064                 printk(KERN_INFO "raid5: reshape will continue\n");
4065                 /* OK, we should be able to continue; */
4066         }
4067
4068
4069         mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4070         if ((conf = mddev->private) == NULL)
4071                 goto abort;
4072         if (mddev->reshape_position == MaxSector) {
4073                 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4074         } else {
4075                 conf->raid_disks = mddev->raid_disks;
4076                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4077         }
4078
4079         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4080                               GFP_KERNEL);
4081         if (!conf->disks)
4082                 goto abort;
4083
4084         conf->mddev = mddev;
4085
4086         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4087                 goto abort;
4088
4089         if (mddev->level == 6) {
4090                 conf->spare_page = alloc_page(GFP_KERNEL);
4091                 if (!conf->spare_page)
4092                         goto abort;
4093         }
4094         spin_lock_init(&conf->device_lock);
4095         init_waitqueue_head(&conf->wait_for_stripe);
4096         init_waitqueue_head(&conf->wait_for_overlap);
4097         INIT_LIST_HEAD(&conf->handle_list);
4098         INIT_LIST_HEAD(&conf->delayed_list);
4099         INIT_LIST_HEAD(&conf->bitmap_list);
4100         INIT_LIST_HEAD(&conf->inactive_list);
4101         atomic_set(&conf->active_stripes, 0);
4102         atomic_set(&conf->preread_active_stripes, 0);
4103         atomic_set(&conf->active_aligned_reads, 0);
4104
4105         pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4106
4107         ITERATE_RDEV(mddev,rdev,tmp) {
4108                 raid_disk = rdev->raid_disk;
4109                 if (raid_disk >= conf->raid_disks
4110                     || raid_disk < 0)
4111                         continue;
4112                 disk = conf->disks + raid_disk;
4113
4114                 disk->rdev = rdev;
4115
4116                 if (test_bit(In_sync, &rdev->flags)) {
4117                         char b[BDEVNAME_SIZE];
4118                         printk(KERN_INFO "raid5: device %s operational as raid"
4119                                 " disk %d\n", bdevname(rdev->bdev,b),
4120                                 raid_disk);
4121                         working_disks++;
4122                 }
4123         }
4124
4125         /*
4126          * 0 for a fully functional array, 1 or 2 for a degraded array.
4127          */
4128         mddev->degraded = conf->raid_disks - working_disks;
4129         conf->mddev = mddev;
4130         conf->chunk_size = mddev->chunk_size;
4131         conf->level = mddev->level;
4132         if (conf->level == 6)
4133                 conf->max_degraded = 2;
4134         else
4135                 conf->max_degraded = 1;
4136         conf->algorithm = mddev->layout;
4137         conf->max_nr_stripes = NR_STRIPES;
4138         conf->expand_progress = mddev->reshape_position;
4139
4140         /* device size must be a multiple of chunk size */
4141         mddev->size &= ~(mddev->chunk_size/1024 -1);
4142         mddev->resync_max_sectors = mddev->size << 1;
4143
4144         if (conf->level == 6 && conf->raid_disks < 4) {
4145                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4146                        mdname(mddev), conf->raid_disks);
4147                 goto abort;
4148         }
4149         if (!conf->chunk_size || conf->chunk_size % 4) {
4150                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4151                         conf->chunk_size, mdname(mddev));
4152                 goto abort;
4153         }
4154         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4155                 printk(KERN_ERR
4156                         "raid5: unsupported parity algorithm %d for %s\n",
4157                         conf->algorithm, mdname(mddev));
4158                 goto abort;
4159         }
4160         if (mddev->degraded > conf->max_degraded) {
4161                 printk(KERN_ERR "raid5: not enough operational devices for %s"
4162                         " (%d/%d failed)\n",
4163                         mdname(mddev), mddev->degraded, conf->raid_disks);
4164                 goto abort;
4165         }
4166
4167         if (mddev->degraded > 0 &&
4168             mddev->recovery_cp != MaxSector) {
4169                 if (mddev->ok_start_degraded)
4170                         printk(KERN_WARNING
4171                                "raid5: starting dirty degraded array: %s"
4172                                "- data corruption possible.\n",
4173                                mdname(mddev));
4174                 else {
4175                         printk(KERN_ERR
4176                                "raid5: cannot start dirty degraded array for %s\n",
4177                                mdname(mddev));
4178                         goto abort;
4179                 }
4180         }
4181
4182         {
4183                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4184                 if (!mddev->thread) {
4185                         printk(KERN_ERR
4186                                 "raid5: couldn't allocate thread for %s\n",
4187                                 mdname(mddev));
4188                         goto abort;
4189                 }
4190         }
4191         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4192                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4193         if (grow_stripes(conf, conf->max_nr_stripes)) {
4194                 printk(KERN_ERR
4195                         "raid5: couldn't allocate %dkB for buffers\n", memory);
4196                 shrink_stripes(conf);
4197                 md_unregister_thread(mddev->thread);
4198                 goto abort;
4199         } else
4200                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4201                         memory, mdname(mddev));
4202
4203         if (mddev->degraded == 0)
4204                 printk("raid5: raid level %d set %s active with %d out of %d"
4205                         " devices, algorithm %d\n", conf->level, mdname(mddev),
4206                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4207                         conf->algorithm);
4208         else
4209                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4210                         " out of %d devices, algorithm %d\n", conf->level,
4211                         mdname(mddev), mddev->raid_disks - mddev->degraded,
4212                         mddev->raid_disks, conf->algorithm);
4213
4214         print_raid5_conf(conf);
4215
4216         if (conf->expand_progress != MaxSector) {
4217                 printk("...ok start reshape thread\n");
4218                 conf->expand_lo = conf->expand_progress;
4219                 atomic_set(&conf->reshape_stripes, 0);
4220                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4221                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4222                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4223                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4224                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4225                                                         "%s_reshape");
4226         }
4227
4228         /* read-ahead size must cover two whole stripes, which is
4229          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4230          */
4231         {
4232                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4233                 int stripe = data_disks *
4234                         (mddev->chunk_size / PAGE_SIZE);
4235                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4236                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4237         }
4238
4239         /* Ok, everything is just fine now */
4240         if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4241                 printk(KERN_WARNING
4242                        "raid5: failed to create sysfs attributes for %s\n",
4243                        mdname(mddev));
4244
4245         mddev->queue->unplug_fn = raid5_unplug_device;
4246         mddev->queue->backing_dev_info.congested_data = mddev;
4247         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4248
4249         mddev->array_size =  mddev->size * (conf->previous_raid_disks -
4250                                             conf->max_degraded);
4251
4252         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4253
4254         return 0;
4255 abort:
4256         if (conf) {
4257                 print_raid5_conf(conf);
4258                 safe_put_page(conf->spare_page);
4259                 kfree(conf->disks);
4260                 kfree(conf->stripe_hashtbl);
4261                 kfree(conf);
4262         }
4263         mddev->private = NULL;
4264         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4265         return -EIO;
4266 }
4267
4268
4269
4270 static int stop(mddev_t *mddev)
4271 {
4272         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4273
4274         md_unregister_thread(mddev->thread);
4275         mddev->thread = NULL;
4276         shrink_stripes(conf);
4277         kfree(conf->stripe_hashtbl);
4278         mddev->queue->backing_dev_info.congested_fn = NULL;
4279         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4280         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4281         kfree(conf->disks);
4282         kfree(conf);
4283         mddev->private = NULL;
4284         return 0;
4285 }
4286
4287 #ifdef DEBUG
4288 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
4289 {
4290         int i;
4291
4292         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4293                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4294         seq_printf(seq, "sh %llu,  count %d.\n",
4295                    (unsigned long long)sh->sector, atomic_read(&sh->count));
4296         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4297         for (i = 0; i < sh->disks; i++) {
4298                 seq_printf(seq, "(cache%d: %p %ld) ",
4299                            i, sh->dev[i].page, sh->dev[i].flags);
4300         }
4301         seq_printf(seq, "\n");
4302 }
4303
4304 static void printall (struct seq_file *seq, raid5_conf_t *conf)
4305 {
4306         struct stripe_head *sh;
4307         struct hlist_node *hn;
4308         int i;
4309
4310         spin_lock_irq(&conf->device_lock);
4311         for (i = 0; i < NR_HASH; i++) {
4312                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4313                         if (sh->raid_conf != conf)
4314                                 continue;
4315                         print_sh(seq, sh);
4316                 }
4317         }
4318         spin_unlock_irq(&conf->device_lock);
4319 }
4320 #endif
4321
4322 static void status (struct seq_file *seq, mddev_t *mddev)
4323 {
4324         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4325         int i;
4326
4327         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4328         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4329         for (i = 0; i < conf->raid_disks; i++)
4330                 seq_printf (seq, "%s",
4331                                conf->disks[i].rdev &&
4332                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4333         seq_printf (seq, "]");
4334 #ifdef DEBUG
4335         seq_printf (seq, "\n");
4336         printall(seq, conf);
4337 #endif
4338 }
4339
4340 static void print_raid5_conf (raid5_conf_t *conf)
4341 {
4342         int i;
4343         struct disk_info *tmp;
4344
4345         printk("RAID5 conf printout:\n");
4346         if (!conf) {
4347                 printk("(conf==NULL)\n");
4348                 return;
4349         }
4350         printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4351                  conf->raid_disks - conf->mddev->degraded);
4352
4353         for (i = 0; i < conf->raid_disks; i++) {
4354                 char b[BDEVNAME_SIZE];
4355                 tmp = conf->disks + i;
4356                 if (tmp->rdev)
4357                 printk(" disk %d, o:%d, dev:%s\n",
4358                         i, !test_bit(Faulty, &tmp->rdev->flags),
4359                         bdevname(tmp->rdev->bdev,b));
4360         }
4361 }
4362
4363 static int raid5_spare_active(mddev_t *mddev)
4364 {
4365         int i;
4366         raid5_conf_t *conf = mddev->private;
4367         struct disk_info *tmp;
4368
4369         for (i = 0; i < conf->raid_disks; i++) {
4370                 tmp = conf->disks + i;
4371                 if (tmp->rdev
4372                     && !test_bit(Faulty, &tmp->rdev->flags)
4373                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4374                         unsigned long flags;
4375                         spin_lock_irqsave(&conf->device_lock, flags);
4376                         mddev->degraded--;
4377                         spin_unlock_irqrestore(&conf->device_lock, flags);
4378                 }
4379         }
4380         print_raid5_conf(conf);
4381         return 0;
4382 }
4383
4384 static int raid5_remove_disk(mddev_t *mddev, int number)
4385 {
4386         raid5_conf_t *conf = mddev->private;
4387         int err = 0;
4388         mdk_rdev_t *rdev;
4389         struct disk_info *p = conf->disks + number;
4390
4391         print_raid5_conf(conf);
4392         rdev = p->rdev;
4393         if (rdev) {
4394                 if (test_bit(In_sync, &rdev->flags) ||
4395                     atomic_read(&rdev->nr_pending)) {
4396                         err = -EBUSY;
4397                         goto abort;
4398                 }
4399                 p->rdev = NULL;
4400                 synchronize_rcu();
4401                 if (atomic_read(&rdev->nr_pending)) {
4402                         /* lost the race, try later */
4403                         err = -EBUSY;
4404                         p->rdev = rdev;
4405                 }
4406         }
4407 abort:
4408
4409         print_raid5_conf(conf);
4410         return err;
4411 }
4412
4413 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4414 {
4415         raid5_conf_t *conf = mddev->private;
4416         int found = 0;
4417         int disk;
4418         struct disk_info *p;
4419
4420         if (mddev->degraded > conf->max_degraded)
4421                 /* no point adding a device */
4422                 return 0;
4423
4424         /*
4425          * find the disk ... but prefer rdev->saved_raid_disk
4426          * if possible.
4427          */
4428         if (rdev->saved_raid_disk >= 0 &&
4429             conf->disks[rdev->saved_raid_disk].rdev == NULL)
4430                 disk = rdev->saved_raid_disk;
4431         else
4432                 disk = 0;
4433         for ( ; disk < conf->raid_disks; disk++)
4434                 if ((p=conf->disks + disk)->rdev == NULL) {
4435                         clear_bit(In_sync, &rdev->flags);
4436                         rdev->raid_disk = disk;
4437                         found = 1;
4438                         if (rdev->saved_raid_disk != disk)
4439                                 conf->fullsync = 1;
4440                         rcu_assign_pointer(p->rdev, rdev);
4441                         break;
4442                 }
4443         print_raid5_conf(conf);
4444         return found;
4445 }
4446
4447 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4448 {
4449         /* no resync is happening, and there is enough space
4450          * on all devices, so we can resize.
4451          * We need to make sure resync covers any new space.
4452          * If the array is shrinking we should possibly wait until
4453          * any io in the removed space completes, but it hardly seems
4454          * worth it.
4455          */
4456         raid5_conf_t *conf = mddev_to_conf(mddev);
4457
4458         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4459         mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
4460         set_capacity(mddev->gendisk, mddev->array_size << 1);
4461         mddev->changed = 1;
4462         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
4463                 mddev->recovery_cp = mddev->size << 1;
4464                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4465         }
4466         mddev->size = sectors /2;
4467         mddev->resync_max_sectors = sectors;
4468         return 0;
4469 }
4470
4471 #ifdef CONFIG_MD_RAID5_RESHAPE
4472 static int raid5_check_reshape(mddev_t *mddev)
4473 {
4474         raid5_conf_t *conf = mddev_to_conf(mddev);
4475         int err;
4476
4477         if (mddev->delta_disks < 0 ||
4478             mddev->new_level != mddev->level)
4479                 return -EINVAL; /* Cannot shrink array or change level yet */
4480         if (mddev->delta_disks == 0)
4481                 return 0; /* nothing to do */
4482
4483         /* Can only proceed if there are plenty of stripe_heads.
4484          * We need a minimum of one full stripe,, and for sensible progress
4485          * it is best to have about 4 times that.
4486          * If we require 4 times, then the default 256 4K stripe_heads will
4487          * allow for chunk sizes up to 256K, which is probably OK.
4488          * If the chunk size is greater, user-space should request more
4489          * stripe_heads first.
4490          */
4491         if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4492             (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4493                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
4494                        (mddev->chunk_size / STRIPE_SIZE)*4);
4495                 return -ENOSPC;
4496         }
4497
4498         err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4499         if (err)
4500                 return err;
4501
4502         if (mddev->degraded > conf->max_degraded)
4503                 return -EINVAL;
4504         /* looks like we might be able to manage this */
4505         return 0;
4506 }
4507
4508 static int raid5_start_reshape(mddev_t *mddev)
4509 {
4510         raid5_conf_t *conf = mddev_to_conf(mddev);
4511         mdk_rdev_t *rdev;
4512         struct list_head *rtmp;
4513         int spares = 0;
4514         int added_devices = 0;
4515         unsigned long flags;
4516
4517         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4518                 return -EBUSY;
4519
4520         ITERATE_RDEV(mddev, rdev, rtmp)
4521                 if (rdev->raid_disk < 0 &&
4522                     !test_bit(Faulty, &rdev->flags))
4523                         spares++;
4524
4525         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4526                 /* Not enough devices even to make a degraded array
4527                  * of that size
4528                  */
4529                 return -EINVAL;
4530
4531         atomic_set(&conf->reshape_stripes, 0);
4532         spin_lock_irq(&conf->device_lock);
4533         conf->previous_raid_disks = conf->raid_disks;
4534         conf->raid_disks += mddev->delta_disks;
4535         conf->expand_progress = 0;
4536         conf->expand_lo = 0;
4537         spin_unlock_irq(&conf->device_lock);
4538
4539         /* Add some new drives, as many as will fit.
4540          * We know there are enough to make the newly sized array work.
4541          */
4542         ITERATE_RDEV(mddev, rdev, rtmp)
4543                 if (rdev->raid_disk < 0 &&
4544                     !test_bit(Faulty, &rdev->flags)) {
4545                         if (raid5_add_disk(mddev, rdev)) {
4546                                 char nm[20];
4547                                 set_bit(In_sync, &rdev->flags);
4548                                 added_devices++;
4549                                 rdev->recovery_offset = 0;
4550                                 sprintf(nm, "rd%d", rdev->raid_disk);
4551                                 if (sysfs_create_link(&mddev->kobj,
4552                                                       &rdev->kobj, nm))
4553                                         printk(KERN_WARNING
4554                                                "raid5: failed to create "
4555                                                " link %s for %s\n",
4556                                                nm, mdname(mddev));
4557                         } else
4558                                 break;
4559                 }
4560
4561         spin_lock_irqsave(&conf->device_lock, flags);
4562         mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4563         spin_unlock_irqrestore(&conf->device_lock, flags);
4564         mddev->raid_disks = conf->raid_disks;
4565         mddev->reshape_position = 0;
4566         set_bit(MD_CHANGE_DEVS, &mddev->flags);
4567
4568         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4569         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4570         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4571         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4572         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4573                                                 "%s_reshape");
4574         if (!mddev->sync_thread) {
4575                 mddev->recovery = 0;
4576                 spin_lock_irq(&conf->device_lock);
4577                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4578                 conf->expand_progress = MaxSector;
4579                 spin_unlock_irq(&conf->device_lock);
4580                 return -EAGAIN;
4581         }
4582         md_wakeup_thread(mddev->sync_thread);
4583         md_new_event(mddev);
4584         return 0;
4585 }
4586 #endif
4587
4588 static void end_reshape(raid5_conf_t *conf)
4589 {
4590         struct block_device *bdev;
4591
4592         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4593                 conf->mddev->array_size = conf->mddev->size *
4594                         (conf->raid_disks - conf->max_degraded);
4595                 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
4596                 conf->mddev->changed = 1;
4597
4598                 bdev = bdget_disk(conf->mddev->gendisk, 0);
4599                 if (bdev) {
4600                         mutex_lock(&bdev->bd_inode->i_mutex);
4601                         i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
4602                         mutex_unlock(&bdev->bd_inode->i_mutex);
4603                         bdput(bdev);
4604                 }
4605                 spin_lock_irq(&conf->device_lock);
4606                 conf->expand_progress = MaxSector;
4607                 spin_unlock_irq(&conf->device_lock);
4608                 conf->mddev->reshape_position = MaxSector;
4609
4610                 /* read-ahead size must cover two whole stripes, which is
4611                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4612                  */
4613                 {
4614                         int data_disks = conf->previous_raid_disks - conf->max_degraded;
4615                         int stripe = data_disks *
4616                                 (conf->mddev->chunk_size / PAGE_SIZE);
4617                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4618                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4619                 }
4620         }
4621 }
4622
4623 static void raid5_quiesce(mddev_t *mddev, int state)
4624 {
4625         raid5_conf_t *conf = mddev_to_conf(mddev);
4626
4627         switch(state) {
4628         case 2: /* resume for a suspend */
4629                 wake_up(&conf->wait_for_overlap);
4630                 break;
4631
4632         case 1: /* stop all writes */
4633                 spin_lock_irq(&conf->device_lock);
4634                 conf->quiesce = 1;
4635                 wait_event_lock_irq(conf->wait_for_stripe,
4636                                     atomic_read(&conf->active_stripes) == 0 &&
4637                                     atomic_read(&conf->active_aligned_reads) == 0,
4638                                     conf->device_lock, /* nothing */);
4639                 spin_unlock_irq(&conf->device_lock);
4640                 break;
4641
4642         case 0: /* re-enable writes */
4643                 spin_lock_irq(&conf->device_lock);
4644                 conf->quiesce = 0;
4645                 wake_up(&conf->wait_for_stripe);
4646                 wake_up(&conf->wait_for_overlap);
4647                 spin_unlock_irq(&conf->device_lock);
4648                 break;
4649         }
4650 }
4651
4652 static struct mdk_personality raid6_personality =
4653 {
4654         .name           = "raid6",
4655         .level          = 6,
4656         .owner          = THIS_MODULE,
4657         .make_request   = make_request,
4658         .run            = run,
4659         .stop           = stop,
4660         .status         = status,
4661         .error_handler  = error,
4662         .hot_add_disk   = raid5_add_disk,
4663         .hot_remove_disk= raid5_remove_disk,
4664         .spare_active   = raid5_spare_active,
4665         .sync_request   = sync_request,
4666         .resize         = raid5_resize,
4667 #ifdef CONFIG_MD_RAID5_RESHAPE
4668         .check_reshape  = raid5_check_reshape,
4669         .start_reshape  = raid5_start_reshape,
4670 #endif
4671         .quiesce        = raid5_quiesce,
4672 };
4673 static struct mdk_personality raid5_personality =
4674 {
4675         .name           = "raid5",
4676         .level          = 5,
4677         .owner          = THIS_MODULE,
4678         .make_request   = make_request,
4679         .run            = run,
4680         .stop           = stop,
4681         .status         = status,
4682         .error_handler  = error,
4683         .hot_add_disk   = raid5_add_disk,
4684         .hot_remove_disk= raid5_remove_disk,
4685         .spare_active   = raid5_spare_active,
4686         .sync_request   = sync_request,
4687         .resize         = raid5_resize,
4688 #ifdef CONFIG_MD_RAID5_RESHAPE
4689         .check_reshape  = raid5_check_reshape,
4690         .start_reshape  = raid5_start_reshape,
4691 #endif
4692         .quiesce        = raid5_quiesce,
4693 };
4694
4695 static struct mdk_personality raid4_personality =
4696 {
4697         .name           = "raid4",
4698         .level          = 4,
4699         .owner          = THIS_MODULE,
4700         .make_request   = make_request,
4701         .run            = run,
4702         .stop           = stop,
4703         .status         = status,
4704         .error_handler  = error,
4705         .hot_add_disk   = raid5_add_disk,
4706         .hot_remove_disk= raid5_remove_disk,
4707         .spare_active   = raid5_spare_active,
4708         .sync_request   = sync_request,
4709         .resize         = raid5_resize,
4710 #ifdef CONFIG_MD_RAID5_RESHAPE
4711         .check_reshape  = raid5_check_reshape,
4712         .start_reshape  = raid5_start_reshape,
4713 #endif
4714         .quiesce        = raid5_quiesce,
4715 };
4716
4717 static int __init raid5_init(void)
4718 {
4719         int e;
4720
4721         e = raid6_select_algo();
4722         if ( e )
4723                 return e;
4724         register_md_personality(&raid6_personality);
4725         register_md_personality(&raid5_personality);
4726         register_md_personality(&raid4_personality);
4727         return 0;
4728 }
4729
4730 static void raid5_exit(void)
4731 {
4732         unregister_md_personality(&raid6_personality);
4733         unregister_md_personality(&raid5_personality);
4734         unregister_md_personality(&raid4_personality);
4735 }
4736
4737 module_init(raid5_init);
4738 module_exit(raid5_exit);
4739 MODULE_LICENSE("GPL");
4740 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4741 MODULE_ALIAS("md-raid5");
4742 MODULE_ALIAS("md-raid4");
4743 MODULE_ALIAS("md-level-5");
4744 MODULE_ALIAS("md-level-4");
4745 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4746 MODULE_ALIAS("md-raid6");
4747 MODULE_ALIAS("md-level-6");
4748
4749 /* This used to be two separate modules, they were: */
4750 MODULE_ALIAS("raid5");
4751 MODULE_ALIAS("raid6");