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                 int bytes = bi->bi_size;
 112
 113                 return_bi = bi->bi_next;
 114                 bi->bi_next = NULL;
 115                 bi->bi_size = 0;
 116                 bi->bi_end_io(bi, bytes,
 117                               test_bit(BIO_UPTODATE, &bi->bi_flags)
 118                                 ? 0 : -EIO);
 119                 bi = return_bi;
 120         }
 121 }
 122
 123 static void print_raid5_conf (raid5_conf_t *conf);
 124
 125 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
 126 {
 127         if (atomic_dec_and_test(&sh->count)) {
 128                 BUG_ON(!list_empty(&sh->lru));
 129                 BUG_ON(atomic_read(&conf->active_stripes)==0);
 130                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
 131                         if (test_bit(STRIPE_DELAYED, &sh->state)) {
 132                                 list_add_tail(&sh->lru, &conf->delayed_list);
 133                                 blk_plug_device(conf->mddev->queue);
 134                         } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 135                                    sh->bm_seq - conf->seq_write > 0) {
 136                                 list_add_tail(&sh->lru, &conf->bitmap_list);
 137                                 blk_plug_device(conf->mddev->queue);
 138                         } else {
 139                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
 140                                 list_add_tail(&sh->lru, &conf->handle_list);
 141                         }
 142                         md_wakeup_thread(conf->mddev->thread);
 143                 } else {
 144                         BUG_ON(sh->ops.pending);
 145                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
 146                                 atomic_dec(&conf->preread_active_stripes);
 147                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
 148                                         md_wakeup_thread(conf->mddev->thread);
 149                         }
 150                         atomic_dec(&conf->active_stripes);
 151                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
 152                                 list_add_tail(&sh->lru, &conf->inactive_list);
 153                                 wake_up(&conf->wait_for_stripe);
 154                                 if (conf->retry_read_aligned)
 155                                         md_wakeup_thread(conf->mddev->thread);
 156                         }
 157                 }
 158         }
 159 }
 160 static void release_stripe(struct stripe_head *sh)
 161 {
 162         raid5_conf_t *conf = sh->raid_conf;
 163         unsigned long flags;
 164
 165         spin_lock_irqsave(&conf->device_lock, flags);
 166         __release_stripe(conf, sh);
 167         spin_unlock_irqrestore(&conf->device_lock, flags);
 168 }
 169
 170 static inline void remove_hash(struct stripe_head *sh)
 171 {
 172         pr_debug("remove_hash(), stripe %llu\n",
 173                 (unsigned long long)sh->sector);
 174
 175         hlist_del_init(&sh->hash);
 176 }
 177
 178 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
 179 {
 180         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 181
 182         pr_debug("insert_hash(), stripe %llu\n",
 183                 (unsigned long long)sh->sector);
 184
 185         CHECK_DEVLOCK();
 186         hlist_add_head(&sh->hash, hp);
 187 }
 188
 189
 190 /* find an idle stripe, make sure it is unhashed, and return it. */
 191 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
 192 {
 193         struct stripe_head *sh = NULL;
 194         struct list_head *first;
 195
 196         CHECK_DEVLOCK();
 197         if (list_empty(&conf->inactive_list))
 198                 goto out;
 199         first = conf->inactive_list.next;
 200         sh = list_entry(first, struct stripe_head, lru);
 201         list_del_init(first);
 202         remove_hash(sh);
 203         atomic_inc(&conf->active_stripes);
 204 out:
 205         return sh;
 206 }
 207
 208 static void shrink_buffers(struct stripe_head *sh, int num)
 209 {
 210         struct page *p;
 211         int i;
 212
 213         for (i=0; i<num ; i++) {
 214                 p = sh->dev[i].page;
 215                 if (!p)
 216                         continue;
 217                 sh->dev[i].page = NULL;
 218                 put_page(p);
 219         }
 220 }
 221
 222 static int grow_buffers(struct stripe_head *sh, int num)
 223 {
 224         int i;
 225
 226         for (i=0; i<num; i++) {
 227                 struct page *page;
 228
 229                 if (!(page = alloc_page(GFP_KERNEL))) {
 230                         return 1;
 231                 }
 232                 sh->dev[i].page = page;
 233         }
 234         return 0;
 235 }
 236
 237 static void raid5_build_block (struct stripe_head *sh, int i);
 238
 239 static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int disks)
 240 {
 241         raid5_conf_t *conf = sh->raid_conf;
 242         int i;
 243
 244         BUG_ON(atomic_read(&sh->count) != 0);
 245         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 246         BUG_ON(sh->ops.pending || sh->ops.ack || sh->ops.complete);
 247
 248         CHECK_DEVLOCK();
 249         pr_debug("init_stripe called, stripe %llu\n",
 250                 (unsigned long long)sh->sector);
 251
 252         remove_hash(sh);
 253
 254         sh->sector = sector;
 255         sh->pd_idx = pd_idx;
 256         sh->state = 0;
 257
 258         sh->disks = disks;
 259
 260         for (i = sh->disks; i--; ) {
 261                 struct r5dev *dev = &sh->dev[i];
 262
 263                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 264                     test_bit(R5_LOCKED, &dev->flags)) {
 265                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
 266                                (unsigned long long)sh->sector, i, dev->toread,
 267                                dev->read, dev->towrite, dev->written,
 268                                test_bit(R5_LOCKED, &dev->flags));
 269                         BUG();
 270                 }
 271                 dev->flags = 0;
 272                 raid5_build_block(sh, i);
 273         }
 274         insert_hash(conf, sh);
 275 }
 276
 277 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
 278 {
 279         struct stripe_head *sh;
 280         struct hlist_node *hn;
 281
 282         CHECK_DEVLOCK();
 283         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 284         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
 285                 if (sh->sector == sector && sh->disks == disks)
 286                         return sh;
 287         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 288         return NULL;
 289 }
 290
 291 static void unplug_slaves(mddev_t *mddev);
 292 static void raid5_unplug_device(struct request_queue *q);
 293
 294 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector, int disks,
 295                                              int pd_idx, int noblock)
 296 {
 297         struct stripe_head *sh;
 298
 299         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 300
 301         spin_lock_irq(&conf->device_lock);
 302
 303         do {
 304                 wait_event_lock_irq(conf->wait_for_stripe,
 305                                     conf->quiesce == 0,
 306                                     conf->device_lock, /* nothing */);
 307                 sh = __find_stripe(conf, sector, disks);
 308                 if (!sh) {
 309                         if (!conf->inactive_blocked)
 310                                 sh = get_free_stripe(conf);
 311                         if (noblock && sh == NULL)
 312                                 break;
 313                         if (!sh) {
 314                                 conf->inactive_blocked = 1;
 315                                 wait_event_lock_irq(conf->wait_for_stripe,
 316                                                     !list_empty(&conf->inactive_list) &&
 317                                                     (atomic_read(&conf->active_stripes)
 318                                                      < (conf->max_nr_stripes *3/4)
 319                                                      || !conf->inactive_blocked),
 320                                                     conf->device_lock,
 321                                                     raid5_unplug_device(conf->mddev->queue)
 322                                         );
 323                                 conf->inactive_blocked = 0;
 324                         } else
 325                                 init_stripe(sh, sector, pd_idx, disks);
 326                 } else {
 327                         if (atomic_read(&sh->count)) {
 328                           BUG_ON(!list_empty(&sh->lru));
 329                         } else {
 330                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 331                                         atomic_inc(&conf->active_stripes);
 332                                 if (list_empty(&sh->lru) &&
 333                                     !test_bit(STRIPE_EXPANDING, &sh->state))
 334                                         BUG();
 335                                 list_del_init(&sh->lru);
 336                         }
 337                 }
 338         } while (sh == NULL);
 339
 340         if (sh)
 341                 atomic_inc(&sh->count);
 342
 343         spin_unlock_irq(&conf->device_lock);
 344         return sh;
 345 }
 346
 347 /* test_and_ack_op() ensures that we only dequeue an operation once */
 348 #define test_and_ack_op(op, pend) \
 349 do {                                                    \
 350         if (test_bit(op, &sh->ops.pending) &&           \
 351                 !test_bit(op, &sh->ops.complete)) {     \
 352                 if (test_and_set_bit(op, &sh->ops.ack)) \
 353                         clear_bit(op, &pend);           \
 354                 else                                    \
 355                         ack++;                          \
 356         } else                                          \
 357                 clear_bit(op, &pend);                   \
 358 } while (0)
 359
 360 /* find new work to run, do not resubmit work that is already
 361  * in flight
 362  */
 363 static unsigned long get_stripe_work(struct stripe_head *sh)
 364 {
 365         unsigned long pending;
 366         int ack = 0;
 367
 368         pending = sh->ops.pending;
 369
 370         test_and_ack_op(STRIPE_OP_BIOFILL, pending);
 371         test_and_ack_op(STRIPE_OP_COMPUTE_BLK, pending);
 372         test_and_ack_op(STRIPE_OP_PREXOR, pending);
 373         test_and_ack_op(STRIPE_OP_BIODRAIN, pending);
 374         test_and_ack_op(STRIPE_OP_POSTXOR, pending);
 375         test_and_ack_op(STRIPE_OP_CHECK, pending);
 376         if (test_and_clear_bit(STRIPE_OP_IO, &sh->ops.pending))
 377                 ack++;
 378
 379         sh->ops.count -= ack;
 380         BUG_ON(sh->ops.count < 0);
 381
 382         return pending;
 383 }
 384
 385 static int
 386 raid5_end_read_request(struct bio *bi, unsigned int bytes_done, int error);
 387 static int
 388 raid5_end_write_request (struct bio *bi, unsigned int bytes_done, int error);
 389
 390 static void ops_run_io(struct stripe_head *sh)
 391 {
 392         raid5_conf_t *conf = sh->raid_conf;
 393         int i, disks = sh->disks;
 394
 395         might_sleep();
 396
 397         for (i = disks; i--; ) {
 398                 int rw;
 399                 struct bio *bi;
 400                 mdk_rdev_t *rdev;
 401                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
 402                         rw = WRITE;
 403                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
 404                         rw = READ;
 405                 else
 406                         continue;
 407
 408                 bi = &sh->dev[i].req;
 409
 410                 bi->bi_rw = rw;
 411                 if (rw == WRITE)
 412                         bi->bi_end_io = raid5_end_write_request;
 413                 else
 414                         bi->bi_end_io = raid5_end_read_request;
 415
 416                 rcu_read_lock();
 417                 rdev = rcu_dereference(conf->disks[i].rdev);
 418                 if (rdev && test_bit(Faulty, &rdev->flags))
 419                         rdev = NULL;
 420                 if (rdev)
 421                         atomic_inc(&rdev->nr_pending);
 422                 rcu_read_unlock();
 423
 424                 if (rdev) {
 425                         if (test_bit(STRIPE_SYNCING, &sh->state) ||
 426                                 test_bit(STRIPE_EXPAND_SOURCE, &sh->state) ||
 427                                 test_bit(STRIPE_EXPAND_READY, &sh->state))
 428                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
 429
 430                         bi->bi_bdev = rdev->bdev;
 431                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
 432                                 __FUNCTION__, (unsigned long long)sh->sector,
 433                                 bi->bi_rw, i);
 434                         atomic_inc(&sh->count);
 435                         bi->bi_sector = sh->sector + rdev->data_offset;
 436                         bi->bi_flags = 1 << BIO_UPTODATE;
 437                         bi->bi_vcnt = 1;
 438                         bi->bi_max_vecs = 1;
 439                         bi->bi_idx = 0;
 440                         bi->bi_io_vec = &sh->dev[i].vec;
 441                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
 442                         bi->bi_io_vec[0].bv_offset = 0;
 443                         bi->bi_size = STRIPE_SIZE;
 444                         bi->bi_next = NULL;
 445                         if (rw == WRITE &&
 446                             test_bit(R5_ReWrite, &sh->dev[i].flags))
 447                                 atomic_add(STRIPE_SECTORS,
 448                                         &rdev->corrected_errors);
 449                         generic_make_request(bi);
 450                 } else {
 451                         if (rw == WRITE)
 452                                 set_bit(STRIPE_DEGRADED, &sh->state);
 453                         pr_debug("skip op %ld on disc %d for sector %llu\n",
 454                                 bi->bi_rw, i, (unsigned long long)sh->sector);
 455                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 456                         set_bit(STRIPE_HANDLE, &sh->state);
 457                 }
 458         }
 459 }
 460
 461 static struct dma_async_tx_descriptor *
 462 async_copy_data(int frombio, struct bio *bio, struct page *page,
 463         sector_t sector, struct dma_async_tx_descriptor *tx)
 464 {
 465         struct bio_vec *bvl;
 466         struct page *bio_page;
 467         int i;
 468         int page_offset;
 469
 470         if (bio->bi_sector >= sector)
 471                 page_offset = (signed)(bio->bi_sector - sector) * 512;
 472         else
 473                 page_offset = (signed)(sector - bio->bi_sector) * -512;
 474         bio_for_each_segment(bvl, bio, i) {
 475                 int len = bio_iovec_idx(bio, i)->bv_len;
 476                 int clen;
 477                 int b_offset = 0;
 478
 479                 if (page_offset < 0) {
 480                         b_offset = -page_offset;
 481                         page_offset += b_offset;
 482                         len -= b_offset;
 483                 }
 484
 485                 if (len > 0 && page_offset + len > STRIPE_SIZE)
 486                         clen = STRIPE_SIZE - page_offset;
 487                 else
 488                         clen = len;
 489
 490                 if (clen > 0) {
 491                         b_offset += bio_iovec_idx(bio, i)->bv_offset;
 492                         bio_page = bio_iovec_idx(bio, i)->bv_page;
 493                         if (frombio)
 494                                 tx = async_memcpy(page, bio_page, page_offset,
 495                                         b_offset, clen,
 496                                         ASYNC_TX_DEP_ACK,
 497                                         tx, NULL, NULL);
 498                         else
 499                                 tx = async_memcpy(bio_page, page, b_offset,
 500                                         page_offset, clen,
 501                                         ASYNC_TX_DEP_ACK,
 502                                         tx, NULL, NULL);
 503                 }
 504                 if (clen < len) /* hit end of page */
 505                         break;
 506                 page_offset +=  len;
 507         }
 508
 509         return tx;
 510 }
 511
 512 static void ops_complete_biofill(void *stripe_head_ref)
 513 {
 514         struct stripe_head *sh = stripe_head_ref;
 515         struct bio *return_bi = NULL;
 516         raid5_conf_t *conf = sh->raid_conf;
 517         int i, more_to_read = 0;
 518
 519         pr_debug("%s: stripe %llu\n", __FUNCTION__,
 520                 (unsigned long long)sh->sector);
 521
 522         /* clear completed biofills */
 523         for (i = sh->disks; i--; ) {
 524                 struct r5dev *dev = &sh->dev[i];
 525                 /* check if this stripe has new incoming reads */
 526                 if (dev->toread)
 527                         more_to_read++;
 528
 529                 /* acknowledge completion of a biofill operation */
 530                 /* and check if we need to reply to a read request
 531                 */
 532                 if (test_bit(R5_Wantfill, &dev->flags) && !dev->toread) {
 533                         struct bio *rbi, *rbi2;
 534                         clear_bit(R5_Wantfill, &dev->flags);
 535
 536                         /* The access to dev->read is outside of the
 537                          * spin_lock_irq(&conf->device_lock), but is protected
 538                          * by the STRIPE_OP_BIOFILL pending bit
 539                          */
 540                         BUG_ON(!dev->read);
 541                         rbi = dev->read;
 542                         dev->read = NULL;
 543                         while (rbi && rbi->bi_sector <
 544                                 dev->sector + STRIPE_SECTORS) {
 545                                 rbi2 = r5_next_bio(rbi, dev->sector);
 546                                 spin_lock_irq(&conf->device_lock);
 547                                 if (--rbi->bi_phys_segments == 0) {
 548                                         rbi->bi_next = return_bi;
 549                                         return_bi = rbi;
 550                                 }
 551                                 spin_unlock_irq(&conf->device_lock);
 552                                 rbi = rbi2;
 553                         }
 554                 }
 555         }
 556         clear_bit(STRIPE_OP_BIOFILL, &sh->ops.ack);
 557         clear_bit(STRIPE_OP_BIOFILL, &sh->ops.pending);
 558
 559         return_io(return_bi);
 560
 561         if (more_to_read)
 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 int raid5_end_read_request(struct bio * bi, unsigned int bytes_done,
1117                                    int error)
1118 {
1119         struct stripe_head *sh = bi->bi_private;
1120         raid5_conf_t *conf = sh->raid_conf;
1121         int disks = sh->disks, i;
1122         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1123         char b[BDEVNAME_SIZE];
1124         mdk_rdev_t *rdev;
1125
1126         if (bi->bi_size)
1127                 return 1;
1128
1129         for (i=0 ; i<disks; i++)
1130                 if (bi == &sh->dev[i].req)
1131                         break;
1132
1133         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1134                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1135                 uptodate);
1136         if (i == disks) {
1137                 BUG();
1138                 return 0;
1139         }
1140
1141         if (uptodate) {
1142                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1143                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1144                         rdev = conf->disks[i].rdev;
1145                         printk(KERN_INFO "raid5:%s: read error corrected (%lu sectors at %llu on %s)\n",
1146                                mdname(conf->mddev), STRIPE_SECTORS,
1147                                (unsigned long long)sh->sector + rdev->data_offset,
1148                                bdevname(rdev->bdev, b));
1149                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1150                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1151                 }
1152                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1153                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1154         } else {
1155                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1156                 int retry = 0;
1157                 rdev = conf->disks[i].rdev;
1158
1159                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1160                 atomic_inc(&rdev->read_errors);
1161                 if (conf->mddev->degraded)
1162                         printk(KERN_WARNING "raid5:%s: read error not correctable (sector %llu on %s).\n",
1163                                mdname(conf->mddev),
1164                                (unsigned long long)sh->sector + rdev->data_offset,
1165                                bdn);
1166                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1167                         /* Oh, no!!! */
1168                         printk(KERN_WARNING "raid5:%s: read error NOT corrected!! (sector %llu on %s).\n",
1169                                mdname(conf->mddev),
1170                                (unsigned long long)sh->sector + rdev->data_offset,
1171                                bdn);
1172                 else if (atomic_read(&rdev->read_errors)
1173                          > conf->max_nr_stripes)
1174                         printk(KERN_WARNING
1175                                "raid5:%s: Too many read errors, failing device %s.\n",
1176                                mdname(conf->mddev), bdn);
1177                 else
1178                         retry = 1;
1179                 if (retry)
1180                         set_bit(R5_ReadError, &sh->dev[i].flags);
1181                 else {
1182                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1183                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1184                         md_error(conf->mddev, rdev);
1185                 }
1186         }
1187         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1188         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1189         set_bit(STRIPE_HANDLE, &sh->state);
1190         release_stripe(sh);
1191         return 0;
1192 }
1193
1194 static int raid5_end_write_request (struct bio *bi, unsigned int bytes_done,
1195                                     int error)
1196 {
1197         struct stripe_head *sh = bi->bi_private;
1198         raid5_conf_t *conf = sh->raid_conf;
1199         int disks = sh->disks, i;
1200         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1201
1202         if (bi->bi_size)
1203                 return 1;
1204
1205         for (i=0 ; i<disks; i++)
1206                 if (bi == &sh->dev[i].req)
1207                         break;
1208
1209         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1210                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1211                 uptodate);
1212         if (i == disks) {
1213                 BUG();
1214                 return 0;
1215         }
1216
1217         if (!uptodate)
1218                 md_error(conf->mddev, conf->disks[i].rdev);
1219
1220         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1221
1222         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1223         set_bit(STRIPE_HANDLE, &sh->state);
1224         release_stripe(sh);
1225         return 0;
1226 }
1227
1228
1229 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1230
1231 static void raid5_build_block (struct stripe_head *sh, int i)
1232 {
1233         struct r5dev *dev = &sh->dev[i];
1234
1235         bio_init(&dev->req);
1236         dev->req.bi_io_vec = &dev->vec;
1237         dev->req.bi_vcnt++;
1238         dev->req.bi_max_vecs++;
1239         dev->vec.bv_page = dev->page;
1240         dev->vec.bv_len = STRIPE_SIZE;
1241         dev->vec.bv_offset = 0;
1242
1243         dev->req.bi_sector = sh->sector;
1244         dev->req.bi_private = sh;
1245
1246         dev->flags = 0;
1247         dev->sector = compute_blocknr(sh, i);
1248 }
1249
1250 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1251 {
1252         char b[BDEVNAME_SIZE];
1253         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1254         pr_debug("raid5: error called\n");
1255
1256         if (!test_bit(Faulty, &rdev->flags)) {
1257                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1258                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1259                         unsigned long flags;
1260                         spin_lock_irqsave(&conf->device_lock, flags);
1261                         mddev->degraded++;
1262                         spin_unlock_irqrestore(&conf->device_lock, flags);
1263                         /*
1264                          * if recovery was running, make sure it aborts.
1265                          */
1266                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
1267                 }
1268                 set_bit(Faulty, &rdev->flags);
1269                 printk (KERN_ALERT
1270                         "raid5: Disk failure on %s, disabling device."
1271                         " Operation continuing on %d devices\n",
1272                         bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1273         }
1274 }
1275
1276 /*
1277  * Input: a 'big' sector number,
1278  * Output: index of the data and parity disk, and the sector # in them.
1279  */
1280 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
1281                         unsigned int data_disks, unsigned int * dd_idx,
1282                         unsigned int * pd_idx, raid5_conf_t *conf)
1283 {
1284         long stripe;
1285         unsigned long chunk_number;
1286         unsigned int chunk_offset;
1287         sector_t new_sector;
1288         int sectors_per_chunk = conf->chunk_size >> 9;
1289
1290         /* First compute the information on this sector */
1291
1292         /*
1293          * Compute the chunk number and the sector offset inside the chunk
1294          */
1295         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1296         chunk_number = r_sector;
1297         BUG_ON(r_sector != chunk_number);
1298
1299         /*
1300          * Compute the stripe number
1301          */
1302         stripe = chunk_number / data_disks;
1303
1304         /*
1305          * Compute the data disk and parity disk indexes inside the stripe
1306          */
1307         *dd_idx = chunk_number % data_disks;
1308
1309         /*
1310          * Select the parity disk based on the user selected algorithm.
1311          */
1312         switch(conf->level) {
1313         case 4:
1314                 *pd_idx = data_disks;
1315                 break;
1316         case 5:
1317                 switch (conf->algorithm) {
1318                 case ALGORITHM_LEFT_ASYMMETRIC:
1319                         *pd_idx = data_disks - stripe % raid_disks;
1320                         if (*dd_idx >= *pd_idx)
1321                                 (*dd_idx)++;
1322                         break;
1323                 case ALGORITHM_RIGHT_ASYMMETRIC:
1324                         *pd_idx = stripe % raid_disks;
1325                         if (*dd_idx >= *pd_idx)
1326                                 (*dd_idx)++;
1327                         break;
1328                 case ALGORITHM_LEFT_SYMMETRIC:
1329                         *pd_idx = data_disks - stripe % raid_disks;
1330                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1331                         break;
1332                 case ALGORITHM_RIGHT_SYMMETRIC:
1333                         *pd_idx = stripe % raid_disks;
1334                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1335                         break;
1336                 default:
1337                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1338                                 conf->algorithm);
1339                 }
1340                 break;
1341         case 6:
1342
1343                 /**** FIX THIS ****/
1344                 switch (conf->algorithm) {
1345                 case ALGORITHM_LEFT_ASYMMETRIC:
1346                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1347                         if (*pd_idx == raid_disks-1)
1348                                 (*dd_idx)++;    /* Q D D D P */
1349                         else if (*dd_idx >= *pd_idx)
1350                                 (*dd_idx) += 2; /* D D P Q D */
1351                         break;
1352                 case ALGORITHM_RIGHT_ASYMMETRIC:
1353                         *pd_idx = stripe % raid_disks;
1354                         if (*pd_idx == raid_disks-1)
1355                                 (*dd_idx)++;    /* Q D D D P */
1356                         else if (*dd_idx >= *pd_idx)
1357                                 (*dd_idx) += 2; /* D D P Q D */
1358                         break;
1359                 case ALGORITHM_LEFT_SYMMETRIC:
1360                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1361                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1362                         break;
1363                 case ALGORITHM_RIGHT_SYMMETRIC:
1364                         *pd_idx = stripe % raid_disks;
1365                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1366                         break;
1367                 default:
1368                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1369                                 conf->algorithm);
1370                 }
1371                 break;
1372         }
1373
1374         /*
1375          * Finally, compute the new sector number
1376          */
1377         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1378         return new_sector;
1379 }
1380
1381
1382 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1383 {
1384         raid5_conf_t *conf = sh->raid_conf;
1385         int raid_disks = sh->disks;
1386         int data_disks = raid_disks - conf->max_degraded;
1387         sector_t new_sector = sh->sector, check;
1388         int sectors_per_chunk = conf->chunk_size >> 9;
1389         sector_t stripe;
1390         int chunk_offset;
1391         int chunk_number, dummy1, dummy2, dd_idx = i;
1392         sector_t r_sector;
1393
1394
1395         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1396         stripe = new_sector;
1397         BUG_ON(new_sector != stripe);
1398
1399         if (i == sh->pd_idx)
1400                 return 0;
1401         switch(conf->level) {
1402         case 4: break;
1403         case 5:
1404                 switch (conf->algorithm) {
1405                 case ALGORITHM_LEFT_ASYMMETRIC:
1406                 case ALGORITHM_RIGHT_ASYMMETRIC:
1407                         if (i > sh->pd_idx)
1408                                 i--;
1409                         break;
1410                 case ALGORITHM_LEFT_SYMMETRIC:
1411                 case ALGORITHM_RIGHT_SYMMETRIC:
1412                         if (i < sh->pd_idx)
1413                                 i += raid_disks;
1414                         i -= (sh->pd_idx + 1);
1415                         break;
1416                 default:
1417                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1418                                conf->algorithm);
1419                 }
1420                 break;
1421         case 6:
1422                 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
1423                         return 0; /* It is the Q disk */
1424                 switch (conf->algorithm) {
1425                 case ALGORITHM_LEFT_ASYMMETRIC:
1426                 case ALGORITHM_RIGHT_ASYMMETRIC:
1427                         if (sh->pd_idx == raid_disks-1)
1428                                 i--;    /* Q D D D P */
1429                         else if (i > sh->pd_idx)
1430                                 i -= 2; /* D D P Q D */
1431                         break;
1432                 case ALGORITHM_LEFT_SYMMETRIC:
1433                 case ALGORITHM_RIGHT_SYMMETRIC:
1434                         if (sh->pd_idx == raid_disks-1)
1435                                 i--; /* Q D D D P */
1436                         else {
1437                                 /* D D P Q D */
1438                                 if (i < sh->pd_idx)
1439                                         i += raid_disks;
1440                                 i -= (sh->pd_idx + 2);
1441                         }
1442                         break;
1443                 default:
1444                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1445                                 conf->algorithm);
1446                 }
1447                 break;
1448         }
1449
1450         chunk_number = stripe * data_disks + i;
1451         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1452
1453         check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
1454         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
1455                 printk(KERN_ERR "compute_blocknr: map not correct\n");
1456                 return 0;
1457         }
1458         return r_sector;
1459 }
1460
1461
1462
1463 /*
1464  * Copy data between a page in the stripe cache, and one or more bion
1465  * The page could align with the middle of the bio, or there could be
1466  * several bion, each with several bio_vecs, which cover part of the page
1467  * Multiple bion are linked together on bi_next.  There may be extras
1468  * at the end of this list.  We ignore them.
1469  */
1470 static void copy_data(int frombio, struct bio *bio,
1471                      struct page *page,
1472                      sector_t sector)
1473 {
1474         char *pa = page_address(page);
1475         struct bio_vec *bvl;
1476         int i;
1477         int page_offset;
1478
1479         if (bio->bi_sector >= sector)
1480                 page_offset = (signed)(bio->bi_sector - sector) * 512;
1481         else
1482                 page_offset = (signed)(sector - bio->bi_sector) * -512;
1483         bio_for_each_segment(bvl, bio, i) {
1484                 int len = bio_iovec_idx(bio,i)->bv_len;
1485                 int clen;
1486                 int b_offset = 0;
1487
1488                 if (page_offset < 0) {
1489                         b_offset = -page_offset;
1490                         page_offset += b_offset;
1491                         len -= b_offset;
1492                 }
1493
1494                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1495                         clen = STRIPE_SIZE - page_offset;
1496                 else clen = len;
1497
1498                 if (clen > 0) {
1499                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1500                         if (frombio)
1501                                 memcpy(pa+page_offset, ba+b_offset, clen);
1502                         else
1503                                 memcpy(ba+b_offset, pa+page_offset, clen);
1504                         __bio_kunmap_atomic(ba, KM_USER0);
1505                 }
1506                 if (clen < len) /* hit end of page */
1507                         break;
1508                 page_offset +=  len;
1509         }
1510 }
1511
1512 #define check_xor()     do {                                              \
1513                                 if (count == MAX_XOR_BLOCKS) {            \
1514                                 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1515                                 count = 0;                                \
1516                            }                                              \
1517                         } while(0)
1518
1519 static void compute_parity6(struct stripe_head *sh, int method)
1520 {
1521         raid6_conf_t *conf = sh->raid_conf;
1522         int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1523         struct bio *chosen;
1524         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1525         void *ptrs[disks];
1526
1527         qd_idx = raid6_next_disk(pd_idx, disks);
1528         d0_idx = raid6_next_disk(qd_idx, disks);
1529
1530         pr_debug("compute_parity, stripe %llu, method %d\n",
1531                 (unsigned long long)sh->sector, method);
1532
1533         switch(method) {
1534         case READ_MODIFY_WRITE:
1535                 BUG();          /* READ_MODIFY_WRITE N/A for RAID-6 */
1536         case RECONSTRUCT_WRITE:
1537                 for (i= disks; i-- ;)
1538                         if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1539                                 chosen = sh->dev[i].towrite;
1540                                 sh->dev[i].towrite = NULL;
1541
1542                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1543                                         wake_up(&conf->wait_for_overlap);
1544
1545                                 BUG_ON(sh->dev[i].written);
1546                                 sh->dev[i].written = chosen;
1547                         }
1548                 break;
1549         case CHECK_PARITY:
1550                 BUG();          /* Not implemented yet */
1551         }
1552
1553         for (i = disks; i--;)
1554                 if (sh->dev[i].written) {
1555                         sector_t sector = sh->dev[i].sector;
1556                         struct bio *wbi = sh->dev[i].written;
1557                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1558                                 copy_data(1, wbi, sh->dev[i].page, sector);
1559                                 wbi = r5_next_bio(wbi, sector);
1560                         }
1561
1562                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1563                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
1564                 }
1565
1566 //      switch(method) {
1567 //      case RECONSTRUCT_WRITE:
1568 //      case CHECK_PARITY:
1569 //      case UPDATE_PARITY:
1570                 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1571                 /* FIX: Is this ordering of drives even remotely optimal? */
1572                 count = 0;
1573                 i = d0_idx;
1574                 do {
1575                         ptrs[count++] = page_address(sh->dev[i].page);
1576                         if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1577                                 printk("block %d/%d not uptodate on parity calc\n", i,count);
1578                         i = raid6_next_disk(i, disks);
1579                 } while ( i != d0_idx );
1580 //              break;
1581 //      }
1582
1583         raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1584
1585         switch(method) {
1586         case RECONSTRUCT_WRITE:
1587                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1588                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1589                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
1590                 set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
1591                 break;
1592         case UPDATE_PARITY:
1593                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1594                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1595                 break;
1596         }
1597 }
1598
1599
1600 /* Compute one missing block */
1601 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1602 {
1603         int i, count, disks = sh->disks;
1604         void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1605         int pd_idx = sh->pd_idx;
1606         int qd_idx = raid6_next_disk(pd_idx, disks);
1607
1608         pr_debug("compute_block_1, stripe %llu, idx %d\n",
1609                 (unsigned long long)sh->sector, dd_idx);
1610
1611         if ( dd_idx == qd_idx ) {
1612                 /* We're actually computing the Q drive */
1613                 compute_parity6(sh, UPDATE_PARITY);
1614         } else {
1615                 dest = page_address(sh->dev[dd_idx].page);
1616                 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1617                 count = 0;
1618                 for (i = disks ; i--; ) {
1619                         if (i == dd_idx || i == qd_idx)
1620                                 continue;
1621                         p = page_address(sh->dev[i].page);
1622                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1623                                 ptr[count++] = p;
1624                         else
1625                                 printk("compute_block() %d, stripe %llu, %d"
1626                                        " not present\n", dd_idx,
1627                                        (unsigned long long)sh->sector, i);
1628
1629                         check_xor();
1630                 }
1631                 if (count)
1632                         xor_blocks(count, STRIPE_SIZE, dest, ptr);
1633                 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1634                 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1635         }
1636 }
1637
1638 /* Compute two missing blocks */
1639 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1640 {
1641         int i, count, disks = sh->disks;
1642         int pd_idx = sh->pd_idx;
1643         int qd_idx = raid6_next_disk(pd_idx, disks);
1644         int d0_idx = raid6_next_disk(qd_idx, disks);
1645         int faila, failb;
1646
1647         /* faila and failb are disk numbers relative to d0_idx */
1648         /* pd_idx become disks-2 and qd_idx become disks-1 */
1649         faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1650         failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1651
1652         BUG_ON(faila == failb);
1653         if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1654
1655         pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1656                (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1657
1658         if ( failb == disks-1 ) {
1659                 /* Q disk is one of the missing disks */
1660                 if ( faila == disks-2 ) {
1661                         /* Missing P+Q, just recompute */
1662                         compute_parity6(sh, UPDATE_PARITY);
1663                         return;
1664                 } else {
1665                         /* We're missing D+Q; recompute D from P */
1666                         compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1667                         compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1668                         return;
1669                 }
1670         }
1671
1672         /* We're missing D+P or D+D; build pointer table */
1673         {
1674                 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1675                 void *ptrs[disks];
1676
1677                 count = 0;
1678                 i = d0_idx;
1679                 do {
1680                         ptrs[count++] = page_address(sh->dev[i].page);
1681                         i = raid6_next_disk(i, disks);
1682                         if (i != dd_idx1 && i != dd_idx2 &&
1683                             !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1684                                 printk("compute_2 with missing block %d/%d\n", count, i);
1685                 } while ( i != d0_idx );
1686
1687                 if ( failb == disks-2 ) {
1688                         /* We're missing D+P. */
1689                         raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1690                 } else {
1691                         /* We're missing D+D. */
1692                         raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1693                 }
1694
1695                 /* Both the above update both missing blocks */
1696                 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1697                 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1698         }
1699 }
1700
1701 static int
1702 handle_write_operations5(struct stripe_head *sh, int rcw, int expand)
1703 {
1704         int i, pd_idx = sh->pd_idx, disks = sh->disks;
1705         int locked = 0;
1706
1707         if (rcw) {
1708                 /* if we are not expanding this is a proper write request, and
1709                  * there will be bios with new data to be drained into the
1710                  * stripe cache
1711                  */
1712                 if (!expand) {
1713                         set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1714                         sh->ops.count++;
1715                 }
1716
1717                 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1718                 sh->ops.count++;
1719
1720                 for (i = disks; i--; ) {
1721                         struct r5dev *dev = &sh->dev[i];
1722
1723                         if (dev->towrite) {
1724                                 set_bit(R5_LOCKED, &dev->flags);
1725                                 if (!expand)
1726                                         clear_bit(R5_UPTODATE, &dev->flags);
1727                                 locked++;
1728                         }
1729                 }
1730         } else {
1731                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1732                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1733
1734                 set_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
1735                 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1736                 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1737
1738                 sh->ops.count += 3;
1739
1740                 for (i = disks; i--; ) {
1741                         struct r5dev *dev = &sh->dev[i];
1742                         if (i == pd_idx)
1743                                 continue;
1744
1745                         /* For a read-modify write there may be blocks that are
1746                          * locked for reading while others are ready to be
1747                          * written so we distinguish these blocks by the
1748                          * R5_Wantprexor bit
1749                          */
1750                         if (dev->towrite &&
1751                             (test_bit(R5_UPTODATE, &dev->flags) ||
1752                             test_bit(R5_Wantcompute, &dev->flags))) {
1753                                 set_bit(R5_Wantprexor, &dev->flags);
1754                                 set_bit(R5_LOCKED, &dev->flags);
1755                                 clear_bit(R5_UPTODATE, &dev->flags);
1756                                 locked++;
1757                         }
1758                 }
1759         }
1760
1761         /* keep the parity disk locked while asynchronous operations
1762          * are in flight
1763          */
1764         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1765         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1766         locked++;
1767
1768         pr_debug("%s: stripe %llu locked: %d pending: %lx\n",
1769                 __FUNCTION__, (unsigned long long)sh->sector,
1770                 locked, sh->ops.pending);
1771
1772         return locked;
1773 }
1774
1775 /*
1776  * Each stripe/dev can have one or more bion attached.
1777  * toread/towrite point to the first in a chain.
1778  * The bi_next chain must be in order.
1779  */
1780 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1781 {
1782         struct bio **bip;
1783         raid5_conf_t *conf = sh->raid_conf;
1784         int firstwrite=0;
1785
1786         pr_debug("adding bh b#%llu to stripe s#%llu\n",
1787                 (unsigned long long)bi->bi_sector,
1788                 (unsigned long long)sh->sector);
1789
1790
1791         spin_lock(&sh->lock);
1792         spin_lock_irq(&conf->device_lock);
1793         if (forwrite) {
1794                 bip = &sh->dev[dd_idx].towrite;
1795                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1796                         firstwrite = 1;
1797         } else
1798                 bip = &sh->dev[dd_idx].toread;
1799         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1800                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1801                         goto overlap;
1802                 bip = & (*bip)->bi_next;
1803         }
1804         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1805                 goto overlap;
1806
1807         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1808         if (*bip)
1809                 bi->bi_next = *bip;
1810         *bip = bi;
1811         bi->bi_phys_segments ++;
1812         spin_unlock_irq(&conf->device_lock);
1813         spin_unlock(&sh->lock);
1814
1815         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1816                 (unsigned long long)bi->bi_sector,
1817                 (unsigned long long)sh->sector, dd_idx);
1818
1819         if (conf->mddev->bitmap && firstwrite) {
1820                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1821                                   STRIPE_SECTORS, 0);
1822                 sh->bm_seq = conf->seq_flush+1;
1823                 set_bit(STRIPE_BIT_DELAY, &sh->state);
1824         }
1825
1826         if (forwrite) {
1827                 /* check if page is covered */
1828                 sector_t sector = sh->dev[dd_idx].sector;
1829                 for (bi=sh->dev[dd_idx].towrite;
1830                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1831                              bi && bi->bi_sector <= sector;
1832                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1833                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1834                                 sector = bi->bi_sector + (bi->bi_size>>9);
1835                 }
1836                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1837                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1838         }
1839         return 1;
1840
1841  overlap:
1842         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1843         spin_unlock_irq(&conf->device_lock);
1844         spin_unlock(&sh->lock);
1845         return 0;
1846 }
1847
1848 static void end_reshape(raid5_conf_t *conf);
1849
1850 static int page_is_zero(struct page *p)
1851 {
1852         char *a = page_address(p);
1853         return ((*(u32*)a) == 0 &&
1854                 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1855 }
1856
1857 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1858 {
1859         int sectors_per_chunk = conf->chunk_size >> 9;
1860         int pd_idx, dd_idx;
1861         int chunk_offset = sector_div(stripe, sectors_per_chunk);
1862
1863         raid5_compute_sector(stripe * (disks - conf->max_degraded)
1864                              *sectors_per_chunk + chunk_offset,
1865                              disks, disks - conf->max_degraded,
1866                              &dd_idx, &pd_idx, conf);
1867         return pd_idx;
1868 }
1869
1870 static void
1871 handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh,
1872                                 struct stripe_head_state *s, int disks,
1873                                 struct bio **return_bi)
1874 {
1875         int i;
1876         for (i = disks; i--; ) {
1877                 struct bio *bi;
1878                 int bitmap_end = 0;
1879
1880                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1881                         mdk_rdev_t *rdev;
1882                         rcu_read_lock();
1883                         rdev = rcu_dereference(conf->disks[i].rdev);
1884                         if (rdev && test_bit(In_sync, &rdev->flags))
1885                                 /* multiple read failures in one stripe */
1886                                 md_error(conf->mddev, rdev);
1887                         rcu_read_unlock();
1888                 }
1889                 spin_lock_irq(&conf->device_lock);
1890                 /* fail all writes first */
1891                 bi = sh->dev[i].towrite;
1892                 sh->dev[i].towrite = NULL;
1893                 if (bi) {
1894                         s->to_write--;
1895                         bitmap_end = 1;
1896                 }
1897
1898                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1899                         wake_up(&conf->wait_for_overlap);
1900
1901                 while (bi && bi->bi_sector <
1902                         sh->dev[i].sector + STRIPE_SECTORS) {
1903                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1904                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1905                         if (--bi->bi_phys_segments == 0) {
1906                                 md_write_end(conf->mddev);
1907                                 bi->bi_next = *return_bi;
1908                                 *return_bi = bi;
1909                         }
1910                         bi = nextbi;
1911                 }
1912                 /* and fail all 'written' */
1913                 bi = sh->dev[i].written;
1914                 sh->dev[i].written = NULL;
1915                 if (bi) bitmap_end = 1;
1916                 while (bi && bi->bi_sector <
1917                        sh->dev[i].sector + STRIPE_SECTORS) {
1918                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1919                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1920                         if (--bi->bi_phys_segments == 0) {
1921                                 md_write_end(conf->mddev);
1922                                 bi->bi_next = *return_bi;
1923                                 *return_bi = bi;
1924                         }
1925                         bi = bi2;
1926                 }
1927
1928                 /* fail any reads if this device is non-operational and
1929                  * the data has not reached the cache yet.
1930                  */
1931                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
1932                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1933                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
1934                         bi = sh->dev[i].toread;
1935                         sh->dev[i].toread = NULL;
1936                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1937                                 wake_up(&conf->wait_for_overlap);
1938                         if (bi) s->to_read--;
1939                         while (bi && bi->bi_sector <
1940                                sh->dev[i].sector + STRIPE_SECTORS) {
1941                                 struct bio *nextbi =
1942                                         r5_next_bio(bi, sh->dev[i].sector);
1943                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1944                                 if (--bi->bi_phys_segments == 0) {
1945                                         bi->bi_next = *return_bi;
1946                                         *return_bi = bi;
1947                                 }
1948                                 bi = nextbi;
1949                         }
1950                 }
1951                 spin_unlock_irq(&conf->device_lock);
1952                 if (bitmap_end)
1953                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1954                                         STRIPE_SECTORS, 0, 0);
1955         }
1956
1957 }
1958
1959 /* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks
1960  * to process
1961  */
1962 static int __handle_issuing_new_read_requests5(struct stripe_head *sh,
1963                         struct stripe_head_state *s, int disk_idx, int disks)
1964 {
1965         struct r5dev *dev = &sh->dev[disk_idx];
1966         struct r5dev *failed_dev = &sh->dev[s->failed_num];
1967
1968         /* don't schedule compute operations or reads on the parity block while
1969          * a check is in flight
1970          */
1971         if ((disk_idx == sh->pd_idx) &&
1972              test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
1973                 return ~0;
1974
1975         /* is the data in this block needed, and can we get it? */
1976         if (!test_bit(R5_LOCKED, &dev->flags) &&
1977             !test_bit(R5_UPTODATE, &dev->flags) && (dev->toread ||
1978             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1979              s->syncing || s->expanding || (s->failed &&
1980              (failed_dev->toread || (failed_dev->towrite &&
1981              !test_bit(R5_OVERWRITE, &failed_dev->flags)
1982              ))))) {
1983                 /* 1/ We would like to get this block, possibly by computing it,
1984                  * but we might not be able to.
1985                  *
1986                  * 2/ Since parity check operations potentially make the parity
1987                  * block !uptodate it will need to be refreshed before any
1988                  * compute operations on data disks are scheduled.
1989                  *
1990                  * 3/ We hold off parity block re-reads until check operations
1991                  * have quiesced.
1992                  */
1993                 if ((s->uptodate == disks - 1) &&
1994                     !test_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
1995                         set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
1996                         set_bit(R5_Wantcompute, &dev->flags);
1997                         sh->ops.target = disk_idx;
1998                         s->req_compute = 1;
1999                         sh->ops.count++;
2000                         /* Careful: from this point on 'uptodate' is in the eye
2001                          * of raid5_run_ops which services 'compute' operations
2002                          * before writes. R5_Wantcompute flags a block that will
2003                          * be R5_UPTODATE by the time it is needed for a
2004                          * subsequent operation.
2005                          */
2006                         s->uptodate++;
2007                         return 0; /* uptodate + compute == disks */
2008                 } else if ((s->uptodate < disks - 1) &&
2009                         test_bit(R5_Insync, &dev->flags)) {
2010                         /* Note: we hold off compute operations while checks are
2011                          * in flight, but we still prefer 'compute' over 'read'
2012                          * hence we only read if (uptodate < * disks-1)
2013                          */
2014                         set_bit(R5_LOCKED, &dev->flags);
2015                         set_bit(R5_Wantread, &dev->flags);
2016                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2017                                 sh->ops.count++;
2018                         s->locked++;
2019                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2020                                 s->syncing);
2021                 }
2022         }
2023
2024         return ~0;
2025 }
2026
2027 static void handle_issuing_new_read_requests5(struct stripe_head *sh,
2028                         struct stripe_head_state *s, int disks)
2029 {
2030         int i;
2031
2032         /* Clear completed compute operations.  Parity recovery
2033          * (STRIPE_OP_MOD_REPAIR_PD) implies a write-back which is handled
2034          * later on in this routine
2035          */
2036         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2037                 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2038                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2039                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2040                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2041         }
2042
2043         /* look for blocks to read/compute, skip this if a compute
2044          * is already in flight, or if the stripe contents are in the
2045          * midst of changing due to a write
2046          */
2047         if (!test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2048                 !test_bit(STRIPE_OP_PREXOR, &sh->ops.pending) &&
2049                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2050                 for (i = disks; i--; )
2051                         if (__handle_issuing_new_read_requests5(
2052                                 sh, s, i, disks) == 0)
2053                                 break;
2054         }
2055         set_bit(STRIPE_HANDLE, &sh->state);
2056 }
2057
2058 static void handle_issuing_new_read_requests6(struct stripe_head *sh,
2059                         struct stripe_head_state *s, struct r6_state *r6s,
2060                         int disks)
2061 {
2062         int i;
2063         for (i = disks; i--; ) {
2064                 struct r5dev *dev = &sh->dev[i];
2065                 if (!test_bit(R5_LOCKED, &dev->flags) &&
2066                     !test_bit(R5_UPTODATE, &dev->flags) &&
2067                     (dev->toread || (dev->towrite &&
2068                      !test_bit(R5_OVERWRITE, &dev->flags)) ||
2069                      s->syncing || s->expanding ||
2070                      (s->failed >= 1 &&
2071                       (sh->dev[r6s->failed_num[0]].toread ||
2072                        s->to_write)) ||
2073                      (s->failed >= 2 &&
2074                       (sh->dev[r6s->failed_num[1]].toread ||
2075                        s->to_write)))) {
2076                         /* we would like to get this block, possibly
2077                          * by computing it, but we might not be able to
2078                          */
2079                         if (s->uptodate == disks-1) {
2080                                 pr_debug("Computing stripe %llu block %d\n",
2081                                        (unsigned long long)sh->sector, i);
2082                                 compute_block_1(sh, i, 0);
2083                                 s->uptodate++;
2084                         } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2085                                 /* Computing 2-failure is *very* expensive; only
2086                                  * do it if failed >= 2
2087                                  */
2088                                 int other;
2089                                 for (other = disks; other--; ) {
2090                                         if (other == i)
2091                                                 continue;
2092                                         if (!test_bit(R5_UPTODATE,
2093                                               &sh->dev[other].flags))
2094                                                 break;
2095                                 }
2096                                 BUG_ON(other < 0);
2097                                 pr_debug("Computing stripe %llu blocks %d,%d\n",
2098                                        (unsigned long long)sh->sector,
2099                                        i, other);
2100                                 compute_block_2(sh, i, other);
2101                                 s->uptodate += 2;
2102                         } else if (test_bit(R5_Insync, &dev->flags)) {
2103                                 set_bit(R5_LOCKED, &dev->flags);
2104                                 set_bit(R5_Wantread, &dev->flags);
2105                                 s->locked++;
2106                                 pr_debug("Reading block %d (sync=%d)\n",
2107                                         i, s->syncing);
2108                         }
2109                 }
2110         }
2111         set_bit(STRIPE_HANDLE, &sh->state);
2112 }
2113
2114
2115 /* handle_completed_write_requests
2116  * any written block on an uptodate or failed drive can be returned.
2117  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2118  * never LOCKED, so we don't need to test 'failed' directly.
2119  */
2120 static void handle_completed_write_requests(raid5_conf_t *conf,
2121         struct stripe_head *sh, int disks, struct bio **return_bi)
2122 {
2123         int i;
2124         struct r5dev *dev;
2125
2126         for (i = disks; i--; )
2127                 if (sh->dev[i].written) {
2128                         dev = &sh->dev[i];
2129                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2130                                 test_bit(R5_UPTODATE, &dev->flags)) {
2131                                 /* We can return any write requests */
2132                                 struct bio *wbi, *wbi2;
2133                                 int bitmap_end = 0;
2134                                 pr_debug("Return write for disc %d\n", i);
2135                                 spin_lock_irq(&conf->device_lock);
2136                                 wbi = dev->written;
2137                                 dev->written = NULL;
2138                                 while (wbi && wbi->bi_sector <
2139                                         dev->sector + STRIPE_SECTORS) {
2140                                         wbi2 = r5_next_bio(wbi, dev->sector);
2141                                         if (--wbi->bi_phys_segments == 0) {
2142                                                 md_write_end(conf->mddev);
2143                                                 wbi->bi_next = *return_bi;
2144                                                 *return_bi = wbi;
2145                                         }
2146                                         wbi = wbi2;
2147                                 }
2148                                 if (dev->towrite == NULL)
2149                                         bitmap_end = 1;
2150                                 spin_unlock_irq(&conf->device_lock);
2151                                 if (bitmap_end)
2152                                         bitmap_endwrite(conf->mddev->bitmap,
2153                                                         sh->sector,
2154                                                         STRIPE_SECTORS,
2155                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2156                                                         0);
2157                         }
2158                 }
2159 }
2160
2161 static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
2162                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2163 {
2164         int rmw = 0, rcw = 0, i;
2165         for (i = disks; i--; ) {
2166                 /* would I have to read this buffer for read_modify_write */
2167                 struct r5dev *dev = &sh->dev[i];
2168                 if ((dev->towrite || i == sh->pd_idx) &&
2169                     !test_bit(R5_LOCKED, &dev->flags) &&
2170                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2171                       test_bit(R5_Wantcompute, &dev->flags))) {
2172                         if (test_bit(R5_Insync, &dev->flags))
2173                                 rmw++;
2174                         else
2175                                 rmw += 2*disks;  /* cannot read it */
2176                 }
2177                 /* Would I have to read this buffer for reconstruct_write */
2178                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2179                     !test_bit(R5_LOCKED, &dev->flags) &&
2180                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2181                     test_bit(R5_Wantcompute, &dev->flags))) {
2182                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2183                         else
2184                                 rcw += 2*disks;
2185                 }
2186         }
2187         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2188                 (unsigned long long)sh->sector, rmw, rcw);
2189         set_bit(STRIPE_HANDLE, &sh->state);
2190         if (rmw < rcw && rmw > 0)
2191                 /* prefer read-modify-write, but need to get some data */
2192                 for (i = disks; i--; ) {
2193                         struct r5dev *dev = &sh->dev[i];
2194                         if ((dev->towrite || i == sh->pd_idx) &&
2195                             !test_bit(R5_LOCKED, &dev->flags) &&
2196                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2197                             test_bit(R5_Wantcompute, &dev->flags)) &&
2198                             test_bit(R5_Insync, &dev->flags)) {
2199                                 if (
2200                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2201                                         pr_debug("Read_old block "
2202                                                 "%d for r-m-w\n", i);
2203                                         set_bit(R5_LOCKED, &dev->flags);
2204                                         set_bit(R5_Wantread, &dev->flags);
2205                                         if (!test_and_set_bit(
2206                                                 STRIPE_OP_IO, &sh->ops.pending))
2207                                                 sh->ops.count++;
2208                                         s->locked++;
2209                                 } else {
2210                                         set_bit(STRIPE_DELAYED, &sh->state);
2211                                         set_bit(STRIPE_HANDLE, &sh->state);
2212                                 }
2213                         }
2214                 }
2215         if (rcw <= rmw && rcw > 0)
2216                 /* want reconstruct write, but need to get some data */
2217                 for (i = disks; i--; ) {
2218                         struct r5dev *dev = &sh->dev[i];
2219                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2220                             i != sh->pd_idx &&
2221                             !test_bit(R5_LOCKED, &dev->flags) &&
2222                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2223                             test_bit(R5_Wantcompute, &dev->flags)) &&
2224                             test_bit(R5_Insync, &dev->flags)) {
2225                                 if (
2226                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2227                                         pr_debug("Read_old block "
2228                                                 "%d for Reconstruct\n", i);
2229                                         set_bit(R5_LOCKED, &dev->flags);
2230                                         set_bit(R5_Wantread, &dev->flags);
2231                                         if (!test_and_set_bit(
2232                                                 STRIPE_OP_IO, &sh->ops.pending))
2233                                                 sh->ops.count++;
2234                                         s->locked++;
2235                                 } else {
2236                                         set_bit(STRIPE_DELAYED, &sh->state);
2237                                         set_bit(STRIPE_HANDLE, &sh->state);
2238                                 }
2239                         }
2240                 }
2241         /* now if nothing is locked, and if we have enough data,
2242          * we can start a write request
2243          */
2244         /* since handle_stripe can be called at any time we need to handle the
2245          * case where a compute block operation has been submitted and then a
2246          * subsequent call wants to start a write request.  raid5_run_ops only
2247          * handles the case where compute block and postxor are requested
2248          * simultaneously.  If this is not the case then new writes need to be
2249          * held off until the compute completes.
2250          */
2251         if ((s->req_compute ||
2252             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) &&
2253                 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2254                 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2255                 s->locked += handle_write_operations5(sh, rcw == 0, 0);
2256 }
2257
2258 static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
2259                 struct stripe_head *sh, struct stripe_head_state *s,
2260                 struct r6_state *r6s, int disks)
2261 {
2262         int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2263         int qd_idx = r6s->qd_idx;
2264         for (i = disks; i--; ) {
2265                 struct r5dev *dev = &sh->dev[i];
2266                 /* Would I have to read this buffer for reconstruct_write */
2267                 if (!test_bit(R5_OVERWRITE, &dev->flags)
2268                     && i != pd_idx && i != qd_idx
2269                     && (!test_bit(R5_LOCKED, &dev->flags)
2270                             ) &&
2271                     !test_bit(R5_UPTODATE, &dev->flags)) {
2272                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2273                         else {
2274                                 pr_debug("raid6: must_compute: "
2275                                         "disk %d flags=%#lx\n", i, dev->flags);
2276                                 must_compute++;
2277                         }
2278                 }
2279         }
2280         pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2281                (unsigned long long)sh->sector, rcw, must_compute);
2282         set_bit(STRIPE_HANDLE, &sh->state);
2283
2284         if (rcw > 0)
2285                 /* want reconstruct write, but need to get some data */
2286                 for (i = disks; i--; ) {
2287                         struct r5dev *dev = &sh->dev[i];
2288                         if (!test_bit(R5_OVERWRITE, &dev->flags)
2289                             && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2290                             && !test_bit(R5_LOCKED, &dev->flags) &&
2291                             !test_bit(R5_UPTODATE, &dev->flags) &&
2292                             test_bit(R5_Insync, &dev->flags)) {
2293                                 if (
2294                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2295                                         pr_debug("Read_old stripe %llu "
2296                                                 "block %d for Reconstruct\n",
2297                                              (unsigned long long)sh->sector, i);
2298                                         set_bit(R5_LOCKED, &dev->flags);
2299                                         set_bit(R5_Wantread, &dev->flags);
2300                                         s->locked++;
2301                                 } else {
2302                                         pr_debug("Request delayed stripe %llu "
2303                                                 "block %d for Reconstruct\n",
2304                                              (unsigned long long)sh->sector, i);
2305                                         set_bit(STRIPE_DELAYED, &sh->state);
2306                                         set_bit(STRIPE_HANDLE, &sh->state);
2307                                 }
2308                         }
2309                 }
2310         /* now if nothing is locked, and if we have enough data, we can start a
2311          * write request
2312          */
2313         if (s->locked == 0 && rcw == 0 &&
2314             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2315                 if (must_compute > 0) {
2316                         /* We have failed blocks and need to compute them */
2317                         switch (s->failed) {
2318                         case 0:
2319                                 BUG();
2320                         case 1:
2321                                 compute_block_1(sh, r6s->failed_num[0], 0);
2322                                 break;
2323                         case 2:
2324                                 compute_block_2(sh, r6s->failed_num[0],
2325                                                 r6s->failed_num[1]);
2326                                 break;
2327                         default: /* This request should have been failed? */
2328                                 BUG();
2329                         }
2330                 }
2331
2332                 pr_debug("Computing parity for stripe %llu\n",
2333                         (unsigned long long)sh->sector);
2334                 compute_parity6(sh, RECONSTRUCT_WRITE);
2335                 /* now every locked buffer is ready to be written */
2336                 for (i = disks; i--; )
2337                         if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2338                                 pr_debug("Writing stripe %llu block %d\n",
2339                                        (unsigned long long)sh->sector, i);
2340                                 s->locked++;
2341                                 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2342                         }
2343                 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2344                 set_bit(STRIPE_INSYNC, &sh->state);
2345
2346                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2347                         atomic_dec(&conf->preread_active_stripes);
2348                         if (atomic_read(&conf->preread_active_stripes) <
2349                             IO_THRESHOLD)
2350                                 md_wakeup_thread(conf->mddev->thread);
2351                 }
2352         }
2353 }
2354
2355 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2356                                 struct stripe_head_state *s, int disks)
2357 {
2358         set_bit(STRIPE_HANDLE, &sh->state);
2359         /* Take one of the following actions:
2360          * 1/ start a check parity operation if (uptodate == disks)
2361          * 2/ finish a check parity operation and act on the result
2362          * 3/ skip to the writeback section if we previously
2363          *    initiated a recovery operation
2364          */
2365         if (s->failed == 0 &&
2366             !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2367                 if (!test_and_set_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
2368                         BUG_ON(s->uptodate != disks);
2369                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2370                         sh->ops.count++;
2371                         s->uptodate--;
2372                 } else if (
2373                        test_and_clear_bit(STRIPE_OP_CHECK, &sh->ops.complete)) {
2374                         clear_bit(STRIPE_OP_CHECK, &sh->ops.ack);
2375                         clear_bit(STRIPE_OP_CHECK, &sh->ops.pending);
2376
2377                         if (sh->ops.zero_sum_result == 0)
2378                                 /* parity is correct (on disc,
2379                                  * not in buffer any more)
2380                                  */
2381                                 set_bit(STRIPE_INSYNC, &sh->state);
2382                         else {
2383                                 conf->mddev->resync_mismatches +=
2384                                         STRIPE_SECTORS;
2385                                 if (test_bit(
2386                                      MD_RECOVERY_CHECK, &conf->mddev->recovery))
2387                                         /* don't try to repair!! */
2388                                         set_bit(STRIPE_INSYNC, &sh->state);
2389                                 else {
2390                                         set_bit(STRIPE_OP_COMPUTE_BLK,
2391                                                 &sh->ops.pending);
2392                                         set_bit(STRIPE_OP_MOD_REPAIR_PD,
2393                                                 &sh->ops.pending);
2394                                         set_bit(R5_Wantcompute,
2395                                                 &sh->dev[sh->pd_idx].flags);
2396                                         sh->ops.target = sh->pd_idx;
2397                                         sh->ops.count++;
2398                                         s->uptodate++;
2399                                 }
2400                         }
2401                 }
2402         }
2403
2404         /* check if we can clear a parity disk reconstruct */
2405         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2406                 test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2407
2408                 clear_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending);
2409                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2410                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2411                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2412         }
2413
2414         /* Wait for check parity and compute block operations to complete
2415          * before write-back
2416          */
2417         if (!test_bit(STRIPE_INSYNC, &sh->state) &&
2418                 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending) &&
2419                 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) {
2420                 struct r5dev *dev;
2421                 /* either failed parity check, or recovery is happening */
2422                 if (s->failed == 0)
2423                         s->failed_num = sh->pd_idx;
2424                 dev = &sh->dev[s->failed_num];
2425                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2426                 BUG_ON(s->uptodate != disks);
2427
2428                 set_bit(R5_LOCKED, &dev->flags);
2429                 set_bit(R5_Wantwrite, &dev->flags);
2430                 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2431                         sh->ops.count++;
2432
2433                 clear_bit(STRIPE_DEGRADED, &sh->state);
2434                 s->locked++;
2435                 set_bit(STRIPE_INSYNC, &sh->state);
2436         }
2437 }
2438
2439
2440 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2441                                 struct stripe_head_state *s,
2442                                 struct r6_state *r6s, struct page *tmp_page,
2443                                 int disks)
2444 {
2445         int update_p = 0, update_q = 0;
2446         struct r5dev *dev;
2447         int pd_idx = sh->pd_idx;
2448         int qd_idx = r6s->qd_idx;
2449
2450         set_bit(STRIPE_HANDLE, &sh->state);
2451
2452         BUG_ON(s->failed > 2);
2453         BUG_ON(s->uptodate < disks);
2454         /* Want to check and possibly repair P and Q.
2455          * However there could be one 'failed' device, in which
2456          * case we can only check one of them, possibly using the
2457          * other to generate missing data
2458          */
2459
2460         /* If !tmp_page, we cannot do the calculations,
2461          * but as we have set STRIPE_HANDLE, we will soon be called
2462          * by stripe_handle with a tmp_page - just wait until then.
2463          */
2464         if (tmp_page) {
2465                 if (s->failed == r6s->q_failed) {
2466                         /* The only possible failed device holds 'Q', so it
2467                          * makes sense to check P (If anything else were failed,
2468                          * we would have used P to recreate it).
2469                          */
2470                         compute_block_1(sh, pd_idx, 1);
2471                         if (!page_is_zero(sh->dev[pd_idx].page)) {
2472                                 compute_block_1(sh, pd_idx, 0);
2473                                 update_p = 1;
2474                         }
2475                 }
2476                 if (!r6s->q_failed && s->failed < 2) {
2477                         /* q is not failed, and we didn't use it to generate
2478                          * anything, so it makes sense to check it
2479                          */
2480                         memcpy(page_address(tmp_page),
2481                                page_address(sh->dev[qd_idx].page),
2482                                STRIPE_SIZE);
2483                         compute_parity6(sh, UPDATE_PARITY);
2484                         if (memcmp(page_address(tmp_page),
2485                                    page_address(sh->dev[qd_idx].page),
2486                                    STRIPE_SIZE) != 0) {
2487                                 clear_bit(STRIPE_INSYNC, &sh->state);
2488                                 update_q = 1;
2489                         }
2490                 }
2491                 if (update_p || update_q) {
2492                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2493                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2494                                 /* don't try to repair!! */
2495                                 update_p = update_q = 0;
2496                 }
2497
2498                 /* now write out any block on a failed drive,
2499                  * or P or Q if they need it
2500                  */
2501
2502                 if (s->failed == 2) {
2503                         dev = &sh->dev[r6s->failed_num[1]];
2504                         s->locked++;
2505                         set_bit(R5_LOCKED, &dev->flags);
2506                         set_bit(R5_Wantwrite, &dev->flags);
2507                 }
2508                 if (s->failed >= 1) {
2509                         dev = &sh->dev[r6s->failed_num[0]];
2510                         s->locked++;
2511                         set_bit(R5_LOCKED, &dev->flags);
2512                         set_bit(R5_Wantwrite, &dev->flags);
2513                 }
2514
2515                 if (update_p) {
2516                         dev = &sh->dev[pd_idx];
2517                         s->locked++;
2518                         set_bit(R5_LOCKED, &dev->flags);
2519                         set_bit(R5_Wantwrite, &dev->flags);
2520                 }
2521                 if (update_q) {
2522                         dev = &sh->dev[qd_idx];
2523                         s->locked++;
2524                         set_bit(R5_LOCKED, &dev->flags);
2525                         set_bit(R5_Wantwrite, &dev->flags);
2526                 }
2527                 clear_bit(STRIPE_DEGRADED, &sh->state);
2528
2529                 set_bit(STRIPE_INSYNC, &sh->state);
2530         }
2531 }
2532
2533 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2534                                 struct r6_state *r6s)
2535 {
2536         int i;
2537
2538         /* We have read all the blocks in this stripe and now we need to
2539          * copy some of them into a target stripe for expand.
2540          */
2541         struct dma_async_tx_descriptor *tx = NULL;
2542         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2543         for (i = 0; i < sh->disks; i++)
2544                 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
2545                         int dd_idx, pd_idx, j;
2546                         struct stripe_head *sh2;
2547
2548                         sector_t bn = compute_blocknr(sh, i);
2549                         sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2550                                                 conf->raid_disks -
2551                                                 conf->max_degraded, &dd_idx,
2552                                                 &pd_idx, conf);
2553                         sh2 = get_active_stripe(conf, s, conf->raid_disks,
2554                                                 pd_idx, 1);
2555                         if (sh2 == NULL)
2556                                 /* so far only the early blocks of this stripe
2557                                  * have been requested.  When later blocks
2558                                  * get requested, we will try again
2559                                  */
2560                                 continue;
2561                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2562                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2563                                 /* must have already done this block */
2564                                 release_stripe(sh2);
2565                                 continue;
2566                         }
2567
2568                         /* place all the copies on one channel */
2569                         tx = async_memcpy(sh2->dev[dd_idx].page,
2570                                 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2571                                 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2572
2573                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2574                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2575                         for (j = 0; j < conf->raid_disks; j++)
2576                                 if (j != sh2->pd_idx &&
2577                                     (!r6s || j != raid6_next_disk(sh2->pd_idx,
2578                                                                  sh2->disks)) &&
2579                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2580                                         break;
2581                         if (j == conf->raid_disks) {
2582                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2583                                 set_bit(STRIPE_HANDLE, &sh2->state);
2584                         }
2585                         release_stripe(sh2);
2586
2587                 }
2588         /* done submitting copies, wait for them to complete */
2589         if (tx) {
2590                 async_tx_ack(tx);
2591                 dma_wait_for_async_tx(tx);
2592         }
2593 }
2594
2595 /*
2596  * handle_stripe - do things to a stripe.
2597  *
2598  * We lock the stripe and then examine the state of various bits
2599  * to see what needs to be done.
2600  * Possible results:
2601  *    return some read request which now have data
2602  *    return some write requests which are safely on disc
2603  *    schedule a read on some buffers
2604  *    schedule a write of some buffers
2605  *    return confirmation of parity correctness
2606  *
2607  * buffers are taken off read_list or write_list, and bh_cache buffers
2608  * get BH_Lock set before the stripe lock is released.
2609  *
2610  */
2611
2612 static void handle_stripe5(struct stripe_head *sh)
2613 {
2614         raid5_conf_t *conf = sh->raid_conf;
2615         int disks = sh->disks, i;
2616         struct bio *return_bi = NULL;
2617         struct stripe_head_state s;
2618         struct r5dev *dev;
2619         unsigned long pending = 0;
2620
2621         memset(&s, 0, sizeof(s));
2622         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2623                 "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
2624                 atomic_read(&sh->count), sh->pd_idx,
2625                 sh->ops.pending, sh->ops.ack, sh->ops.complete);
2626
2627         spin_lock(&sh->lock);
2628         clear_bit(STRIPE_HANDLE, &sh->state);
2629         clear_bit(STRIPE_DELAYED, &sh->state);
2630
2631         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2632         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2633         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2634         /* Now to look around and see what can be done */
2635
2636         rcu_read_lock();
2637         for (i=disks; i--; ) {
2638                 mdk_rdev_t *rdev;
2639                 struct r5dev *dev = &sh->dev[i];
2640                 clear_bit(R5_Insync, &dev->flags);
2641
2642                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2643                         "written %p\n", i, dev->flags, dev->toread, dev->read,
2644                         dev->towrite, dev->written);
2645
2646                 /* maybe we can request a biofill operation
2647                  *
2648                  * new wantfill requests are only permitted while
2649                  * STRIPE_OP_BIOFILL is clear
2650                  */
2651                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2652                         !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2653                         set_bit(R5_Wantfill, &dev->flags);
2654
2655                 /* now count some things */
2656                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2657                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2658                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2659
2660                 if (test_bit(R5_Wantfill, &dev->flags))
2661                         s.to_fill++;
2662                 else if (dev->toread)
2663                         s.to_read++;
2664                 if (dev->towrite) {
2665                         s.to_write++;
2666                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2667                                 s.non_overwrite++;
2668                 }
2669                 if (dev->written)
2670                         s.written++;
2671                 rdev = rcu_dereference(conf->disks[i].rdev);
2672                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2673                         /* The ReadError flag will just be confusing now */
2674                         clear_bit(R5_ReadError, &dev->flags);
2675                         clear_bit(R5_ReWrite, &dev->flags);
2676                 }
2677                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2678                     || test_bit(R5_ReadError, &dev->flags)) {
2679                         s.failed++;
2680                         s.failed_num = i;
2681                 } else
2682                         set_bit(R5_Insync, &dev->flags);
2683         }
2684         rcu_read_unlock();
2685
2686         if (s.to_fill && !test_and_set_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2687                 sh->ops.count++;
2688
2689         pr_debug("locked=%d uptodate=%d to_read=%d"
2690                 " to_write=%d failed=%d failed_num=%d\n",
2691                 s.locked, s.uptodate, s.to_read, s.to_write,
2692                 s.failed, s.failed_num);
2693         /* check if the array has lost two devices and, if so, some requests might
2694          * need to be failed
2695          */
2696         if (s.failed > 1 && s.to_read+s.to_write+s.written)
2697                 handle_requests_to_failed_array(conf, sh, &s, disks,
2698                                                 &return_bi);
2699         if (s.failed > 1 && s.syncing) {
2700                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2701                 clear_bit(STRIPE_SYNCING, &sh->state);
2702                 s.syncing = 0;
2703         }
2704
2705         /* might be able to return some write requests if the parity block
2706          * is safe, or on a failed drive
2707          */
2708         dev = &sh->dev[sh->pd_idx];
2709         if ( s.written &&
2710              ((test_bit(R5_Insync, &dev->flags) &&
2711                !test_bit(R5_LOCKED, &dev->flags) &&
2712                test_bit(R5_UPTODATE, &dev->flags)) ||
2713                (s.failed == 1 && s.failed_num == sh->pd_idx)))
2714                 handle_completed_write_requests(conf, sh, disks, &return_bi);
2715
2716         /* Now we might consider reading some blocks, either to check/generate
2717          * parity, or to satisfy requests
2718          * or to load a block that is being partially written.
2719          */
2720         if (s.to_read || s.non_overwrite ||
2721             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding ||
2722             test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2723                 handle_issuing_new_read_requests5(sh, &s, disks);
2724
2725         /* Now we check to see if any write operations have recently
2726          * completed
2727          */
2728
2729         /* leave prexor set until postxor is done, allows us to distinguish
2730          * a rmw from a rcw during biodrain
2731          */
2732         if (test_bit(STRIPE_OP_PREXOR, &sh->ops.complete) &&
2733                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2734
2735                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
2736                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.ack);
2737                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
2738
2739                 for (i = disks; i--; )
2740                         clear_bit(R5_Wantprexor, &sh->dev[i].flags);
2741         }
2742
2743         /* if only POSTXOR is set then this is an 'expand' postxor */
2744         if (test_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete) &&
2745                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2746
2747                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
2748                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.ack);
2749                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
2750
2751                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2752                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2753                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2754
2755                 /* All the 'written' buffers and the parity block are ready to
2756                  * be written back to disk
2757                  */
2758                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2759                 for (i = disks; i--; ) {
2760                         dev = &sh->dev[i];
2761                         if (test_bit(R5_LOCKED, &dev->flags) &&
2762                                 (i == sh->pd_idx || dev->written)) {
2763                                 pr_debug("Writing block %d\n", i);
2764                                 set_bit(R5_Wantwrite, &dev->flags);
2765                                 if (!test_and_set_bit(
2766                                     STRIPE_OP_IO, &sh->ops.pending))
2767                                         sh->ops.count++;
2768                                 if (!test_bit(R5_Insync, &dev->flags) ||
2769                                     (i == sh->pd_idx && s.failed == 0))
2770                                         set_bit(STRIPE_INSYNC, &sh->state);
2771                         }
2772                 }
2773                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2774                         atomic_dec(&conf->preread_active_stripes);
2775                         if (atomic_read(&conf->preread_active_stripes) <
2776                                 IO_THRESHOLD)
2777                                 md_wakeup_thread(conf->mddev->thread);
2778                 }
2779         }
2780
2781         /* Now to consider new write requests and what else, if anything
2782          * should be read.  We do not handle new writes when:
2783          * 1/ A 'write' operation (copy+xor) is already in flight.
2784          * 2/ A 'check' operation is in flight, as it may clobber the parity
2785          *    block.
2786          */
2787         if (s.to_write && !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending) &&
2788                           !test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
2789                 handle_issuing_new_write_requests5(conf, sh, &s, disks);
2790
2791         /* maybe we need to check and possibly fix the parity for this stripe
2792          * Any reads will already have been scheduled, so we just see if enough
2793          * data is available.  The parity check is held off while parity
2794          * dependent operations are in flight.
2795          */
2796         if ((s.syncing && s.locked == 0 &&
2797              !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2798              !test_bit(STRIPE_INSYNC, &sh->state)) ||
2799               test_bit(STRIPE_OP_CHECK, &sh->ops.pending) ||
2800               test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending))
2801                 handle_parity_checks5(conf, sh, &s, disks);
2802
2803         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2804                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2805                 clear_bit(STRIPE_SYNCING, &sh->state);
2806         }
2807
2808         /* If the failed drive is just a ReadError, then we might need to progress
2809          * the repair/check process
2810          */
2811         if (s.failed == 1 && !conf->mddev->ro &&
2812             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2813             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2814             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2815                 ) {
2816                 dev = &sh->dev[s.failed_num];
2817                 if (!test_bit(R5_ReWrite, &dev->flags)) {
2818                         set_bit(R5_Wantwrite, &dev->flags);
2819                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2820                                 sh->ops.count++;
2821                         set_bit(R5_ReWrite, &dev->flags);
2822                         set_bit(R5_LOCKED, &dev->flags);
2823                         s.locked++;
2824                 } else {
2825                         /* let's read it back */
2826                         set_bit(R5_Wantread, &dev->flags);
2827                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2828                                 sh->ops.count++;
2829                         set_bit(R5_LOCKED, &dev->flags);
2830                         s.locked++;
2831                 }
2832         }
2833
2834         /* Finish postxor operations initiated by the expansion
2835          * process
2836          */
2837         if (test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete) &&
2838                 !test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending)) {
2839
2840                 clear_bit(STRIPE_EXPANDING, &sh->state);
2841
2842                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2843                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2844                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2845
2846                 for (i = conf->raid_disks; i--; ) {
2847                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
2848                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2849                                 sh->ops.count++;
2850                 }
2851         }
2852
2853         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2854                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2855                 /* Need to write out all blocks after computing parity */
2856                 sh->disks = conf->raid_disks;
2857                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2858                         conf->raid_disks);
2859                 s.locked += handle_write_operations5(sh, 1, 1);
2860         } else if (s.expanded &&
2861                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2862                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2863                 atomic_dec(&conf->reshape_stripes);
2864                 wake_up(&conf->wait_for_overlap);
2865                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2866         }
2867
2868         if (s.expanding && s.locked == 0)
2869                 handle_stripe_expansion(conf, sh, NULL);
2870
2871         if (sh->ops.count)
2872                 pending = get_stripe_work(sh);
2873
2874         spin_unlock(&sh->lock);
2875
2876         if (pending)
2877                 raid5_run_ops(sh, pending);
2878
2879         return_io(return_bi);
2880
2881 }
2882
2883 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2884 {
2885         raid6_conf_t *conf = sh->raid_conf;
2886         int disks = sh->disks;
2887         struct bio *return_bi = NULL;
2888         int i, pd_idx = sh->pd_idx;
2889         struct stripe_head_state s;
2890         struct r6_state r6s;
2891         struct r5dev *dev, *pdev, *qdev;
2892
2893         r6s.qd_idx = raid6_next_disk(pd_idx, disks);
2894         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2895                 "pd_idx=%d, qd_idx=%d\n",
2896                (unsigned long long)sh->sector, sh->state,
2897                atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2898         memset(&s, 0, sizeof(s));
2899
2900         spin_lock(&sh->lock);
2901         clear_bit(STRIPE_HANDLE, &sh->state);
2902         clear_bit(STRIPE_DELAYED, &sh->state);
2903
2904         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2905         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2906         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2907         /* Now to look around and see what can be done */
2908
2909         rcu_read_lock();
2910         for (i=disks; i--; ) {
2911                 mdk_rdev_t *rdev;
2912                 dev = &sh->dev[i];
2913                 clear_bit(R5_Insync, &dev->flags);
2914
2915                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2916                         i, dev->flags, dev->toread, dev->towrite, dev->written);
2917                 /* maybe we can reply to a read */
2918                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2919                         struct bio *rbi, *rbi2;
2920                         pr_debug("Return read for disc %d\n", i);
2921                         spin_lock_irq(&conf->device_lock);
2922                         rbi = dev->toread;
2923                         dev->toread = NULL;
2924                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2925                                 wake_up(&conf->wait_for_overlap);
2926                         spin_unlock_irq(&conf->device_lock);
2927                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2928                                 copy_data(0, rbi, dev->page, dev->sector);
2929                                 rbi2 = r5_next_bio(rbi, dev->sector);
2930                                 spin_lock_irq(&conf->device_lock);
2931                                 if (--rbi->bi_phys_segments == 0) {
2932                                         rbi->bi_next = return_bi;
2933                                         return_bi = rbi;
2934                                 }
2935                                 spin_unlock_irq(&conf->device_lock);
2936                                 rbi = rbi2;
2937                         }
2938                 }
2939
2940                 /* now count some things */
2941                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2942                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2943
2944
2945                 if (dev->toread)
2946                         s.to_read++;
2947                 if (dev->towrite) {
2948                         s.to_write++;
2949                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2950                                 s.non_overwrite++;
2951                 }
2952                 if (dev->written)
2953                         s.written++;
2954                 rdev = rcu_dereference(conf->disks[i].rdev);
2955                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2956                         /* The ReadError flag will just be confusing now */
2957                         clear_bit(R5_ReadError, &dev->flags);
2958                         clear_bit(R5_ReWrite, &dev->flags);
2959                 }
2960                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2961                     || test_bit(R5_ReadError, &dev->flags)) {
2962                         if (s.failed < 2)
2963                                 r6s.failed_num[s.failed] = i;
2964                         s.failed++;
2965                 } else
2966                         set_bit(R5_Insync, &dev->flags);
2967         }
2968         rcu_read_unlock();
2969         pr_debug("locked=%d uptodate=%d to_read=%d"
2970                " to_write=%d failed=%d failed_num=%d,%d\n",
2971                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
2972                r6s.failed_num[0], r6s.failed_num[1]);
2973         /* check if the array has lost >2 devices and, if so, some requests
2974          * might need to be failed
2975          */
2976         if (s.failed > 2 && s.to_read+s.to_write+s.written)
2977                 handle_requests_to_failed_array(conf, sh, &s, disks,
2978                                                 &return_bi);
2979         if (s.failed > 2 && s.syncing) {
2980                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2981                 clear_bit(STRIPE_SYNCING, &sh->state);
2982                 s.syncing = 0;
2983         }
2984
2985         /*
2986          * might be able to return some write requests if the parity blocks
2987          * are safe, or on a failed drive
2988          */
2989         pdev = &sh->dev[pd_idx];
2990         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
2991                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
2992         qdev = &sh->dev[r6s.qd_idx];
2993         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
2994                 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
2995
2996         if ( s.written &&
2997              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
2998                              && !test_bit(R5_LOCKED, &pdev->flags)
2999                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3000              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3001                              && !test_bit(R5_LOCKED, &qdev->flags)
3002                              && test_bit(R5_UPTODATE, &qdev->flags)))))
3003                 handle_completed_write_requests(conf, sh, disks, &return_bi);
3004
3005         /* Now we might consider reading some blocks, either to check/generate
3006          * parity, or to satisfy requests
3007          * or to load a block that is being partially written.
3008          */
3009         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3010             (s.syncing && (s.uptodate < disks)) || s.expanding)
3011                 handle_issuing_new_read_requests6(sh, &s, &r6s, disks);
3012
3013         /* now to consider writing and what else, if anything should be read */
3014         if (s.to_write)
3015                 handle_issuing_new_write_requests6(conf, sh, &s, &r6s, disks);
3016
3017         /* maybe we need to check and possibly fix the parity for this stripe
3018          * Any reads will already have been scheduled, so we just see if enough
3019          * data is available
3020          */
3021         if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3022                 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3023
3024         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3025                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3026                 clear_bit(STRIPE_SYNCING, &sh->state);
3027         }
3028
3029         /* If the failed drives are just a ReadError, then we might need
3030          * to progress the repair/check process
3031          */
3032         if (s.failed <= 2 && !conf->mddev->ro)
3033                 for (i = 0; i < s.failed; i++) {
3034                         dev = &sh->dev[r6s.failed_num[i]];
3035                         if (test_bit(R5_ReadError, &dev->flags)
3036                             && !test_bit(R5_LOCKED, &dev->flags)
3037                             && test_bit(R5_UPTODATE, &dev->flags)
3038                                 ) {
3039                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3040                                         set_bit(R5_Wantwrite, &dev->flags);
3041                                         set_bit(R5_ReWrite, &dev->flags);
3042                                         set_bit(R5_LOCKED, &dev->flags);
3043                                 } else {
3044                                         /* let's read it back */
3045                                         set_bit(R5_Wantread, &dev->flags);
3046                                         set_bit(R5_LOCKED, &dev->flags);
3047                                 }
3048                         }
3049                 }
3050
3051         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3052                 /* Need to write out all blocks after computing P&Q */
3053                 sh->disks = conf->raid_disks;
3054                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
3055                                              conf->raid_disks);
3056                 compute_parity6(sh, RECONSTRUCT_WRITE);
3057                 for (i = conf->raid_disks ; i-- ;  ) {
3058                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3059                         s.locked++;
3060                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3061                 }
3062                 clear_bit(STRIPE_EXPANDING, &sh->state);
3063         } else if (s.expanded) {
3064                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3065                 atomic_dec(&conf->reshape_stripes);
3066                 wake_up(&conf->wait_for_overlap);
3067                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3068         }
3069
3070         if (s.expanding && s.locked == 0)
3071                 handle_stripe_expansion(conf, sh, &r6s);
3072
3073         spin_unlock(&sh->lock);
3074
3075         return_io(return_bi);
3076
3077         for (i=disks; i-- ;) {
3078                 int rw;
3079                 struct bio *bi;
3080                 mdk_rdev_t *rdev;
3081                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
3082                         rw = WRITE;
3083                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
3084                         rw = READ;
3085                 else
3086                         continue;
3087
3088                 bi = &sh->dev[i].req;
3089
3090                 bi->bi_rw = rw;
3091                 if (rw == WRITE)
3092                         bi->bi_end_io = raid5_end_write_request;
3093                 else
3094                         bi->bi_end_io = raid5_end_read_request;
3095
3096                 rcu_read_lock();
3097                 rdev = rcu_dereference(conf->disks[i].rdev);
3098                 if (rdev && test_bit(Faulty, &rdev->flags))
3099                         rdev = NULL;
3100                 if (rdev)
3101                         atomic_inc(&rdev->nr_pending);
3102                 rcu_read_unlock();
3103
3104                 if (rdev) {
3105                         if (s.syncing || s.expanding || s.expanded)
3106                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
3107
3108                         bi->bi_bdev = rdev->bdev;
3109                         pr_debug("for %llu schedule op %ld on disc %d\n",
3110                                 (unsigned long long)sh->sector, bi->bi_rw, i);
3111                         atomic_inc(&sh->count);
3112                         bi->bi_sector = sh->sector + rdev->data_offset;
3113                         bi->bi_flags = 1 << BIO_UPTODATE;
3114                         bi->bi_vcnt = 1;
3115                         bi->bi_max_vecs = 1;
3116                         bi->bi_idx = 0;
3117                         bi->bi_io_vec = &sh->dev[i].vec;
3118                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
3119                         bi->bi_io_vec[0].bv_offset = 0;
3120                         bi->bi_size = STRIPE_SIZE;
3121                         bi->bi_next = NULL;
3122                         if (rw == WRITE &&
3123                             test_bit(R5_ReWrite, &sh->dev[i].flags))
3124                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
3125                         generic_make_request(bi);
3126                 } else {
3127                         if (rw == WRITE)
3128                                 set_bit(STRIPE_DEGRADED, &sh->state);
3129                         pr_debug("skip op %ld on disc %d for sector %llu\n",
3130                                 bi->bi_rw, i, (unsigned long long)sh->sector);
3131                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
3132                         set_bit(STRIPE_HANDLE, &sh->state);
3133                 }
3134         }
3135 }
3136
3137 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3138 {
3139         if (sh->raid_conf->level == 6)
3140                 handle_stripe6(sh, tmp_page);
3141         else
3142                 handle_stripe5(sh);
3143 }
3144
3145
3146
3147 static void raid5_activate_delayed(raid5_conf_t *conf)
3148 {
3149         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3150                 while (!list_empty(&conf->delayed_list)) {
3151                         struct list_head *l = conf->delayed_list.next;
3152                         struct stripe_head *sh;
3153                         sh = list_entry(l, struct stripe_head, lru);
3154                         list_del_init(l);
3155                         clear_bit(STRIPE_DELAYED, &sh->state);
3156                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3157                                 atomic_inc(&conf->preread_active_stripes);
3158                         list_add_tail(&sh->lru, &conf->handle_list);
3159                 }
3160         }
3161 }
3162
3163 static void activate_bit_delay(raid5_conf_t *conf)
3164 {
3165         /* device_lock is held */
3166         struct list_head head;
3167         list_add(&head, &conf->bitmap_list);
3168         list_del_init(&conf->bitmap_list);
3169         while (!list_empty(&head)) {
3170                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3171                 list_del_init(&sh->lru);
3172                 atomic_inc(&sh->count);
3173                 __release_stripe(conf, sh);
3174         }
3175 }
3176
3177 static void unplug_slaves(mddev_t *mddev)
3178 {
3179         raid5_conf_t *conf = mddev_to_conf(mddev);
3180         int i;
3181
3182         rcu_read_lock();
3183         for (i=0; i<mddev->raid_disks; i++) {
3184                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3185                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3186                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3187
3188                         atomic_inc(&rdev->nr_pending);
3189                         rcu_read_unlock();
3190
3191                         if (r_queue->unplug_fn)
3192                                 r_queue->unplug_fn(r_queue);
3193
3194                         rdev_dec_pending(rdev, mddev);
3195                         rcu_read_lock();
3196                 }
3197         }
3198         rcu_read_unlock();
3199 }
3200
3201 static void raid5_unplug_device(struct request_queue *q)
3202 {
3203         mddev_t *mddev = q->queuedata;
3204         raid5_conf_t *conf = mddev_to_conf(mddev);
3205         unsigned long flags;
3206
3207         spin_lock_irqsave(&conf->device_lock, flags);
3208
3209         if (blk_remove_plug(q)) {
3210                 conf->seq_flush++;
3211                 raid5_activate_delayed(conf);
3212         }
3213         md_wakeup_thread(mddev->thread);
3214
3215         spin_unlock_irqrestore(&conf->device_lock, flags);
3216
3217         unplug_slaves(mddev);
3218 }
3219
3220 static int raid5_issue_flush(struct request_queue *q, struct gendisk *disk,
3221                              sector_t *error_sector)
3222 {
3223         mddev_t *mddev = q->queuedata;
3224         raid5_conf_t *conf = mddev_to_conf(mddev);
3225         int i, ret = 0;
3226
3227         rcu_read_lock();
3228         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
3229                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3230                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
3231                         struct block_device *bdev = rdev->bdev;
3232                         struct request_queue *r_queue = bdev_get_queue(bdev);
3233
3234                         if (!r_queue->issue_flush_fn)
3235                                 ret = -EOPNOTSUPP;
3236                         else {
3237                                 atomic_inc(&rdev->nr_pending);
3238                                 rcu_read_unlock();
3239                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
3240                                                               error_sector);
3241                                 rdev_dec_pending(rdev, mddev);
3242                                 rcu_read_lock();
3243                         }
3244                 }
3245         }
3246         rcu_read_unlock();
3247         return ret;
3248 }
3249
3250 static int raid5_congested(void *data, int bits)
3251 {
3252         mddev_t *mddev = data;
3253         raid5_conf_t *conf = mddev_to_conf(mddev);
3254
3255         /* No difference between reads and writes.  Just check
3256          * how busy the stripe_cache is
3257          */
3258         if (conf->inactive_blocked)
3259                 return 1;
3260         if (conf->quiesce)
3261                 return 1;
3262         if (list_empty_careful(&conf->inactive_list))
3263                 return 1;
3264
3265         return 0;
3266 }
3267
3268 /* We want read requests to align with chunks where possible,
3269  * but write requests don't need to.
3270  */
3271 static int raid5_mergeable_bvec(struct request_queue *q, struct bio *bio, struct bio_vec *biovec)
3272 {
3273         mddev_t *mddev = q->queuedata;
3274         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3275         int max;
3276         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3277         unsigned int bio_sectors = bio->bi_size >> 9;
3278
3279         if (bio_data_dir(bio) == WRITE)
3280                 return biovec->bv_len; /* always allow writes to be mergeable */
3281
3282         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3283         if (max < 0) max = 0;
3284         if (max <= biovec->bv_len && bio_sectors == 0)
3285                 return biovec->bv_len;
3286         else
3287                 return max;
3288 }
3289
3290
3291 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3292 {
3293         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3294         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3295         unsigned int bio_sectors = bio->bi_size >> 9;
3296
3297         return  chunk_sectors >=
3298                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3299 }
3300
3301 /*
3302  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3303  *  later sampled by raid5d.
3304  */
3305 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3306 {
3307         unsigned long flags;
3308
3309         spin_lock_irqsave(&conf->device_lock, flags);
3310
3311         bi->bi_next = conf->retry_read_aligned_list;
3312         conf->retry_read_aligned_list = bi;
3313
3314         spin_unlock_irqrestore(&conf->device_lock, flags);
3315         md_wakeup_thread(conf->mddev->thread);
3316 }
3317
3318
3319 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3320 {
3321         struct bio *bi;
3322
3323         bi = conf->retry_read_aligned;
3324         if (bi) {
3325                 conf->retry_read_aligned = NULL;
3326                 return bi;
3327         }
3328         bi = conf->retry_read_aligned_list;
3329         if(bi) {
3330                 conf->retry_read_aligned_list = bi->bi_next;
3331                 bi->bi_next = NULL;
3332                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3333                 bi->bi_hw_segments = 0; /* count of processed stripes */
3334         }
3335
3336         return bi;
3337 }
3338
3339
3340 /*
3341  *  The "raid5_align_endio" should check if the read succeeded and if it
3342  *  did, call bio_endio on the original bio (having bio_put the new bio
3343  *  first).
3344  *  If the read failed..
3345  */
3346 static int raid5_align_endio(struct bio *bi, unsigned int bytes, int error)
3347 {
3348         struct bio* raid_bi  = bi->bi_private;
3349         mddev_t *mddev;
3350         raid5_conf_t *conf;
3351         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3352         mdk_rdev_t *rdev;
3353
3354         if (bi->bi_size)
3355                 return 1;
3356         bio_put(bi);
3357
3358         mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3359         conf = mddev_to_conf(mddev);
3360         rdev = (void*)raid_bi->bi_next;
3361         raid_bi->bi_next = NULL;
3362
3363         rdev_dec_pending(rdev, conf->mddev);
3364
3365         if (!error && uptodate) {
3366                 bio_endio(raid_bi, bytes, 0);
3367                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3368                         wake_up(&conf->wait_for_stripe);
3369                 return 0;
3370         }
3371
3372
3373         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3374
3375         add_bio_to_retry(raid_bi, conf);
3376         return 0;
3377 }
3378
3379 static int bio_fits_rdev(struct bio *bi)
3380 {
3381         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3382
3383         if ((bi->bi_size>>9) > q->max_sectors)
3384                 return 0;
3385         blk_recount_segments(q, bi);
3386         if (bi->bi_phys_segments > q->max_phys_segments ||
3387             bi->bi_hw_segments > q->max_hw_segments)
3388                 return 0;
3389
3390         if (q->merge_bvec_fn)
3391                 /* it's too hard to apply the merge_bvec_fn at this stage,
3392                  * just just give up
3393                  */
3394                 return 0;
3395
3396         return 1;
3397 }
3398
3399
3400 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3401 {
3402         mddev_t *mddev = q->queuedata;
3403         raid5_conf_t *conf = mddev_to_conf(mddev);
3404         const unsigned int raid_disks = conf->raid_disks;
3405         const unsigned int data_disks = raid_disks - conf->max_degraded;
3406         unsigned int dd_idx, pd_idx;
3407         struct bio* align_bi;
3408         mdk_rdev_t *rdev;
3409
3410         if (!in_chunk_boundary(mddev, raid_bio)) {
3411                 pr_debug("chunk_aligned_read : non aligned\n");
3412                 return 0;
3413         }
3414         /*
3415          * use bio_clone to make a copy of the bio
3416          */
3417         align_bi = bio_clone(raid_bio, GFP_NOIO);
3418         if (!align_bi)
3419                 return 0;
3420         /*
3421          *   set bi_end_io to a new function, and set bi_private to the
3422          *     original bio.
3423          */
3424         align_bi->bi_end_io  = raid5_align_endio;
3425         align_bi->bi_private = raid_bio;
3426         /*
3427          *      compute position
3428          */
3429         align_bi->bi_sector =  raid5_compute_sector(raid_bio->bi_sector,
3430                                         raid_disks,
3431                                         data_disks,
3432                                         &dd_idx,
3433                                         &pd_idx,
3434                                         conf);
3435
3436         rcu_read_lock();
3437         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3438         if (rdev && test_bit(In_sync, &rdev->flags)) {
3439                 atomic_inc(&rdev->nr_pending);
3440                 rcu_read_unlock();
3441                 raid_bio->bi_next = (void*)rdev;
3442                 align_bi->bi_bdev =  rdev->bdev;
3443                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3444                 align_bi->bi_sector += rdev->data_offset;
3445
3446                 if (!bio_fits_rdev(align_bi)) {
3447                         /* too big in some way */
3448                         bio_put(align_bi);
3449                         rdev_dec_pending(rdev, mddev);
3450                         return 0;
3451                 }
3452
3453                 spin_lock_irq(&conf->device_lock);
3454                 wait_event_lock_irq(conf->wait_for_stripe,
3455                                     conf->quiesce == 0,
3456                                     conf->device_lock, /* nothing */);
3457                 atomic_inc(&conf->active_aligned_reads);
3458                 spin_unlock_irq(&conf->device_lock);
3459
3460                 generic_make_request(align_bi);
3461                 return 1;
3462         } else {
3463                 rcu_read_unlock();
3464                 bio_put(align_bi);
3465                 return 0;
3466         }
3467 }
3468
3469
3470 static int make_request(struct request_queue *q, struct bio * bi)
3471 {
3472         mddev_t *mddev = q->queuedata;
3473         raid5_conf_t *conf = mddev_to_conf(mddev);
3474         unsigned int dd_idx, pd_idx;
3475         sector_t new_sector;
3476         sector_t logical_sector, last_sector;
3477         struct stripe_head *sh;
3478         const int rw = bio_data_dir(bi);
3479         int remaining;
3480
3481         if (unlikely(bio_barrier(bi))) {
3482                 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
3483                 return 0;
3484         }
3485
3486         md_write_start(mddev, bi);
3487
3488         disk_stat_inc(mddev->gendisk, ios[rw]);
3489         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
3490
3491         if (rw == READ &&
3492              mddev->reshape_position == MaxSector &&
3493              chunk_aligned_read(q,bi))
3494                 return 0;
3495
3496         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3497         last_sector = bi->bi_sector + (bi->bi_size>>9);
3498         bi->bi_next = NULL;
3499         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
3500
3501         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3502                 DEFINE_WAIT(w);
3503                 int disks, data_disks;
3504
3505         retry:
3506                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3507                 if (likely(conf->expand_progress == MaxSector))
3508                         disks = conf->raid_disks;
3509                 else {
3510                         /* spinlock is needed as expand_progress may be
3511                          * 64bit on a 32bit platform, and so it might be
3512                          * possible to see a half-updated value
3513                          * Ofcourse expand_progress could change after
3514                          * the lock is dropped, so once we get a reference
3515                          * to the stripe that we think it is, we will have
3516                          * to check again.
3517                          */
3518                         spin_lock_irq(&conf->device_lock);
3519                         disks = conf->raid_disks;
3520                         if (logical_sector >= conf->expand_progress)
3521                                 disks = conf->previous_raid_disks;
3522                         else {
3523                                 if (logical_sector >= conf->expand_lo) {
3524                                         spin_unlock_irq(&conf->device_lock);
3525                                         schedule();
3526                                         goto retry;
3527                                 }
3528                         }
3529                         spin_unlock_irq(&conf->device_lock);
3530                 }
3531                 data_disks = disks - conf->max_degraded;
3532
3533                 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
3534                                                   &dd_idx, &pd_idx, conf);
3535                 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3536                         (unsigned long long)new_sector,
3537                         (unsigned long long)logical_sector);
3538
3539                 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
3540                 if (sh) {
3541                         if (unlikely(conf->expand_progress != MaxSector)) {
3542                                 /* expansion might have moved on while waiting for a
3543                                  * stripe, so we must do the range check again.
3544                                  * Expansion could still move past after this
3545                                  * test, but as we are holding a reference to
3546                                  * 'sh', we know that if that happens,
3547                                  *  STRIPE_EXPANDING will get set and the expansion
3548                                  * won't proceed until we finish with the stripe.
3549                                  */
3550                                 int must_retry = 0;
3551                                 spin_lock_irq(&conf->device_lock);
3552                                 if (logical_sector <  conf->expand_progress &&
3553                                     disks == conf->previous_raid_disks)
3554                                         /* mismatch, need to try again */
3555                                         must_retry = 1;
3556                                 spin_unlock_irq(&conf->device_lock);
3557                                 if (must_retry) {
3558                                         release_stripe(sh);
3559                                         goto retry;
3560                                 }
3561                         }
3562                         /* FIXME what if we get a false positive because these
3563                          * are being updated.
3564                          */
3565                         if (logical_sector >= mddev->suspend_lo &&
3566                             logical_sector < mddev->suspend_hi) {
3567                                 release_stripe(sh);
3568                                 schedule();
3569                                 goto retry;
3570                         }
3571
3572                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3573                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3574                                 /* Stripe is busy expanding or
3575                                  * add failed due to overlap.  Flush everything
3576                                  * and wait a while
3577                                  */
3578                                 raid5_unplug_device(mddev->queue);
3579                                 release_stripe(sh);
3580                                 schedule();
3581                                 goto retry;
3582                         }
3583                         finish_wait(&conf->wait_for_overlap, &w);
3584                         handle_stripe(sh, NULL);
3585                         release_stripe(sh);
3586                 } else {
3587                         /* cannot get stripe for read-ahead, just give-up */
3588                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
3589                         finish_wait(&conf->wait_for_overlap, &w);
3590                         break;
3591                 }
3592
3593         }
3594         spin_lock_irq(&conf->device_lock);
3595         remaining = --bi->bi_phys_segments;
3596         spin_unlock_irq(&conf->device_lock);
3597         if (remaining == 0) {
3598                 int bytes = bi->bi_size;
3599
3600                 if ( rw == WRITE )
3601                         md_write_end(mddev);
3602                 bi->bi_size = 0;
3603                 bi->bi_end_io(bi, bytes,
3604                               test_bit(BIO_UPTODATE, &bi->bi_flags)
3605                                 ? 0 : -EIO);
3606         }
3607         return 0;
3608 }
3609
3610 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3611 {
3612         /* reshaping is quite different to recovery/resync so it is
3613          * handled quite separately ... here.
3614          *
3615          * On each call to sync_request, we gather one chunk worth of
3616          * destination stripes and flag them as expanding.
3617          * Then we find all the source stripes and request reads.
3618          * As the reads complete, handle_stripe will copy the data
3619          * into the destination stripe and release that stripe.
3620          */
3621         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3622         struct stripe_head *sh;
3623         int pd_idx;
3624         sector_t first_sector, last_sector;
3625         int raid_disks = conf->previous_raid_disks;
3626         int data_disks = raid_disks - conf->max_degraded;
3627         int new_data_disks = conf->raid_disks - conf->max_degraded;
3628         int i;
3629         int dd_idx;
3630         sector_t writepos, safepos, gap;
3631
3632         if (sector_nr == 0 &&
3633             conf->expand_progress != 0) {
3634                 /* restarting in the middle, skip the initial sectors */
3635                 sector_nr = conf->expand_progress;
3636                 sector_div(sector_nr, new_data_disks);
3637                 *skipped = 1;
3638                 return sector_nr;
3639         }
3640
3641         /* we update the metadata when there is more than 3Meg
3642          * in the block range (that is rather arbitrary, should
3643          * probably be time based) or when the data about to be
3644          * copied would over-write the source of the data at
3645          * the front of the range.
3646          * i.e. one new_stripe forward from expand_progress new_maps
3647          * to after where expand_lo old_maps to
3648          */
3649         writepos = conf->expand_progress +
3650                 conf->chunk_size/512*(new_data_disks);
3651         sector_div(writepos, new_data_disks);
3652         safepos = conf->expand_lo;
3653         sector_div(safepos, data_disks);
3654         gap = conf->expand_progress - conf->expand_lo;
3655
3656         if (writepos >= safepos ||
3657             gap > (new_data_disks)*3000*2 /*3Meg*/) {
3658                 /* Cannot proceed until we've updated the superblock... */
3659                 wait_event(conf->wait_for_overlap,
3660                            atomic_read(&conf->reshape_stripes)==0);
3661                 mddev->reshape_position = conf->expand_progress;
3662                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3663                 md_wakeup_thread(mddev->thread);
3664                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3665                            kthread_should_stop());
3666                 spin_lock_irq(&conf->device_lock);
3667                 conf->expand_lo = mddev->reshape_position;
3668                 spin_unlock_irq(&conf->device_lock);
3669                 wake_up(&conf->wait_for_overlap);
3670         }
3671
3672         for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3673                 int j;
3674                 int skipped = 0;
3675                 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3676                 sh = get_active_stripe(conf, sector_nr+i,
3677                                        conf->raid_disks, pd_idx, 0);
3678                 set_bit(STRIPE_EXPANDING, &sh->state);
3679                 atomic_inc(&conf->reshape_stripes);
3680                 /* If any of this stripe is beyond the end of the old
3681                  * array, then we need to zero those blocks
3682                  */
3683                 for (j=sh->disks; j--;) {
3684                         sector_t s;
3685                         if (j == sh->pd_idx)
3686                                 continue;
3687                         if (conf->level == 6 &&
3688                             j == raid6_next_disk(sh->pd_idx, sh->disks))
3689                                 continue;
3690                         s = compute_blocknr(sh, j);
3691                         if (s < (mddev->array_size<<1)) {
3692                                 skipped = 1;
3693                                 continue;
3694                         }
3695                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3696                         set_bit(R5_Expanded, &sh->dev[j].flags);
3697                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
3698                 }
3699                 if (!skipped) {
3700                         set_bit(STRIPE_EXPAND_READY, &sh->state);
3701                         set_bit(STRIPE_HANDLE, &sh->state);
3702                 }
3703                 release_stripe(sh);
3704         }
3705         spin_lock_irq(&conf->device_lock);
3706         conf->expand_progress = (sector_nr + i) * new_data_disks;
3707         spin_unlock_irq(&conf->device_lock);
3708         /* Ok, those stripe are ready. We can start scheduling
3709          * reads on the source stripes.
3710          * The source stripes are determined by mapping the first and last
3711          * block on the destination stripes.
3712          */
3713         first_sector =
3714                 raid5_compute_sector(sector_nr*(new_data_disks),
3715                                      raid_disks, data_disks,
3716                                      &dd_idx, &pd_idx, conf);
3717         last_sector =
3718                 raid5_compute_sector((sector_nr+conf->chunk_size/512)
3719                                      *(new_data_disks) -1,
3720                                      raid_disks, data_disks,
3721                                      &dd_idx, &pd_idx, conf);
3722         if (last_sector >= (mddev->size<<1))
3723                 last_sector = (mddev->size<<1)-1;
3724         while (first_sector <= last_sector) {
3725                 pd_idx = stripe_to_pdidx(first_sector, conf,
3726                                          conf->previous_raid_disks);
3727                 sh = get_active_stripe(conf, first_sector,
3728                                        conf->previous_raid_disks, pd_idx, 0);
3729                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3730                 set_bit(STRIPE_HANDLE, &sh->state);
3731                 release_stripe(sh);
3732                 first_sector += STRIPE_SECTORS;
3733         }
3734         return conf->chunk_size>>9;
3735 }
3736
3737 /* FIXME go_faster isn't used */
3738 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3739 {
3740         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3741         struct stripe_head *sh;
3742         int pd_idx;
3743         int raid_disks = conf->raid_disks;
3744         sector_t max_sector = mddev->size << 1;
3745         int sync_blocks;
3746         int still_degraded = 0;
3747         int i;
3748
3749         if (sector_nr >= max_sector) {
3750                 /* just being told to finish up .. nothing much to do */
3751                 unplug_slaves(mddev);
3752                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3753                         end_reshape(conf);
3754                         return 0;
3755                 }
3756
3757                 if (mddev->curr_resync < max_sector) /* aborted */
3758                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3759                                         &sync_blocks, 1);
3760                 else /* completed sync */
3761                         conf->fullsync = 0;
3762                 bitmap_close_sync(mddev->bitmap);
3763
3764                 return 0;
3765         }
3766
3767         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3768                 return reshape_request(mddev, sector_nr, skipped);
3769
3770         /* if there is too many failed drives and we are trying
3771          * to resync, then assert that we are finished, because there is
3772          * nothing we can do.
3773          */
3774         if (mddev->degraded >= conf->max_degraded &&
3775             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3776                 sector_t rv = (mddev->size << 1) - sector_nr;
3777                 *skipped = 1;
3778                 return rv;
3779         }
3780         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3781             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3782             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3783                 /* we can skip this block, and probably more */
3784                 sync_blocks /= STRIPE_SECTORS;
3785                 *skipped = 1;
3786                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3787         }
3788
3789         pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3790         sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3791         if (sh == NULL) {
3792                 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3793                 /* make sure we don't swamp the stripe cache if someone else
3794                  * is trying to get access
3795                  */
3796                 schedule_timeout_uninterruptible(1);
3797         }
3798         /* Need to check if array will still be degraded after recovery/resync
3799          * We don't need to check the 'failed' flag as when that gets set,
3800          * recovery aborts.
3801          */
3802         for (i=0; i<mddev->raid_disks; i++)
3803                 if (conf->disks[i].rdev == NULL)
3804                         still_degraded = 1;
3805
3806         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3807
3808         spin_lock(&sh->lock);
3809         set_bit(STRIPE_SYNCING, &sh->state);
3810         clear_bit(STRIPE_INSYNC, &sh->state);
3811         spin_unlock(&sh->lock);
3812
3813         handle_stripe(sh, NULL);
3814         release_stripe(sh);
3815
3816         return STRIPE_SECTORS;
3817 }
3818
3819 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3820 {
3821         /* We may not be able to submit a whole bio at once as there
3822          * may not be enough stripe_heads available.
3823          * We cannot pre-allocate enough stripe_heads as we may need
3824          * more than exist in the cache (if we allow ever large chunks).
3825          * So we do one stripe head at a time and record in
3826          * ->bi_hw_segments how many have been done.
3827          *
3828          * We *know* that this entire raid_bio is in one chunk, so
3829          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3830          */
3831         struct stripe_head *sh;
3832         int dd_idx, pd_idx;
3833         sector_t sector, logical_sector, last_sector;
3834         int scnt = 0;
3835         int remaining;
3836         int handled = 0;
3837
3838         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3839         sector = raid5_compute_sector(  logical_sector,
3840                                         conf->raid_disks,
3841                                         conf->raid_disks - conf->max_degraded,
3842                                         &dd_idx,
3843                                         &pd_idx,
3844                                         conf);
3845         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3846
3847         for (; logical_sector < last_sector;
3848              logical_sector += STRIPE_SECTORS,
3849                      sector += STRIPE_SECTORS,
3850                      scnt++) {
3851
3852                 if (scnt < raid_bio->bi_hw_segments)
3853                         /* already done this stripe */
3854                         continue;
3855
3856                 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3857
3858                 if (!sh) {
3859                         /* failed to get a stripe - must wait */
3860                         raid_bio->bi_hw_segments = scnt;
3861                         conf->retry_read_aligned = raid_bio;
3862                         return handled;
3863                 }
3864
3865                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3866                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3867                         release_stripe(sh);
3868                         raid_bio->bi_hw_segments = scnt;
3869                         conf->retry_read_aligned = raid_bio;
3870                         return handled;
3871                 }
3872
3873                 handle_stripe(sh, NULL);
3874                 release_stripe(sh);
3875                 handled++;
3876         }
3877         spin_lock_irq(&conf->device_lock);
3878         remaining = --raid_bio->bi_phys_segments;
3879         spin_unlock_irq(&conf->device_lock);
3880         if (remaining == 0) {
3881                 int bytes = raid_bio->bi_size;
3882
3883                 raid_bio->bi_size = 0;
3884                 raid_bio->bi_end_io(raid_bio, bytes,
3885                               test_bit(BIO_UPTODATE, &raid_bio->bi_flags)
3886                                 ? 0 : -EIO);
3887         }
3888         if (atomic_dec_and_test(&conf->active_aligned_reads))
3889                 wake_up(&conf->wait_for_stripe);
3890         return handled;
3891 }
3892
3893
3894
3895 /*
3896  * This is our raid5 kernel thread.
3897  *
3898  * We scan the hash table for stripes which can be handled now.
3899  * During the scan, completed stripes are saved for us by the interrupt
3900  * handler, so that they will not have to wait for our next wakeup.
3901  */
3902 static void raid5d (mddev_t *mddev)
3903 {
3904         struct stripe_head *sh;
3905         raid5_conf_t *conf = mddev_to_conf(mddev);
3906         int handled;
3907
3908         pr_debug("+++ raid5d active\n");
3909
3910         md_check_recovery(mddev);
3911
3912         handled = 0;
3913         spin_lock_irq(&conf->device_lock);
3914         while (1) {
3915                 struct list_head *first;
3916                 struct bio *bio;
3917
3918                 if (conf->seq_flush != conf->seq_write) {
3919                         int seq = conf->seq_flush;
3920                         spin_unlock_irq(&conf->device_lock);
3921                         bitmap_unplug(mddev->bitmap);
3922                         spin_lock_irq(&conf->device_lock);
3923                         conf->seq_write = seq;
3924                         activate_bit_delay(conf);
3925                 }
3926
3927                 if (list_empty(&conf->handle_list) &&
3928                     atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
3929                     !blk_queue_plugged(mddev->queue) &&
3930                     !list_empty(&conf->delayed_list))
3931                         raid5_activate_delayed(conf);
3932
3933                 while ((bio = remove_bio_from_retry(conf))) {
3934                         int ok;
3935                         spin_unlock_irq(&conf->device_lock);
3936                         ok = retry_aligned_read(conf, bio);
3937                         spin_lock_irq(&conf->device_lock);
3938                         if (!ok)
3939                                 break;
3940                         handled++;
3941                 }
3942
3943                 if (list_empty(&conf->handle_list)) {
3944                         async_tx_issue_pending_all();
3945                         break;
3946                 }
3947
3948                 first = conf->handle_list.next;
3949                 sh = list_entry(first, struct stripe_head, lru);
3950
3951                 list_del_init(first);
3952                 atomic_inc(&sh->count);
3953                 BUG_ON(atomic_read(&sh->count)!= 1);
3954                 spin_unlock_irq(&conf->device_lock);
3955
3956                 handled++;
3957                 handle_stripe(sh, conf->spare_page);
3958                 release_stripe(sh);
3959
3960                 spin_lock_irq(&conf->device_lock);
3961         }
3962         pr_debug("%d stripes handled\n", handled);
3963
3964         spin_unlock_irq(&conf->device_lock);
3965
3966         unplug_slaves(mddev);
3967
3968         pr_debug("--- raid5d inactive\n");
3969 }
3970
3971 static ssize_t
3972 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
3973 {
3974         raid5_conf_t *conf = mddev_to_conf(mddev);
3975         if (conf)
3976                 return sprintf(page, "%d\n", conf->max_nr_stripes);
3977         else
3978                 return 0;
3979 }
3980
3981 static ssize_t
3982 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
3983 {
3984         raid5_conf_t *conf = mddev_to_conf(mddev);
3985         char *end;
3986         int new;
3987         if (len >= PAGE_SIZE)
3988                 return -EINVAL;
3989         if (!conf)
3990                 return -ENODEV;
3991
3992         new = simple_strtoul(page, &end, 10);
3993         if (!*page || (*end && *end != '\n') )
3994                 return -EINVAL;
3995         if (new <= 16 || new > 32768)
3996                 return -EINVAL;
3997         while (new < conf->max_nr_stripes) {
3998                 if (drop_one_stripe(conf))
3999                         conf->max_nr_stripes--;
4000                 else
4001                         break;
4002         }
4003         md_allow_write(mddev);
4004         while (new > conf->max_nr_stripes) {
4005                 if (grow_one_stripe(conf))
4006                         conf->max_nr_stripes++;
4007                 else break;
4008         }
4009         return len;
4010 }
4011
4012 static struct md_sysfs_entry
4013 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4014                                 raid5_show_stripe_cache_size,
4015                                 raid5_store_stripe_cache_size);
4016
4017 static ssize_t
4018 stripe_cache_active_show(mddev_t *mddev, char *page)
4019 {
4020         raid5_conf_t *conf = mddev_to_conf(mddev);
4021         if (conf)
4022                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4023         else
4024                 return 0;
4025 }
4026
4027 static struct md_sysfs_entry
4028 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4029
4030 static struct attribute *raid5_attrs[] =  {
4031         &raid5_stripecache_size.attr,
4032         &raid5_stripecache_active.attr,
4033         NULL,
4034 };
4035 static struct attribute_group raid5_attrs_group = {
4036         .name = NULL,
4037         .attrs = raid5_attrs,
4038 };
4039
4040 static int run(mddev_t *mddev)
4041 {
4042         raid5_conf_t *conf;
4043         int raid_disk, memory;
4044         mdk_rdev_t *rdev;
4045         struct disk_info *disk;
4046         struct list_head *tmp;
4047         int working_disks = 0;
4048
4049         if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
4050                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4051                        mdname(mddev), mddev->level);
4052                 return -EIO;
4053         }
4054
4055         if (mddev->reshape_position != MaxSector) {
4056                 /* Check that we can continue the reshape.
4057                  * Currently only disks can change, it must
4058                  * increase, and we must be past the point where
4059                  * a stripe over-writes itself
4060                  */
4061                 sector_t here_new, here_old;
4062                 int old_disks;
4063                 int max_degraded = (mddev->level == 5 ? 1 : 2);
4064
4065                 if (mddev->new_level != mddev->level ||
4066                     mddev->new_layout != mddev->layout ||
4067                     mddev->new_chunk != mddev->chunk_size) {
4068                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4069                                "required - aborting.\n",
4070                                mdname(mddev));
4071                         return -EINVAL;
4072                 }
4073                 if (mddev->delta_disks <= 0) {
4074                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4075                                "(reduce disks) required - aborting.\n",
4076                                mdname(mddev));
4077                         return -EINVAL;
4078                 }
4079                 old_disks = mddev->raid_disks - mddev->delta_disks;
4080                 /* reshape_position must be on a new-stripe boundary, and one
4081                  * further up in new geometry must map after here in old
4082                  * geometry.
4083                  */
4084                 here_new = mddev->reshape_position;
4085                 if (sector_div(here_new, (mddev->chunk_size>>9)*
4086                                (mddev->raid_disks - max_degraded))) {
4087                         printk(KERN_ERR "raid5: reshape_position not "
4088                                "on a stripe boundary\n");
4089                         return -EINVAL;
4090                 }
4091                 /* here_new is the stripe we will write to */
4092                 here_old = mddev->reshape_position;
4093                 sector_div(here_old, (mddev->chunk_size>>9)*
4094                            (old_disks-max_degraded));
4095                 /* here_old is the first stripe that we might need to read
4096                  * from */
4097                 if (here_new >= here_old) {
4098                         /* Reading from the same stripe as writing to - bad */
4099                         printk(KERN_ERR "raid5: reshape_position too early for "
4100                                "auto-recovery - aborting.\n");
4101                         return -EINVAL;
4102                 }
4103                 printk(KERN_INFO "raid5: reshape will continue\n");
4104                 /* OK, we should be able to continue; */
4105         }
4106
4107
4108         mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4109         if ((conf = mddev->private) == NULL)
4110                 goto abort;
4111         if (mddev->reshape_position == MaxSector) {
4112                 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4113         } else {
4114                 conf->raid_disks = mddev->raid_disks;
4115                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4116         }
4117
4118         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4119                               GFP_KERNEL);
4120         if (!conf->disks)
4121                 goto abort;
4122
4123         conf->mddev = mddev;
4124
4125         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4126                 goto abort;
4127
4128         if (mddev->level == 6) {
4129                 conf->spare_page = alloc_page(GFP_KERNEL);
4130                 if (!conf->spare_page)
4131                         goto abort;
4132         }
4133         spin_lock_init(&conf->device_lock);
4134         init_waitqueue_head(&conf->wait_for_stripe);
4135         init_waitqueue_head(&conf->wait_for_overlap);
4136         INIT_LIST_HEAD(&conf->handle_list);
4137         INIT_LIST_HEAD(&conf->delayed_list);
4138         INIT_LIST_HEAD(&conf->bitmap_list);
4139         INIT_LIST_HEAD(&conf->inactive_list);
4140         atomic_set(&conf->active_stripes, 0);
4141         atomic_set(&conf->preread_active_stripes, 0);
4142         atomic_set(&conf->active_aligned_reads, 0);
4143
4144         pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4145
4146         ITERATE_RDEV(mddev,rdev,tmp) {
4147                 raid_disk = rdev->raid_disk;
4148                 if (raid_disk >= conf->raid_disks
4149                     || raid_disk < 0)
4150                         continue;
4151                 disk = conf->disks + raid_disk;
4152
4153                 disk->rdev = rdev;
4154
4155                 if (test_bit(In_sync, &rdev->flags)) {
4156                         char b[BDEVNAME_SIZE];
4157                         printk(KERN_INFO "raid5: device %s operational as raid"
4158                                 " disk %d\n", bdevname(rdev->bdev,b),
4159                                 raid_disk);
4160                         working_disks++;
4161                 }
4162         }
4163
4164         /*
4165          * 0 for a fully functional array, 1 or 2 for a degraded array.
4166          */
4167         mddev->degraded = conf->raid_disks - working_disks;
4168         conf->mddev = mddev;
4169         conf->chunk_size = mddev->chunk_size;
4170         conf->level = mddev->level;
4171         if (conf->level == 6)
4172                 conf->max_degraded = 2;
4173         else
4174                 conf->max_degraded = 1;
4175         conf->algorithm = mddev->layout;
4176         conf->max_nr_stripes = NR_STRIPES;
4177         conf->expand_progress = mddev->reshape_position;
4178
4179         /* device size must be a multiple of chunk size */
4180         mddev->size &= ~(mddev->chunk_size/1024 -1);
4181         mddev->resync_max_sectors = mddev->size << 1;
4182
4183         if (conf->level == 6 && conf->raid_disks < 4) {
4184                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4185                        mdname(mddev), conf->raid_disks);
4186                 goto abort;
4187         }
4188         if (!conf->chunk_size || conf->chunk_size % 4) {
4189                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4190                         conf->chunk_size, mdname(mddev));
4191                 goto abort;
4192         }
4193         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4194                 printk(KERN_ERR
4195                         "raid5: unsupported parity algorithm %d for %s\n",
4196                         conf->algorithm, mdname(mddev));
4197                 goto abort;
4198         }
4199         if (mddev->degraded > conf->max_degraded) {
4200                 printk(KERN_ERR "raid5: not enough operational devices for %s"
4201                         " (%d/%d failed)\n",
4202                         mdname(mddev), mddev->degraded, conf->raid_disks);
4203                 goto abort;
4204         }
4205
4206         if (mddev->degraded > 0 &&
4207             mddev->recovery_cp != MaxSector) {
4208                 if (mddev->ok_start_degraded)
4209                         printk(KERN_WARNING
4210                                "raid5: starting dirty degraded array: %s"
4211                                "- data corruption possible.\n",
4212                                mdname(mddev));
4213                 else {
4214                         printk(KERN_ERR
4215                                "raid5: cannot start dirty degraded array for %s\n",
4216                                mdname(mddev));
4217                         goto abort;
4218                 }
4219         }
4220
4221         {
4222                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4223                 if (!mddev->thread) {
4224                         printk(KERN_ERR
4225                                 "raid5: couldn't allocate thread for %s\n",
4226                                 mdname(mddev));
4227                         goto abort;
4228                 }
4229         }
4230         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4231                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4232         if (grow_stripes(conf, conf->max_nr_stripes)) {
4233                 printk(KERN_ERR
4234                         "raid5: couldn't allocate %dkB for buffers\n", memory);
4235                 shrink_stripes(conf);
4236                 md_unregister_thread(mddev->thread);
4237                 goto abort;
4238         } else
4239                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4240                         memory, mdname(mddev));
4241
4242         if (mddev->degraded == 0)
4243                 printk("raid5: raid level %d set %s active with %d out of %d"
4244                         " devices, algorithm %d\n", conf->level, mdname(mddev),
4245                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4246                         conf->algorithm);
4247         else
4248                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4249                         " out of %d devices, algorithm %d\n", conf->level,
4250                         mdname(mddev), mddev->raid_disks - mddev->degraded,
4251                         mddev->raid_disks, conf->algorithm);
4252
4253         print_raid5_conf(conf);
4254
4255         if (conf->expand_progress != MaxSector) {
4256                 printk("...ok start reshape thread\n");
4257                 conf->expand_lo = conf->expand_progress;
4258                 atomic_set(&conf->reshape_stripes, 0);
4259                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4260                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4261                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4262                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4263                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4264                                                         "%s_reshape");
4265         }
4266
4267         /* read-ahead size must cover two whole stripes, which is
4268          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4269          */
4270         {
4271                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4272                 int stripe = data_disks *
4273                         (mddev->chunk_size / PAGE_SIZE);
4274                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4275                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4276         }
4277
4278         /* Ok, everything is just fine now */
4279         if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4280                 printk(KERN_WARNING
4281                        "raid5: failed to create sysfs attributes for %s\n",
4282                        mdname(mddev));
4283
4284         mddev->queue->unplug_fn = raid5_unplug_device;
4285         mddev->queue->issue_flush_fn = raid5_issue_flush;
4286         mddev->queue->backing_dev_info.congested_data = mddev;
4287         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4288
4289         mddev->array_size =  mddev->size * (conf->previous_raid_disks -
4290                                             conf->max_degraded);
4291
4292         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4293
4294         return 0;
4295 abort:
4296         if (conf) {
4297                 print_raid5_conf(conf);
4298                 safe_put_page(conf->spare_page);
4299                 kfree(conf->disks);
4300                 kfree(conf->stripe_hashtbl);
4301                 kfree(conf);
4302         }
4303         mddev->private = NULL;
4304         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4305         return -EIO;
4306 }
4307
4308
4309
4310 static int stop(mddev_t *mddev)
4311 {
4312         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4313
4314         md_unregister_thread(mddev->thread);
4315         mddev->thread = NULL;
4316         shrink_stripes(conf);
4317         kfree(conf->stripe_hashtbl);
4318         mddev->queue->backing_dev_info.congested_fn = NULL;
4319         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4320         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4321         kfree(conf->disks);
4322         kfree(conf);
4323         mddev->private = NULL;
4324         return 0;
4325 }
4326
4327 #ifdef DEBUG
4328 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
4329 {
4330         int i;
4331
4332         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4333                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4334         seq_printf(seq, "sh %llu,  count %d.\n",
4335                    (unsigned long long)sh->sector, atomic_read(&sh->count));
4336         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4337         for (i = 0; i < sh->disks; i++) {
4338                 seq_printf(seq, "(cache%d: %p %ld) ",
4339                            i, sh->dev[i].page, sh->dev[i].flags);
4340         }
4341         seq_printf(seq, "\n");
4342 }
4343
4344 static void printall (struct seq_file *seq, raid5_conf_t *conf)
4345 {
4346         struct stripe_head *sh;
4347         struct hlist_node *hn;
4348         int i;
4349
4350         spin_lock_irq(&conf->device_lock);
4351         for (i = 0; i < NR_HASH; i++) {
4352                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4353                         if (sh->raid_conf != conf)
4354                                 continue;
4355                         print_sh(seq, sh);
4356                 }
4357         }
4358         spin_unlock_irq(&conf->device_lock);
4359 }
4360 #endif
4361
4362 static void status (struct seq_file *seq, mddev_t *mddev)
4363 {
4364         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4365         int i;
4366
4367         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4368         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4369         for (i = 0; i < conf->raid_disks; i++)
4370                 seq_printf (seq, "%s",
4371                                conf->disks[i].rdev &&
4372                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4373         seq_printf (seq, "]");
4374 #ifdef DEBUG
4375         seq_printf (seq, "\n");
4376         printall(seq, conf);
4377 #endif
4378 }
4379
4380 static void print_raid5_conf (raid5_conf_t *conf)
4381 {
4382         int i;
4383         struct disk_info *tmp;
4384
4385         printk("RAID5 conf printout:\n");
4386         if (!conf) {
4387                 printk("(conf==NULL)\n");
4388                 return;
4389         }
4390         printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4391                  conf->raid_disks - conf->mddev->degraded);
4392
4393         for (i = 0; i < conf->raid_disks; i++) {
4394                 char b[BDEVNAME_SIZE];
4395                 tmp = conf->disks + i;
4396                 if (tmp->rdev)
4397                 printk(" disk %d, o:%d, dev:%s\n",
4398                         i, !test_bit(Faulty, &tmp->rdev->flags),
4399                         bdevname(tmp->rdev->bdev,b));
4400         }
4401 }
4402
4403 static int raid5_spare_active(mddev_t *mddev)
4404 {
4405         int i;
4406         raid5_conf_t *conf = mddev->private;
4407         struct disk_info *tmp;
4408
4409         for (i = 0; i < conf->raid_disks; i++) {
4410                 tmp = conf->disks + i;
4411                 if (tmp->rdev
4412                     && !test_bit(Faulty, &tmp->rdev->flags)
4413                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4414                         unsigned long flags;
4415                         spin_lock_irqsave(&conf->device_lock, flags);
4416                         mddev->degraded--;
4417                         spin_unlock_irqrestore(&conf->device_lock, flags);
4418                 }
4419         }
4420         print_raid5_conf(conf);
4421         return 0;
4422 }
4423
4424 static int raid5_remove_disk(mddev_t *mddev, int number)
4425 {
4426         raid5_conf_t *conf = mddev->private;
4427         int err = 0;
4428         mdk_rdev_t *rdev;
4429         struct disk_info *p = conf->disks + number;
4430
4431         print_raid5_conf(conf);
4432         rdev = p->rdev;
4433         if (rdev) {
4434                 if (test_bit(In_sync, &rdev->flags) ||
4435                     atomic_read(&rdev->nr_pending)) {
4436                         err = -EBUSY;
4437                         goto abort;
4438                 }
4439                 p->rdev = NULL;
4440                 synchronize_rcu();
4441                 if (atomic_read(&rdev->nr_pending)) {
4442                         /* lost the race, try later */
4443                         err = -EBUSY;
4444                         p->rdev = rdev;
4445                 }
4446         }
4447 abort:
4448
4449         print_raid5_conf(conf);
4450         return err;
4451 }
4452
4453 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4454 {
4455         raid5_conf_t *conf = mddev->private;
4456         int found = 0;
4457         int disk;
4458         struct disk_info *p;
4459
4460         if (mddev->degraded > conf->max_degraded)
4461                 /* no point adding a device */
4462                 return 0;
4463
4464         /*
4465          * find the disk ... but prefer rdev->saved_raid_disk
4466          * if possible.
4467          */
4468         if (rdev->saved_raid_disk >= 0 &&
4469             conf->disks[rdev->saved_raid_disk].rdev == NULL)
4470                 disk = rdev->saved_raid_disk;
4471         else
4472                 disk = 0;
4473         for ( ; disk < conf->raid_disks; disk++)
4474                 if ((p=conf->disks + disk)->rdev == NULL) {
4475                         clear_bit(In_sync, &rdev->flags);
4476                         rdev->raid_disk = disk;
4477                         found = 1;
4478                         if (rdev->saved_raid_disk != disk)
4479                                 conf->fullsync = 1;
4480                         rcu_assign_pointer(p->rdev, rdev);
4481                         break;
4482                 }
4483         print_raid5_conf(conf);
4484         return found;
4485 }
4486
4487 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4488 {
4489         /* no resync is happening, and there is enough space
4490          * on all devices, so we can resize.
4491          * We need to make sure resync covers any new space.
4492          * If the array is shrinking we should possibly wait until
4493          * any io in the removed space completes, but it hardly seems
4494          * worth it.
4495          */
4496         raid5_conf_t *conf = mddev_to_conf(mddev);
4497
4498         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4499         mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
4500         set_capacity(mddev->gendisk, mddev->array_size << 1);
4501         mddev->changed = 1;
4502         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
4503                 mddev->recovery_cp = mddev->size << 1;
4504                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4505         }
4506         mddev->size = sectors /2;
4507         mddev->resync_max_sectors = sectors;
4508         return 0;
4509 }
4510
4511 #ifdef CONFIG_MD_RAID5_RESHAPE
4512 static int raid5_check_reshape(mddev_t *mddev)
4513 {
4514         raid5_conf_t *conf = mddev_to_conf(mddev);
4515         int err;
4516
4517         if (mddev->delta_disks < 0 ||
4518             mddev->new_level != mddev->level)
4519                 return -EINVAL; /* Cannot shrink array or change level yet */
4520         if (mddev->delta_disks == 0)
4521                 return 0; /* nothing to do */
4522
4523         /* Can only proceed if there are plenty of stripe_heads.
4524          * We need a minimum of one full stripe,, and for sensible progress
4525          * it is best to have about 4 times that.
4526          * If we require 4 times, then the default 256 4K stripe_heads will
4527          * allow for chunk sizes up to 256K, which is probably OK.
4528          * If the chunk size is greater, user-space should request more
4529          * stripe_heads first.
4530          */
4531         if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4532             (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4533                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
4534                        (mddev->chunk_size / STRIPE_SIZE)*4);
4535                 return -ENOSPC;
4536         }
4537
4538         err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4539         if (err)
4540                 return err;
4541
4542         if (mddev->degraded > conf->max_degraded)
4543                 return -EINVAL;
4544         /* looks like we might be able to manage this */
4545         return 0;
4546 }
4547
4548 static int raid5_start_reshape(mddev_t *mddev)
4549 {
4550         raid5_conf_t *conf = mddev_to_conf(mddev);
4551         mdk_rdev_t *rdev;
4552         struct list_head *rtmp;
4553         int spares = 0;
4554         int added_devices = 0;
4555         unsigned long flags;
4556
4557         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4558                 return -EBUSY;
4559
4560         ITERATE_RDEV(mddev, rdev, rtmp)
4561                 if (rdev->raid_disk < 0 &&
4562                     !test_bit(Faulty, &rdev->flags))
4563                         spares++;
4564
4565         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4566                 /* Not enough devices even to make a degraded array
4567                  * of that size
4568                  */
4569                 return -EINVAL;
4570
4571         atomic_set(&conf->reshape_stripes, 0);
4572         spin_lock_irq(&conf->device_lock);
4573         conf->previous_raid_disks = conf->raid_disks;
4574         conf->raid_disks += mddev->delta_disks;
4575         conf->expand_progress = 0;
4576         conf->expand_lo = 0;
4577         spin_unlock_irq(&conf->device_lock);
4578
4579         /* Add some new drives, as many as will fit.
4580          * We know there are enough to make the newly sized array work.
4581          */
4582         ITERATE_RDEV(mddev, rdev, rtmp)
4583                 if (rdev->raid_disk < 0 &&
4584                     !test_bit(Faulty, &rdev->flags)) {
4585                         if (raid5_add_disk(mddev, rdev)) {
4586                                 char nm[20];
4587                                 set_bit(In_sync, &rdev->flags);
4588                                 added_devices++;
4589                                 rdev->recovery_offset = 0;
4590                                 sprintf(nm, "rd%d", rdev->raid_disk);
4591                                 if (sysfs_create_link(&mddev->kobj,
4592                                                       &rdev->kobj, nm))
4593                                         printk(KERN_WARNING
4594                                                "raid5: failed to create "
4595                                                " link %s for %s\n",
4596                                                nm, mdname(mddev));
4597                         } else
4598                                 break;
4599                 }
4600
4601         spin_lock_irqsave(&conf->device_lock, flags);
4602         mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4603         spin_unlock_irqrestore(&conf->device_lock, flags);
4604         mddev->raid_disks = conf->raid_disks;
4605         mddev->reshape_position = 0;
4606         set_bit(MD_CHANGE_DEVS, &mddev->flags);
4607
4608         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4609         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4610         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4611         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4612         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4613                                                 "%s_reshape");
4614         if (!mddev->sync_thread) {
4615                 mddev->recovery = 0;
4616                 spin_lock_irq(&conf->device_lock);
4617                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4618                 conf->expand_progress = MaxSector;
4619                 spin_unlock_irq(&conf->device_lock);
4620                 return -EAGAIN;
4621         }
4622         md_wakeup_thread(mddev->sync_thread);
4623         md_new_event(mddev);
4624         return 0;
4625 }
4626 #endif
4627
4628 static void end_reshape(raid5_conf_t *conf)
4629 {
4630         struct block_device *bdev;
4631
4632         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4633                 conf->mddev->array_size = conf->mddev->size *
4634                         (conf->raid_disks - conf->max_degraded);
4635                 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
4636                 conf->mddev->changed = 1;
4637
4638                 bdev = bdget_disk(conf->mddev->gendisk, 0);
4639                 if (bdev) {
4640                         mutex_lock(&bdev->bd_inode->i_mutex);
4641                         i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
4642                         mutex_unlock(&bdev->bd_inode->i_mutex);
4643                         bdput(bdev);
4644                 }
4645                 spin_lock_irq(&conf->device_lock);
4646                 conf->expand_progress = MaxSector;
4647                 spin_unlock_irq(&conf->device_lock);
4648                 conf->mddev->reshape_position = MaxSector;
4649
4650                 /* read-ahead size must cover two whole stripes, which is
4651                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4652                  */
4653                 {
4654                         int data_disks = conf->previous_raid_disks - conf->max_degraded;
4655                         int stripe = data_disks *
4656                                 (conf->mddev->chunk_size / PAGE_SIZE);
4657                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4658                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4659                 }
4660         }
4661 }
4662
4663 static void raid5_quiesce(mddev_t *mddev, int state)
4664 {
4665         raid5_conf_t *conf = mddev_to_conf(mddev);
4666
4667         switch(state) {
4668         case 2: /* resume for a suspend */
4669                 wake_up(&conf->wait_for_overlap);
4670                 break;
4671
4672         case 1: /* stop all writes */
4673                 spin_lock_irq(&conf->device_lock);
4674                 conf->quiesce = 1;
4675                 wait_event_lock_irq(conf->wait_for_stripe,
4676                                     atomic_read(&conf->active_stripes) == 0 &&
4677                                     atomic_read(&conf->active_aligned_reads) == 0,
4678                                     conf->device_lock, /* nothing */);
4679                 spin_unlock_irq(&conf->device_lock);
4680                 break;
4681
4682         case 0: /* re-enable writes */
4683                 spin_lock_irq(&conf->device_lock);
4684                 conf->quiesce = 0;
4685                 wake_up(&conf->wait_for_stripe);
4686                 wake_up(&conf->wait_for_overlap);
4687                 spin_unlock_irq(&conf->device_lock);
4688                 break;
4689         }
4690 }
4691
4692 static struct mdk_personality raid6_personality =
4693 {
4694         .name           = "raid6",
4695         .level          = 6,
4696         .owner          = THIS_MODULE,
4697         .make_request   = make_request,
4698         .run            = run,
4699         .stop           = stop,
4700         .status         = status,
4701         .error_handler  = error,
4702         .hot_add_disk   = raid5_add_disk,
4703         .hot_remove_disk= raid5_remove_disk,
4704         .spare_active   = raid5_spare_active,
4705         .sync_request   = sync_request,
4706         .resize         = raid5_resize,
4707 #ifdef CONFIG_MD_RAID5_RESHAPE
4708         .check_reshape  = raid5_check_reshape,
4709         .start_reshape  = raid5_start_reshape,
4710 #endif
4711         .quiesce        = raid5_quiesce,
4712 };
4713 static struct mdk_personality raid5_personality =
4714 {
4715         .name           = "raid5",
4716         .level          = 5,
4717         .owner          = THIS_MODULE,
4718         .make_request   = make_request,
4719         .run            = run,
4720         .stop           = stop,
4721         .status         = status,
4722         .error_handler  = error,
4723         .hot_add_disk   = raid5_add_disk,
4724         .hot_remove_disk= raid5_remove_disk,
4725         .spare_active   = raid5_spare_active,
4726         .sync_request   = sync_request,
4727         .resize         = raid5_resize,
4728 #ifdef CONFIG_MD_RAID5_RESHAPE
4729         .check_reshape  = raid5_check_reshape,
4730         .start_reshape  = raid5_start_reshape,
4731 #endif
4732         .quiesce        = raid5_quiesce,
4733 };
4734
4735 static struct mdk_personality raid4_personality =
4736 {
4737         .name           = "raid4",
4738         .level          = 4,
4739         .owner          = THIS_MODULE,
4740         .make_request   = make_request,
4741         .run            = run,
4742         .stop           = stop,
4743         .status         = status,
4744         .error_handler  = error,
4745         .hot_add_disk   = raid5_add_disk,
4746         .hot_remove_disk= raid5_remove_disk,
4747         .spare_active   = raid5_spare_active,
4748         .sync_request   = sync_request,
4749         .resize         = raid5_resize,
4750 #ifdef CONFIG_MD_RAID5_RESHAPE
4751         .check_reshape  = raid5_check_reshape,
4752         .start_reshape  = raid5_start_reshape,
4753 #endif
4754         .quiesce        = raid5_quiesce,
4755 };
4756
4757 static int __init raid5_init(void)
4758 {
4759         int e;
4760
4761         e = raid6_select_algo();
4762         if ( e )
4763                 return e;
4764         register_md_personality(&raid6_personality);
4765         register_md_personality(&raid5_personality);
4766         register_md_personality(&raid4_personality);
4767         return 0;
4768 }
4769
4770 static void raid5_exit(void)
4771 {
4772         unregister_md_personality(&raid6_personality);
4773         unregister_md_personality(&raid5_personality);
4774         unregister_md_personality(&raid4_personality);
4775 }
4776
4777 module_init(raid5_init);
4778 module_exit(raid5_exit);
4779 MODULE_LICENSE("GPL");
4780 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4781 MODULE_ALIAS("md-raid5");
4782 MODULE_ALIAS("md-raid4");
4783 MODULE_ALIAS("md-level-5");
4784 MODULE_ALIAS("md-level-4");
4785 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4786 MODULE_ALIAS("md-raid6");
4787 MODULE_ALIAS("md-level-6");
4788
4789 /* This used to be two separate modules, they were: */
4790 MODULE_ALIAS("raid5");
4791 MODULE_ALIAS("raid6");