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/blkdev.h>
  47 #include <linux/kthread.h>
  48 #include <linux/raid/pq.h>
  49 #include <linux/async_tx.h>
  50 #include <linux/async.h>
  51 #include <linux/seq_file.h>
  52 #include <linux/cpu.h>
  53 #include <linux/slab.h>
  54 #include "md.h"
  55 #include "raid5.h"
  56 #include "raid0.h"
  57 #include "bitmap.h"
  58
  59 /*
  60  * Stripe cache
  61  */
  62
  63 #define NR_STRIPES              256
  64 #define STRIPE_SIZE             PAGE_SIZE
  65 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
  66 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
  67 #define IO_THRESHOLD            1
  68 #define BYPASS_THRESHOLD        1
  69 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
  70 #define HASH_MASK               (NR_HASH - 1)
  71
  72 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
  73
  74 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
  75  * order without overlap.  There may be several bio's per stripe+device, and
  76  * a bio could span several devices.
  77  * When walking this list for a particular stripe+device, we must never proceed
  78  * beyond a bio that extends past this device, as the next bio might no longer
  79  * be valid.
  80  * This macro is used to determine the 'next' bio in the list, given the sector
  81  * of the current stripe+device
  82  */
  83 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
  84 /*
  85  * The following can be used to debug the driver
  86  */
  87 #define RAID5_PARANOIA  1
  88 #if RAID5_PARANOIA && defined(CONFIG_SMP)
  89 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
  90 #else
  91 # define CHECK_DEVLOCK()
  92 #endif
  93
  94 #ifdef DEBUG
  95 #define inline
  96 #define __inline__
  97 #endif
  98
  99 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
 100
 101 /*
 102  * We maintain a biased count of active stripes in the bottom 16 bits of
 103  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
 104  */
 105 static inline int raid5_bi_phys_segments(struct bio *bio)
 106 {
 107         return bio->bi_phys_segments & 0xffff;
 108 }
 109
 110 static inline int raid5_bi_hw_segments(struct bio *bio)
 111 {
 112         return (bio->bi_phys_segments >> 16) & 0xffff;
 113 }
 114
 115 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
 116 {
 117         --bio->bi_phys_segments;
 118         return raid5_bi_phys_segments(bio);
 119 }
 120
 121 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
 122 {
 123         unsigned short val = raid5_bi_hw_segments(bio);
 124
 125         --val;
 126         bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
 127         return val;
 128 }
 129
 130 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
 131 {
 132         bio->bi_phys_segments = raid5_bi_phys_segments(bio) || (cnt << 16);
 133 }
 134
 135 /* Find first data disk in a raid6 stripe */
 136 static inline int raid6_d0(struct stripe_head *sh)
 137 {
 138         if (sh->ddf_layout)
 139                 /* ddf always start from first device */
 140                 return 0;
 141         /* md starts just after Q block */
 142         if (sh->qd_idx == sh->disks - 1)
 143                 return 0;
 144         else
 145                 return sh->qd_idx + 1;
 146 }
 147 static inline int raid6_next_disk(int disk, int raid_disks)
 148 {
 149         disk++;
 150         return (disk < raid_disks) ? disk : 0;
 151 }
 152
 153 /* When walking through the disks in a raid5, starting at raid6_d0,
 154  * We need to map each disk to a 'slot', where the data disks are slot
 155  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 156  * is raid_disks-1.  This help does that mapping.
 157  */
 158 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 159                              int *count, int syndrome_disks)
 160 {
 161         int slot = *count;
 162
 163         if (sh->ddf_layout)
 164                 (*count)++;
 165         if (idx == sh->pd_idx)
 166                 return syndrome_disks;
 167         if (idx == sh->qd_idx)
 168                 return syndrome_disks + 1;
 169         if (!sh->ddf_layout)
 170                 (*count)++;
 171         return slot;
 172 }
 173
 174 static void return_io(struct bio *return_bi)
 175 {
 176         struct bio *bi = return_bi;
 177         while (bi) {
 178
 179                 return_bi = bi->bi_next;
 180                 bi->bi_next = NULL;
 181                 bi->bi_size = 0;
 182                 bio_endio(bi, 0);
 183                 bi = return_bi;
 184         }
 185 }
 186
 187 static void print_raid5_conf (raid5_conf_t *conf);
 188
 189 static int stripe_operations_active(struct stripe_head *sh)
 190 {
 191         return sh->check_state || sh->reconstruct_state ||
 192                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 193                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 194 }
 195
 196 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
 197 {
 198         if (atomic_dec_and_test(&sh->count)) {
 199                 BUG_ON(!list_empty(&sh->lru));
 200                 BUG_ON(atomic_read(&conf->active_stripes)==0);
 201                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
 202                         if (test_bit(STRIPE_DELAYED, &sh->state)) {
 203                                 list_add_tail(&sh->lru, &conf->delayed_list);
 204                                 blk_plug_device(conf->mddev->queue);
 205                         } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 206                                    sh->bm_seq - conf->seq_write > 0) {
 207                                 list_add_tail(&sh->lru, &conf->bitmap_list);
 208                                 blk_plug_device(conf->mddev->queue);
 209                         } else {
 210                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
 211                                 list_add_tail(&sh->lru, &conf->handle_list);
 212                         }
 213                         md_wakeup_thread(conf->mddev->thread);
 214                 } else {
 215                         BUG_ON(stripe_operations_active(sh));
 216                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
 217                                 atomic_dec(&conf->preread_active_stripes);
 218                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
 219                                         md_wakeup_thread(conf->mddev->thread);
 220                         }
 221                         atomic_dec(&conf->active_stripes);
 222                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
 223                                 list_add_tail(&sh->lru, &conf->inactive_list);
 224                                 wake_up(&conf->wait_for_stripe);
 225                                 if (conf->retry_read_aligned)
 226                                         md_wakeup_thread(conf->mddev->thread);
 227                         }
 228                 }
 229         }
 230 }
 231
 232 static void release_stripe(struct stripe_head *sh)
 233 {
 234         raid5_conf_t *conf = sh->raid_conf;
 235         unsigned long flags;
 236
 237         spin_lock_irqsave(&conf->device_lock, flags);
 238         __release_stripe(conf, sh);
 239         spin_unlock_irqrestore(&conf->device_lock, flags);
 240 }
 241
 242 static inline void remove_hash(struct stripe_head *sh)
 243 {
 244         pr_debug("remove_hash(), stripe %llu\n",
 245                 (unsigned long long)sh->sector);
 246
 247         hlist_del_init(&sh->hash);
 248 }
 249
 250 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
 251 {
 252         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 253
 254         pr_debug("insert_hash(), stripe %llu\n",
 255                 (unsigned long long)sh->sector);
 256
 257         CHECK_DEVLOCK();
 258         hlist_add_head(&sh->hash, hp);
 259 }
 260
 261
 262 /* find an idle stripe, make sure it is unhashed, and return it. */
 263 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
 264 {
 265         struct stripe_head *sh = NULL;
 266         struct list_head *first;
 267
 268         CHECK_DEVLOCK();
 269         if (list_empty(&conf->inactive_list))
 270                 goto out;
 271         first = conf->inactive_list.next;
 272         sh = list_entry(first, struct stripe_head, lru);
 273         list_del_init(first);
 274         remove_hash(sh);
 275         atomic_inc(&conf->active_stripes);
 276 out:
 277         return sh;
 278 }
 279
 280 static void shrink_buffers(struct stripe_head *sh)
 281 {
 282         struct page *p;
 283         int i;
 284         int num = sh->raid_conf->pool_size;
 285
 286         for (i = 0; i < num ; i++) {
 287                 p = sh->dev[i].page;
 288                 if (!p)
 289                         continue;
 290                 sh->dev[i].page = NULL;
 291                 put_page(p);
 292         }
 293 }
 294
 295 static int grow_buffers(struct stripe_head *sh)
 296 {
 297         int i;
 298         int num = sh->raid_conf->pool_size;
 299
 300         for (i = 0; i < num; i++) {
 301                 struct page *page;
 302
 303                 if (!(page = alloc_page(GFP_KERNEL))) {
 304                         return 1;
 305                 }
 306                 sh->dev[i].page = page;
 307         }
 308         return 0;
 309 }
 310
 311 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
 312 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
 313                             struct stripe_head *sh);
 314
 315 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 316 {
 317         raid5_conf_t *conf = sh->raid_conf;
 318         int i;
 319
 320         BUG_ON(atomic_read(&sh->count) != 0);
 321         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 322         BUG_ON(stripe_operations_active(sh));
 323
 324         CHECK_DEVLOCK();
 325         pr_debug("init_stripe called, stripe %llu\n",
 326                 (unsigned long long)sh->sector);
 327
 328         remove_hash(sh);
 329
 330         sh->generation = conf->generation - previous;
 331         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 332         sh->sector = sector;
 333         stripe_set_idx(sector, conf, previous, sh);
 334         sh->state = 0;
 335
 336
 337         for (i = sh->disks; i--; ) {
 338                 struct r5dev *dev = &sh->dev[i];
 339
 340                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 341                     test_bit(R5_LOCKED, &dev->flags)) {
 342                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
 343                                (unsigned long long)sh->sector, i, dev->toread,
 344                                dev->read, dev->towrite, dev->written,
 345                                test_bit(R5_LOCKED, &dev->flags));
 346                         BUG();
 347                 }
 348                 dev->flags = 0;
 349                 raid5_build_block(sh, i, previous);
 350         }
 351         insert_hash(conf, sh);
 352 }
 353
 354 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
 355                                          short generation)
 356 {
 357         struct stripe_head *sh;
 358         struct hlist_node *hn;
 359
 360         CHECK_DEVLOCK();
 361         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 362         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
 363                 if (sh->sector == sector && sh->generation == generation)
 364                         return sh;
 365         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 366         return NULL;
 367 }
 368
 369 static void unplug_slaves(mddev_t *mddev);
 370 static void raid5_unplug_device(struct request_queue *q);
 371
 372 static struct stripe_head *
 373 get_active_stripe(raid5_conf_t *conf, sector_t sector,
 374                   int previous, int noblock, int noquiesce)
 375 {
 376         struct stripe_head *sh;
 377
 378         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 379
 380         spin_lock_irq(&conf->device_lock);
 381
 382         do {
 383                 wait_event_lock_irq(conf->wait_for_stripe,
 384                                     conf->quiesce == 0 || noquiesce,
 385                                     conf->device_lock, /* nothing */);
 386                 sh = __find_stripe(conf, sector, conf->generation - previous);
 387                 if (!sh) {
 388                         if (!conf->inactive_blocked)
 389                                 sh = get_free_stripe(conf);
 390                         if (noblock && sh == NULL)
 391                                 break;
 392                         if (!sh) {
 393                                 conf->inactive_blocked = 1;
 394                                 wait_event_lock_irq(conf->wait_for_stripe,
 395                                                     !list_empty(&conf->inactive_list) &&
 396                                                     (atomic_read(&conf->active_stripes)
 397                                                      < (conf->max_nr_stripes *3/4)
 398                                                      || !conf->inactive_blocked),
 399                                                     conf->device_lock,
 400                                                     raid5_unplug_device(conf->mddev->queue)
 401                                         );
 402                                 conf->inactive_blocked = 0;
 403                         } else
 404                                 init_stripe(sh, sector, previous);
 405                 } else {
 406                         if (atomic_read(&sh->count)) {
 407                                 BUG_ON(!list_empty(&sh->lru)
 408                                     && !test_bit(STRIPE_EXPANDING, &sh->state));
 409                         } else {
 410                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 411                                         atomic_inc(&conf->active_stripes);
 412                                 if (list_empty(&sh->lru) &&
 413                                     !test_bit(STRIPE_EXPANDING, &sh->state))
 414                                         BUG();
 415                                 list_del_init(&sh->lru);
 416                         }
 417                 }
 418         } while (sh == NULL);
 419
 420         if (sh)
 421                 atomic_inc(&sh->count);
 422
 423         spin_unlock_irq(&conf->device_lock);
 424         return sh;
 425 }
 426
 427 static void
 428 raid5_end_read_request(struct bio *bi, int error);
 429 static void
 430 raid5_end_write_request(struct bio *bi, int error);
 431
 432 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
 433 {
 434         raid5_conf_t *conf = sh->raid_conf;
 435         int i, disks = sh->disks;
 436
 437         might_sleep();
 438
 439         for (i = disks; i--; ) {
 440                 int rw;
 441                 struct bio *bi;
 442                 mdk_rdev_t *rdev;
 443                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
 444                         rw = WRITE;
 445                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
 446                         rw = READ;
 447                 else
 448                         continue;
 449
 450                 bi = &sh->dev[i].req;
 451
 452                 bi->bi_rw = rw;
 453                 if (rw == WRITE)
 454                         bi->bi_end_io = raid5_end_write_request;
 455                 else
 456                         bi->bi_end_io = raid5_end_read_request;
 457
 458                 rcu_read_lock();
 459                 rdev = rcu_dereference(conf->disks[i].rdev);
 460                 if (rdev && test_bit(Faulty, &rdev->flags))
 461                         rdev = NULL;
 462                 if (rdev)
 463                         atomic_inc(&rdev->nr_pending);
 464                 rcu_read_unlock();
 465
 466                 if (rdev) {
 467                         if (s->syncing || s->expanding || s->expanded)
 468                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
 469
 470                         set_bit(STRIPE_IO_STARTED, &sh->state);
 471
 472                         bi->bi_bdev = rdev->bdev;
 473                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
 474                                 __func__, (unsigned long long)sh->sector,
 475                                 bi->bi_rw, i);
 476                         atomic_inc(&sh->count);
 477                         bi->bi_sector = sh->sector + rdev->data_offset;
 478                         bi->bi_flags = 1 << BIO_UPTODATE;
 479                         bi->bi_vcnt = 1;
 480                         bi->bi_max_vecs = 1;
 481                         bi->bi_idx = 0;
 482                         bi->bi_io_vec = &sh->dev[i].vec;
 483                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
 484                         bi->bi_io_vec[0].bv_offset = 0;
 485                         bi->bi_size = STRIPE_SIZE;
 486                         bi->bi_next = NULL;
 487                         if (rw == WRITE &&
 488                             test_bit(R5_ReWrite, &sh->dev[i].flags))
 489                                 atomic_add(STRIPE_SECTORS,
 490                                         &rdev->corrected_errors);
 491                         generic_make_request(bi);
 492                 } else {
 493                         if (rw == WRITE)
 494                                 set_bit(STRIPE_DEGRADED, &sh->state);
 495                         pr_debug("skip op %ld on disc %d for sector %llu\n",
 496                                 bi->bi_rw, i, (unsigned long long)sh->sector);
 497                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 498                         set_bit(STRIPE_HANDLE, &sh->state);
 499                 }
 500         }
 501 }
 502
 503 static struct dma_async_tx_descriptor *
 504 async_copy_data(int frombio, struct bio *bio, struct page *page,
 505         sector_t sector, struct dma_async_tx_descriptor *tx)
 506 {
 507         struct bio_vec *bvl;
 508         struct page *bio_page;
 509         int i;
 510         int page_offset;
 511         struct async_submit_ctl submit;
 512         enum async_tx_flags flags = 0;
 513
 514         if (bio->bi_sector >= sector)
 515                 page_offset = (signed)(bio->bi_sector - sector) * 512;
 516         else
 517                 page_offset = (signed)(sector - bio->bi_sector) * -512;
 518
 519         if (frombio)
 520                 flags |= ASYNC_TX_FENCE;
 521         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
 522
 523         bio_for_each_segment(bvl, bio, i) {
 524                 int len = bio_iovec_idx(bio, i)->bv_len;
 525                 int clen;
 526                 int b_offset = 0;
 527
 528                 if (page_offset < 0) {
 529                         b_offset = -page_offset;
 530                         page_offset += b_offset;
 531                         len -= b_offset;
 532                 }
 533
 534                 if (len > 0 && page_offset + len > STRIPE_SIZE)
 535                         clen = STRIPE_SIZE - page_offset;
 536                 else
 537                         clen = len;
 538
 539                 if (clen > 0) {
 540                         b_offset += bio_iovec_idx(bio, i)->bv_offset;
 541                         bio_page = bio_iovec_idx(bio, i)->bv_page;
 542                         if (frombio)
 543                                 tx = async_memcpy(page, bio_page, page_offset,
 544                                                   b_offset, clen, &submit);
 545                         else
 546                                 tx = async_memcpy(bio_page, page, b_offset,
 547                                                   page_offset, clen, &submit);
 548                 }
 549                 /* chain the operations */
 550                 submit.depend_tx = tx;
 551
 552                 if (clen < len) /* hit end of page */
 553                         break;
 554                 page_offset +=  len;
 555         }
 556
 557         return tx;
 558 }
 559
 560 static void ops_complete_biofill(void *stripe_head_ref)
 561 {
 562         struct stripe_head *sh = stripe_head_ref;
 563         struct bio *return_bi = NULL;
 564         raid5_conf_t *conf = sh->raid_conf;
 565         int i;
 566
 567         pr_debug("%s: stripe %llu\n", __func__,
 568                 (unsigned long long)sh->sector);
 569
 570         /* clear completed biofills */
 571         spin_lock_irq(&conf->device_lock);
 572         for (i = sh->disks; i--; ) {
 573                 struct r5dev *dev = &sh->dev[i];
 574
 575                 /* acknowledge completion of a biofill operation */
 576                 /* and check if we need to reply to a read request,
 577                  * new R5_Wantfill requests are held off until
 578                  * !STRIPE_BIOFILL_RUN
 579                  */
 580                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
 581                         struct bio *rbi, *rbi2;
 582
 583                         BUG_ON(!dev->read);
 584                         rbi = dev->read;
 585                         dev->read = NULL;
 586                         while (rbi && rbi->bi_sector <
 587                                 dev->sector + STRIPE_SECTORS) {
 588                                 rbi2 = r5_next_bio(rbi, dev->sector);
 589                                 if (!raid5_dec_bi_phys_segments(rbi)) {
 590                                         rbi->bi_next = return_bi;
 591                                         return_bi = rbi;
 592                                 }
 593                                 rbi = rbi2;
 594                         }
 595                 }
 596         }
 597         spin_unlock_irq(&conf->device_lock);
 598         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
 599
 600         return_io(return_bi);
 601
 602         set_bit(STRIPE_HANDLE, &sh->state);
 603         release_stripe(sh);
 604 }
 605
 606 static void ops_run_biofill(struct stripe_head *sh)
 607 {
 608         struct dma_async_tx_descriptor *tx = NULL;
 609         raid5_conf_t *conf = sh->raid_conf;
 610         struct async_submit_ctl submit;
 611         int i;
 612
 613         pr_debug("%s: stripe %llu\n", __func__,
 614                 (unsigned long long)sh->sector);
 615
 616         for (i = sh->disks; i--; ) {
 617                 struct r5dev *dev = &sh->dev[i];
 618                 if (test_bit(R5_Wantfill, &dev->flags)) {
 619                         struct bio *rbi;
 620                         spin_lock_irq(&conf->device_lock);
 621                         dev->read = rbi = dev->toread;
 622                         dev->toread = NULL;
 623                         spin_unlock_irq(&conf->device_lock);
 624                         while (rbi && rbi->bi_sector <
 625                                 dev->sector + STRIPE_SECTORS) {
 626                                 tx = async_copy_data(0, rbi, dev->page,
 627                                         dev->sector, tx);
 628                                 rbi = r5_next_bio(rbi, dev->sector);
 629                         }
 630                 }
 631         }
 632
 633         atomic_inc(&sh->count);
 634         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
 635         async_trigger_callback(&submit);
 636 }
 637
 638 static void mark_target_uptodate(struct stripe_head *sh, int target)
 639 {
 640         struct r5dev *tgt;
 641
 642         if (target < 0)
 643                 return;
 644
 645         tgt = &sh->dev[target];
 646         set_bit(R5_UPTODATE, &tgt->flags);
 647         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
 648         clear_bit(R5_Wantcompute, &tgt->flags);
 649 }
 650
 651 static void ops_complete_compute(void *stripe_head_ref)
 652 {
 653         struct stripe_head *sh = stripe_head_ref;
 654
 655         pr_debug("%s: stripe %llu\n", __func__,
 656                 (unsigned long long)sh->sector);
 657
 658         /* mark the computed target(s) as uptodate */
 659         mark_target_uptodate(sh, sh->ops.target);
 660         mark_target_uptodate(sh, sh->ops.target2);
 661
 662         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
 663         if (sh->check_state == check_state_compute_run)
 664                 sh->check_state = check_state_compute_result;
 665         set_bit(STRIPE_HANDLE, &sh->state);
 666         release_stripe(sh);
 667 }
 668
 669 /* return a pointer to the address conversion region of the scribble buffer */
 670 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
 671                                  struct raid5_percpu *percpu)
 672 {
 673         return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
 674 }
 675
 676 static struct dma_async_tx_descriptor *
 677 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
 678 {
 679         int disks = sh->disks;
 680         struct page **xor_srcs = percpu->scribble;
 681         int target = sh->ops.target;
 682         struct r5dev *tgt = &sh->dev[target];
 683         struct page *xor_dest = tgt->page;
 684         int count = 0;
 685         struct dma_async_tx_descriptor *tx;
 686         struct async_submit_ctl submit;
 687         int i;
 688
 689         pr_debug("%s: stripe %llu block: %d\n",
 690                 __func__, (unsigned long long)sh->sector, target);
 691         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
 692
 693         for (i = disks; i--; )
 694                 if (i != target)
 695                         xor_srcs[count++] = sh->dev[i].page;
 696
 697         atomic_inc(&sh->count);
 698
 699         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
 700                           ops_complete_compute, sh, to_addr_conv(sh, percpu));
 701         if (unlikely(count == 1))
 702                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
 703         else
 704                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
 705
 706         return tx;
 707 }
 708
 709 /* set_syndrome_sources - populate source buffers for gen_syndrome
 710  * @srcs - (struct page *) array of size sh->disks
 711  * @sh - stripe_head to parse
 712  *
 713  * Populates srcs in proper layout order for the stripe and returns the
 714  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
 715  * destination buffer is recorded in srcs[count] and the Q destination
 716  * is recorded in srcs[count+1]].
 717  */
 718 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
 719 {
 720         int disks = sh->disks;
 721         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
 722         int d0_idx = raid6_d0(sh);
 723         int count;
 724         int i;
 725
 726         for (i = 0; i < disks; i++)
 727                 srcs[i] = NULL;
 728
 729         count = 0;
 730         i = d0_idx;
 731         do {
 732                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
 733
 734                 srcs[slot] = sh->dev[i].page;
 735                 i = raid6_next_disk(i, disks);
 736         } while (i != d0_idx);
 737
 738         return syndrome_disks;
 739 }
 740
 741 static struct dma_async_tx_descriptor *
 742 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
 743 {
 744         int disks = sh->disks;
 745         struct page **blocks = percpu->scribble;
 746         int target;
 747         int qd_idx = sh->qd_idx;
 748         struct dma_async_tx_descriptor *tx;
 749         struct async_submit_ctl submit;
 750         struct r5dev *tgt;
 751         struct page *dest;
 752         int i;
 753         int count;
 754
 755         if (sh->ops.target < 0)
 756                 target = sh->ops.target2;
 757         else if (sh->ops.target2 < 0)
 758                 target = sh->ops.target;
 759         else
 760                 /* we should only have one valid target */
 761                 BUG();
 762         BUG_ON(target < 0);
 763         pr_debug("%s: stripe %llu block: %d\n",
 764                 __func__, (unsigned long long)sh->sector, target);
 765
 766         tgt = &sh->dev[target];
 767         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
 768         dest = tgt->page;
 769
 770         atomic_inc(&sh->count);
 771
 772         if (target == qd_idx) {
 773                 count = set_syndrome_sources(blocks, sh);
 774                 blocks[count] = NULL; /* regenerating p is not necessary */
 775                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
 776                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
 777                                   ops_complete_compute, sh,
 778                                   to_addr_conv(sh, percpu));
 779                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
 780         } else {
 781                 /* Compute any data- or p-drive using XOR */
 782                 count = 0;
 783                 for (i = disks; i-- ; ) {
 784                         if (i == target || i == qd_idx)
 785                                 continue;
 786                         blocks[count++] = sh->dev[i].page;
 787                 }
 788
 789                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
 790                                   NULL, ops_complete_compute, sh,
 791                                   to_addr_conv(sh, percpu));
 792                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
 793         }
 794
 795         return tx;
 796 }
 797
 798 static struct dma_async_tx_descriptor *
 799 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
 800 {
 801         int i, count, disks = sh->disks;
 802         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
 803         int d0_idx = raid6_d0(sh);
 804         int faila = -1, failb = -1;
 805         int target = sh->ops.target;
 806         int target2 = sh->ops.target2;
 807         struct r5dev *tgt = &sh->dev[target];
 808         struct r5dev *tgt2 = &sh->dev[target2];
 809         struct dma_async_tx_descriptor *tx;
 810         struct page **blocks = percpu->scribble;
 811         struct async_submit_ctl submit;
 812
 813         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
 814                  __func__, (unsigned long long)sh->sector, target, target2);
 815         BUG_ON(target < 0 || target2 < 0);
 816         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
 817         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
 818
 819         /* we need to open-code set_syndrome_sources to handle the
 820          * slot number conversion for 'faila' and 'failb'
 821          */
 822         for (i = 0; i < disks ; i++)
 823                 blocks[i] = NULL;
 824         count = 0;
 825         i = d0_idx;
 826         do {
 827                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
 828
 829                 blocks[slot] = sh->dev[i].page;
 830
 831                 if (i == target)
 832                         faila = slot;
 833                 if (i == target2)
 834                         failb = slot;
 835                 i = raid6_next_disk(i, disks);
 836         } while (i != d0_idx);
 837
 838         BUG_ON(faila == failb);
 839         if (failb < faila)
 840                 swap(faila, failb);
 841         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
 842                  __func__, (unsigned long long)sh->sector, faila, failb);
 843
 844         atomic_inc(&sh->count);
 845
 846         if (failb == syndrome_disks+1) {
 847                 /* Q disk is one of the missing disks */
 848                 if (faila == syndrome_disks) {
 849                         /* Missing P+Q, just recompute */
 850                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
 851                                           ops_complete_compute, sh,
 852                                           to_addr_conv(sh, percpu));
 853                         return async_gen_syndrome(blocks, 0, syndrome_disks+2,
 854                                                   STRIPE_SIZE, &submit);
 855                 } else {
 856                         struct page *dest;
 857                         int data_target;
 858                         int qd_idx = sh->qd_idx;
 859
 860                         /* Missing D+Q: recompute D from P, then recompute Q */
 861                         if (target == qd_idx)
 862                                 data_target = target2;
 863                         else
 864                                 data_target = target;
 865
 866                         count = 0;
 867                         for (i = disks; i-- ; ) {
 868                                 if (i == data_target || i == qd_idx)
 869                                         continue;
 870                                 blocks[count++] = sh->dev[i].page;
 871                         }
 872                         dest = sh->dev[data_target].page;
 873                         init_async_submit(&submit,
 874                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
 875                                           NULL, NULL, NULL,
 876                                           to_addr_conv(sh, percpu));
 877                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
 878                                        &submit);
 879
 880                         count = set_syndrome_sources(blocks, sh);
 881                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
 882                                           ops_complete_compute, sh,
 883                                           to_addr_conv(sh, percpu));
 884                         return async_gen_syndrome(blocks, 0, count+2,
 885                                                   STRIPE_SIZE, &submit);
 886                 }
 887         } else {
 888                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
 889                                   ops_complete_compute, sh,
 890                                   to_addr_conv(sh, percpu));
 891                 if (failb == syndrome_disks) {
 892                         /* We're missing D+P. */
 893                         return async_raid6_datap_recov(syndrome_disks+2,
 894                                                        STRIPE_SIZE, faila,
 895                                                        blocks, &submit);
 896                 } else {
 897                         /* We're missing D+D. */
 898                         return async_raid6_2data_recov(syndrome_disks+2,
 899                                                        STRIPE_SIZE, faila, failb,
 900                                                        blocks, &submit);
 901                 }
 902         }
 903 }
 904
 905
 906 static void ops_complete_prexor(void *stripe_head_ref)
 907 {
 908         struct stripe_head *sh = stripe_head_ref;
 909
 910         pr_debug("%s: stripe %llu\n", __func__,
 911                 (unsigned long long)sh->sector);
 912 }
 913
 914 static struct dma_async_tx_descriptor *
 915 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
 916                struct dma_async_tx_descriptor *tx)
 917 {
 918         int disks = sh->disks;
 919         struct page **xor_srcs = percpu->scribble;
 920         int count = 0, pd_idx = sh->pd_idx, i;
 921         struct async_submit_ctl submit;
 922
 923         /* existing parity data subtracted */
 924         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
 925
 926         pr_debug("%s: stripe %llu\n", __func__,
 927                 (unsigned long long)sh->sector);
 928
 929         for (i = disks; i--; ) {
 930                 struct r5dev *dev = &sh->dev[i];
 931                 /* Only process blocks that are known to be uptodate */
 932                 if (test_bit(R5_Wantdrain, &dev->flags))
 933                         xor_srcs[count++] = dev->page;
 934         }
 935
 936         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
 937                           ops_complete_prexor, sh, to_addr_conv(sh, percpu));
 938         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
 939
 940         return tx;
 941 }
 942
 943 static struct dma_async_tx_descriptor *
 944 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
 945 {
 946         int disks = sh->disks;
 947         int i;
 948
 949         pr_debug("%s: stripe %llu\n", __func__,
 950                 (unsigned long long)sh->sector);
 951
 952         for (i = disks; i--; ) {
 953                 struct r5dev *dev = &sh->dev[i];
 954                 struct bio *chosen;
 955
 956                 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
 957                         struct bio *wbi;
 958
 959                         spin_lock(&sh->lock);
 960                         chosen = dev->towrite;
 961                         dev->towrite = NULL;
 962                         BUG_ON(dev->written);
 963                         wbi = dev->written = chosen;
 964                         spin_unlock(&sh->lock);
 965
 966                         while (wbi && wbi->bi_sector <
 967                                 dev->sector + STRIPE_SECTORS) {
 968                                 tx = async_copy_data(1, wbi, dev->page,
 969                                         dev->sector, tx);
 970                                 wbi = r5_next_bio(wbi, dev->sector);
 971                         }
 972                 }
 973         }
 974
 975         return tx;
 976 }
 977
 978 static void ops_complete_reconstruct(void *stripe_head_ref)
 979 {
 980         struct stripe_head *sh = stripe_head_ref;
 981         int disks = sh->disks;
 982         int pd_idx = sh->pd_idx;
 983         int qd_idx = sh->qd_idx;
 984         int i;
 985
 986         pr_debug("%s: stripe %llu\n", __func__,
 987                 (unsigned long long)sh->sector);
 988
 989         for (i = disks; i--; ) {
 990                 struct r5dev *dev = &sh->dev[i];
 991
 992                 if (dev->written || i == pd_idx || i == qd_idx)
 993                         set_bit(R5_UPTODATE, &dev->flags);
 994         }
 995
 996         if (sh->reconstruct_state == reconstruct_state_drain_run)
 997                 sh->reconstruct_state = reconstruct_state_drain_result;
 998         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
 999                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1000         else {
1001                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1002                 sh->reconstruct_state = reconstruct_state_result;
1003         }
1004
1005         set_bit(STRIPE_HANDLE, &sh->state);
1006         release_stripe(sh);
1007 }
1008
1009 static void
1010 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1011                      struct dma_async_tx_descriptor *tx)
1012 {
1013         int disks = sh->disks;
1014         struct page **xor_srcs = percpu->scribble;
1015         struct async_submit_ctl submit;
1016         int count = 0, pd_idx = sh->pd_idx, i;
1017         struct page *xor_dest;
1018         int prexor = 0;
1019         unsigned long flags;
1020
1021         pr_debug("%s: stripe %llu\n", __func__,
1022                 (unsigned long long)sh->sector);
1023
1024         /* check if prexor is active which means only process blocks
1025          * that are part of a read-modify-write (written)
1026          */
1027         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1028                 prexor = 1;
1029                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1030                 for (i = disks; i--; ) {
1031                         struct r5dev *dev = &sh->dev[i];
1032                         if (dev->written)
1033                                 xor_srcs[count++] = dev->page;
1034                 }
1035         } else {
1036                 xor_dest = sh->dev[pd_idx].page;
1037                 for (i = disks; i--; ) {
1038                         struct r5dev *dev = &sh->dev[i];
1039                         if (i != pd_idx)
1040                                 xor_srcs[count++] = dev->page;
1041                 }
1042         }
1043
1044         /* 1/ if we prexor'd then the dest is reused as a source
1045          * 2/ if we did not prexor then we are redoing the parity
1046          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1047          * for the synchronous xor case
1048          */
1049         flags = ASYNC_TX_ACK |
1050                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1051
1052         atomic_inc(&sh->count);
1053
1054         init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1055                           to_addr_conv(sh, percpu));
1056         if (unlikely(count == 1))
1057                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1058         else
1059                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1060 }
1061
1062 static void
1063 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1064                      struct dma_async_tx_descriptor *tx)
1065 {
1066         struct async_submit_ctl submit;
1067         struct page **blocks = percpu->scribble;
1068         int count;
1069
1070         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1071
1072         count = set_syndrome_sources(blocks, sh);
1073
1074         atomic_inc(&sh->count);
1075
1076         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1077                           sh, to_addr_conv(sh, percpu));
1078         async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1079 }
1080
1081 static void ops_complete_check(void *stripe_head_ref)
1082 {
1083         struct stripe_head *sh = stripe_head_ref;
1084
1085         pr_debug("%s: stripe %llu\n", __func__,
1086                 (unsigned long long)sh->sector);
1087
1088         sh->check_state = check_state_check_result;
1089         set_bit(STRIPE_HANDLE, &sh->state);
1090         release_stripe(sh);
1091 }
1092
1093 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1094 {
1095         int disks = sh->disks;
1096         int pd_idx = sh->pd_idx;
1097         int qd_idx = sh->qd_idx;
1098         struct page *xor_dest;
1099         struct page **xor_srcs = percpu->scribble;
1100         struct dma_async_tx_descriptor *tx;
1101         struct async_submit_ctl submit;
1102         int count;
1103         int i;
1104
1105         pr_debug("%s: stripe %llu\n", __func__,
1106                 (unsigned long long)sh->sector);
1107
1108         count = 0;
1109         xor_dest = sh->dev[pd_idx].page;
1110         xor_srcs[count++] = xor_dest;
1111         for (i = disks; i--; ) {
1112                 if (i == pd_idx || i == qd_idx)
1113                         continue;
1114                 xor_srcs[count++] = sh->dev[i].page;
1115         }
1116
1117         init_async_submit(&submit, 0, NULL, NULL, NULL,
1118                           to_addr_conv(sh, percpu));
1119         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1120                            &sh->ops.zero_sum_result, &submit);
1121
1122         atomic_inc(&sh->count);
1123         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1124         tx = async_trigger_callback(&submit);
1125 }
1126
1127 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1128 {
1129         struct page **srcs = percpu->scribble;
1130         struct async_submit_ctl submit;
1131         int count;
1132
1133         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1134                 (unsigned long long)sh->sector, checkp);
1135
1136         count = set_syndrome_sources(srcs, sh);
1137         if (!checkp)
1138                 srcs[count] = NULL;
1139
1140         atomic_inc(&sh->count);
1141         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1142                           sh, to_addr_conv(sh, percpu));
1143         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1144                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1145 }
1146
1147 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1148 {
1149         int overlap_clear = 0, i, disks = sh->disks;
1150         struct dma_async_tx_descriptor *tx = NULL;
1151         raid5_conf_t *conf = sh->raid_conf;
1152         int level = conf->level;
1153         struct raid5_percpu *percpu;
1154         unsigned long cpu;
1155
1156         cpu = get_cpu();
1157         percpu = per_cpu_ptr(conf->percpu, cpu);
1158         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1159                 ops_run_biofill(sh);
1160                 overlap_clear++;
1161         }
1162
1163         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1164                 if (level < 6)
1165                         tx = ops_run_compute5(sh, percpu);
1166                 else {
1167                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
1168                                 tx = ops_run_compute6_1(sh, percpu);
1169                         else
1170                                 tx = ops_run_compute6_2(sh, percpu);
1171                 }
1172                 /* terminate the chain if reconstruct is not set to be run */
1173                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1174                         async_tx_ack(tx);
1175         }
1176
1177         if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1178                 tx = ops_run_prexor(sh, percpu, tx);
1179
1180         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1181                 tx = ops_run_biodrain(sh, tx);
1182                 overlap_clear++;
1183         }
1184
1185         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1186                 if (level < 6)
1187                         ops_run_reconstruct5(sh, percpu, tx);
1188                 else
1189                         ops_run_reconstruct6(sh, percpu, tx);
1190         }
1191
1192         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1193                 if (sh->check_state == check_state_run)
1194                         ops_run_check_p(sh, percpu);
1195                 else if (sh->check_state == check_state_run_q)
1196                         ops_run_check_pq(sh, percpu, 0);
1197                 else if (sh->check_state == check_state_run_pq)
1198                         ops_run_check_pq(sh, percpu, 1);
1199                 else
1200                         BUG();
1201         }
1202
1203         if (overlap_clear)
1204                 for (i = disks; i--; ) {
1205                         struct r5dev *dev = &sh->dev[i];
1206                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1207                                 wake_up(&sh->raid_conf->wait_for_overlap);
1208                 }
1209         put_cpu();
1210 }
1211
1212 #ifdef CONFIG_MULTICORE_RAID456
1213 static void async_run_ops(void *param, async_cookie_t cookie)
1214 {
1215         struct stripe_head *sh = param;
1216         unsigned long ops_request = sh->ops.request;
1217
1218         clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1219         wake_up(&sh->ops.wait_for_ops);
1220
1221         __raid_run_ops(sh, ops_request);
1222         release_stripe(sh);
1223 }
1224
1225 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1226 {
1227         /* since handle_stripe can be called outside of raid5d context
1228          * we need to ensure sh->ops.request is de-staged before another
1229          * request arrives
1230          */
1231         wait_event(sh->ops.wait_for_ops,
1232                    !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1233         sh->ops.request = ops_request;
1234
1235         atomic_inc(&sh->count);
1236         async_schedule(async_run_ops, sh);
1237 }
1238 #else
1239 #define raid_run_ops __raid_run_ops
1240 #endif
1241
1242 static int grow_one_stripe(raid5_conf_t *conf)
1243 {
1244         struct stripe_head *sh;
1245         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
1246         if (!sh)
1247                 return 0;
1248         memset(sh, 0, sizeof(*sh) + (conf->pool_size-1)*sizeof(struct r5dev));
1249         sh->raid_conf = conf;
1250         spin_lock_init(&sh->lock);
1251         #ifdef CONFIG_MULTICORE_RAID456
1252         init_waitqueue_head(&sh->ops.wait_for_ops);
1253         #endif
1254
1255         if (grow_buffers(sh)) {
1256                 shrink_buffers(sh);
1257                 kmem_cache_free(conf->slab_cache, sh);
1258                 return 0;
1259         }
1260         /* we just created an active stripe so... */
1261         atomic_set(&sh->count, 1);
1262         atomic_inc(&conf->active_stripes);
1263         INIT_LIST_HEAD(&sh->lru);
1264         release_stripe(sh);
1265         return 1;
1266 }
1267
1268 static int grow_stripes(raid5_conf_t *conf, int num)
1269 {
1270         struct kmem_cache *sc;
1271         int devs = max(conf->raid_disks, conf->previous_raid_disks);
1272
1273         sprintf(conf->cache_name[0],
1274                 "raid%d-%s", conf->level, mdname(conf->mddev));
1275         sprintf(conf->cache_name[1],
1276                 "raid%d-%s-alt", conf->level, mdname(conf->mddev));
1277         conf->active_name = 0;
1278         sc = kmem_cache_create(conf->cache_name[conf->active_name],
1279                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1280                                0, 0, NULL);
1281         if (!sc)
1282                 return 1;
1283         conf->slab_cache = sc;
1284         conf->pool_size = devs;
1285         while (num--)
1286                 if (!grow_one_stripe(conf))
1287                         return 1;
1288         return 0;
1289 }
1290
1291 /**
1292  * scribble_len - return the required size of the scribble region
1293  * @num - total number of disks in the array
1294  *
1295  * The size must be enough to contain:
1296  * 1/ a struct page pointer for each device in the array +2
1297  * 2/ room to convert each entry in (1) to its corresponding dma
1298  *    (dma_map_page()) or page (page_address()) address.
1299  *
1300  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1301  * calculate over all devices (not just the data blocks), using zeros in place
1302  * of the P and Q blocks.
1303  */
1304 static size_t scribble_len(int num)
1305 {
1306         size_t len;
1307
1308         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1309
1310         return len;
1311 }
1312
1313 static int resize_stripes(raid5_conf_t *conf, int newsize)
1314 {
1315         /* Make all the stripes able to hold 'newsize' devices.
1316          * New slots in each stripe get 'page' set to a new page.
1317          *
1318          * This happens in stages:
1319          * 1/ create a new kmem_cache and allocate the required number of
1320          *    stripe_heads.
1321          * 2/ gather all the old stripe_heads and tranfer the pages across
1322          *    to the new stripe_heads.  This will have the side effect of
1323          *    freezing the array as once all stripe_heads have been collected,
1324          *    no IO will be possible.  Old stripe heads are freed once their
1325          *    pages have been transferred over, and the old kmem_cache is
1326          *    freed when all stripes are done.
1327          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
1328          *    we simple return a failre status - no need to clean anything up.
1329          * 4/ allocate new pages for the new slots in the new stripe_heads.
1330          *    If this fails, we don't bother trying the shrink the
1331          *    stripe_heads down again, we just leave them as they are.
1332          *    As each stripe_head is processed the new one is released into
1333          *    active service.
1334          *
1335          * Once step2 is started, we cannot afford to wait for a write,
1336          * so we use GFP_NOIO allocations.
1337          */
1338         struct stripe_head *osh, *nsh;
1339         LIST_HEAD(newstripes);
1340         struct disk_info *ndisks;
1341         unsigned long cpu;
1342         int err;
1343         struct kmem_cache *sc;
1344         int i;
1345
1346         if (newsize <= conf->pool_size)
1347                 return 0; /* never bother to shrink */
1348
1349         err = md_allow_write(conf->mddev);
1350         if (err)
1351                 return err;
1352
1353         /* Step 1 */
1354         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1355                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1356                                0, 0, NULL);
1357         if (!sc)
1358                 return -ENOMEM;
1359
1360         for (i = conf->max_nr_stripes; i; i--) {
1361                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1362                 if (!nsh)
1363                         break;
1364
1365                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1366
1367                 nsh->raid_conf = conf;
1368                 spin_lock_init(&nsh->lock);
1369                 #ifdef CONFIG_MULTICORE_RAID456
1370                 init_waitqueue_head(&nsh->ops.wait_for_ops);
1371                 #endif
1372
1373                 list_add(&nsh->lru, &newstripes);
1374         }
1375         if (i) {
1376                 /* didn't get enough, give up */
1377                 while (!list_empty(&newstripes)) {
1378                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1379                         list_del(&nsh->lru);
1380                         kmem_cache_free(sc, nsh);
1381                 }
1382                 kmem_cache_destroy(sc);
1383                 return -ENOMEM;
1384         }
1385         /* Step 2 - Must use GFP_NOIO now.
1386          * OK, we have enough stripes, start collecting inactive
1387          * stripes and copying them over
1388          */
1389         list_for_each_entry(nsh, &newstripes, lru) {
1390                 spin_lock_irq(&conf->device_lock);
1391                 wait_event_lock_irq(conf->wait_for_stripe,
1392                                     !list_empty(&conf->inactive_list),
1393                                     conf->device_lock,
1394                                     unplug_slaves(conf->mddev)
1395                         );
1396                 osh = get_free_stripe(conf);
1397                 spin_unlock_irq(&conf->device_lock);
1398                 atomic_set(&nsh->count, 1);
1399                 for(i=0; i<conf->pool_size; i++)
1400                         nsh->dev[i].page = osh->dev[i].page;
1401                 for( ; i<newsize; i++)
1402                         nsh->dev[i].page = NULL;
1403                 kmem_cache_free(conf->slab_cache, osh);
1404         }
1405         kmem_cache_destroy(conf->slab_cache);
1406
1407         /* Step 3.
1408          * At this point, we are holding all the stripes so the array
1409          * is completely stalled, so now is a good time to resize
1410          * conf->disks and the scribble region
1411          */
1412         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1413         if (ndisks) {
1414                 for (i=0; i<conf->raid_disks; i++)
1415                         ndisks[i] = conf->disks[i];
1416                 kfree(conf->disks);
1417                 conf->disks = ndisks;
1418         } else
1419                 err = -ENOMEM;
1420
1421         get_online_cpus();
1422         conf->scribble_len = scribble_len(newsize);
1423         for_each_present_cpu(cpu) {
1424                 struct raid5_percpu *percpu;
1425                 void *scribble;
1426
1427                 percpu = per_cpu_ptr(conf->percpu, cpu);
1428                 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1429
1430                 if (scribble) {
1431                         kfree(percpu->scribble);
1432                         percpu->scribble = scribble;
1433                 } else {
1434                         err = -ENOMEM;
1435                         break;
1436                 }
1437         }
1438         put_online_cpus();
1439
1440         /* Step 4, return new stripes to service */
1441         while(!list_empty(&newstripes)) {
1442                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1443                 list_del_init(&nsh->lru);
1444
1445                 for (i=conf->raid_disks; i < newsize; i++)
1446                         if (nsh->dev[i].page == NULL) {
1447                                 struct page *p = alloc_page(GFP_NOIO);
1448                                 nsh->dev[i].page = p;
1449                                 if (!p)
1450                                         err = -ENOMEM;
1451                         }
1452                 release_stripe(nsh);
1453         }
1454         /* critical section pass, GFP_NOIO no longer needed */
1455
1456         conf->slab_cache = sc;
1457         conf->active_name = 1-conf->active_name;
1458         conf->pool_size = newsize;
1459         return err;
1460 }
1461
1462 static int drop_one_stripe(raid5_conf_t *conf)
1463 {
1464         struct stripe_head *sh;
1465
1466         spin_lock_irq(&conf->device_lock);
1467         sh = get_free_stripe(conf);
1468         spin_unlock_irq(&conf->device_lock);
1469         if (!sh)
1470                 return 0;
1471         BUG_ON(atomic_read(&sh->count));
1472         shrink_buffers(sh);
1473         kmem_cache_free(conf->slab_cache, sh);
1474         atomic_dec(&conf->active_stripes);
1475         return 1;
1476 }
1477
1478 static void shrink_stripes(raid5_conf_t *conf)
1479 {
1480         while (drop_one_stripe(conf))
1481                 ;
1482
1483         if (conf->slab_cache)
1484                 kmem_cache_destroy(conf->slab_cache);
1485         conf->slab_cache = NULL;
1486 }
1487
1488 static void raid5_end_read_request(struct bio * bi, int error)
1489 {
1490         struct stripe_head *sh = bi->bi_private;
1491         raid5_conf_t *conf = sh->raid_conf;
1492         int disks = sh->disks, i;
1493         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1494         char b[BDEVNAME_SIZE];
1495         mdk_rdev_t *rdev;
1496
1497
1498         for (i=0 ; i<disks; i++)
1499                 if (bi == &sh->dev[i].req)
1500                         break;
1501
1502         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1503                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1504                 uptodate);
1505         if (i == disks) {
1506                 BUG();
1507                 return;
1508         }
1509
1510         if (uptodate) {
1511                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1512                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1513                         rdev = conf->disks[i].rdev;
1514                         printk_rl(KERN_INFO "md/raid:%s: read error corrected"
1515                                   " (%lu sectors at %llu on %s)\n",
1516                                   mdname(conf->mddev), STRIPE_SECTORS,
1517                                   (unsigned long long)(sh->sector
1518                                                        + rdev->data_offset),
1519                                   bdevname(rdev->bdev, b));
1520                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1521                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1522                 }
1523                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1524                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1525         } else {
1526                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1527                 int retry = 0;
1528                 rdev = conf->disks[i].rdev;
1529
1530                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1531                 atomic_inc(&rdev->read_errors);
1532                 if (conf->mddev->degraded >= conf->max_degraded)
1533                         printk_rl(KERN_WARNING
1534                                   "md/raid:%s: read error not correctable "
1535                                   "(sector %llu on %s).\n",
1536                                   mdname(conf->mddev),
1537                                   (unsigned long long)(sh->sector
1538                                                        + rdev->data_offset),
1539                                   bdn);
1540                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1541                         /* Oh, no!!! */
1542                         printk_rl(KERN_WARNING
1543                                   "md/raid:%s: read error NOT corrected!! "
1544                                   "(sector %llu on %s).\n",
1545                                   mdname(conf->mddev),
1546                                   (unsigned long long)(sh->sector
1547                                                        + rdev->data_offset),
1548                                   bdn);
1549                 else if (atomic_read(&rdev->read_errors)
1550                          > conf->max_nr_stripes)
1551                         printk(KERN_WARNING
1552                                "md/raid:%s: Too many read errors, failing device %s.\n",
1553                                mdname(conf->mddev), bdn);
1554                 else
1555                         retry = 1;
1556                 if (retry)
1557                         set_bit(R5_ReadError, &sh->dev[i].flags);
1558                 else {
1559                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1560                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1561                         md_error(conf->mddev, rdev);
1562                 }
1563         }
1564         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1565         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1566         set_bit(STRIPE_HANDLE, &sh->state);
1567         release_stripe(sh);
1568 }
1569
1570 static void raid5_end_write_request(struct bio *bi, int error)
1571 {
1572         struct stripe_head *sh = bi->bi_private;
1573         raid5_conf_t *conf = sh->raid_conf;
1574         int disks = sh->disks, i;
1575         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1576
1577         for (i=0 ; i<disks; i++)
1578                 if (bi == &sh->dev[i].req)
1579                         break;
1580
1581         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1582                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1583                 uptodate);
1584         if (i == disks) {
1585                 BUG();
1586                 return;
1587         }
1588
1589         if (!uptodate)
1590                 md_error(conf->mddev, conf->disks[i].rdev);
1591
1592         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1593
1594         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1595         set_bit(STRIPE_HANDLE, &sh->state);
1596         release_stripe(sh);
1597 }
1598
1599
1600 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1601
1602 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1603 {
1604         struct r5dev *dev = &sh->dev[i];
1605
1606         bio_init(&dev->req);
1607         dev->req.bi_io_vec = &dev->vec;
1608         dev->req.bi_vcnt++;
1609         dev->req.bi_max_vecs++;
1610         dev->vec.bv_page = dev->page;
1611         dev->vec.bv_len = STRIPE_SIZE;
1612         dev->vec.bv_offset = 0;
1613
1614         dev->req.bi_sector = sh->sector;
1615         dev->req.bi_private = sh;
1616
1617         dev->flags = 0;
1618         dev->sector = compute_blocknr(sh, i, previous);
1619 }
1620
1621 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1622 {
1623         char b[BDEVNAME_SIZE];
1624         raid5_conf_t *conf = mddev->private;
1625         pr_debug("raid456: error called\n");
1626
1627         if (!test_bit(Faulty, &rdev->flags)) {
1628                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1629                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1630                         unsigned long flags;
1631                         spin_lock_irqsave(&conf->device_lock, flags);
1632                         mddev->degraded++;
1633                         spin_unlock_irqrestore(&conf->device_lock, flags);
1634                         /*
1635                          * if recovery was running, make sure it aborts.
1636                          */
1637                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1638                 }
1639                 set_bit(Faulty, &rdev->flags);
1640                 printk(KERN_ALERT
1641                        "md/raid:%s: Disk failure on %s, disabling device.\n"
1642                        KERN_ALERT
1643                        "md/raid:%s: Operation continuing on %d devices.\n",
1644                        mdname(mddev),
1645                        bdevname(rdev->bdev, b),
1646                        mdname(mddev),
1647                        conf->raid_disks - mddev->degraded);
1648         }
1649 }
1650
1651 /*
1652  * Input: a 'big' sector number,
1653  * Output: index of the data and parity disk, and the sector # in them.
1654  */
1655 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1656                                      int previous, int *dd_idx,
1657                                      struct stripe_head *sh)
1658 {
1659         sector_t stripe, stripe2;
1660         sector_t chunk_number;
1661         unsigned int chunk_offset;
1662         int pd_idx, qd_idx;
1663         int ddf_layout = 0;
1664         sector_t new_sector;
1665         int algorithm = previous ? conf->prev_algo
1666                                  : conf->algorithm;
1667         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1668                                          : conf->chunk_sectors;
1669         int raid_disks = previous ? conf->previous_raid_disks
1670                                   : conf->raid_disks;
1671         int data_disks = raid_disks - conf->max_degraded;
1672
1673         /* First compute the information on this sector */
1674
1675         /*
1676          * Compute the chunk number and the sector offset inside the chunk
1677          */
1678         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1679         chunk_number = r_sector;
1680
1681         /*
1682          * Compute the stripe number
1683          */
1684         stripe = chunk_number;
1685         *dd_idx = sector_div(stripe, data_disks);
1686         stripe2 = stripe;
1687         /*
1688          * Select the parity disk based on the user selected algorithm.
1689          */
1690         pd_idx = qd_idx = ~0;
1691         switch(conf->level) {
1692         case 4:
1693                 pd_idx = data_disks;
1694                 break;
1695         case 5:
1696                 switch (algorithm) {
1697                 case ALGORITHM_LEFT_ASYMMETRIC:
1698                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
1699                         if (*dd_idx >= pd_idx)
1700                                 (*dd_idx)++;
1701                         break;
1702                 case ALGORITHM_RIGHT_ASYMMETRIC:
1703                         pd_idx = sector_div(stripe2, raid_disks);
1704                         if (*dd_idx >= pd_idx)
1705                                 (*dd_idx)++;
1706                         break;
1707                 case ALGORITHM_LEFT_SYMMETRIC:
1708                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
1709                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1710                         break;
1711                 case ALGORITHM_RIGHT_SYMMETRIC:
1712                         pd_idx = sector_div(stripe2, raid_disks);
1713                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1714                         break;
1715                 case ALGORITHM_PARITY_0:
1716                         pd_idx = 0;
1717                         (*dd_idx)++;
1718                         break;
1719                 case ALGORITHM_PARITY_N:
1720                         pd_idx = data_disks;
1721                         break;
1722                 default:
1723                         BUG();
1724                 }
1725                 break;
1726         case 6:
1727
1728                 switch (algorithm) {
1729                 case ALGORITHM_LEFT_ASYMMETRIC:
1730                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1731                         qd_idx = pd_idx + 1;
1732                         if (pd_idx == raid_disks-1) {
1733                                 (*dd_idx)++;    /* Q D D D P */
1734                                 qd_idx = 0;
1735                         } else if (*dd_idx >= pd_idx)
1736                                 (*dd_idx) += 2; /* D D P Q D */
1737                         break;
1738                 case ALGORITHM_RIGHT_ASYMMETRIC:
1739                         pd_idx = sector_div(stripe2, raid_disks);
1740                         qd_idx = pd_idx + 1;
1741                         if (pd_idx == raid_disks-1) {
1742                                 (*dd_idx)++;    /* Q D D D P */
1743                                 qd_idx = 0;
1744                         } else if (*dd_idx >= pd_idx)
1745                                 (*dd_idx) += 2; /* D D P Q D */
1746                         break;
1747                 case ALGORITHM_LEFT_SYMMETRIC:
1748                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1749                         qd_idx = (pd_idx + 1) % raid_disks;
1750                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1751                         break;
1752                 case ALGORITHM_RIGHT_SYMMETRIC:
1753                         pd_idx = sector_div(stripe2, raid_disks);
1754                         qd_idx = (pd_idx + 1) % raid_disks;
1755                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1756                         break;
1757
1758                 case ALGORITHM_PARITY_0:
1759                         pd_idx = 0;
1760                         qd_idx = 1;
1761                         (*dd_idx) += 2;
1762                         break;
1763                 case ALGORITHM_PARITY_N:
1764                         pd_idx = data_disks;
1765                         qd_idx = data_disks + 1;
1766                         break;
1767
1768                 case ALGORITHM_ROTATING_ZERO_RESTART:
1769                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
1770                          * of blocks for computing Q is different.
1771                          */
1772                         pd_idx = sector_div(stripe2, raid_disks);
1773                         qd_idx = pd_idx + 1;
1774                         if (pd_idx == raid_disks-1) {
1775                                 (*dd_idx)++;    /* Q D D D P */
1776                                 qd_idx = 0;
1777                         } else if (*dd_idx >= pd_idx)
1778                                 (*dd_idx) += 2; /* D D P Q D */
1779                         ddf_layout = 1;
1780                         break;
1781
1782                 case ALGORITHM_ROTATING_N_RESTART:
1783                         /* Same a left_asymmetric, by first stripe is
1784                          * D D D P Q  rather than
1785                          * Q D D D P
1786                          */
1787                         stripe2 += 1;
1788                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1789                         qd_idx = pd_idx + 1;
1790                         if (pd_idx == raid_disks-1) {
1791                                 (*dd_idx)++;    /* Q D D D P */
1792                                 qd_idx = 0;
1793                         } else if (*dd_idx >= pd_idx)
1794                                 (*dd_idx) += 2; /* D D P Q D */
1795                         ddf_layout = 1;
1796                         break;
1797
1798                 case ALGORITHM_ROTATING_N_CONTINUE:
1799                         /* Same as left_symmetric but Q is before P */
1800                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1801                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1802                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1803                         ddf_layout = 1;
1804                         break;
1805
1806                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1807                         /* RAID5 left_asymmetric, with Q on last device */
1808                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1809                         if (*dd_idx >= pd_idx)
1810                                 (*dd_idx)++;
1811                         qd_idx = raid_disks - 1;
1812                         break;
1813
1814                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1815                         pd_idx = sector_div(stripe2, raid_disks-1);
1816                         if (*dd_idx >= pd_idx)
1817                                 (*dd_idx)++;
1818                         qd_idx = raid_disks - 1;
1819                         break;
1820
1821                 case ALGORITHM_LEFT_SYMMETRIC_6:
1822                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1823                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1824                         qd_idx = raid_disks - 1;
1825                         break;
1826
1827                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1828                         pd_idx = sector_div(stripe2, raid_disks-1);
1829                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1830                         qd_idx = raid_disks - 1;
1831                         break;
1832
1833                 case ALGORITHM_PARITY_0_6:
1834                         pd_idx = 0;
1835                         (*dd_idx)++;
1836                         qd_idx = raid_disks - 1;
1837                         break;
1838
1839                 default:
1840                         BUG();
1841                 }
1842                 break;
1843         }
1844
1845         if (sh) {
1846                 sh->pd_idx = pd_idx;
1847                 sh->qd_idx = qd_idx;
1848                 sh->ddf_layout = ddf_layout;
1849         }
1850         /*
1851          * Finally, compute the new sector number
1852          */
1853         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1854         return new_sector;
1855 }
1856
1857
1858 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1859 {
1860         raid5_conf_t *conf = sh->raid_conf;
1861         int raid_disks = sh->disks;
1862         int data_disks = raid_disks - conf->max_degraded;
1863         sector_t new_sector = sh->sector, check;
1864         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1865                                          : conf->chunk_sectors;
1866         int algorithm = previous ? conf->prev_algo
1867                                  : conf->algorithm;
1868         sector_t stripe;
1869         int chunk_offset;
1870         sector_t chunk_number;
1871         int dummy1, dd_idx = i;
1872         sector_t r_sector;
1873         struct stripe_head sh2;
1874
1875
1876         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1877         stripe = new_sector;
1878
1879         if (i == sh->pd_idx)
1880                 return 0;
1881         switch(conf->level) {
1882         case 4: break;
1883         case 5:
1884                 switch (algorithm) {
1885                 case ALGORITHM_LEFT_ASYMMETRIC:
1886                 case ALGORITHM_RIGHT_ASYMMETRIC:
1887                         if (i > sh->pd_idx)
1888                                 i--;
1889                         break;
1890                 case ALGORITHM_LEFT_SYMMETRIC:
1891                 case ALGORITHM_RIGHT_SYMMETRIC:
1892                         if (i < sh->pd_idx)
1893                                 i += raid_disks;
1894                         i -= (sh->pd_idx + 1);
1895                         break;
1896                 case ALGORITHM_PARITY_0:
1897                         i -= 1;
1898                         break;
1899                 case ALGORITHM_PARITY_N:
1900                         break;
1901                 default:
1902                         BUG();
1903                 }
1904                 break;
1905         case 6:
1906                 if (i == sh->qd_idx)
1907                         return 0; /* It is the Q disk */
1908                 switch (algorithm) {
1909                 case ALGORITHM_LEFT_ASYMMETRIC:
1910                 case ALGORITHM_RIGHT_ASYMMETRIC:
1911                 case ALGORITHM_ROTATING_ZERO_RESTART:
1912                 case ALGORITHM_ROTATING_N_RESTART:
1913                         if (sh->pd_idx == raid_disks-1)
1914                                 i--;    /* Q D D D P */
1915                         else if (i > sh->pd_idx)
1916                                 i -= 2; /* D D P Q D */
1917                         break;
1918                 case ALGORITHM_LEFT_SYMMETRIC:
1919                 case ALGORITHM_RIGHT_SYMMETRIC:
1920                         if (sh->pd_idx == raid_disks-1)
1921                                 i--; /* Q D D D P */
1922                         else {
1923                                 /* D D P Q D */
1924                                 if (i < sh->pd_idx)
1925                                         i += raid_disks;
1926                                 i -= (sh->pd_idx + 2);
1927                         }
1928                         break;
1929                 case ALGORITHM_PARITY_0:
1930                         i -= 2;
1931                         break;
1932                 case ALGORITHM_PARITY_N:
1933                         break;
1934                 case ALGORITHM_ROTATING_N_CONTINUE:
1935                         /* Like left_symmetric, but P is before Q */
1936                         if (sh->pd_idx == 0)
1937                                 i--;    /* P D D D Q */
1938                         else {
1939                                 /* D D Q P D */
1940                                 if (i < sh->pd_idx)
1941                                         i += raid_disks;
1942                                 i -= (sh->pd_idx + 1);
1943                         }
1944                         break;
1945                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1946                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1947                         if (i > sh->pd_idx)
1948                                 i--;
1949                         break;
1950                 case ALGORITHM_LEFT_SYMMETRIC_6:
1951                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1952                         if (i < sh->pd_idx)
1953                                 i += data_disks + 1;
1954                         i -= (sh->pd_idx + 1);
1955                         break;
1956                 case ALGORITHM_PARITY_0_6:
1957                         i -= 1;
1958                         break;
1959                 default:
1960                         BUG();
1961                 }
1962                 break;
1963         }
1964
1965         chunk_number = stripe * data_disks + i;
1966         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
1967
1968         check = raid5_compute_sector(conf, r_sector,
1969                                      previous, &dummy1, &sh2);
1970         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
1971                 || sh2.qd_idx != sh->qd_idx) {
1972                 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
1973                        mdname(conf->mddev));
1974                 return 0;
1975         }
1976         return r_sector;
1977 }
1978
1979
1980 static void
1981 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
1982                          int rcw, int expand)
1983 {
1984         int i, pd_idx = sh->pd_idx, disks = sh->disks;
1985         raid5_conf_t *conf = sh->raid_conf;
1986         int level = conf->level;
1987
1988         if (rcw) {
1989                 /* if we are not expanding this is a proper write request, and
1990                  * there will be bios with new data to be drained into the
1991                  * stripe cache
1992                  */
1993                 if (!expand) {
1994                         sh->reconstruct_state = reconstruct_state_drain_run;
1995                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1996                 } else
1997                         sh->reconstruct_state = reconstruct_state_run;
1998
1999                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2000
2001                 for (i = disks; i--; ) {
2002                         struct r5dev *dev = &sh->dev[i];
2003
2004                         if (dev->towrite) {
2005                                 set_bit(R5_LOCKED, &dev->flags);
2006                                 set_bit(R5_Wantdrain, &dev->flags);
2007                                 if (!expand)
2008                                         clear_bit(R5_UPTODATE, &dev->flags);
2009                                 s->locked++;
2010                         }
2011                 }
2012                 if (s->locked + conf->max_degraded == disks)
2013                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2014                                 atomic_inc(&conf->pending_full_writes);
2015         } else {
2016                 BUG_ON(level == 6);
2017                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2018                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2019
2020                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2021                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2022                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2023                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2024
2025                 for (i = disks; i--; ) {
2026                         struct r5dev *dev = &sh->dev[i];
2027                         if (i == pd_idx)
2028                                 continue;
2029
2030                         if (dev->towrite &&
2031                             (test_bit(R5_UPTODATE, &dev->flags) ||
2032                              test_bit(R5_Wantcompute, &dev->flags))) {
2033                                 set_bit(R5_Wantdrain, &dev->flags);
2034                                 set_bit(R5_LOCKED, &dev->flags);
2035                                 clear_bit(R5_UPTODATE, &dev->flags);
2036                                 s->locked++;
2037                         }
2038                 }
2039         }
2040
2041         /* keep the parity disk(s) locked while asynchronous operations
2042          * are in flight
2043          */
2044         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2045         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2046         s->locked++;
2047
2048         if (level == 6) {
2049                 int qd_idx = sh->qd_idx;
2050                 struct r5dev *dev = &sh->dev[qd_idx];
2051
2052                 set_bit(R5_LOCKED, &dev->flags);
2053                 clear_bit(R5_UPTODATE, &dev->flags);
2054                 s->locked++;
2055         }
2056
2057         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2058                 __func__, (unsigned long long)sh->sector,
2059                 s->locked, s->ops_request);
2060 }
2061
2062 /*
2063  * Each stripe/dev can have one or more bion attached.
2064  * toread/towrite point to the first in a chain.
2065  * The bi_next chain must be in order.
2066  */
2067 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2068 {
2069         struct bio **bip;
2070         raid5_conf_t *conf = sh->raid_conf;
2071         int firstwrite=0;
2072
2073         pr_debug("adding bh b#%llu to stripe s#%llu\n",
2074                 (unsigned long long)bi->bi_sector,
2075                 (unsigned long long)sh->sector);
2076
2077
2078         spin_lock(&sh->lock);
2079         spin_lock_irq(&conf->device_lock);
2080         if (forwrite) {
2081                 bip = &sh->dev[dd_idx].towrite;
2082                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2083                         firstwrite = 1;
2084         } else
2085                 bip = &sh->dev[dd_idx].toread;
2086         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2087                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2088                         goto overlap;
2089                 bip = & (*bip)->bi_next;
2090         }
2091         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2092                 goto overlap;
2093
2094         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2095         if (*bip)
2096                 bi->bi_next = *bip;
2097         *bip = bi;
2098         bi->bi_phys_segments++;
2099         spin_unlock_irq(&conf->device_lock);
2100         spin_unlock(&sh->lock);
2101
2102         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2103                 (unsigned long long)bi->bi_sector,
2104                 (unsigned long long)sh->sector, dd_idx);
2105
2106         if (conf->mddev->bitmap && firstwrite) {
2107                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2108                                   STRIPE_SECTORS, 0);
2109                 sh->bm_seq = conf->seq_flush+1;
2110                 set_bit(STRIPE_BIT_DELAY, &sh->state);
2111         }
2112
2113         if (forwrite) {
2114                 /* check if page is covered */
2115                 sector_t sector = sh->dev[dd_idx].sector;
2116                 for (bi=sh->dev[dd_idx].towrite;
2117                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2118                              bi && bi->bi_sector <= sector;
2119                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2120                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2121                                 sector = bi->bi_sector + (bi->bi_size>>9);
2122                 }
2123                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2124                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2125         }
2126         return 1;
2127
2128  overlap:
2129         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2130         spin_unlock_irq(&conf->device_lock);
2131         spin_unlock(&sh->lock);
2132         return 0;
2133 }
2134
2135 static void end_reshape(raid5_conf_t *conf);
2136
2137 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2138                             struct stripe_head *sh)
2139 {
2140         int sectors_per_chunk =
2141                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2142         int dd_idx;
2143         int chunk_offset = sector_div(stripe, sectors_per_chunk);
2144         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2145
2146         raid5_compute_sector(conf,
2147                              stripe * (disks - conf->max_degraded)
2148                              *sectors_per_chunk + chunk_offset,
2149                              previous,
2150                              &dd_idx, sh);
2151 }
2152
2153 static void
2154 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2155                                 struct stripe_head_state *s, int disks,
2156                                 struct bio **return_bi)
2157 {
2158         int i;
2159         for (i = disks; i--; ) {
2160                 struct bio *bi;
2161                 int bitmap_end = 0;
2162
2163                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2164                         mdk_rdev_t *rdev;
2165                         rcu_read_lock();
2166                         rdev = rcu_dereference(conf->disks[i].rdev);
2167                         if (rdev && test_bit(In_sync, &rdev->flags))
2168                                 /* multiple read failures in one stripe */
2169                                 md_error(conf->mddev, rdev);
2170                         rcu_read_unlock();
2171                 }
2172                 spin_lock_irq(&conf->device_lock);
2173                 /* fail all writes first */
2174                 bi = sh->dev[i].towrite;
2175                 sh->dev[i].towrite = NULL;
2176                 if (bi) {
2177                         s->to_write--;
2178                         bitmap_end = 1;
2179                 }
2180
2181                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2182                         wake_up(&conf->wait_for_overlap);
2183
2184                 while (bi && bi->bi_sector <
2185                         sh->dev[i].sector + STRIPE_SECTORS) {
2186                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2187                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2188                         if (!raid5_dec_bi_phys_segments(bi)) {
2189                                 md_write_end(conf->mddev);
2190                                 bi->bi_next = *return_bi;
2191                                 *return_bi = bi;
2192                         }
2193                         bi = nextbi;
2194                 }
2195                 /* and fail all 'written' */
2196                 bi = sh->dev[i].written;
2197                 sh->dev[i].written = NULL;
2198                 if (bi) bitmap_end = 1;
2199                 while (bi && bi->bi_sector <
2200                        sh->dev[i].sector + STRIPE_SECTORS) {
2201                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2202                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2203                         if (!raid5_dec_bi_phys_segments(bi)) {
2204                                 md_write_end(conf->mddev);
2205                                 bi->bi_next = *return_bi;
2206                                 *return_bi = bi;
2207                         }
2208                         bi = bi2;
2209                 }
2210
2211                 /* fail any reads if this device is non-operational and
2212                  * the data has not reached the cache yet.
2213                  */
2214                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2215                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2216                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
2217                         bi = sh->dev[i].toread;
2218                         sh->dev[i].toread = NULL;
2219                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2220                                 wake_up(&conf->wait_for_overlap);
2221                         if (bi) s->to_read--;
2222                         while (bi && bi->bi_sector <
2223                                sh->dev[i].sector + STRIPE_SECTORS) {
2224                                 struct bio *nextbi =
2225                                         r5_next_bio(bi, sh->dev[i].sector);
2226                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2227                                 if (!raid5_dec_bi_phys_segments(bi)) {
2228                                         bi->bi_next = *return_bi;
2229                                         *return_bi = bi;
2230                                 }
2231                                 bi = nextbi;
2232                         }
2233                 }
2234                 spin_unlock_irq(&conf->device_lock);
2235                 if (bitmap_end)
2236                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2237                                         STRIPE_SECTORS, 0, 0);
2238         }
2239
2240         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2241                 if (atomic_dec_and_test(&conf->pending_full_writes))
2242                         md_wakeup_thread(conf->mddev->thread);
2243 }
2244
2245 /* fetch_block5 - checks the given member device to see if its data needs
2246  * to be read or computed to satisfy a request.
2247  *
2248  * Returns 1 when no more member devices need to be checked, otherwise returns
2249  * 0 to tell the loop in handle_stripe_fill5 to continue
2250  */
2251 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2252                         int disk_idx, int disks)
2253 {
2254         struct r5dev *dev = &sh->dev[disk_idx];
2255         struct r5dev *failed_dev = &sh->dev[s->failed_num];
2256
2257         /* is the data in this block needed, and can we get it? */
2258         if (!test_bit(R5_LOCKED, &dev->flags) &&
2259             !test_bit(R5_UPTODATE, &dev->flags) &&
2260             (dev->toread ||
2261              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2262              s->syncing || s->expanding ||
2263              (s->failed &&
2264               (failed_dev->toread ||
2265                (failed_dev->towrite &&
2266                 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2267                 /* We would like to get this block, possibly by computing it,
2268                  * otherwise read it if the backing disk is insync
2269                  */
2270                 if ((s->uptodate == disks - 1) &&
2271                     (s->failed && disk_idx == s->failed_num)) {
2272                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2273                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2274                         set_bit(R5_Wantcompute, &dev->flags);
2275                         sh->ops.target = disk_idx;
2276                         sh->ops.target2 = -1;
2277                         s->req_compute = 1;
2278                         /* Careful: from this point on 'uptodate' is in the eye
2279                          * of raid_run_ops which services 'compute' operations
2280                          * before writes. R5_Wantcompute flags a block that will
2281                          * be R5_UPTODATE by the time it is needed for a
2282                          * subsequent operation.
2283                          */
2284                         s->uptodate++;
2285                         return 1; /* uptodate + compute == disks */
2286                 } else if (test_bit(R5_Insync, &dev->flags)) {
2287                         set_bit(R5_LOCKED, &dev->flags);
2288                         set_bit(R5_Wantread, &dev->flags);
2289                         s->locked++;
2290                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2291                                 s->syncing);
2292                 }
2293         }
2294
2295         return 0;
2296 }
2297
2298 /**
2299  * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2300  */
2301 static void handle_stripe_fill5(struct stripe_head *sh,
2302                         struct stripe_head_state *s, int disks)
2303 {
2304         int i;
2305
2306         /* look for blocks to read/compute, skip this if a compute
2307          * is already in flight, or if the stripe contents are in the
2308          * midst of changing due to a write
2309          */
2310         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2311             !sh->reconstruct_state)
2312                 for (i = disks; i--; )
2313                         if (fetch_block5(sh, s, i, disks))
2314                                 break;
2315         set_bit(STRIPE_HANDLE, &sh->state);
2316 }
2317
2318 /* fetch_block6 - checks the given member device to see if its data needs
2319  * to be read or computed to satisfy a request.
2320  *
2321  * Returns 1 when no more member devices need to be checked, otherwise returns
2322  * 0 to tell the loop in handle_stripe_fill6 to continue
2323  */
2324 static int fetch_block6(struct stripe_head *sh, struct stripe_head_state *s,
2325                          struct r6_state *r6s, int disk_idx, int disks)
2326 {
2327         struct r5dev *dev = &sh->dev[disk_idx];
2328         struct r5dev *fdev[2] = { &sh->dev[r6s->failed_num[0]],
2329                                   &sh->dev[r6s->failed_num[1]] };
2330
2331         if (!test_bit(R5_LOCKED, &dev->flags) &&
2332             !test_bit(R5_UPTODATE, &dev->flags) &&
2333             (dev->toread ||
2334              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2335              s->syncing || s->expanding ||
2336              (s->failed >= 1 &&
2337               (fdev[0]->toread || s->to_write)) ||
2338              (s->failed >= 2 &&
2339               (fdev[1]->toread || s->to_write)))) {
2340                 /* we would like to get this block, possibly by computing it,
2341                  * otherwise read it if the backing disk is insync
2342                  */
2343                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2344                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2345                 if ((s->uptodate == disks - 1) &&
2346                     (s->failed && (disk_idx == r6s->failed_num[0] ||
2347                                    disk_idx == r6s->failed_num[1]))) {
2348                         /* have disk failed, and we're requested to fetch it;
2349                          * do compute it
2350                          */
2351                         pr_debug("Computing stripe %llu block %d\n",
2352                                (unsigned long long)sh->sector, disk_idx);
2353                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2354                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2355                         set_bit(R5_Wantcompute, &dev->flags);
2356                         sh->ops.target = disk_idx;
2357                         sh->ops.target2 = -1; /* no 2nd target */
2358                         s->req_compute = 1;
2359                         s->uptodate++;
2360                         return 1;
2361                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2362                         /* Computing 2-failure is *very* expensive; only
2363                          * do it if failed >= 2
2364                          */
2365                         int other;
2366                         for (other = disks; other--; ) {
2367                                 if (other == disk_idx)
2368                                         continue;
2369                                 if (!test_bit(R5_UPTODATE,
2370                                       &sh->dev[other].flags))
2371                                         break;
2372                         }
2373                         BUG_ON(other < 0);
2374                         pr_debug("Computing stripe %llu blocks %d,%d\n",
2375                                (unsigned long long)sh->sector,
2376                                disk_idx, other);
2377                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2378                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2379                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2380                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
2381                         sh->ops.target = disk_idx;
2382                         sh->ops.target2 = other;
2383                         s->uptodate += 2;
2384                         s->req_compute = 1;
2385                         return 1;
2386                 } else if (test_bit(R5_Insync, &dev->flags)) {
2387                         set_bit(R5_LOCKED, &dev->flags);
2388                         set_bit(R5_Wantread, &dev->flags);
2389                         s->locked++;
2390                         pr_debug("Reading block %d (sync=%d)\n",
2391                                 disk_idx, s->syncing);
2392                 }
2393         }
2394
2395         return 0;
2396 }
2397
2398 /**
2399  * handle_stripe_fill6 - read or compute data to satisfy pending requests.
2400  */
2401 static void handle_stripe_fill6(struct stripe_head *sh,
2402                         struct stripe_head_state *s, struct r6_state *r6s,
2403                         int disks)
2404 {
2405         int i;
2406
2407         /* look for blocks to read/compute, skip this if a compute
2408          * is already in flight, or if the stripe contents are in the
2409          * midst of changing due to a write
2410          */
2411         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2412             !sh->reconstruct_state)
2413                 for (i = disks; i--; )
2414                         if (fetch_block6(sh, s, r6s, i, disks))
2415                                 break;
2416         set_bit(STRIPE_HANDLE, &sh->state);
2417 }
2418
2419
2420 /* handle_stripe_clean_event
2421  * any written block on an uptodate or failed drive can be returned.
2422  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2423  * never LOCKED, so we don't need to test 'failed' directly.
2424  */
2425 static void handle_stripe_clean_event(raid5_conf_t *conf,
2426         struct stripe_head *sh, int disks, struct bio **return_bi)
2427 {
2428         int i;
2429         struct r5dev *dev;
2430
2431         for (i = disks; i--; )
2432                 if (sh->dev[i].written) {
2433                         dev = &sh->dev[i];
2434                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2435                                 test_bit(R5_UPTODATE, &dev->flags)) {
2436                                 /* We can return any write requests */
2437                                 struct bio *wbi, *wbi2;
2438                                 int bitmap_end = 0;
2439                                 pr_debug("Return write for disc %d\n", i);
2440                                 spin_lock_irq(&conf->device_lock);
2441                                 wbi = dev->written;
2442                                 dev->written = NULL;
2443                                 while (wbi && wbi->bi_sector <
2444                                         dev->sector + STRIPE_SECTORS) {
2445                                         wbi2 = r5_next_bio(wbi, dev->sector);
2446                                         if (!raid5_dec_bi_phys_segments(wbi)) {
2447                                                 md_write_end(conf->mddev);
2448                                                 wbi->bi_next = *return_bi;
2449                                                 *return_bi = wbi;
2450                                         }
2451                                         wbi = wbi2;
2452                                 }
2453                                 if (dev->towrite == NULL)
2454                                         bitmap_end = 1;
2455                                 spin_unlock_irq(&conf->device_lock);
2456                                 if (bitmap_end)
2457                                         bitmap_endwrite(conf->mddev->bitmap,
2458                                                         sh->sector,
2459                                                         STRIPE_SECTORS,
2460                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2461                                                         0);
2462                         }
2463                 }
2464
2465         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2466                 if (atomic_dec_and_test(&conf->pending_full_writes))
2467                         md_wakeup_thread(conf->mddev->thread);
2468 }
2469
2470 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2471                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2472 {
2473         int rmw = 0, rcw = 0, i;
2474         for (i = disks; i--; ) {
2475                 /* would I have to read this buffer for read_modify_write */
2476                 struct r5dev *dev = &sh->dev[i];
2477                 if ((dev->towrite || i == sh->pd_idx) &&
2478                     !test_bit(R5_LOCKED, &dev->flags) &&
2479                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2480                       test_bit(R5_Wantcompute, &dev->flags))) {
2481                         if (test_bit(R5_Insync, &dev->flags))
2482                                 rmw++;
2483                         else
2484                                 rmw += 2*disks;  /* cannot read it */
2485                 }
2486                 /* Would I have to read this buffer for reconstruct_write */
2487                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2488                     !test_bit(R5_LOCKED, &dev->flags) &&
2489                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2490                     test_bit(R5_Wantcompute, &dev->flags))) {
2491                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2492                         else
2493                                 rcw += 2*disks;
2494                 }
2495         }
2496         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2497                 (unsigned long long)sh->sector, rmw, rcw);
2498         set_bit(STRIPE_HANDLE, &sh->state);
2499         if (rmw < rcw && rmw > 0)
2500                 /* prefer read-modify-write, but need to get some data */
2501                 for (i = disks; i--; ) {
2502                         struct r5dev *dev = &sh->dev[i];
2503                         if ((dev->towrite || i == sh->pd_idx) &&
2504                             !test_bit(R5_LOCKED, &dev->flags) &&
2505                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2506                             test_bit(R5_Wantcompute, &dev->flags)) &&
2507                             test_bit(R5_Insync, &dev->flags)) {
2508                                 if (
2509                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2510                                         pr_debug("Read_old block "
2511                                                 "%d for r-m-w\n", i);
2512                                         set_bit(R5_LOCKED, &dev->flags);
2513                                         set_bit(R5_Wantread, &dev->flags);
2514                                         s->locked++;
2515                                 } else {
2516                                         set_bit(STRIPE_DELAYED, &sh->state);
2517                                         set_bit(STRIPE_HANDLE, &sh->state);
2518                                 }
2519                         }
2520                 }
2521         if (rcw <= rmw && rcw > 0)
2522                 /* want reconstruct write, but need to get some data */
2523                 for (i = disks; i--; ) {
2524                         struct r5dev *dev = &sh->dev[i];
2525                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2526                             i != sh->pd_idx &&
2527                             !test_bit(R5_LOCKED, &dev->flags) &&
2528                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2529                             test_bit(R5_Wantcompute, &dev->flags)) &&
2530                             test_bit(R5_Insync, &dev->flags)) {
2531                                 if (
2532                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2533                                         pr_debug("Read_old block "
2534                                                 "%d for Reconstruct\n", i);
2535                                         set_bit(R5_LOCKED, &dev->flags);
2536                                         set_bit(R5_Wantread, &dev->flags);
2537                                         s->locked++;
2538                                 } else {
2539                                         set_bit(STRIPE_DELAYED, &sh->state);
2540                                         set_bit(STRIPE_HANDLE, &sh->state);
2541                                 }
2542                         }
2543                 }
2544         /* now if nothing is locked, and if we have enough data,
2545          * we can start a write request
2546          */
2547         /* since handle_stripe can be called at any time we need to handle the
2548          * case where a compute block operation has been submitted and then a
2549          * subsequent call wants to start a write request.  raid_run_ops only
2550          * handles the case where compute block and reconstruct are requested
2551          * simultaneously.  If this is not the case then new writes need to be
2552          * held off until the compute completes.
2553          */
2554         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2555             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2556             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2557                 schedule_reconstruction(sh, s, rcw == 0, 0);
2558 }
2559
2560 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2561                 struct stripe_head *sh, struct stripe_head_state *s,
2562                 struct r6_state *r6s, int disks)
2563 {
2564         int rcw = 0, pd_idx = sh->pd_idx, i;
2565         int qd_idx = sh->qd_idx;
2566
2567         set_bit(STRIPE_HANDLE, &sh->state);
2568         for (i = disks; i--; ) {
2569                 struct r5dev *dev = &sh->dev[i];
2570                 /* check if we haven't enough data */
2571                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2572                     i != pd_idx && i != qd_idx &&
2573                     !test_bit(R5_LOCKED, &dev->flags) &&
2574                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2575                       test_bit(R5_Wantcompute, &dev->flags))) {
2576                         rcw++;
2577                         if (!test_bit(R5_Insync, &dev->flags))
2578                                 continue; /* it's a failed drive */
2579
2580                         if (
2581                           test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2582                                 pr_debug("Read_old stripe %llu "
2583                                         "block %d for Reconstruct\n",
2584                                      (unsigned long long)sh->sector, i);
2585                                 set_bit(R5_LOCKED, &dev->flags);
2586                                 set_bit(R5_Wantread, &dev->flags);
2587                                 s->locked++;
2588                         } else {
2589                                 pr_debug("Request delayed stripe %llu "
2590                                         "block %d for Reconstruct\n",
2591                                      (unsigned long long)sh->sector, i);
2592                                 set_bit(STRIPE_DELAYED, &sh->state);
2593                                 set_bit(STRIPE_HANDLE, &sh->state);
2594                         }
2595                 }
2596         }
2597         /* now if nothing is locked, and if we have enough data, we can start a
2598          * write request
2599          */
2600         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2601             s->locked == 0 && rcw == 0 &&
2602             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2603                 schedule_reconstruction(sh, s, 1, 0);
2604         }
2605 }
2606
2607 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2608                                 struct stripe_head_state *s, int disks)
2609 {
2610         struct r5dev *dev = NULL;
2611
2612         set_bit(STRIPE_HANDLE, &sh->state);
2613
2614         switch (sh->check_state) {
2615         case check_state_idle:
2616                 /* start a new check operation if there are no failures */
2617                 if (s->failed == 0) {
2618                         BUG_ON(s->uptodate != disks);
2619                         sh->check_state = check_state_run;
2620                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2621                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2622                         s->uptodate--;
2623                         break;
2624                 }
2625                 dev = &sh->dev[s->failed_num];
2626                 /* fall through */
2627         case check_state_compute_result:
2628                 sh->check_state = check_state_idle;
2629                 if (!dev)
2630                         dev = &sh->dev[sh->pd_idx];
2631
2632                 /* check that a write has not made the stripe insync */
2633                 if (test_bit(STRIPE_INSYNC, &sh->state))
2634                         break;
2635
2636                 /* either failed parity check, or recovery is happening */
2637                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2638                 BUG_ON(s->uptodate != disks);
2639
2640                 set_bit(R5_LOCKED, &dev->flags);
2641                 s->locked++;
2642                 set_bit(R5_Wantwrite, &dev->flags);
2643
2644                 clear_bit(STRIPE_DEGRADED, &sh->state);
2645                 set_bit(STRIPE_INSYNC, &sh->state);
2646                 break;
2647         case check_state_run:
2648                 break; /* we will be called again upon completion */
2649         case check_state_check_result:
2650                 sh->check_state = check_state_idle;
2651
2652                 /* if a failure occurred during the check operation, leave
2653                  * STRIPE_INSYNC not set and let the stripe be handled again
2654                  */
2655                 if (s->failed)
2656                         break;
2657
2658                 /* handle a successful check operation, if parity is correct
2659                  * we are done.  Otherwise update the mismatch count and repair
2660                  * parity if !MD_RECOVERY_CHECK
2661                  */
2662                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2663                         /* parity is correct (on disc,
2664                          * not in buffer any more)
2665                          */
2666                         set_bit(STRIPE_INSYNC, &sh->state);
2667                 else {
2668                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2669                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2670                                 /* don't try to repair!! */
2671                                 set_bit(STRIPE_INSYNC, &sh->state);
2672                         else {
2673                                 sh->check_state = check_state_compute_run;
2674                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2675                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2676                                 set_bit(R5_Wantcompute,
2677                                         &sh->dev[sh->pd_idx].flags);
2678                                 sh->ops.target = sh->pd_idx;
2679                                 sh->ops.target2 = -1;
2680                                 s->uptodate++;
2681                         }
2682                 }
2683                 break;
2684         case check_state_compute_run:
2685                 break;
2686         default:
2687                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2688                        __func__, sh->check_state,
2689                        (unsigned long long) sh->sector);
2690                 BUG();
2691         }
2692 }
2693
2694
2695 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2696                                   struct stripe_head_state *s,
2697                                   struct r6_state *r6s, int disks)
2698 {
2699         int pd_idx = sh->pd_idx;
2700         int qd_idx = sh->qd_idx;
2701         struct r5dev *dev;
2702
2703         set_bit(STRIPE_HANDLE, &sh->state);
2704
2705         BUG_ON(s->failed > 2);
2706
2707         /* Want to check and possibly repair P and Q.
2708          * However there could be one 'failed' device, in which
2709          * case we can only check one of them, possibly using the
2710          * other to generate missing data
2711          */
2712
2713         switch (sh->check_state) {
2714         case check_state_idle:
2715                 /* start a new check operation if there are < 2 failures */
2716                 if (s->failed == r6s->q_failed) {
2717                         /* The only possible failed device holds Q, so it
2718                          * makes sense to check P (If anything else were failed,
2719                          * we would have used P to recreate it).
2720                          */
2721                         sh->check_state = check_state_run;
2722                 }
2723                 if (!r6s->q_failed && s->failed < 2) {
2724                         /* Q is not failed, and we didn't use it to generate
2725                          * anything, so it makes sense to check it
2726                          */
2727                         if (sh->check_state == check_state_run)
2728                                 sh->check_state = check_state_run_pq;
2729                         else
2730                                 sh->check_state = check_state_run_q;
2731                 }
2732
2733                 /* discard potentially stale zero_sum_result */
2734                 sh->ops.zero_sum_result = 0;
2735
2736                 if (sh->check_state == check_state_run) {
2737                         /* async_xor_zero_sum destroys the contents of P */
2738                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2739                         s->uptodate--;
2740                 }
2741                 if (sh->check_state >= check_state_run &&
2742                     sh->check_state <= check_state_run_pq) {
2743                         /* async_syndrome_zero_sum preserves P and Q, so
2744                          * no need to mark them !uptodate here
2745                          */
2746                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2747                         break;
2748                 }
2749
2750                 /* we have 2-disk failure */
2751                 BUG_ON(s->failed != 2);
2752                 /* fall through */
2753         case check_state_compute_result:
2754                 sh->check_state = check_state_idle;
2755
2756                 /* check that a write has not made the stripe insync */
2757                 if (test_bit(STRIPE_INSYNC, &sh->state))
2758                         break;
2759
2760                 /* now write out any block on a failed drive,
2761                  * or P or Q if they were recomputed
2762                  */
2763                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2764                 if (s->failed == 2) {
2765                         dev = &sh->dev[r6s->failed_num[1]];
2766                         s->locked++;
2767                         set_bit(R5_LOCKED, &dev->flags);
2768                         set_bit(R5_Wantwrite, &dev->flags);
2769                 }
2770                 if (s->failed >= 1) {
2771                         dev = &sh->dev[r6s->failed_num[0]];
2772                         s->locked++;
2773                         set_bit(R5_LOCKED, &dev->flags);
2774                         set_bit(R5_Wantwrite, &dev->flags);
2775                 }
2776                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2777                         dev = &sh->dev[pd_idx];
2778                         s->locked++;
2779                         set_bit(R5_LOCKED, &dev->flags);
2780                         set_bit(R5_Wantwrite, &dev->flags);
2781                 }
2782                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2783                         dev = &sh->dev[qd_idx];
2784                         s->locked++;
2785                         set_bit(R5_LOCKED, &dev->flags);
2786                         set_bit(R5_Wantwrite, &dev->flags);
2787                 }
2788                 clear_bit(STRIPE_DEGRADED, &sh->state);
2789
2790                 set_bit(STRIPE_INSYNC, &sh->state);
2791                 break;
2792         case check_state_run:
2793         case check_state_run_q:
2794         case check_state_run_pq:
2795                 break; /* we will be called again upon completion */
2796         case check_state_check_result:
2797                 sh->check_state = check_state_idle;
2798
2799                 /* handle a successful check operation, if parity is correct
2800                  * we are done.  Otherwise update the mismatch count and repair
2801                  * parity if !MD_RECOVERY_CHECK
2802                  */
2803                 if (sh->ops.zero_sum_result == 0) {
2804                         /* both parities are correct */
2805                         if (!s->failed)
2806                                 set_bit(STRIPE_INSYNC, &sh->state);
2807                         else {
2808                                 /* in contrast to the raid5 case we can validate
2809                                  * parity, but still have a failure to write
2810                                  * back
2811                                  */
2812                                 sh->check_state = check_state_compute_result;
2813                                 /* Returning at this point means that we may go
2814                                  * off and bring p and/or q uptodate again so
2815                                  * we make sure to check zero_sum_result again
2816                                  * to verify if p or q need writeback
2817                                  */
2818                         }
2819                 } else {
2820                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2821                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2822                                 /* don't try to repair!! */
2823                                 set_bit(STRIPE_INSYNC, &sh->state);
2824                         else {
2825                                 int *target = &sh->ops.target;
2826
2827                                 sh->ops.target = -1;
2828                                 sh->ops.target2 = -1;
2829                                 sh->check_state = check_state_compute_run;
2830                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2831                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2832                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2833                                         set_bit(R5_Wantcompute,
2834                                                 &sh->dev[pd_idx].flags);
2835                                         *target = pd_idx;
2836                                         target = &sh->ops.target2;
2837                                         s->uptodate++;
2838                                 }
2839                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2840                                         set_bit(R5_Wantcompute,
2841                                                 &sh->dev[qd_idx].flags);
2842                                         *target = qd_idx;
2843                                         s->uptodate++;
2844                                 }
2845                         }
2846                 }
2847                 break;
2848         case check_state_compute_run:
2849                 break;
2850         default:
2851                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2852                        __func__, sh->check_state,
2853                        (unsigned long long) sh->sector);
2854                 BUG();
2855         }
2856 }
2857
2858 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2859                                 struct r6_state *r6s)
2860 {
2861         int i;
2862
2863         /* We have read all the blocks in this stripe and now we need to
2864          * copy some of them into a target stripe for expand.
2865          */
2866         struct dma_async_tx_descriptor *tx = NULL;
2867         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2868         for (i = 0; i < sh->disks; i++)
2869                 if (i != sh->pd_idx && i != sh->qd_idx) {
2870                         int dd_idx, j;
2871                         struct stripe_head *sh2;
2872                         struct async_submit_ctl submit;
2873
2874                         sector_t bn = compute_blocknr(sh, i, 1);
2875                         sector_t s = raid5_compute_sector(conf, bn, 0,
2876                                                           &dd_idx, NULL);
2877                         sh2 = get_active_stripe(conf, s, 0, 1, 1);
2878                         if (sh2 == NULL)
2879                                 /* so far only the early blocks of this stripe
2880                                  * have been requested.  When later blocks
2881                                  * get requested, we will try again
2882                                  */
2883                                 continue;
2884                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2885                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2886                                 /* must have already done this block */
2887                                 release_stripe(sh2);
2888                                 continue;
2889                         }
2890
2891                         /* place all the copies on one channel */
2892                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2893                         tx = async_memcpy(sh2->dev[dd_idx].page,
2894                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
2895                                           &submit);
2896
2897                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2898                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2899                         for (j = 0; j < conf->raid_disks; j++)
2900                                 if (j != sh2->pd_idx &&
2901                                     (!r6s || j != sh2->qd_idx) &&
2902                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2903                                         break;
2904                         if (j == conf->raid_disks) {
2905                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2906                                 set_bit(STRIPE_HANDLE, &sh2->state);
2907                         }
2908                         release_stripe(sh2);
2909
2910                 }
2911         /* done submitting copies, wait for them to complete */
2912         if (tx) {
2913                 async_tx_ack(tx);
2914                 dma_wait_for_async_tx(tx);
2915         }
2916 }
2917
2918
2919 /*
2920  * handle_stripe - do things to a stripe.
2921  *
2922  * We lock the stripe and then examine the state of various bits
2923  * to see what needs to be done.
2924  * Possible results:
2925  *    return some read request which now have data
2926  *    return some write requests which are safely on disc
2927  *    schedule a read on some buffers
2928  *    schedule a write of some buffers
2929  *    return confirmation of parity correctness
2930  *
2931  * buffers are taken off read_list or write_list, and bh_cache buffers
2932  * get BH_Lock set before the stripe lock is released.
2933  *
2934  */
2935
2936 static void handle_stripe5(struct stripe_head *sh)
2937 {
2938         raid5_conf_t *conf = sh->raid_conf;
2939         int disks = sh->disks, i;
2940         struct bio *return_bi = NULL;
2941         struct stripe_head_state s;
2942         struct r5dev *dev;
2943         mdk_rdev_t *blocked_rdev = NULL;
2944         int prexor;
2945         int dec_preread_active = 0;
2946
2947         memset(&s, 0, sizeof(s));
2948         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2949                  "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2950                  atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2951                  sh->reconstruct_state);
2952
2953         spin_lock(&sh->lock);
2954         clear_bit(STRIPE_HANDLE, &sh->state);
2955         clear_bit(STRIPE_DELAYED, &sh->state);
2956
2957         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2958         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2959         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2960
2961         /* Now to look around and see what can be done */
2962         rcu_read_lock();
2963         for (i=disks; i--; ) {
2964                 mdk_rdev_t *rdev;
2965
2966                 dev = &sh->dev[i];
2967                 clear_bit(R5_Insync, &dev->flags);
2968
2969                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2970                         "written %p\n", i, dev->flags, dev->toread, dev->read,
2971                         dev->towrite, dev->written);
2972
2973                 /* maybe we can request a biofill operation
2974                  *
2975                  * new wantfill requests are only permitted while
2976                  * ops_complete_biofill is guaranteed to be inactive
2977                  */
2978                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2979                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2980                         set_bit(R5_Wantfill, &dev->flags);
2981
2982                 /* now count some things */
2983                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2984                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2985                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2986
2987                 if (test_bit(R5_Wantfill, &dev->flags))
2988                         s.to_fill++;
2989                 else if (dev->toread)
2990                         s.to_read++;
2991                 if (dev->towrite) {
2992                         s.to_write++;
2993                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2994                                 s.non_overwrite++;
2995                 }
2996                 if (dev->written)
2997                         s.written++;
2998                 rdev = rcu_dereference(conf->disks[i].rdev);
2999                 if (blocked_rdev == NULL &&
3000                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3001                         blocked_rdev = rdev;
3002                         atomic_inc(&rdev->nr_pending);
3003                 }
3004                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3005                         /* The ReadError flag will just be confusing now */
3006                         clear_bit(R5_ReadError, &dev->flags);
3007                         clear_bit(R5_ReWrite, &dev->flags);
3008                 }
3009                 if (!rdev || !test_bit(In_sync, &rdev->flags)
3010                     || test_bit(R5_ReadError, &dev->flags)) {
3011                         s.failed++;
3012                         s.failed_num = i;
3013                 } else
3014                         set_bit(R5_Insync, &dev->flags);
3015         }
3016         rcu_read_unlock();
3017
3018         if (unlikely(blocked_rdev)) {
3019                 if (s.syncing || s.expanding || s.expanded ||
3020                     s.to_write || s.written) {
3021                         set_bit(STRIPE_HANDLE, &sh->state);
3022                         goto unlock;
3023                 }
3024                 /* There is nothing for the blocked_rdev to block */
3025                 rdev_dec_pending(blocked_rdev, conf->mddev);
3026                 blocked_rdev = NULL;
3027         }
3028
3029         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3030                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3031                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3032         }
3033
3034         pr_debug("locked=%d uptodate=%d to_read=%d"
3035                 " to_write=%d failed=%d failed_num=%d\n",
3036                 s.locked, s.uptodate, s.to_read, s.to_write,
3037                 s.failed, s.failed_num);
3038         /* check if the array has lost two devices and, if so, some requests might
3039          * need to be failed
3040          */
3041         if (s.failed > 1 && s.to_read+s.to_write+s.written)
3042                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3043         if (s.failed > 1 && s.syncing) {
3044                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3045                 clear_bit(STRIPE_SYNCING, &sh->state);
3046                 s.syncing = 0;
3047         }
3048
3049         /* might be able to return some write requests if the parity block
3050          * is safe, or on a failed drive
3051          */
3052         dev = &sh->dev[sh->pd_idx];
3053         if ( s.written &&
3054              ((test_bit(R5_Insync, &dev->flags) &&
3055                !test_bit(R5_LOCKED, &dev->flags) &&
3056                test_bit(R5_UPTODATE, &dev->flags)) ||
3057                (s.failed == 1 && s.failed_num == sh->pd_idx)))
3058                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3059
3060         /* Now we might consider reading some blocks, either to check/generate
3061          * parity, or to satisfy requests
3062          * or to load a block that is being partially written.
3063          */
3064         if (s.to_read || s.non_overwrite ||
3065             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3066                 handle_stripe_fill5(sh, &s, disks);
3067
3068         /* Now we check to see if any write operations have recently
3069          * completed
3070          */
3071         prexor = 0;
3072         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3073                 prexor = 1;
3074         if (sh->reconstruct_state == reconstruct_state_drain_result ||
3075             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3076                 sh->reconstruct_state = reconstruct_state_idle;
3077
3078                 /* All the 'written' buffers and the parity block are ready to
3079                  * be written back to disk
3080                  */
3081                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3082                 for (i = disks; i--; ) {
3083                         dev = &sh->dev[i];
3084                         if (test_bit(R5_LOCKED, &dev->flags) &&
3085                                 (i == sh->pd_idx || dev->written)) {
3086                                 pr_debug("Writing block %d\n", i);
3087                                 set_bit(R5_Wantwrite, &dev->flags);
3088                                 if (prexor)
3089                                         continue;
3090                                 if (!test_bit(R5_Insync, &dev->flags) ||
3091                                     (i == sh->pd_idx && s.failed == 0))
3092                                         set_bit(STRIPE_INSYNC, &sh->state);
3093                         }
3094                 }
3095                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3096                         dec_preread_active = 1;
3097         }
3098
3099         /* Now to consider new write requests and what else, if anything
3100          * should be read.  We do not handle new writes when:
3101          * 1/ A 'write' operation (copy+xor) is already in flight.
3102          * 2/ A 'check' operation is in flight, as it may clobber the parity
3103          *    block.
3104          */
3105         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3106                 handle_stripe_dirtying5(conf, sh, &s, disks);
3107
3108         /* maybe we need to check and possibly fix the parity for this stripe
3109          * Any reads will already have been scheduled, so we just see if enough
3110          * data is available.  The parity check is held off while parity
3111          * dependent operations are in flight.
3112          */
3113         if (sh->check_state ||
3114             (s.syncing && s.locked == 0 &&
3115              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3116              !test_bit(STRIPE_INSYNC, &sh->state)))
3117                 handle_parity_checks5(conf, sh, &s, disks);
3118
3119         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3120                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3121                 clear_bit(STRIPE_SYNCING, &sh->state);
3122         }
3123
3124         /* If the failed drive is just a ReadError, then we might need to progress
3125          * the repair/check process
3126          */
3127         if (s.failed == 1 && !conf->mddev->ro &&
3128             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
3129             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
3130             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
3131                 ) {
3132                 dev = &sh->dev[s.failed_num];
3133                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3134                         set_bit(R5_Wantwrite, &dev->flags);
3135                         set_bit(R5_ReWrite, &dev->flags);
3136                         set_bit(R5_LOCKED, &dev->flags);
3137                         s.locked++;
3138                 } else {
3139                         /* let's read it back */
3140                         set_bit(R5_Wantread, &dev->flags);
3141                         set_bit(R5_LOCKED, &dev->flags);
3142                         s.locked++;
3143                 }
3144         }
3145
3146         /* Finish reconstruct operations initiated by the expansion process */
3147         if (sh->reconstruct_state == reconstruct_state_result) {
3148                 struct stripe_head *sh2
3149                         = get_active_stripe(conf, sh->sector, 1, 1, 1);
3150                 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3151                         /* sh cannot be written until sh2 has been read.
3152                          * so arrange for sh to be delayed a little
3153                          */
3154                         set_bit(STRIPE_DELAYED, &sh->state);
3155                         set_bit(STRIPE_HANDLE, &sh->state);
3156                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3157                                               &sh2->state))
3158                                 atomic_inc(&conf->preread_active_stripes);
3159                         release_stripe(sh2);
3160                         goto unlock;
3161                 }
3162                 if (sh2)
3163                         release_stripe(sh2);
3164
3165                 sh->reconstruct_state = reconstruct_state_idle;
3166                 clear_bit(STRIPE_EXPANDING, &sh->state);
3167                 for (i = conf->raid_disks; i--; ) {
3168                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3169                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3170                         s.locked++;
3171                 }
3172         }
3173
3174         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3175             !sh->reconstruct_state) {
3176                 /* Need to write out all blocks after computing parity */
3177                 sh->disks = conf->raid_disks;
3178                 stripe_set_idx(sh->sector, conf, 0, sh);
3179                 schedule_reconstruction(sh, &s, 1, 1);
3180         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3181                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3182                 atomic_dec(&conf->reshape_stripes);
3183                 wake_up(&conf->wait_for_overlap);
3184                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3185         }
3186
3187         if (s.expanding && s.locked == 0 &&
3188             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3189                 handle_stripe_expansion(conf, sh, NULL);
3190
3191  unlock:
3192         spin_unlock(&sh->lock);
3193
3194         /* wait for this device to become unblocked */
3195         if (unlikely(blocked_rdev))
3196                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3197
3198         if (s.ops_request)
3199                 raid_run_ops(sh, s.ops_request);
3200
3201         ops_run_io(sh, &s);
3202
3203         if (dec_preread_active) {
3204                 /* We delay this until after ops_run_io so that if make_request
3205                  * is waiting on a barrier, it won't continue until the writes
3206                  * have actually been submitted.
3207                  */
3208                 atomic_dec(&conf->preread_active_stripes);
3209                 if (atomic_read(&conf->preread_active_stripes) <
3210                     IO_THRESHOLD)
3211                         md_wakeup_thread(conf->mddev->thread);
3212         }
3213         return_io(return_bi);
3214 }
3215
3216 static void handle_stripe6(struct stripe_head *sh)
3217 {
3218         raid5_conf_t *conf = sh->raid_conf;
3219         int disks = sh->disks;
3220         struct bio *return_bi = NULL;
3221         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3222         struct stripe_head_state s;
3223         struct r6_state r6s;
3224         struct r5dev *dev, *pdev, *qdev;
3225         mdk_rdev_t *blocked_rdev = NULL;
3226         int dec_preread_active = 0;
3227
3228         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3229                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3230                (unsigned long long)sh->sector, sh->state,
3231                atomic_read(&sh->count), pd_idx, qd_idx,
3232                sh->check_state, sh->reconstruct_state);
3233         memset(&s, 0, sizeof(s));
3234
3235         spin_lock(&sh->lock);
3236         clear_bit(STRIPE_HANDLE, &sh->state);
3237         clear_bit(STRIPE_DELAYED, &sh->state);
3238
3239         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3240         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3241         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3242         /* Now to look around and see what can be done */
3243
3244         rcu_read_lock();
3245         for (i=disks; i--; ) {
3246                 mdk_rdev_t *rdev;
3247                 dev = &sh->dev[i];
3248                 clear_bit(R5_Insync, &dev->flags);
3249
3250                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3251                         i, dev->flags, dev->toread, dev->towrite, dev->written);
3252                 /* maybe we can reply to a read
3253                  *
3254                  * new wantfill requests are only permitted while
3255                  * ops_complete_biofill is guaranteed to be inactive
3256                  */
3257                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3258                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3259                         set_bit(R5_Wantfill, &dev->flags);
3260
3261                 /* now count some things */
3262                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3263                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3264                 if (test_bit(R5_Wantcompute, &dev->flags)) {
3265                         s.compute++;
3266                         BUG_ON(s.compute > 2);
3267                 }
3268
3269                 if (test_bit(R5_Wantfill, &dev->flags)) {
3270                         s.to_fill++;
3271                 } else if (dev->toread)
3272                         s.to_read++;
3273                 if (dev->towrite) {
3274                         s.to_write++;
3275                         if (!test_bit(R5_OVERWRITE, &dev->flags))
3276                                 s.non_overwrite++;
3277                 }
3278                 if (dev->written)
3279                         s.written++;
3280                 rdev = rcu_dereference(conf->disks[i].rdev);
3281                 if (blocked_rdev == NULL &&
3282                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3283                         blocked_rdev = rdev;
3284                         atomic_inc(&rdev->nr_pending);
3285                 }
3286                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3287                         /* The ReadError flag will just be confusing now */
3288                         clear_bit(R5_ReadError, &dev->flags);
3289                         clear_bit(R5_ReWrite, &dev->flags);
3290                 }
3291                 if (!rdev || !test_bit(In_sync, &rdev->flags)
3292                     || test_bit(R5_ReadError, &dev->flags)) {
3293                         if (s.failed < 2)
3294                                 r6s.failed_num[s.failed] = i;
3295                         s.failed++;
3296                 } else
3297                         set_bit(R5_Insync, &dev->flags);
3298         }
3299         rcu_read_unlock();
3300
3301         if (unlikely(blocked_rdev)) {
3302                 if (s.syncing || s.expanding || s.expanded ||
3303                     s.to_write || s.written) {
3304                         set_bit(STRIPE_HANDLE, &sh->state);
3305                         goto unlock;
3306                 }
3307                 /* There is nothing for the blocked_rdev to block */
3308                 rdev_dec_pending(blocked_rdev, conf->mddev);
3309                 blocked_rdev = NULL;
3310         }
3311
3312         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3313                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3314                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3315         }
3316
3317         pr_debug("locked=%d uptodate=%d to_read=%d"
3318                " to_write=%d failed=%d failed_num=%d,%d\n",
3319                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3320                r6s.failed_num[0], r6s.failed_num[1]);
3321         /* check if the array has lost >2 devices and, if so, some requests
3322          * might need to be failed
3323          */
3324         if (s.failed > 2 && s.to_read+s.to_write+s.written)
3325                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3326         if (s.failed > 2 && s.syncing) {
3327                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3328                 clear_bit(STRIPE_SYNCING, &sh->state);
3329                 s.syncing = 0;
3330         }
3331
3332         /*
3333          * might be able to return some write requests if the parity blocks
3334          * are safe, or on a failed drive
3335          */
3336         pdev = &sh->dev[pd_idx];
3337         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3338                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3339         qdev = &sh->dev[qd_idx];
3340         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == qd_idx)
3341                 || (s.failed >= 2 && r6s.failed_num[1] == qd_idx);
3342
3343         if ( s.written &&
3344              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3345                              && !test_bit(R5_LOCKED, &pdev->flags)
3346                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3347              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3348                              && !test_bit(R5_LOCKED, &qdev->flags)
3349                              && test_bit(R5_UPTODATE, &qdev->flags)))))
3350                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3351
3352         /* Now we might consider reading some blocks, either to check/generate
3353          * parity, or to satisfy requests
3354          * or to load a block that is being partially written.
3355          */
3356         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3357             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3358                 handle_stripe_fill6(sh, &s, &r6s, disks);
3359
3360         /* Now we check to see if any write operations have recently
3361          * completed
3362          */
3363         if (sh->reconstruct_state == reconstruct_state_drain_result) {
3364
3365                 sh->reconstruct_state = reconstruct_state_idle;
3366                 /* All the 'written' buffers and the parity blocks are ready to
3367                  * be written back to disk
3368                  */
3369                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3370                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags));
3371                 for (i = disks; i--; ) {
3372                         dev = &sh->dev[i];
3373                         if (test_bit(R5_LOCKED, &dev->flags) &&
3374                             (i == sh->pd_idx || i == qd_idx ||
3375                              dev->written)) {
3376                                 pr_debug("Writing block %d\n", i);
3377                                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3378                                 set_bit(R5_Wantwrite, &dev->flags);
3379                                 if (!test_bit(R5_Insync, &dev->flags) ||
3380                                     ((i == sh->pd_idx || i == qd_idx) &&
3381                                       s.failed == 0))
3382                                         set_bit(STRIPE_INSYNC, &sh->state);
3383                         }
3384                 }
3385                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3386                         dec_preread_active = 1;
3387         }
3388
3389         /* Now to consider new write requests and what else, if anything
3390          * should be read.  We do not handle new writes when:
3391          * 1/ A 'write' operation (copy+gen_syndrome) is already in flight.
3392          * 2/ A 'check' operation is in flight, as it may clobber the parity
3393          *    block.
3394          */
3395         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3396                 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3397
3398         /* maybe we need to check and possibly fix the parity for this stripe
3399          * Any reads will already have been scheduled, so we just see if enough
3400          * data is available.  The parity check is held off while parity
3401          * dependent operations are in flight.
3402          */
3403         if (sh->check_state ||
3404             (s.syncing && s.locked == 0 &&
3405              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3406              !test_bit(STRIPE_INSYNC, &sh->state)))
3407                 handle_parity_checks6(conf, sh, &s, &r6s, disks);
3408
3409         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3410                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3411                 clear_bit(STRIPE_SYNCING, &sh->state);
3412         }
3413
3414         /* If the failed drives are just a ReadError, then we might need
3415          * to progress the repair/check process
3416          */
3417         if (s.failed <= 2 && !conf->mddev->ro)
3418                 for (i = 0; i < s.failed; i++) {
3419                         dev = &sh->dev[r6s.failed_num[i]];
3420                         if (test_bit(R5_ReadError, &dev->flags)
3421                             && !test_bit(R5_LOCKED, &dev->flags)
3422                             && test_bit(R5_UPTODATE, &dev->flags)
3423                                 ) {
3424                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3425                                         set_bit(R5_Wantwrite, &dev->flags);
3426                                         set_bit(R5_ReWrite, &dev->flags);
3427                                         set_bit(R5_LOCKED, &dev->flags);
3428                                         s.locked++;
3429                                 } else {
3430                                         /* let's read it back */
3431                                         set_bit(R5_Wantread, &dev->flags);
3432                                         set_bit(R5_LOCKED, &dev->flags);
3433                                         s.locked++;
3434                                 }
3435                         }
3436                 }
3437
3438         /* Finish reconstruct operations initiated by the expansion process */
3439         if (sh->reconstruct_state == reconstruct_state_result) {
3440                 sh->reconstruct_state = reconstruct_state_idle;
3441                 clear_bit(STRIPE_EXPANDING, &sh->state);
3442                 for (i = conf->raid_disks; i--; ) {
3443                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3444                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3445                         s.locked++;
3446                 }
3447         }
3448
3449         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3450             !sh->reconstruct_state) {
3451                 struct stripe_head *sh2
3452                         = get_active_stripe(conf, sh->sector, 1, 1, 1);
3453                 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3454                         /* sh cannot be written until sh2 has been read.
3455                          * so arrange for sh to be delayed a little
3456                          */
3457                         set_bit(STRIPE_DELAYED, &sh->state);
3458                         set_bit(STRIPE_HANDLE, &sh->state);
3459                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3460                                               &sh2->state))
3461                                 atomic_inc(&conf->preread_active_stripes);
3462                         release_stripe(sh2);
3463                         goto unlock;
3464                 }
3465                 if (sh2)
3466                         release_stripe(sh2);
3467
3468                 /* Need to write out all blocks after computing P&Q */
3469                 sh->disks = conf->raid_disks;
3470                 stripe_set_idx(sh->sector, conf, 0, sh);
3471                 schedule_reconstruction(sh, &s, 1, 1);
3472         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3473                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3474                 atomic_dec(&conf->reshape_stripes);
3475                 wake_up(&conf->wait_for_overlap);
3476                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3477         }
3478
3479         if (s.expanding && s.locked == 0 &&
3480             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3481                 handle_stripe_expansion(conf, sh, &r6s);
3482
3483  unlock:
3484         spin_unlock(&sh->lock);
3485
3486         /* wait for this device to become unblocked */
3487         if (unlikely(blocked_rdev))
3488                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3489
3490         if (s.ops_request)
3491                 raid_run_ops(sh, s.ops_request);
3492
3493         ops_run_io(sh, &s);
3494
3495
3496         if (dec_preread_active) {
3497                 /* We delay this until after ops_run_io so that if make_request
3498                  * is waiting on a barrier, it won't continue until the writes
3499                  * have actually been submitted.
3500                  */
3501                 atomic_dec(&conf->preread_active_stripes);
3502                 if (atomic_read(&conf->preread_active_stripes) <
3503                     IO_THRESHOLD)
3504                         md_wakeup_thread(conf->mddev->thread);
3505         }
3506
3507         return_io(return_bi);
3508 }
3509
3510 static void handle_stripe(struct stripe_head *sh)
3511 {
3512         if (sh->raid_conf->level == 6)
3513                 handle_stripe6(sh);
3514         else
3515                 handle_stripe5(sh);
3516 }
3517
3518 static void raid5_activate_delayed(raid5_conf_t *conf)
3519 {
3520         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3521                 while (!list_empty(&conf->delayed_list)) {
3522                         struct list_head *l = conf->delayed_list.next;
3523                         struct stripe_head *sh;
3524                         sh = list_entry(l, struct stripe_head, lru);
3525                         list_del_init(l);
3526                         clear_bit(STRIPE_DELAYED, &sh->state);
3527                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3528                                 atomic_inc(&conf->preread_active_stripes);
3529                         list_add_tail(&sh->lru, &conf->hold_list);
3530                 }
3531         } else
3532                 blk_plug_device(conf->mddev->queue);
3533 }
3534
3535 static void activate_bit_delay(raid5_conf_t *conf)
3536 {
3537         /* device_lock is held */
3538         struct list_head head;
3539         list_add(&head, &conf->bitmap_list);
3540         list_del_init(&conf->bitmap_list);
3541         while (!list_empty(&head)) {
3542                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3543                 list_del_init(&sh->lru);
3544                 atomic_inc(&sh->count);
3545                 __release_stripe(conf, sh);
3546         }
3547 }
3548
3549 static void unplug_slaves(mddev_t *mddev)
3550 {
3551         raid5_conf_t *conf = mddev->private;
3552         int i;
3553         int devs = max(conf->raid_disks, conf->previous_raid_disks);
3554
3555         rcu_read_lock();
3556         for (i = 0; i < devs; i++) {
3557                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3558                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3559                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3560
3561                         atomic_inc(&rdev->nr_pending);
3562                         rcu_read_unlock();
3563
3564                         blk_unplug(r_queue);
3565
3566                         rdev_dec_pending(rdev, mddev);
3567                         rcu_read_lock();
3568                 }
3569         }
3570         rcu_read_unlock();
3571 }
3572
3573 static void raid5_unplug_device(struct request_queue *q)
3574 {
3575         mddev_t *mddev = q->queuedata;
3576         raid5_conf_t *conf = mddev->private;
3577         unsigned long flags;
3578
3579         spin_lock_irqsave(&conf->device_lock, flags);
3580
3581         if (blk_remove_plug(q)) {
3582                 conf->seq_flush++;
3583                 raid5_activate_delayed(conf);
3584         }
3585         md_wakeup_thread(mddev->thread);
3586
3587         spin_unlock_irqrestore(&conf->device_lock, flags);
3588
3589         unplug_slaves(mddev);
3590 }
3591
3592 static int raid5_congested(void *data, int bits)
3593 {
3594         mddev_t *mddev = data;
3595         raid5_conf_t *conf = mddev->private;
3596
3597         /* No difference between reads and writes.  Just check
3598          * how busy the stripe_cache is
3599          */
3600
3601         if (mddev_congested(mddev, bits))
3602                 return 1;
3603         if (conf->inactive_blocked)
3604                 return 1;
3605         if (conf->quiesce)
3606                 return 1;
3607         if (list_empty_careful(&conf->inactive_list))
3608                 return 1;
3609
3610         return 0;
3611 }
3612
3613 /* We want read requests to align with chunks where possible,
3614  * but write requests don't need to.
3615  */
3616 static int raid5_mergeable_bvec(struct request_queue *q,
3617                                 struct bvec_merge_data *bvm,
3618                                 struct bio_vec *biovec)
3619 {
3620         mddev_t *mddev = q->queuedata;
3621         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3622         int max;
3623         unsigned int chunk_sectors = mddev->chunk_sectors;
3624         unsigned int bio_sectors = bvm->bi_size >> 9;
3625
3626         if ((bvm->bi_rw & 1) == WRITE)
3627                 return biovec->bv_len; /* always allow writes to be mergeable */
3628
3629         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3630                 chunk_sectors = mddev->new_chunk_sectors;
3631         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3632         if (max < 0) max = 0;
3633         if (max <= biovec->bv_len && bio_sectors == 0)
3634                 return biovec->bv_len;
3635         else
3636                 return max;
3637 }
3638
3639
3640 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3641 {
3642         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3643         unsigned int chunk_sectors = mddev->chunk_sectors;
3644         unsigned int bio_sectors = bio->bi_size >> 9;
3645
3646         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3647                 chunk_sectors = mddev->new_chunk_sectors;
3648         return  chunk_sectors >=
3649                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3650 }
3651
3652 /*
3653  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3654  *  later sampled by raid5d.
3655  */
3656 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3657 {
3658         unsigned long flags;
3659
3660         spin_lock_irqsave(&conf->device_lock, flags);
3661
3662         bi->bi_next = conf->retry_read_aligned_list;
3663         conf->retry_read_aligned_list = bi;
3664
3665         spin_unlock_irqrestore(&conf->device_lock, flags);
3666         md_wakeup_thread(conf->mddev->thread);
3667 }
3668
3669
3670 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3671 {
3672         struct bio *bi;
3673
3674         bi = conf->retry_read_aligned;
3675         if (bi) {
3676                 conf->retry_read_aligned = NULL;
3677                 return bi;
3678         }
3679         bi = conf->retry_read_aligned_list;
3680         if(bi) {
3681                 conf->retry_read_aligned_list = bi->bi_next;
3682                 bi->bi_next = NULL;
3683                 /*
3684                  * this sets the active strip count to 1 and the processed
3685                  * strip count to zero (upper 8 bits)
3686                  */
3687                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3688         }
3689
3690         return bi;
3691 }
3692
3693
3694 /*
3695  *  The "raid5_align_endio" should check if the read succeeded and if it
3696  *  did, call bio_endio on the original bio (having bio_put the new bio
3697  *  first).
3698  *  If the read failed..
3699  */
3700 static void raid5_align_endio(struct bio *bi, int error)
3701 {
3702         struct bio* raid_bi  = bi->bi_private;
3703         mddev_t *mddev;
3704         raid5_conf_t *conf;
3705         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3706         mdk_rdev_t *rdev;
3707
3708         bio_put(bi);
3709
3710         rdev = (void*)raid_bi->bi_next;
3711         raid_bi->bi_next = NULL;
3712         mddev = rdev->mddev;
3713         conf = mddev->private;
3714
3715         rdev_dec_pending(rdev, conf->mddev);
3716
3717         if (!error && uptodate) {
3718                 bio_endio(raid_bi, 0);
3719                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3720                         wake_up(&conf->wait_for_stripe);
3721                 return;
3722         }
3723
3724
3725         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3726
3727         add_bio_to_retry(raid_bi, conf);
3728 }
3729
3730 static int bio_fits_rdev(struct bio *bi)
3731 {
3732         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3733
3734         if ((bi->bi_size>>9) > queue_max_sectors(q))
3735                 return 0;
3736         blk_recount_segments(q, bi);
3737         if (bi->bi_phys_segments > queue_max_segments(q))
3738                 return 0;
3739
3740         if (q->merge_bvec_fn)
3741                 /* it's too hard to apply the merge_bvec_fn at this stage,
3742                  * just just give up
3743                  */
3744                 return 0;
3745
3746         return 1;
3747 }
3748
3749
3750 static int chunk_aligned_read(mddev_t *mddev, struct bio * raid_bio)
3751 {
3752         raid5_conf_t *conf = mddev->private;
3753         int dd_idx;
3754         struct bio* align_bi;
3755         mdk_rdev_t *rdev;
3756
3757         if (!in_chunk_boundary(mddev, raid_bio)) {
3758                 pr_debug("chunk_aligned_read : non aligned\n");
3759                 return 0;
3760         }
3761         /*
3762          * use bio_clone to make a copy of the bio
3763          */
3764         align_bi = bio_clone(raid_bio, GFP_NOIO);
3765         if (!align_bi)
3766                 return 0;
3767         /*
3768          *   set bi_end_io to a new function, and set bi_private to the
3769          *     original bio.
3770          */
3771         align_bi->bi_end_io  = raid5_align_endio;
3772         align_bi->bi_private = raid_bio;
3773         /*
3774          *      compute position
3775          */
3776         align_bi->bi_sector =  raid5_compute_sector(conf, raid_bio->bi_sector,
3777                                                     0,
3778                                                     &dd_idx, NULL);
3779
3780         rcu_read_lock();
3781         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3782         if (rdev && test_bit(In_sync, &rdev->flags)) {
3783                 atomic_inc(&rdev->nr_pending);
3784                 rcu_read_unlock();
3785                 raid_bio->bi_next = (void*)rdev;
3786                 align_bi->bi_bdev =  rdev->bdev;
3787                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3788                 align_bi->bi_sector += rdev->data_offset;
3789
3790                 if (!bio_fits_rdev(align_bi)) {
3791                         /* too big in some way */
3792                         bio_put(align_bi);
3793                         rdev_dec_pending(rdev, mddev);
3794                         return 0;
3795                 }
3796
3797                 spin_lock_irq(&conf->device_lock);
3798                 wait_event_lock_irq(conf->wait_for_stripe,
3799                                     conf->quiesce == 0,
3800                                     conf->device_lock, /* nothing */);
3801                 atomic_inc(&conf->active_aligned_reads);
3802                 spin_unlock_irq(&conf->device_lock);
3803
3804                 generic_make_request(align_bi);
3805                 return 1;
3806         } else {
3807                 rcu_read_unlock();
3808                 bio_put(align_bi);
3809                 return 0;
3810         }
3811 }
3812
3813 /* __get_priority_stripe - get the next stripe to process
3814  *
3815  * Full stripe writes are allowed to pass preread active stripes up until
3816  * the bypass_threshold is exceeded.  In general the bypass_count
3817  * increments when the handle_list is handled before the hold_list; however, it
3818  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3819  * stripe with in flight i/o.  The bypass_count will be reset when the
3820  * head of the hold_list has changed, i.e. the head was promoted to the
3821  * handle_list.
3822  */
3823 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3824 {
3825         struct stripe_head *sh;
3826
3827         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3828                   __func__,
3829                   list_empty(&conf->handle_list) ? "empty" : "busy",
3830                   list_empty(&conf->hold_list) ? "empty" : "busy",
3831                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
3832
3833         if (!list_empty(&conf->handle_list)) {
3834                 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3835
3836                 if (list_empty(&conf->hold_list))
3837                         conf->bypass_count = 0;
3838                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3839                         if (conf->hold_list.next == conf->last_hold)
3840                                 conf->bypass_count++;
3841                         else {
3842                                 conf->last_hold = conf->hold_list.next;
3843                                 conf->bypass_count -= conf->bypass_threshold;
3844                                 if (conf->bypass_count < 0)
3845                                         conf->bypass_count = 0;
3846                         }
3847                 }
3848         } else if (!list_empty(&conf->hold_list) &&
3849                    ((conf->bypass_threshold &&
3850                      conf->bypass_count > conf->bypass_threshold) ||
3851                     atomic_read(&conf->pending_full_writes) == 0)) {
3852                 sh = list_entry(conf->hold_list.next,
3853                                 typeof(*sh), lru);
3854                 conf->bypass_count -= conf->bypass_threshold;
3855                 if (conf->bypass_count < 0)
3856                         conf->bypass_count = 0;
3857         } else
3858                 return NULL;
3859
3860         list_del_init(&sh->lru);
3861         atomic_inc(&sh->count);
3862         BUG_ON(atomic_read(&sh->count) != 1);
3863         return sh;
3864 }
3865
3866 static int make_request(mddev_t *mddev, struct bio * bi)
3867 {
3868         raid5_conf_t *conf = mddev->private;
3869         int dd_idx;
3870         sector_t new_sector;
3871         sector_t logical_sector, last_sector;
3872         struct stripe_head *sh;
3873         const int rw = bio_data_dir(bi);
3874         int remaining;
3875
3876         if (unlikely(bio_rw_flagged(bi, BIO_RW_BARRIER))) {
3877                 /* Drain all pending writes.  We only really need
3878                  * to ensure they have been submitted, but this is
3879                  * easier.
3880                  */
3881                 mddev->pers->quiesce(mddev, 1);
3882                 mddev->pers->quiesce(mddev, 0);
3883                 md_barrier_request(mddev, bi);
3884                 return 0;
3885         }
3886
3887         md_write_start(mddev, bi);
3888
3889         if (rw == READ &&
3890              mddev->reshape_position == MaxSector &&
3891              chunk_aligned_read(mddev,bi))
3892                 return 0;
3893
3894         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3895         last_sector = bi->bi_sector + (bi->bi_size>>9);
3896         bi->bi_next = NULL;
3897         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
3898
3899         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3900                 DEFINE_WAIT(w);
3901                 int disks, data_disks;
3902                 int previous;
3903
3904         retry:
3905                 previous = 0;
3906                 disks = conf->raid_disks;
3907                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3908                 if (unlikely(conf->reshape_progress != MaxSector)) {
3909                         /* spinlock is needed as reshape_progress may be
3910                          * 64bit on a 32bit platform, and so it might be
3911                          * possible to see a half-updated value
3912                          * Ofcourse reshape_progress could change after
3913                          * the lock is dropped, so once we get a reference
3914                          * to the stripe that we think it is, we will have
3915                          * to check again.
3916                          */
3917                         spin_lock_irq(&conf->device_lock);
3918                         if (mddev->delta_disks < 0
3919                             ? logical_sector < conf->reshape_progress
3920                             : logical_sector >= conf->reshape_progress) {
3921                                 disks = conf->previous_raid_disks;
3922                                 previous = 1;
3923                         } else {
3924                                 if (mddev->delta_disks < 0
3925                                     ? logical_sector < conf->reshape_safe
3926                                     : logical_sector >= conf->reshape_safe) {
3927                                         spin_unlock_irq(&conf->device_lock);
3928                                         schedule();
3929                                         goto retry;
3930                                 }
3931                         }
3932                         spin_unlock_irq(&conf->device_lock);
3933                 }
3934                 data_disks = disks - conf->max_degraded;
3935
3936                 new_sector = raid5_compute_sector(conf, logical_sector,
3937                                                   previous,
3938                                                   &dd_idx, NULL);
3939                 pr_debug("raid456: make_request, sector %llu logical %llu\n",
3940                         (unsigned long long)new_sector,
3941                         (unsigned long long)logical_sector);
3942
3943                 sh = get_active_stripe(conf, new_sector, previous,
3944                                        (bi->bi_rw&RWA_MASK), 0);
3945                 if (sh) {
3946                         if (unlikely(previous)) {
3947                                 /* expansion might have moved on while waiting for a
3948                                  * stripe, so we must do the range check again.
3949                                  * Expansion could still move past after this
3950                                  * test, but as we are holding a reference to
3951                                  * 'sh', we know that if that happens,
3952                                  *  STRIPE_EXPANDING will get set and the expansion
3953                                  * won't proceed until we finish with the stripe.
3954                                  */
3955                                 int must_retry = 0;
3956                                 spin_lock_irq(&conf->device_lock);
3957                                 if (mddev->delta_disks < 0
3958                                     ? logical_sector >= conf->reshape_progress
3959                                     : logical_sector < conf->reshape_progress)
3960                                         /* mismatch, need to try again */
3961                                         must_retry = 1;
3962                                 spin_unlock_irq(&conf->device_lock);
3963                                 if (must_retry) {
3964                                         release_stripe(sh);
3965                                         schedule();
3966                                         goto retry;
3967                                 }
3968                         }
3969
3970                         if (bio_data_dir(bi) == WRITE &&
3971                             logical_sector >= mddev->suspend_lo &&
3972                             logical_sector < mddev->suspend_hi) {
3973                                 release_stripe(sh);
3974                                 /* As the suspend_* range is controlled by
3975                                  * userspace, we want an interruptible
3976                                  * wait.
3977                                  */
3978                                 flush_signals(current);
3979                                 prepare_to_wait(&conf->wait_for_overlap,
3980                                                 &w, TASK_INTERRUPTIBLE);
3981                                 if (logical_sector >= mddev->suspend_lo &&
3982                                     logical_sector < mddev->suspend_hi)
3983                                         schedule();
3984                                 goto retry;
3985                         }
3986
3987                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3988                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3989                                 /* Stripe is busy expanding or
3990                                  * add failed due to overlap.  Flush everything
3991                                  * and wait a while
3992                                  */
3993                                 raid5_unplug_device(mddev->queue);
3994                                 release_stripe(sh);
3995                                 schedule();
3996                                 goto retry;
3997                         }
3998                         finish_wait(&conf->wait_for_overlap, &w);
3999                         set_bit(STRIPE_HANDLE, &sh->state);
4000                         clear_bit(STRIPE_DELAYED, &sh->state);
4001                         if (mddev->barrier &&
4002                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4003                                 atomic_inc(&conf->preread_active_stripes);
4004                         release_stripe(sh);
4005                 } else {
4006                         /* cannot get stripe for read-ahead, just give-up */
4007                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
4008                         finish_wait(&conf->wait_for_overlap, &w);
4009                         break;
4010                 }
4011
4012         }
4013         spin_lock_irq(&conf->device_lock);
4014         remaining = raid5_dec_bi_phys_segments(bi);
4015         spin_unlock_irq(&conf->device_lock);
4016         if (remaining == 0) {
4017
4018                 if ( rw == WRITE )
4019                         md_write_end(mddev);
4020
4021                 bio_endio(bi, 0);
4022         }
4023
4024         if (mddev->barrier) {
4025                 /* We need to wait for the stripes to all be handled.
4026                  * So: wait for preread_active_stripes to drop to 0.
4027                  */
4028                 wait_event(mddev->thread->wqueue,
4029                            atomic_read(&conf->preread_active_stripes) == 0);
4030         }
4031         return 0;
4032 }
4033
4034 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
4035
4036 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
4037 {
4038         /* reshaping is quite different to recovery/resync so it is
4039          * handled quite separately ... here.
4040          *
4041          * On each call to sync_request, we gather one chunk worth of
4042          * destination stripes and flag them as expanding.
4043          * Then we find all the source stripes and request reads.
4044          * As the reads complete, handle_stripe will copy the data
4045          * into the destination stripe and release that stripe.
4046          */
4047         raid5_conf_t *conf = mddev->private;
4048         struct stripe_head *sh;
4049         sector_t first_sector, last_sector;
4050         int raid_disks = conf->previous_raid_disks;
4051         int data_disks = raid_disks - conf->max_degraded;
4052         int new_data_disks = conf->raid_disks - conf->max_degraded;
4053         int i;
4054         int dd_idx;
4055         sector_t writepos, readpos, safepos;
4056         sector_t stripe_addr;
4057         int reshape_sectors;
4058         struct list_head stripes;
4059
4060         if (sector_nr == 0) {
4061                 /* If restarting in the middle, skip the initial sectors */
4062                 if (mddev->delta_disks < 0 &&
4063                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4064                         sector_nr = raid5_size(mddev, 0, 0)
4065                                 - conf->reshape_progress;
4066                 } else if (mddev->delta_disks >= 0 &&
4067                            conf->reshape_progress > 0)
4068                         sector_nr = conf->reshape_progress;
4069                 sector_div(sector_nr, new_data_disks);
4070                 if (sector_nr) {
4071                         mddev->curr_resync_completed = sector_nr;
4072                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4073                         *skipped = 1;
4074                         return sector_nr;
4075                 }
4076         }
4077
4078         /* We need to process a full chunk at a time.
4079          * If old and new chunk sizes differ, we need to process the
4080          * largest of these
4081          */
4082         if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4083                 reshape_sectors = mddev->new_chunk_sectors;
4084         else
4085                 reshape_sectors = mddev->chunk_sectors;
4086
4087         /* we update the metadata when there is more than 3Meg
4088          * in the block range (that is rather arbitrary, should
4089          * probably be time based) or when the data about to be
4090          * copied would over-write the source of the data at
4091          * the front of the range.
4092          * i.e. one new_stripe along from reshape_progress new_maps
4093          * to after where reshape_safe old_maps to
4094          */
4095         writepos = conf->reshape_progress;
4096         sector_div(writepos, new_data_disks);
4097         readpos = conf->reshape_progress;
4098         sector_div(readpos, data_disks);
4099         safepos = conf->reshape_safe;
4100         sector_div(safepos, data_disks);
4101         if (mddev->delta_disks < 0) {
4102                 writepos -= min_t(sector_t, reshape_sectors, writepos);
4103                 readpos += reshape_sectors;
4104                 safepos += reshape_sectors;
4105         } else {
4106                 writepos += reshape_sectors;
4107                 readpos -= min_t(sector_t, reshape_sectors, readpos);
4108                 safepos -= min_t(sector_t, reshape_sectors, safepos);
4109         }
4110
4111         /* 'writepos' is the most advanced device address we might write.
4112          * 'readpos' is the least advanced device address we might read.
4113          * 'safepos' is the least address recorded in the metadata as having
4114          *     been reshaped.
4115          * If 'readpos' is behind 'writepos', then there is no way that we can
4116          * ensure safety in the face of a crash - that must be done by userspace
4117          * making a backup of the data.  So in that case there is no particular
4118          * rush to update metadata.
4119          * Otherwise if 'safepos' is behind 'writepos', then we really need to
4120          * update the metadata to advance 'safepos' to match 'readpos' so that
4121          * we can be safe in the event of a crash.
4122          * So we insist on updating metadata if safepos is behind writepos and
4123          * readpos is beyond writepos.
4124          * In any case, update the metadata every 10 seconds.
4125          * Maybe that number should be configurable, but I'm not sure it is
4126          * worth it.... maybe it could be a multiple of safemode_delay???
4127          */
4128         if ((mddev->delta_disks < 0
4129              ? (safepos > writepos && readpos < writepos)
4130              : (safepos < writepos && readpos > writepos)) ||
4131             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4132                 /* Cannot proceed until we've updated the superblock... */
4133                 wait_event(conf->wait_for_overlap,
4134                            atomic_read(&conf->reshape_stripes)==0);
4135                 mddev->reshape_position = conf->reshape_progress;
4136                 mddev->curr_resync_completed = mddev->curr_resync;
4137                 conf->reshape_checkpoint = jiffies;
4138                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4139                 md_wakeup_thread(mddev->thread);
4140                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4141                            kthread_should_stop());
4142                 spin_lock_irq(&conf->device_lock);
4143                 conf->reshape_safe = mddev->reshape_position;
4144                 spin_unlock_irq(&conf->device_lock);
4145                 wake_up(&conf->wait_for_overlap);
4146                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4147         }
4148
4149         if (mddev->delta_disks < 0) {
4150                 BUG_ON(conf->reshape_progress == 0);
4151                 stripe_addr = writepos;
4152                 BUG_ON((mddev->dev_sectors &
4153                         ~((sector_t)reshape_sectors - 1))
4154                        - reshape_sectors - stripe_addr
4155                        != sector_nr);
4156         } else {
4157                 BUG_ON(writepos != sector_nr + reshape_sectors);
4158                 stripe_addr = sector_nr;
4159         }
4160         INIT_LIST_HEAD(&stripes);
4161         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4162                 int j;
4163                 int skipped_disk = 0;
4164                 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4165                 set_bit(STRIPE_EXPANDING, &sh->state);
4166                 atomic_inc(&conf->reshape_stripes);
4167                 /* If any of this stripe is beyond the end of the old
4168                  * array, then we need to zero those blocks
4169                  */
4170                 for (j=sh->disks; j--;) {
4171                         sector_t s;
4172                         if (j == sh->pd_idx)
4173                                 continue;
4174                         if (conf->level == 6 &&
4175                             j == sh->qd_idx)
4176                                 continue;
4177                         s = compute_blocknr(sh, j, 0);
4178                         if (s < raid5_size(mddev, 0, 0)) {
4179                                 skipped_disk = 1;
4180                                 continue;
4181                         }
4182                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4183                         set_bit(R5_Expanded, &sh->dev[j].flags);
4184                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
4185                 }
4186                 if (!skipped_disk) {
4187                         set_bit(STRIPE_EXPAND_READY, &sh->state);
4188                         set_bit(STRIPE_HANDLE, &sh->state);
4189                 }
4190                 list_add(&sh->lru, &stripes);
4191         }
4192         spin_lock_irq(&conf->device_lock);
4193         if (mddev->delta_disks < 0)
4194                 conf->reshape_progress -= reshape_sectors * new_data_disks;
4195         else
4196                 conf->reshape_progress += reshape_sectors * new_data_disks;
4197         spin_unlock_irq(&conf->device_lock);
4198         /* Ok, those stripe are ready. We can start scheduling
4199          * reads on the source stripes.
4200          * The source stripes are determined by mapping the first and last
4201          * block on the destination stripes.
4202          */
4203         first_sector =
4204                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4205                                      1, &dd_idx, NULL);
4206         last_sector =
4207                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4208                                             * new_data_disks - 1),
4209                                      1, &dd_idx, NULL);
4210         if (last_sector >= mddev->dev_sectors)
4211                 last_sector = mddev->dev_sectors - 1;
4212         while (first_sector <= last_sector) {
4213                 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4214                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4215                 set_bit(STRIPE_HANDLE, &sh->state);
4216                 release_stripe(sh);
4217                 first_sector += STRIPE_SECTORS;
4218         }
4219         /* Now that the sources are clearly marked, we can release
4220          * the destination stripes
4221          */
4222         while (!list_empty(&stripes)) {
4223                 sh = list_entry(stripes.next, struct stripe_head, lru);
4224                 list_del_init(&sh->lru);
4225                 release_stripe(sh);
4226         }
4227         /* If this takes us to the resync_max point where we have to pause,
4228          * then we need to write out the superblock.
4229          */
4230         sector_nr += reshape_sectors;
4231         if ((sector_nr - mddev->curr_resync_completed) * 2
4232             >= mddev->resync_max - mddev->curr_resync_completed) {
4233                 /* Cannot proceed until we've updated the superblock... */
4234                 wait_event(conf->wait_for_overlap,
4235                            atomic_read(&conf->reshape_stripes) == 0);
4236                 mddev->reshape_position = conf->reshape_progress;
4237                 mddev->curr_resync_completed = mddev->curr_resync + reshape_sectors;
4238                 conf->reshape_checkpoint = jiffies;
4239                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4240                 md_wakeup_thread(mddev->thread);
4241                 wait_event(mddev->sb_wait,
4242                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4243                            || kthread_should_stop());
4244                 spin_lock_irq(&conf->device_lock);
4245                 conf->reshape_safe = mddev->reshape_position;
4246                 spin_unlock_irq(&conf->device_lock);
4247                 wake_up(&conf->wait_for_overlap);
4248                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4249         }
4250         return reshape_sectors;
4251 }
4252
4253 /* FIXME go_faster isn't used */
4254 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
4255 {
4256         raid5_conf_t *conf = mddev->private;
4257         struct stripe_head *sh;
4258         sector_t max_sector = mddev->dev_sectors;
4259         int sync_blocks;
4260         int still_degraded = 0;
4261         int i;
4262
4263         if (sector_nr >= max_sector) {
4264                 /* just being told to finish up .. nothing much to do */
4265                 unplug_slaves(mddev);
4266
4267                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4268                         end_reshape(conf);
4269                         return 0;
4270                 }
4271
4272                 if (mddev->curr_resync < max_sector) /* aborted */
4273                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4274                                         &sync_blocks, 1);
4275                 else /* completed sync */
4276                         conf->fullsync = 0;
4277                 bitmap_close_sync(mddev->bitmap);
4278
4279                 return 0;
4280         }
4281
4282         /* Allow raid5_quiesce to complete */
4283         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4284
4285         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4286                 return reshape_request(mddev, sector_nr, skipped);
4287
4288         /* No need to check resync_max as we never do more than one
4289          * stripe, and as resync_max will always be on a chunk boundary,
4290          * if the check in md_do_sync didn't fire, there is no chance
4291          * of overstepping resync_max here
4292          */
4293
4294         /* if there is too many failed drives and we are trying
4295          * to resync, then assert that we are finished, because there is
4296          * nothing we can do.
4297          */
4298         if (mddev->degraded >= conf->max_degraded &&
4299             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4300                 sector_t rv = mddev->dev_sectors - sector_nr;
4301                 *skipped = 1;
4302                 return rv;
4303         }
4304         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4305             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4306             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4307                 /* we can skip this block, and probably more */
4308                 sync_blocks /= STRIPE_SECTORS;
4309                 *skipped = 1;
4310                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4311         }
4312
4313
4314         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4315
4316         sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4317         if (sh == NULL) {
4318                 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4319                 /* make sure we don't swamp the stripe cache if someone else
4320                  * is trying to get access
4321                  */
4322                 schedule_timeout_uninterruptible(1);
4323         }
4324         /* Need to check if array will still be degraded after recovery/resync
4325          * We don't need to check the 'failed' flag as when that gets set,
4326          * recovery aborts.
4327          */
4328         for (i = 0; i < conf->raid_disks; i++)
4329                 if (conf->disks[i].rdev == NULL)
4330                         still_degraded = 1;
4331
4332         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4333
4334         spin_lock(&sh->lock);
4335         set_bit(STRIPE_SYNCING, &sh->state);
4336         clear_bit(STRIPE_INSYNC, &sh->state);
4337         spin_unlock(&sh->lock);
4338
4339         handle_stripe(sh);
4340         release_stripe(sh);
4341
4342         return STRIPE_SECTORS;
4343 }
4344
4345 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4346 {
4347         /* We may not be able to submit a whole bio at once as there
4348          * may not be enough stripe_heads available.
4349          * We cannot pre-allocate enough stripe_heads as we may need
4350          * more than exist in the cache (if we allow ever large chunks).
4351          * So we do one stripe head at a time and record in
4352          * ->bi_hw_segments how many have been done.
4353          *
4354          * We *know* that this entire raid_bio is in one chunk, so
4355          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4356          */
4357         struct stripe_head *sh;
4358         int dd_idx;
4359         sector_t sector, logical_sector, last_sector;
4360         int scnt = 0;
4361         int remaining;
4362         int handled = 0;
4363
4364         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4365         sector = raid5_compute_sector(conf, logical_sector,
4366                                       0, &dd_idx, NULL);
4367         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4368
4369         for (; logical_sector < last_sector;
4370              logical_sector += STRIPE_SECTORS,
4371                      sector += STRIPE_SECTORS,
4372                      scnt++) {
4373
4374                 if (scnt < raid5_bi_hw_segments(raid_bio))
4375                         /* already done this stripe */
4376                         continue;
4377
4378                 sh = get_active_stripe(conf, sector, 0, 1, 0);
4379
4380                 if (!sh) {
4381                         /* failed to get a stripe - must wait */
4382                         raid5_set_bi_hw_segments(raid_bio, scnt);
4383                         conf->retry_read_aligned = raid_bio;
4384                         return handled;
4385                 }
4386
4387                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4388                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4389                         release_stripe(sh);
4390                         raid5_set_bi_hw_segments(raid_bio, scnt);
4391                         conf->retry_read_aligned = raid_bio;
4392                         return handled;
4393                 }
4394
4395                 handle_stripe(sh);
4396                 release_stripe(sh);
4397                 handled++;
4398         }
4399         spin_lock_irq(&conf->device_lock);
4400         remaining = raid5_dec_bi_phys_segments(raid_bio);
4401         spin_unlock_irq(&conf->device_lock);
4402         if (remaining == 0)
4403                 bio_endio(raid_bio, 0);
4404         if (atomic_dec_and_test(&conf->active_aligned_reads))
4405                 wake_up(&conf->wait_for_stripe);
4406         return handled;
4407 }
4408
4409
4410 /*
4411  * This is our raid5 kernel thread.
4412  *
4413  * We scan the hash table for stripes which can be handled now.
4414  * During the scan, completed stripes are saved for us by the interrupt
4415  * handler, so that they will not have to wait for our next wakeup.
4416  */
4417 static void raid5d(mddev_t *mddev)
4418 {
4419         struct stripe_head *sh;
4420         raid5_conf_t *conf = mddev->private;
4421         int handled;
4422
4423         pr_debug("+++ raid5d active\n");
4424
4425         md_check_recovery(mddev);
4426
4427         handled = 0;
4428         spin_lock_irq(&conf->device_lock);
4429         while (1) {
4430                 struct bio *bio;
4431
4432                 if (conf->seq_flush != conf->seq_write) {
4433                         int seq = conf->seq_flush;
4434                         spin_unlock_irq(&conf->device_lock);
4435                         bitmap_unplug(mddev->bitmap);
4436                         spin_lock_irq(&conf->device_lock);
4437                         conf->seq_write = seq;
4438                         activate_bit_delay(conf);
4439                 }
4440
4441                 while ((bio = remove_bio_from_retry(conf))) {
4442                         int ok;
4443                         spin_unlock_irq(&conf->device_lock);
4444                         ok = retry_aligned_read(conf, bio);
4445                         spin_lock_irq(&conf->device_lock);
4446                         if (!ok)
4447                                 break;
4448                         handled++;
4449                 }
4450
4451                 sh = __get_priority_stripe(conf);
4452
4453                 if (!sh)
4454                         break;
4455                 spin_unlock_irq(&conf->device_lock);
4456
4457                 handled++;
4458                 handle_stripe(sh);
4459                 release_stripe(sh);
4460                 cond_resched();
4461
4462                 spin_lock_irq(&conf->device_lock);
4463         }
4464         pr_debug("%d stripes handled\n", handled);
4465
4466         spin_unlock_irq(&conf->device_lock);
4467
4468         async_tx_issue_pending_all();
4469         unplug_slaves(mddev);
4470
4471         pr_debug("--- raid5d inactive\n");
4472 }
4473
4474 static ssize_t
4475 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4476 {
4477         raid5_conf_t *conf = mddev->private;
4478         if (conf)
4479                 return sprintf(page, "%d\n", conf->max_nr_stripes);
4480         else
4481                 return 0;
4482 }
4483
4484 static ssize_t
4485 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4486 {
4487         raid5_conf_t *conf = mddev->private;
4488         unsigned long new;
4489         int err;
4490
4491         if (len >= PAGE_SIZE)
4492                 return -EINVAL;
4493         if (!conf)
4494                 return -ENODEV;
4495
4496         if (strict_strtoul(page, 10, &new))
4497                 return -EINVAL;
4498         if (new <= 16 || new > 32768)
4499                 return -EINVAL;
4500         while (new < conf->max_nr_stripes) {
4501                 if (drop_one_stripe(conf))
4502                         conf->max_nr_stripes--;
4503                 else
4504                         break;
4505         }
4506         err = md_allow_write(mddev);
4507         if (err)
4508                 return err;
4509         while (new > conf->max_nr_stripes) {
4510                 if (grow_one_stripe(conf))
4511                         conf->max_nr_stripes++;
4512                 else break;
4513         }
4514         return len;
4515 }
4516
4517 static struct md_sysfs_entry
4518 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4519                                 raid5_show_stripe_cache_size,
4520                                 raid5_store_stripe_cache_size);
4521
4522 static ssize_t
4523 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4524 {
4525         raid5_conf_t *conf = mddev->private;
4526         if (conf)
4527                 return sprintf(page, "%d\n", conf->bypass_threshold);
4528         else
4529                 return 0;
4530 }
4531
4532 static ssize_t
4533 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4534 {
4535         raid5_conf_t *conf = mddev->private;
4536         unsigned long new;
4537         if (len >= PAGE_SIZE)
4538                 return -EINVAL;
4539         if (!conf)
4540                 return -ENODEV;
4541
4542         if (strict_strtoul(page, 10, &new))
4543                 return -EINVAL;
4544         if (new > conf->max_nr_stripes)
4545                 return -EINVAL;
4546         conf->bypass_threshold = new;
4547         return len;
4548 }
4549
4550 static struct md_sysfs_entry
4551 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4552                                         S_IRUGO | S_IWUSR,
4553                                         raid5_show_preread_threshold,
4554                                         raid5_store_preread_threshold);
4555
4556 static ssize_t
4557 stripe_cache_active_show(mddev_t *mddev, char *page)
4558 {
4559         raid5_conf_t *conf = mddev->private;
4560         if (conf)
4561                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4562         else
4563                 return 0;
4564 }
4565
4566 static struct md_sysfs_entry
4567 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4568
4569 static struct attribute *raid5_attrs[] =  {
4570         &raid5_stripecache_size.attr,
4571         &raid5_stripecache_active.attr,
4572         &raid5_preread_bypass_threshold.attr,
4573         NULL,
4574 };
4575 static struct attribute_group raid5_attrs_group = {
4576         .name = NULL,
4577         .attrs = raid5_attrs,
4578 };
4579
4580 static sector_t
4581 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4582 {
4583         raid5_conf_t *conf = mddev->private;
4584
4585         if (!sectors)
4586                 sectors = mddev->dev_sectors;
4587         if (!raid_disks)
4588                 /* size is defined by the smallest of previous and new size */
4589                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4590
4591         sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4592         sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4593         return sectors * (raid_disks - conf->max_degraded);
4594 }
4595
4596 static void raid5_free_percpu(raid5_conf_t *conf)
4597 {
4598         struct raid5_percpu *percpu;
4599         unsigned long cpu;
4600
4601         if (!conf->percpu)
4602                 return;
4603
4604         get_online_cpus();
4605         for_each_possible_cpu(cpu) {
4606                 percpu = per_cpu_ptr(conf->percpu, cpu);
4607                 safe_put_page(percpu->spare_page);
4608                 kfree(percpu->scribble);
4609         }
4610 #ifdef CONFIG_HOTPLUG_CPU
4611         unregister_cpu_notifier(&conf->cpu_notify);
4612 #endif
4613         put_online_cpus();
4614
4615         free_percpu(conf->percpu);
4616 }
4617
4618 static void free_conf(raid5_conf_t *conf)
4619 {
4620         shrink_stripes(conf);
4621         raid5_free_percpu(conf);
4622         kfree(conf->disks);
4623         kfree(conf->stripe_hashtbl);
4624         kfree(conf);
4625 }
4626
4627 #ifdef CONFIG_HOTPLUG_CPU
4628 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4629                               void *hcpu)
4630 {
4631         raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
4632         long cpu = (long)hcpu;
4633         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4634
4635         switch (action) {
4636         case CPU_UP_PREPARE:
4637         case CPU_UP_PREPARE_FROZEN:
4638                 if (conf->level == 6 && !percpu->spare_page)
4639                         percpu->spare_page = alloc_page(GFP_KERNEL);
4640                 if (!percpu->scribble)
4641                         percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4642
4643                 if (!percpu->scribble ||
4644                     (conf->level == 6 && !percpu->spare_page)) {
4645                         safe_put_page(percpu->spare_page);
4646                         kfree(percpu->scribble);
4647                         pr_err("%s: failed memory allocation for cpu%ld\n",
4648                                __func__, cpu);
4649                         return notifier_from_errno(-ENOMEM);
4650                 }
4651                 break;
4652         case CPU_DEAD:
4653         case CPU_DEAD_FROZEN:
4654                 safe_put_page(percpu->spare_page);
4655                 kfree(percpu->scribble);
4656                 percpu->spare_page = NULL;
4657                 percpu->scribble = NULL;
4658                 break;
4659         default:
4660                 break;
4661         }
4662         return NOTIFY_OK;
4663 }
4664 #endif
4665
4666 static int raid5_alloc_percpu(raid5_conf_t *conf)
4667 {
4668         unsigned long cpu;
4669         struct page *spare_page;
4670         struct raid5_percpu __percpu *allcpus;
4671         void *scribble;
4672         int err;
4673
4674         allcpus = alloc_percpu(struct raid5_percpu);
4675         if (!allcpus)
4676                 return -ENOMEM;
4677         conf->percpu = allcpus;
4678
4679         get_online_cpus();
4680         err = 0;
4681         for_each_present_cpu(cpu) {
4682                 if (conf->level == 6) {
4683                         spare_page = alloc_page(GFP_KERNEL);
4684                         if (!spare_page) {
4685                                 err = -ENOMEM;
4686                                 break;
4687                         }
4688                         per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4689                 }
4690                 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4691                 if (!scribble) {
4692                         err = -ENOMEM;
4693                         break;
4694                 }
4695                 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4696         }
4697 #ifdef CONFIG_HOTPLUG_CPU
4698         conf->cpu_notify.notifier_call = raid456_cpu_notify;
4699         conf->cpu_notify.priority = 0;
4700         if (err == 0)
4701                 err = register_cpu_notifier(&conf->cpu_notify);
4702 #endif
4703         put_online_cpus();
4704
4705         return err;
4706 }
4707
4708 static raid5_conf_t *setup_conf(mddev_t *mddev)
4709 {
4710         raid5_conf_t *conf;
4711         int raid_disk, memory, max_disks;
4712         mdk_rdev_t *rdev;
4713         struct disk_info *disk;
4714
4715         if (mddev->new_level != 5
4716             && mddev->new_level != 4
4717             && mddev->new_level != 6) {
4718                 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4719                        mdname(mddev), mddev->new_level);
4720                 return ERR_PTR(-EIO);
4721         }
4722         if ((mddev->new_level == 5
4723              && !algorithm_valid_raid5(mddev->new_layout)) ||
4724             (mddev->new_level == 6
4725              && !algorithm_valid_raid6(mddev->new_layout))) {
4726                 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4727                        mdname(mddev), mddev->new_layout);
4728                 return ERR_PTR(-EIO);
4729         }
4730         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4731                 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4732                        mdname(mddev), mddev->raid_disks);
4733                 return ERR_PTR(-EINVAL);
4734         }
4735
4736         if (!mddev->new_chunk_sectors ||
4737             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4738             !is_power_of_2(mddev->new_chunk_sectors)) {
4739                 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4740                        mdname(mddev), mddev->new_chunk_sectors << 9);
4741                 return ERR_PTR(-EINVAL);
4742         }
4743
4744         conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4745         if (conf == NULL)
4746                 goto abort;
4747         spin_lock_init(&conf->device_lock);
4748         init_waitqueue_head(&conf->wait_for_stripe);
4749         init_waitqueue_head(&conf->wait_for_overlap);
4750         INIT_LIST_HEAD(&conf->handle_list);
4751         INIT_LIST_HEAD(&conf->hold_list);
4752         INIT_LIST_HEAD(&conf->delayed_list);
4753         INIT_LIST_HEAD(&conf->bitmap_list);
4754         INIT_LIST_HEAD(&conf->inactive_list);
4755         atomic_set(&conf->active_stripes, 0);
4756         atomic_set(&conf->preread_active_stripes, 0);
4757         atomic_set(&conf->active_aligned_reads, 0);
4758         conf->bypass_threshold = BYPASS_THRESHOLD;
4759
4760         conf->raid_disks = mddev->raid_disks;
4761         if (mddev->reshape_position == MaxSector)
4762                 conf->previous_raid_disks = mddev->raid_disks;
4763         else
4764                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4765         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4766         conf->scribble_len = scribble_len(max_disks);
4767
4768         conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4769                               GFP_KERNEL);
4770         if (!conf->disks)
4771                 goto abort;
4772
4773         conf->mddev = mddev;
4774
4775         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4776                 goto abort;
4777
4778         conf->level = mddev->new_level;
4779         if (raid5_alloc_percpu(conf) != 0)
4780                 goto abort;
4781
4782         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4783
4784         list_for_each_entry(rdev, &mddev->disks, same_set) {
4785                 raid_disk = rdev->raid_disk;
4786                 if (raid_disk >= max_disks
4787                     || raid_disk < 0)
4788                         continue;
4789                 disk = conf->disks + raid_disk;
4790
4791                 disk->rdev = rdev;
4792
4793                 if (test_bit(In_sync, &rdev->flags)) {
4794                         char b[BDEVNAME_SIZE];
4795                         printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4796                                " disk %d\n",
4797                                mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4798                 } else
4799                         /* Cannot rely on bitmap to complete recovery */
4800                         conf->fullsync = 1;
4801         }
4802
4803         conf->chunk_sectors = mddev->new_chunk_sectors;
4804         conf->level = mddev->new_level;
4805         if (conf->level == 6)
4806                 conf->max_degraded = 2;
4807         else
4808                 conf->max_degraded = 1;
4809         conf->algorithm = mddev->new_layout;
4810         conf->max_nr_stripes = NR_STRIPES;
4811         conf->reshape_progress = mddev->reshape_position;
4812         if (conf->reshape_progress != MaxSector) {
4813                 conf->prev_chunk_sectors = mddev->chunk_sectors;
4814                 conf->prev_algo = mddev->layout;
4815         }
4816
4817         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4818                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4819         if (grow_stripes(conf, conf->max_nr_stripes)) {
4820                 printk(KERN_ERR
4821                        "md/raid:%s: couldn't allocate %dkB for buffers\n",
4822                        mdname(mddev), memory);
4823                 goto abort;
4824         } else
4825                 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4826                        mdname(mddev), memory);
4827
4828         conf->thread = md_register_thread(raid5d, mddev, NULL);
4829         if (!conf->thread) {
4830                 printk(KERN_ERR
4831                        "md/raid:%s: couldn't allocate thread.\n",
4832                        mdname(mddev));
4833                 goto abort;
4834         }
4835
4836         return conf;
4837
4838  abort:
4839         if (conf) {
4840                 free_conf(conf);
4841                 return ERR_PTR(-EIO);
4842         } else
4843                 return ERR_PTR(-ENOMEM);
4844 }
4845
4846
4847 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4848 {
4849         switch (algo) {
4850         case ALGORITHM_PARITY_0:
4851                 if (raid_disk < max_degraded)
4852                         return 1;
4853                 break;
4854         case ALGORITHM_PARITY_N:
4855                 if (raid_disk >= raid_disks - max_degraded)
4856                         return 1;
4857                 break;
4858         case ALGORITHM_PARITY_0_6:
4859                 if (raid_disk == 0 ||
4860                     raid_disk == raid_disks - 1)
4861                         return 1;
4862                 break;
4863         case ALGORITHM_LEFT_ASYMMETRIC_6:
4864         case ALGORITHM_RIGHT_ASYMMETRIC_6:
4865         case ALGORITHM_LEFT_SYMMETRIC_6:
4866         case ALGORITHM_RIGHT_SYMMETRIC_6:
4867                 if (raid_disk == raid_disks - 1)
4868                         return 1;
4869         }
4870         return 0;
4871 }
4872
4873 static int run(mddev_t *mddev)
4874 {
4875         raid5_conf_t *conf;
4876         int working_disks = 0, chunk_size;
4877         int dirty_parity_disks = 0;
4878         mdk_rdev_t *rdev;
4879         sector_t reshape_offset = 0;
4880
4881         if (mddev->recovery_cp != MaxSector)
4882                 printk(KERN_NOTICE "md/raid:%s: not clean"
4883                        " -- starting background reconstruction\n",
4884                        mdname(mddev));
4885         if (mddev->reshape_position != MaxSector) {
4886                 /* Check that we can continue the reshape.
4887                  * Currently only disks can change, it must
4888                  * increase, and we must be past the point where
4889                  * a stripe over-writes itself
4890                  */
4891                 sector_t here_new, here_old;
4892                 int old_disks;
4893                 int max_degraded = (mddev->level == 6 ? 2 : 1);
4894
4895                 if (mddev->new_level != mddev->level) {
4896                         printk(KERN_ERR "md/raid:%s: unsupported reshape "
4897                                "required - aborting.\n",
4898                                mdname(mddev));
4899                         return -EINVAL;
4900                 }
4901                 old_disks = mddev->raid_disks - mddev->delta_disks;
4902                 /* reshape_position must be on a new-stripe boundary, and one
4903                  * further up in new geometry must map after here in old
4904                  * geometry.
4905                  */
4906                 here_new = mddev->reshape_position;
4907                 if (sector_div(here_new, mddev->new_chunk_sectors *
4908                                (mddev->raid_disks - max_degraded))) {
4909                         printk(KERN_ERR "md/raid:%s: reshape_position not "
4910                                "on a stripe boundary\n", mdname(mddev));
4911                         return -EINVAL;
4912                 }
4913                 reshape_offset = here_new * mddev->new_chunk_sectors;
4914                 /* here_new is the stripe we will write to */
4915                 here_old = mddev->reshape_position;
4916                 sector_div(here_old, mddev->chunk_sectors *
4917                            (old_disks-max_degraded));
4918                 /* here_old is the first stripe that we might need to read
4919                  * from */
4920                 if (mddev->delta_disks == 0) {
4921                         /* We cannot be sure it is safe to start an in-place
4922                          * reshape.  It is only safe if user-space if monitoring
4923                          * and taking constant backups.
4924                          * mdadm always starts a situation like this in
4925                          * readonly mode so it can take control before
4926                          * allowing any writes.  So just check for that.
4927                          */
4928                         if ((here_new * mddev->new_chunk_sectors !=
4929                              here_old * mddev->chunk_sectors) ||
4930                             mddev->ro == 0) {
4931                                 printk(KERN_ERR "md/raid:%s: in-place reshape must be started"
4932                                        " in read-only mode - aborting\n",
4933                                        mdname(mddev));
4934                                 return -EINVAL;
4935                         }
4936                 } else if (mddev->delta_disks < 0
4937                     ? (here_new * mddev->new_chunk_sectors <=
4938                        here_old * mddev->chunk_sectors)
4939                     : (here_new * mddev->new_chunk_sectors >=
4940                        here_old * mddev->chunk_sectors)) {
4941                         /* Reading from the same stripe as writing to - bad */
4942                         printk(KERN_ERR "md/raid:%s: reshape_position too early for "
4943                                "auto-recovery - aborting.\n",
4944                                mdname(mddev));
4945                         return -EINVAL;
4946                 }
4947                 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
4948                        mdname(mddev));
4949                 /* OK, we should be able to continue; */
4950         } else {
4951                 BUG_ON(mddev->level != mddev->new_level);
4952                 BUG_ON(mddev->layout != mddev->new_layout);
4953                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4954                 BUG_ON(mddev->delta_disks != 0);
4955         }
4956
4957         if (mddev->private == NULL)
4958                 conf = setup_conf(mddev);
4959         else
4960                 conf = mddev->private;
4961
4962         if (IS_ERR(conf))
4963                 return PTR_ERR(conf);
4964
4965         mddev->thread = conf->thread;
4966         conf->thread = NULL;
4967         mddev->private = conf;
4968
4969         /*
4970          * 0 for a fully functional array, 1 or 2 for a degraded array.
4971          */
4972         list_for_each_entry(rdev, &mddev->disks, same_set) {
4973                 if (rdev->raid_disk < 0)
4974                         continue;
4975                 if (test_bit(In_sync, &rdev->flags))
4976                         working_disks++;
4977                 /* This disc is not fully in-sync.  However if it
4978                  * just stored parity (beyond the recovery_offset),
4979                  * when we don't need to be concerned about the
4980                  * array being dirty.
4981                  * When reshape goes 'backwards', we never have
4982                  * partially completed devices, so we only need
4983                  * to worry about reshape going forwards.
4984                  */
4985                 /* Hack because v0.91 doesn't store recovery_offset properly. */
4986                 if (mddev->major_version == 0 &&
4987                     mddev->minor_version > 90)
4988                         rdev->recovery_offset = reshape_offset;
4989
4990                 if (rdev->recovery_offset < reshape_offset) {
4991                         /* We need to check old and new layout */
4992                         if (!only_parity(rdev->raid_disk,
4993                                          conf->algorithm,
4994                                          conf->raid_disks,
4995                                          conf->max_degraded))
4996                                 continue;
4997                 }
4998                 if (!only_parity(rdev->raid_disk,
4999                                  conf->prev_algo,
5000                                  conf->previous_raid_disks,
5001                                  conf->max_degraded))
5002                         continue;
5003                 dirty_parity_disks++;
5004         }
5005
5006         mddev->degraded = (max(conf->raid_disks, conf->previous_raid_disks)
5007                            - working_disks);
5008
5009         if (mddev->degraded > conf->max_degraded) {
5010                 printk(KERN_ERR "md/raid:%s: not enough operational devices"
5011                         " (%d/%d failed)\n",
5012                         mdname(mddev), mddev->degraded, conf->raid_disks);
5013                 goto abort;
5014         }
5015
5016         /* device size must be a multiple of chunk size */
5017         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5018         mddev->resync_max_sectors = mddev->dev_sectors;
5019
5020         if (mddev->degraded > dirty_parity_disks &&
5021             mddev->recovery_cp != MaxSector) {
5022                 if (mddev->ok_start_degraded)
5023                         printk(KERN_WARNING
5024                                "md/raid:%s: starting dirty degraded array"
5025                                " - data corruption possible.\n",
5026                                mdname(mddev));
5027                 else {
5028                         printk(KERN_ERR
5029                                "md/raid:%s: cannot start dirty degraded array.\n",
5030                                mdname(mddev));
5031                         goto abort;
5032                 }
5033         }
5034
5035         if (mddev->degraded == 0)
5036                 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5037                        " devices, algorithm %d\n", mdname(mddev), conf->level,
5038                        mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5039                        mddev->new_layout);
5040         else
5041                 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5042                        " out of %d devices, algorithm %d\n",
5043                        mdname(mddev), conf->level,
5044                        mddev->raid_disks - mddev->degraded,
5045                        mddev->raid_disks, mddev->new_layout);
5046
5047         print_raid5_conf(conf);
5048
5049         if (conf->reshape_progress != MaxSector) {
5050                 conf->reshape_safe = conf->reshape_progress;
5051                 atomic_set(&conf->reshape_stripes, 0);
5052                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5053                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5054                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5055                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5056                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5057                                                         "reshape");
5058         }
5059
5060         /* read-ahead size must cover two whole stripes, which is
5061          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5062          */
5063         {
5064                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5065                 int stripe = data_disks *
5066                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5067                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5068                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5069         }
5070
5071         /* Ok, everything is just fine now */
5072         if (mddev->to_remove == &raid5_attrs_group)
5073                 mddev->to_remove = NULL;
5074         else if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5075                 printk(KERN_WARNING
5076                        "md/raid:%s: failed to create sysfs attributes.\n",
5077                        mdname(mddev));
5078
5079         mddev->queue->queue_lock = &conf->device_lock;
5080
5081         mddev->queue->unplug_fn = raid5_unplug_device;
5082         mddev->queue->backing_dev_info.congested_data = mddev;
5083         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5084
5085         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5086
5087         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5088         chunk_size = mddev->chunk_sectors << 9;
5089         blk_queue_io_min(mddev->queue, chunk_size);
5090         blk_queue_io_opt(mddev->queue, chunk_size *
5091                          (conf->raid_disks - conf->max_degraded));
5092
5093         list_for_each_entry(rdev, &mddev->disks, same_set)
5094                 disk_stack_limits(mddev->gendisk, rdev->bdev,
5095                                   rdev->data_offset << 9);
5096
5097         return 0;
5098 abort:
5099         md_unregister_thread(mddev->thread);
5100         mddev->thread = NULL;
5101         if (conf) {
5102                 print_raid5_conf(conf);
5103                 free_conf(conf);
5104         }
5105         mddev->private = NULL;
5106         printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5107         return -EIO;
5108 }
5109
5110 static int stop(mddev_t *mddev)
5111 {
5112         raid5_conf_t *conf = mddev->private;
5113
5114         md_unregister_thread(mddev->thread);
5115         mddev->thread = NULL;
5116         mddev->queue->backing_dev_info.congested_fn = NULL;
5117         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
5118         free_conf(conf);
5119         mddev->private = NULL;
5120         mddev->to_remove = &raid5_attrs_group;
5121         return 0;
5122 }
5123
5124 #ifdef DEBUG
5125 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
5126 {
5127         int i;
5128
5129         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
5130                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
5131         seq_printf(seq, "sh %llu,  count %d.\n",
5132                    (unsigned long long)sh->sector, atomic_read(&sh->count));
5133         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
5134         for (i = 0; i < sh->disks; i++) {
5135                 seq_printf(seq, "(cache%d: %p %ld) ",
5136                            i, sh->dev[i].page, sh->dev[i].flags);
5137         }
5138         seq_printf(seq, "\n");
5139 }
5140
5141 static void printall(struct seq_file *seq, raid5_conf_t *conf)
5142 {
5143         struct stripe_head *sh;
5144         struct hlist_node *hn;
5145         int i;
5146
5147         spin_lock_irq(&conf->device_lock);
5148         for (i = 0; i < NR_HASH; i++) {
5149                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
5150                         if (sh->raid_conf != conf)
5151                                 continue;
5152                         print_sh(seq, sh);
5153                 }
5154         }
5155         spin_unlock_irq(&conf->device_lock);
5156 }
5157 #endif
5158
5159 static void status(struct seq_file *seq, mddev_t *mddev)
5160 {
5161         raid5_conf_t *conf = mddev->private;
5162         int i;
5163
5164         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5165                 mddev->chunk_sectors / 2, mddev->layout);
5166         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5167         for (i = 0; i < conf->raid_disks; i++)
5168                 seq_printf (seq, "%s",
5169                                conf->disks[i].rdev &&
5170                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5171         seq_printf (seq, "]");
5172 #ifdef DEBUG
5173         seq_printf (seq, "\n");
5174         printall(seq, conf);
5175 #endif
5176 }
5177
5178 static void print_raid5_conf (raid5_conf_t *conf)
5179 {
5180         int i;
5181         struct disk_info *tmp;
5182
5183         printk(KERN_DEBUG "RAID conf printout:\n");
5184         if (!conf) {
5185                 printk("(conf==NULL)\n");
5186                 return;
5187         }
5188         printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5189                conf->raid_disks,
5190                conf->raid_disks - conf->mddev->degraded);
5191
5192         for (i = 0; i < conf->raid_disks; i++) {
5193                 char b[BDEVNAME_SIZE];
5194                 tmp = conf->disks + i;
5195                 if (tmp->rdev)
5196                         printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5197                                i, !test_bit(Faulty, &tmp->rdev->flags),
5198                                bdevname(tmp->rdev->bdev, b));
5199         }
5200 }
5201
5202 static int raid5_spare_active(mddev_t *mddev)
5203 {
5204         int i;
5205         raid5_conf_t *conf = mddev->private;
5206         struct disk_info *tmp;
5207
5208         for (i = 0; i < conf->raid_disks; i++) {
5209                 tmp = conf->disks + i;
5210                 if (tmp->rdev
5211                     && tmp->rdev->recovery_offset == MaxSector
5212                     && !test_bit(Faulty, &tmp->rdev->flags)
5213                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5214                         unsigned long flags;
5215                         spin_lock_irqsave(&conf->device_lock, flags);
5216                         mddev->degraded--;
5217                         spin_unlock_irqrestore(&conf->device_lock, flags);
5218                 }
5219         }
5220         print_raid5_conf(conf);
5221         return 0;
5222 }
5223
5224 static int raid5_remove_disk(mddev_t *mddev, int number)
5225 {
5226         raid5_conf_t *conf = mddev->private;
5227         int err = 0;
5228         mdk_rdev_t *rdev;
5229         struct disk_info *p = conf->disks + number;
5230
5231         print_raid5_conf(conf);
5232         rdev = p->rdev;
5233         if (rdev) {
5234                 if (number >= conf->raid_disks &&
5235                     conf->reshape_progress == MaxSector)
5236                         clear_bit(In_sync, &rdev->flags);
5237
5238                 if (test_bit(In_sync, &rdev->flags) ||
5239                     atomic_read(&rdev->nr_pending)) {
5240                         err = -EBUSY;
5241                         goto abort;
5242                 }
5243                 /* Only remove non-faulty devices if recovery
5244                  * isn't possible.
5245                  */
5246                 if (!test_bit(Faulty, &rdev->flags) &&
5247                     mddev->degraded <= conf->max_degraded &&
5248                     number < conf->raid_disks) {
5249                         err = -EBUSY;
5250                         goto abort;
5251                 }
5252                 p->rdev = NULL;
5253                 synchronize_rcu();
5254                 if (atomic_read(&rdev->nr_pending)) {
5255                         /* lost the race, try later */
5256                         err = -EBUSY;
5257                         p->rdev = rdev;
5258                 }
5259         }
5260 abort:
5261
5262         print_raid5_conf(conf);
5263         return err;
5264 }
5265
5266 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
5267 {
5268         raid5_conf_t *conf = mddev->private;
5269         int err = -EEXIST;
5270         int disk;
5271         struct disk_info *p;
5272         int first = 0;
5273         int last = conf->raid_disks - 1;
5274
5275         if (mddev->degraded > conf->max_degraded)
5276                 /* no point adding a device */
5277                 return -EINVAL;
5278
5279         if (rdev->raid_disk >= 0)
5280                 first = last = rdev->raid_disk;
5281
5282         /*
5283          * find the disk ... but prefer rdev->saved_raid_disk
5284          * if possible.
5285          */
5286         if (rdev->saved_raid_disk >= 0 &&
5287             rdev->saved_raid_disk >= first &&
5288             conf->disks[rdev->saved_raid_disk].rdev == NULL)
5289                 disk = rdev->saved_raid_disk;
5290         else
5291                 disk = first;
5292         for ( ; disk <= last ; disk++)
5293                 if ((p=conf->disks + disk)->rdev == NULL) {
5294                         clear_bit(In_sync, &rdev->flags);
5295                         rdev->raid_disk = disk;
5296                         err = 0;
5297                         if (rdev->saved_raid_disk != disk)
5298                                 conf->fullsync = 1;
5299                         rcu_assign_pointer(p->rdev, rdev);
5300                         break;
5301                 }
5302         print_raid5_conf(conf);
5303         return err;
5304 }
5305
5306 static int raid5_resize(mddev_t *mddev, sector_t sectors)
5307 {
5308         /* no resync is happening, and there is enough space
5309          * on all devices, so we can resize.
5310          * We need to make sure resync covers any new space.
5311          * If the array is shrinking we should possibly wait until
5312          * any io in the removed space completes, but it hardly seems
5313          * worth it.
5314          */
5315         sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5316         md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5317                                                mddev->raid_disks));
5318         if (mddev->array_sectors >
5319             raid5_size(mddev, sectors, mddev->raid_disks))
5320                 return -EINVAL;
5321         set_capacity(mddev->gendisk, mddev->array_sectors);
5322         revalidate_disk(mddev->gendisk);
5323         if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) {
5324                 mddev->recovery_cp = mddev->dev_sectors;
5325                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5326         }
5327         mddev->dev_sectors = sectors;
5328         mddev->resync_max_sectors = sectors;
5329         return 0;
5330 }
5331
5332 static int check_stripe_cache(mddev_t *mddev)
5333 {
5334         /* Can only proceed if there are plenty of stripe_heads.
5335          * We need a minimum of one full stripe,, and for sensible progress
5336          * it is best to have about 4 times that.
5337          * If we require 4 times, then the default 256 4K stripe_heads will
5338          * allow for chunk sizes up to 256K, which is probably OK.
5339          * If the chunk size is greater, user-space should request more
5340          * stripe_heads first.
5341          */
5342         raid5_conf_t *conf = mddev->private;
5343         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5344             > conf->max_nr_stripes ||
5345             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5346             > conf->max_nr_stripes) {
5347                 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
5348                        mdname(mddev),
5349                        ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5350                         / STRIPE_SIZE)*4);
5351                 return 0;
5352         }
5353         return 1;
5354 }
5355
5356 static int check_reshape(mddev_t *mddev)
5357 {
5358         raid5_conf_t *conf = mddev->private;
5359
5360         if (mddev->delta_disks == 0 &&
5361             mddev->new_layout == mddev->layout &&
5362             mddev->new_chunk_sectors == mddev->chunk_sectors)
5363                 return 0; /* nothing to do */
5364         if (mddev->bitmap)
5365                 /* Cannot grow a bitmap yet */
5366                 return -EBUSY;
5367         if (mddev->degraded > conf->max_degraded)
5368                 return -EINVAL;
5369         if (mddev->delta_disks < 0) {
5370                 /* We might be able to shrink, but the devices must
5371                  * be made bigger first.
5372                  * For raid6, 4 is the minimum size.
5373                  * Otherwise 2 is the minimum
5374                  */
5375                 int min = 2;
5376                 if (mddev->level == 6)
5377                         min = 4;
5378                 if (mddev->raid_disks + mddev->delta_disks < min)
5379                         return -EINVAL;
5380         }
5381
5382         if (!check_stripe_cache(mddev))
5383                 return -ENOSPC;
5384
5385         return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5386 }
5387
5388 static int raid5_start_reshape(mddev_t *mddev)
5389 {
5390         raid5_conf_t *conf = mddev->private;
5391         mdk_rdev_t *rdev;
5392         int spares = 0;
5393         int added_devices = 0;
5394         unsigned long flags;
5395
5396         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5397                 return -EBUSY;
5398
5399         if (!check_stripe_cache(mddev))
5400                 return -ENOSPC;
5401
5402         list_for_each_entry(rdev, &mddev->disks, same_set)
5403                 if (rdev->raid_disk < 0 &&
5404                     !test_bit(Faulty, &rdev->flags))
5405                         spares++;
5406
5407         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5408                 /* Not enough devices even to make a degraded array
5409                  * of that size
5410                  */
5411                 return -EINVAL;
5412
5413         /* Refuse to reduce size of the array.  Any reductions in
5414          * array size must be through explicit setting of array_size
5415          * attribute.
5416          */
5417         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5418             < mddev->array_sectors) {
5419                 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5420                        "before number of disks\n", mdname(mddev));
5421                 return -EINVAL;
5422         }
5423
5424         atomic_set(&conf->reshape_stripes, 0);
5425         spin_lock_irq(&conf->device_lock);
5426         conf->previous_raid_disks = conf->raid_disks;
5427         conf->raid_disks += mddev->delta_disks;
5428         conf->prev_chunk_sectors = conf->chunk_sectors;
5429         conf->chunk_sectors = mddev->new_chunk_sectors;
5430         conf->prev_algo = conf->algorithm;
5431         conf->algorithm = mddev->new_layout;
5432         if (mddev->delta_disks < 0)
5433                 conf->reshape_progress = raid5_size(mddev, 0, 0);
5434         else
5435                 conf->reshape_progress = 0;
5436         conf->reshape_safe = conf->reshape_progress;
5437         conf->generation++;
5438         spin_unlock_irq(&conf->device_lock);
5439
5440         /* Add some new drives, as many as will fit.
5441          * We know there are enough to make the newly sized array work.
5442          */
5443         list_for_each_entry(rdev, &mddev->disks, same_set)
5444                 if (rdev->raid_disk < 0 &&
5445                     !test_bit(Faulty, &rdev->flags)) {
5446                         if (raid5_add_disk(mddev, rdev) == 0) {
5447                                 char nm[20];
5448                                 if (rdev->raid_disk >= conf->previous_raid_disks) {
5449                                         set_bit(In_sync, &rdev->flags);
5450                                         added_devices++;
5451                                 } else
5452                                         rdev->recovery_offset = 0;
5453                                 sprintf(nm, "rd%d", rdev->raid_disk);
5454                                 if (sysfs_create_link(&mddev->kobj,
5455                                                       &rdev->kobj, nm))
5456                                         printk(KERN_WARNING
5457                                                "md/raid:%s: failed to create "
5458                                                " link %s\n",
5459                                                mdname(mddev), nm);
5460                         } else
5461                                 break;
5462                 }
5463
5464         /* When a reshape changes the number of devices, ->degraded
5465          * is measured against the large of the pre and post number of
5466          * devices.*/
5467         if (mddev->delta_disks > 0) {
5468                 spin_lock_irqsave(&conf->device_lock, flags);
5469                 mddev->degraded += (conf->raid_disks - conf->previous_raid_disks)
5470                         - added_devices;
5471                 spin_unlock_irqrestore(&conf->device_lock, flags);
5472         }
5473         mddev->raid_disks = conf->raid_disks;
5474         mddev->reshape_position = conf->reshape_progress;
5475         set_bit(MD_CHANGE_DEVS, &mddev->flags);
5476
5477         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5478         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5479         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5480         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5481         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5482                                                 "reshape");
5483         if (!mddev->sync_thread) {
5484                 mddev->recovery = 0;
5485                 spin_lock_irq(&conf->device_lock);
5486                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5487                 conf->reshape_progress = MaxSector;
5488                 spin_unlock_irq(&conf->device_lock);
5489                 return -EAGAIN;
5490         }
5491         conf->reshape_checkpoint = jiffies;
5492         md_wakeup_thread(mddev->sync_thread);
5493         md_new_event(mddev);
5494         return 0;
5495 }
5496
5497 /* This is called from the reshape thread and should make any
5498  * changes needed in 'conf'
5499  */
5500 static void end_reshape(raid5_conf_t *conf)
5501 {
5502
5503         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5504
5505                 spin_lock_irq(&conf->device_lock);
5506                 conf->previous_raid_disks = conf->raid_disks;
5507                 conf->reshape_progress = MaxSector;
5508                 spin_unlock_irq(&conf->device_lock);
5509                 wake_up(&conf->wait_for_overlap);
5510
5511                 /* read-ahead size must cover two whole stripes, which is
5512                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5513                  */
5514                 {
5515                         int data_disks = conf->raid_disks - conf->max_degraded;
5516                         int stripe = data_disks * ((conf->chunk_sectors << 9)
5517                                                    / PAGE_SIZE);
5518                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5519                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5520                 }
5521         }
5522 }
5523
5524 /* This is called from the raid5d thread with mddev_lock held.
5525  * It makes config changes to the device.
5526  */
5527 static void raid5_finish_reshape(mddev_t *mddev)
5528 {
5529         raid5_conf_t *conf = mddev->private;
5530
5531         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5532
5533                 if (mddev->delta_disks > 0) {
5534                         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5535                         set_capacity(mddev->gendisk, mddev->array_sectors);
5536                         revalidate_disk(mddev->gendisk);
5537                 } else {
5538                         int d;
5539                         mddev->degraded = conf->raid_disks;
5540                         for (d = 0; d < conf->raid_disks ; d++)
5541                                 if (conf->disks[d].rdev &&
5542                                     test_bit(In_sync,
5543                                              &conf->disks[d].rdev->flags))
5544                                         mddev->degraded--;
5545                         for (d = conf->raid_disks ;
5546                              d < conf->raid_disks - mddev->delta_disks;
5547                              d++) {
5548                                 mdk_rdev_t *rdev = conf->disks[d].rdev;
5549                                 if (rdev && raid5_remove_disk(mddev, d) == 0) {
5550                                         char nm[20];
5551                                         sprintf(nm, "rd%d", rdev->raid_disk);
5552                                         sysfs_remove_link(&mddev->kobj, nm);
5553                                         rdev->raid_disk = -1;
5554                                 }
5555                         }
5556                 }
5557                 mddev->layout = conf->algorithm;
5558                 mddev->chunk_sectors = conf->chunk_sectors;
5559                 mddev->reshape_position = MaxSector;
5560                 mddev->delta_disks = 0;
5561         }
5562 }
5563
5564 static void raid5_quiesce(mddev_t *mddev, int state)
5565 {
5566         raid5_conf_t *conf = mddev->private;
5567
5568         switch(state) {
5569         case 2: /* resume for a suspend */
5570                 wake_up(&conf->wait_for_overlap);
5571                 break;
5572
5573         case 1: /* stop all writes */
5574                 spin_lock_irq(&conf->device_lock);
5575                 /* '2' tells resync/reshape to pause so that all
5576                  * active stripes can drain
5577                  */
5578                 conf->quiesce = 2;
5579                 wait_event_lock_irq(conf->wait_for_stripe,
5580                                     atomic_read(&conf->active_stripes) == 0 &&
5581                                     atomic_read(&conf->active_aligned_reads) == 0,
5582                                     conf->device_lock, /* nothing */);
5583                 conf->quiesce = 1;
5584                 spin_unlock_irq(&conf->device_lock);
5585                 /* allow reshape to continue */
5586                 wake_up(&conf->wait_for_overlap);
5587                 break;
5588
5589         case 0: /* re-enable writes */
5590                 spin_lock_irq(&conf->device_lock);
5591                 conf->quiesce = 0;
5592                 wake_up(&conf->wait_for_stripe);
5593                 wake_up(&conf->wait_for_overlap);
5594                 spin_unlock_irq(&conf->device_lock);
5595                 break;
5596         }
5597 }
5598
5599
5600 static void *raid45_takeover_raid0(mddev_t *mddev, int level)
5601 {
5602         struct raid0_private_data *raid0_priv = mddev->private;
5603
5604         /* for raid0 takeover only one zone is supported */
5605         if (raid0_priv->nr_strip_zones > 1) {
5606                 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5607                        mdname(mddev));
5608                 return ERR_PTR(-EINVAL);
5609         }
5610
5611         mddev->new_level = level;
5612         mddev->new_layout = ALGORITHM_PARITY_N;
5613         mddev->new_chunk_sectors = mddev->chunk_sectors;
5614         mddev->raid_disks += 1;
5615         mddev->delta_disks = 1;
5616         /* make sure it will be not marked as dirty */
5617         mddev->recovery_cp = MaxSector;
5618
5619         return setup_conf(mddev);
5620 }
5621
5622
5623 static void *raid5_takeover_raid1(mddev_t *mddev)
5624 {
5625         int chunksect;
5626
5627         if (mddev->raid_disks != 2 ||
5628             mddev->degraded > 1)
5629                 return ERR_PTR(-EINVAL);
5630
5631         /* Should check if there are write-behind devices? */
5632
5633         chunksect = 64*2; /* 64K by default */
5634
5635         /* The array must be an exact multiple of chunksize */
5636         while (chunksect && (mddev->array_sectors & (chunksect-1)))
5637                 chunksect >>= 1;
5638
5639         if ((chunksect<<9) < STRIPE_SIZE)
5640                 /* array size does not allow a suitable chunk size */
5641                 return ERR_PTR(-EINVAL);
5642
5643         mddev->new_level = 5;
5644         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5645         mddev->new_chunk_sectors = chunksect;
5646
5647         return setup_conf(mddev);
5648 }
5649
5650 static void *raid5_takeover_raid6(mddev_t *mddev)
5651 {
5652         int new_layout;
5653
5654         switch (mddev->layout) {
5655         case ALGORITHM_LEFT_ASYMMETRIC_6:
5656                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5657                 break;
5658         case ALGORITHM_RIGHT_ASYMMETRIC_6:
5659                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5660                 break;
5661         case ALGORITHM_LEFT_SYMMETRIC_6:
5662                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5663                 break;
5664         case ALGORITHM_RIGHT_SYMMETRIC_6:
5665                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5666                 break;
5667         case ALGORITHM_PARITY_0_6:
5668                 new_layout = ALGORITHM_PARITY_0;
5669                 break;
5670         case ALGORITHM_PARITY_N:
5671                 new_layout = ALGORITHM_PARITY_N;
5672                 break;
5673         default:
5674                 return ERR_PTR(-EINVAL);
5675         }
5676         mddev->new_level = 5;
5677         mddev->new_layout = new_layout;
5678         mddev->delta_disks = -1;
5679         mddev->raid_disks -= 1;
5680         return setup_conf(mddev);
5681 }
5682
5683
5684 static int raid5_check_reshape(mddev_t *mddev)
5685 {
5686         /* For a 2-drive array, the layout and chunk size can be changed
5687          * immediately as not restriping is needed.
5688          * For larger arrays we record the new value - after validation
5689          * to be used by a reshape pass.
5690          */
5691         raid5_conf_t *conf = mddev->private;
5692         int new_chunk = mddev->new_chunk_sectors;
5693
5694         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5695                 return -EINVAL;
5696         if (new_chunk > 0) {
5697                 if (!is_power_of_2(new_chunk))
5698                         return -EINVAL;
5699                 if (new_chunk < (PAGE_SIZE>>9))
5700                         return -EINVAL;
5701                 if (mddev->array_sectors & (new_chunk-1))
5702                         /* not factor of array size */
5703                         return -EINVAL;
5704         }
5705
5706         /* They look valid */
5707
5708         if (mddev->raid_disks == 2) {
5709                 /* can make the change immediately */
5710                 if (mddev->new_layout >= 0) {
5711                         conf->algorithm = mddev->new_layout;
5712                         mddev->layout = mddev->new_layout;
5713                 }
5714                 if (new_chunk > 0) {
5715                         conf->chunk_sectors = new_chunk ;
5716                         mddev->chunk_sectors = new_chunk;
5717                 }
5718                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5719                 md_wakeup_thread(mddev->thread);
5720         }
5721         return check_reshape(mddev);
5722 }
5723
5724 static int raid6_check_reshape(mddev_t *mddev)
5725 {
5726         int new_chunk = mddev->new_chunk_sectors;
5727
5728         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5729                 return -EINVAL;
5730         if (new_chunk > 0) {
5731                 if (!is_power_of_2(new_chunk))
5732                         return -EINVAL;
5733                 if (new_chunk < (PAGE_SIZE >> 9))
5734                         return -EINVAL;
5735                 if (mddev->array_sectors & (new_chunk-1))
5736                         /* not factor of array size */
5737                         return -EINVAL;
5738         }
5739
5740         /* They look valid */
5741         return check_reshape(mddev);
5742 }
5743
5744 static void *raid5_takeover(mddev_t *mddev)
5745 {
5746         /* raid5 can take over:
5747          *  raid0 - if there is only one strip zone - make it a raid4 layout
5748          *  raid1 - if there are two drives.  We need to know the chunk size
5749          *  raid4 - trivial - just use a raid4 layout.
5750          *  raid6 - Providing it is a *_6 layout
5751          */
5752         if (mddev->level == 0)
5753                 return raid45_takeover_raid0(mddev, 5);
5754         if (mddev->level == 1)
5755                 return raid5_takeover_raid1(mddev);
5756         if (mddev->level == 4) {
5757                 mddev->new_layout = ALGORITHM_PARITY_N;
5758                 mddev->new_level = 5;
5759                 return setup_conf(mddev);
5760         }
5761         if (mddev->level == 6)
5762                 return raid5_takeover_raid6(mddev);
5763
5764         return ERR_PTR(-EINVAL);
5765 }
5766
5767 static void *raid4_takeover(mddev_t *mddev)
5768 {
5769         /* raid4 can take over:
5770          *  raid0 - if there is only one strip zone
5771          *  raid5 - if layout is right
5772          */
5773         if (mddev->level == 0)
5774                 return raid45_takeover_raid0(mddev, 4);
5775         if (mddev->level == 5 &&
5776             mddev->layout == ALGORITHM_PARITY_N) {
5777                 mddev->new_layout = 0;
5778                 mddev->new_level = 4;
5779                 return setup_conf(mddev);
5780         }
5781         return ERR_PTR(-EINVAL);
5782 }
5783
5784 static struct mdk_personality raid5_personality;
5785
5786 static void *raid6_takeover(mddev_t *mddev)
5787 {
5788         /* Currently can only take over a raid5.  We map the
5789          * personality to an equivalent raid6 personality
5790          * with the Q block at the end.
5791          */
5792         int new_layout;
5793
5794         if (mddev->pers != &raid5_personality)
5795                 return ERR_PTR(-EINVAL);
5796         if (mddev->degraded > 1)
5797                 return ERR_PTR(-EINVAL);
5798         if (mddev->raid_disks > 253)
5799                 return ERR_PTR(-EINVAL);
5800         if (mddev->raid_disks < 3)
5801                 return ERR_PTR(-EINVAL);
5802
5803         switch (mddev->layout) {
5804         case ALGORITHM_LEFT_ASYMMETRIC:
5805                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5806                 break;
5807         case ALGORITHM_RIGHT_ASYMMETRIC:
5808                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5809                 break;
5810         case ALGORITHM_LEFT_SYMMETRIC:
5811                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5812                 break;
5813         case ALGORITHM_RIGHT_SYMMETRIC:
5814                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5815                 break;
5816         case ALGORITHM_PARITY_0:
5817                 new_layout = ALGORITHM_PARITY_0_6;
5818                 break;
5819         case ALGORITHM_PARITY_N:
5820                 new_layout = ALGORITHM_PARITY_N;
5821                 break;
5822         default:
5823                 return ERR_PTR(-EINVAL);
5824         }
5825         mddev->new_level = 6;
5826         mddev->new_layout = new_layout;
5827         mddev->delta_disks = 1;
5828         mddev->raid_disks += 1;
5829         return setup_conf(mddev);
5830 }
5831
5832
5833 static struct mdk_personality raid6_personality =
5834 {
5835         .name           = "raid6",
5836         .level          = 6,
5837         .owner          = THIS_MODULE,
5838         .make_request   = make_request,
5839         .run            = run,
5840         .stop           = stop,
5841         .status         = status,
5842         .error_handler  = error,
5843         .hot_add_disk   = raid5_add_disk,
5844         .hot_remove_disk= raid5_remove_disk,
5845         .spare_active   = raid5_spare_active,
5846         .sync_request   = sync_request,
5847         .resize         = raid5_resize,
5848         .size           = raid5_size,
5849         .check_reshape  = raid6_check_reshape,
5850         .start_reshape  = raid5_start_reshape,
5851         .finish_reshape = raid5_finish_reshape,
5852         .quiesce        = raid5_quiesce,
5853         .takeover       = raid6_takeover,
5854 };
5855 static struct mdk_personality raid5_personality =
5856 {
5857         .name           = "raid5",
5858         .level          = 5,
5859         .owner          = THIS_MODULE,
5860         .make_request   = make_request,
5861         .run            = run,
5862         .stop           = stop,
5863         .status         = status,
5864         .error_handler  = error,
5865         .hot_add_disk   = raid5_add_disk,
5866         .hot_remove_disk= raid5_remove_disk,
5867         .spare_active   = raid5_spare_active,
5868         .sync_request   = sync_request,
5869         .resize         = raid5_resize,
5870         .size           = raid5_size,
5871         .check_reshape  = raid5_check_reshape,
5872         .start_reshape  = raid5_start_reshape,
5873         .finish_reshape = raid5_finish_reshape,
5874         .quiesce        = raid5_quiesce,
5875         .takeover       = raid5_takeover,
5876 };
5877
5878 static struct mdk_personality raid4_personality =
5879 {
5880         .name           = "raid4",
5881         .level          = 4,
5882         .owner          = THIS_MODULE,
5883         .make_request   = make_request,
5884         .run            = run,
5885         .stop           = stop,
5886         .status         = status,
5887         .error_handler  = error,
5888         .hot_add_disk   = raid5_add_disk,
5889         .hot_remove_disk= raid5_remove_disk,
5890         .spare_active   = raid5_spare_active,
5891         .sync_request   = sync_request,
5892         .resize         = raid5_resize,
5893         .size           = raid5_size,
5894         .check_reshape  = raid5_check_reshape,
5895         .start_reshape  = raid5_start_reshape,
5896         .finish_reshape = raid5_finish_reshape,
5897         .quiesce        = raid5_quiesce,
5898         .takeover       = raid4_takeover,
5899 };
5900
5901 static int __init raid5_init(void)
5902 {
5903         register_md_personality(&raid6_personality);
5904         register_md_personality(&raid5_personality);
5905         register_md_personality(&raid4_personality);
5906         return 0;
5907 }
5908
5909 static void raid5_exit(void)
5910 {
5911         unregister_md_personality(&raid6_personality);
5912         unregister_md_personality(&raid5_personality);
5913         unregister_md_personality(&raid4_personality);
5914 }
5915
5916 module_init(raid5_init);
5917 module_exit(raid5_exit);
5918 MODULE_LICENSE("GPL");
5919 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
5920 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5921 MODULE_ALIAS("md-raid5");
5922 MODULE_ALIAS("md-raid4");
5923 MODULE_ALIAS("md-level-5");
5924 MODULE_ALIAS("md-level-4");
5925 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5926 MODULE_ALIAS("md-raid6");
5927 MODULE_ALIAS("md-level-6");
5928
5929 /* This used to be two separate modules, they were: */
5930 MODULE_ALIAS("raid5");
5931 MODULE_ALIAS("raid6");