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 "md.h"
  54 #include "raid5.h"
  55 #include "bitmap.h"
  56
  57 /*
  58  * Stripe cache
  59  */
  60
  61 #define NR_STRIPES              256
  62 #define STRIPE_SIZE             PAGE_SIZE
  63 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
  64 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
  65 #define IO_THRESHOLD            1
  66 #define BYPASS_THRESHOLD        1
  67 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
  68 #define HASH_MASK               (NR_HASH - 1)
  69
  70 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
  71
  72 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
  73  * order without overlap.  There may be several bio's per stripe+device, and
  74  * a bio could span several devices.
  75  * When walking this list for a particular stripe+device, we must never proceed
  76  * beyond a bio that extends past this device, as the next bio might no longer
  77  * be valid.
  78  * This macro is used to determine the 'next' bio in the list, given the sector
  79  * of the current stripe+device
  80  */
  81 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
  82 /*
  83  * The following can be used to debug the driver
  84  */
  85 #define RAID5_PARANOIA  1
  86 #if RAID5_PARANOIA && defined(CONFIG_SMP)
  87 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
  88 #else
  89 # define CHECK_DEVLOCK()
  90 #endif
  91
  92 #ifdef DEBUG
  93 #define inline
  94 #define __inline__
  95 #endif
  96
  97 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
  98
  99 /*
 100  * We maintain a biased count of active stripes in the bottom 16 bits of
 101  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
 102  */
 103 static inline int raid5_bi_phys_segments(struct bio *bio)
 104 {
 105         return bio->bi_phys_segments & 0xffff;
 106 }
 107
 108 static inline int raid5_bi_hw_segments(struct bio *bio)
 109 {
 110         return (bio->bi_phys_segments >> 16) & 0xffff;
 111 }
 112
 113 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
 114 {
 115         --bio->bi_phys_segments;
 116         return raid5_bi_phys_segments(bio);
 117 }
 118
 119 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
 120 {
 121         unsigned short val = raid5_bi_hw_segments(bio);
 122
 123         --val;
 124         bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
 125         return val;
 126 }
 127
 128 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
 129 {
 130         bio->bi_phys_segments = raid5_bi_phys_segments(bio) || (cnt << 16);
 131 }
 132
 133 /* Find first data disk in a raid6 stripe */
 134 static inline int raid6_d0(struct stripe_head *sh)
 135 {
 136         if (sh->ddf_layout)
 137                 /* ddf always start from first device */
 138                 return 0;
 139         /* md starts just after Q block */
 140         if (sh->qd_idx == sh->disks - 1)
 141                 return 0;
 142         else
 143                 return sh->qd_idx + 1;
 144 }
 145 static inline int raid6_next_disk(int disk, int raid_disks)
 146 {
 147         disk++;
 148         return (disk < raid_disks) ? disk : 0;
 149 }
 150
 151 /* When walking through the disks in a raid5, starting at raid6_d0,
 152  * We need to map each disk to a 'slot', where the data disks are slot
 153  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 154  * is raid_disks-1.  This help does that mapping.
 155  */
 156 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 157                              int *count, int syndrome_disks)
 158 {
 159         int slot;
 160
 161         if (idx == sh->pd_idx)
 162                 return syndrome_disks;
 163         if (idx == sh->qd_idx)
 164                 return syndrome_disks + 1;
 165         slot = (*count)++;
 166         return slot;
 167 }
 168
 169 static void return_io(struct bio *return_bi)
 170 {
 171         struct bio *bi = return_bi;
 172         while (bi) {
 173
 174                 return_bi = bi->bi_next;
 175                 bi->bi_next = NULL;
 176                 bi->bi_size = 0;
 177                 bio_endio(bi, 0);
 178                 bi = return_bi;
 179         }
 180 }
 181
 182 static void print_raid5_conf (raid5_conf_t *conf);
 183
 184 static int stripe_operations_active(struct stripe_head *sh)
 185 {
 186         return sh->check_state || sh->reconstruct_state ||
 187                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 188                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 189 }
 190
 191 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
 192 {
 193         if (atomic_dec_and_test(&sh->count)) {
 194                 BUG_ON(!list_empty(&sh->lru));
 195                 BUG_ON(atomic_read(&conf->active_stripes)==0);
 196                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
 197                         if (test_bit(STRIPE_DELAYED, &sh->state)) {
 198                                 list_add_tail(&sh->lru, &conf->delayed_list);
 199                                 blk_plug_device(conf->mddev->queue);
 200                         } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 201                                    sh->bm_seq - conf->seq_write > 0) {
 202                                 list_add_tail(&sh->lru, &conf->bitmap_list);
 203                                 blk_plug_device(conf->mddev->queue);
 204                         } else {
 205                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
 206                                 list_add_tail(&sh->lru, &conf->handle_list);
 207                         }
 208                         md_wakeup_thread(conf->mddev->thread);
 209                 } else {
 210                         BUG_ON(stripe_operations_active(sh));
 211                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
 212                                 atomic_dec(&conf->preread_active_stripes);
 213                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
 214                                         md_wakeup_thread(conf->mddev->thread);
 215                         }
 216                         atomic_dec(&conf->active_stripes);
 217                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
 218                                 list_add_tail(&sh->lru, &conf->inactive_list);
 219                                 wake_up(&conf->wait_for_stripe);
 220                                 if (conf->retry_read_aligned)
 221                                         md_wakeup_thread(conf->mddev->thread);
 222                         }
 223                 }
 224         }
 225 }
 226
 227 static void release_stripe(struct stripe_head *sh)
 228 {
 229         raid5_conf_t *conf = sh->raid_conf;
 230         unsigned long flags;
 231
 232         spin_lock_irqsave(&conf->device_lock, flags);
 233         __release_stripe(conf, sh);
 234         spin_unlock_irqrestore(&conf->device_lock, flags);
 235 }
 236
 237 static inline void remove_hash(struct stripe_head *sh)
 238 {
 239         pr_debug("remove_hash(), stripe %llu\n",
 240                 (unsigned long long)sh->sector);
 241
 242         hlist_del_init(&sh->hash);
 243 }
 244
 245 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
 246 {
 247         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 248
 249         pr_debug("insert_hash(), stripe %llu\n",
 250                 (unsigned long long)sh->sector);
 251
 252         CHECK_DEVLOCK();
 253         hlist_add_head(&sh->hash, hp);
 254 }
 255
 256
 257 /* find an idle stripe, make sure it is unhashed, and return it. */
 258 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
 259 {
 260         struct stripe_head *sh = NULL;
 261         struct list_head *first;
 262
 263         CHECK_DEVLOCK();
 264         if (list_empty(&conf->inactive_list))
 265                 goto out;
 266         first = conf->inactive_list.next;
 267         sh = list_entry(first, struct stripe_head, lru);
 268         list_del_init(first);
 269         remove_hash(sh);
 270         atomic_inc(&conf->active_stripes);
 271 out:
 272         return sh;
 273 }
 274
 275 static void shrink_buffers(struct stripe_head *sh, int num)
 276 {
 277         struct page *p;
 278         int i;
 279
 280         for (i=0; i<num ; i++) {
 281                 p = sh->dev[i].page;
 282                 if (!p)
 283                         continue;
 284                 sh->dev[i].page = NULL;
 285                 put_page(p);
 286         }
 287 }
 288
 289 static int grow_buffers(struct stripe_head *sh, int num)
 290 {
 291         int i;
 292
 293         for (i=0; i<num; i++) {
 294                 struct page *page;
 295
 296                 if (!(page = alloc_page(GFP_KERNEL))) {
 297                         return 1;
 298                 }
 299                 sh->dev[i].page = page;
 300         }
 301         return 0;
 302 }
 303
 304 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
 305 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
 306                             struct stripe_head *sh);
 307
 308 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 309 {
 310         raid5_conf_t *conf = sh->raid_conf;
 311         int i;
 312
 313         BUG_ON(atomic_read(&sh->count) != 0);
 314         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 315         BUG_ON(stripe_operations_active(sh));
 316
 317         CHECK_DEVLOCK();
 318         pr_debug("init_stripe called, stripe %llu\n",
 319                 (unsigned long long)sh->sector);
 320
 321         remove_hash(sh);
 322
 323         sh->generation = conf->generation - previous;
 324         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 325         sh->sector = sector;
 326         stripe_set_idx(sector, conf, previous, sh);
 327         sh->state = 0;
 328
 329
 330         for (i = sh->disks; i--; ) {
 331                 struct r5dev *dev = &sh->dev[i];
 332
 333                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 334                     test_bit(R5_LOCKED, &dev->flags)) {
 335                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
 336                                (unsigned long long)sh->sector, i, dev->toread,
 337                                dev->read, dev->towrite, dev->written,
 338                                test_bit(R5_LOCKED, &dev->flags));
 339                         BUG();
 340                 }
 341                 dev->flags = 0;
 342                 raid5_build_block(sh, i, previous);
 343         }
 344         insert_hash(conf, sh);
 345 }
 346
 347 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
 348                                          short generation)
 349 {
 350         struct stripe_head *sh;
 351         struct hlist_node *hn;
 352
 353         CHECK_DEVLOCK();
 354         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 355         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
 356                 if (sh->sector == sector && sh->generation == generation)
 357                         return sh;
 358         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 359         return NULL;
 360 }
 361
 362 static void unplug_slaves(mddev_t *mddev);
 363 static void raid5_unplug_device(struct request_queue *q);
 364
 365 static struct stripe_head *
 366 get_active_stripe(raid5_conf_t *conf, sector_t sector,
 367                   int previous, int noblock, int noquiesce)
 368 {
 369         struct stripe_head *sh;
 370
 371         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 372
 373         spin_lock_irq(&conf->device_lock);
 374
 375         do {
 376                 wait_event_lock_irq(conf->wait_for_stripe,
 377                                     conf->quiesce == 0 || noquiesce,
 378                                     conf->device_lock, /* nothing */);
 379                 sh = __find_stripe(conf, sector, conf->generation - previous);
 380                 if (!sh) {
 381                         if (!conf->inactive_blocked)
 382                                 sh = get_free_stripe(conf);
 383                         if (noblock && sh == NULL)
 384                                 break;
 385                         if (!sh) {
 386                                 conf->inactive_blocked = 1;
 387                                 wait_event_lock_irq(conf->wait_for_stripe,
 388                                                     !list_empty(&conf->inactive_list) &&
 389                                                     (atomic_read(&conf->active_stripes)
 390                                                      < (conf->max_nr_stripes *3/4)
 391                                                      || !conf->inactive_blocked),
 392                                                     conf->device_lock,
 393                                                     raid5_unplug_device(conf->mddev->queue)
 394                                         );
 395                                 conf->inactive_blocked = 0;
 396                         } else
 397                                 init_stripe(sh, sector, previous);
 398                 } else {
 399                         if (atomic_read(&sh->count)) {
 400                                 BUG_ON(!list_empty(&sh->lru)
 401                                     && !test_bit(STRIPE_EXPANDING, &sh->state));
 402                         } else {
 403                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 404                                         atomic_inc(&conf->active_stripes);
 405                                 if (list_empty(&sh->lru) &&
 406                                     !test_bit(STRIPE_EXPANDING, &sh->state))
 407                                         BUG();
 408                                 list_del_init(&sh->lru);
 409                         }
 410                 }
 411         } while (sh == NULL);
 412
 413         if (sh)
 414                 atomic_inc(&sh->count);
 415
 416         spin_unlock_irq(&conf->device_lock);
 417         return sh;
 418 }
 419
 420 static void
 421 raid5_end_read_request(struct bio *bi, int error);
 422 static void
 423 raid5_end_write_request(struct bio *bi, int error);
 424
 425 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
 426 {
 427         raid5_conf_t *conf = sh->raid_conf;
 428         int i, disks = sh->disks;
 429
 430         might_sleep();
 431
 432         for (i = disks; i--; ) {
 433                 int rw;
 434                 struct bio *bi;
 435                 mdk_rdev_t *rdev;
 436                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
 437                         rw = WRITE;
 438                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
 439                         rw = READ;
 440                 else
 441                         continue;
 442
 443                 bi = &sh->dev[i].req;
 444
 445                 bi->bi_rw = rw;
 446                 if (rw == WRITE)
 447                         bi->bi_end_io = raid5_end_write_request;
 448                 else
 449                         bi->bi_end_io = raid5_end_read_request;
 450
 451                 rcu_read_lock();
 452                 rdev = rcu_dereference(conf->disks[i].rdev);
 453                 if (rdev && test_bit(Faulty, &rdev->flags))
 454                         rdev = NULL;
 455                 if (rdev)
 456                         atomic_inc(&rdev->nr_pending);
 457                 rcu_read_unlock();
 458
 459                 if (rdev) {
 460                         if (s->syncing || s->expanding || s->expanded)
 461                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
 462
 463                         set_bit(STRIPE_IO_STARTED, &sh->state);
 464
 465                         bi->bi_bdev = rdev->bdev;
 466                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
 467                                 __func__, (unsigned long long)sh->sector,
 468                                 bi->bi_rw, i);
 469                         atomic_inc(&sh->count);
 470                         bi->bi_sector = sh->sector + rdev->data_offset;
 471                         bi->bi_flags = 1 << BIO_UPTODATE;
 472                         bi->bi_vcnt = 1;
 473                         bi->bi_max_vecs = 1;
 474                         bi->bi_idx = 0;
 475                         bi->bi_io_vec = &sh->dev[i].vec;
 476                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
 477                         bi->bi_io_vec[0].bv_offset = 0;
 478                         bi->bi_size = STRIPE_SIZE;
 479                         bi->bi_next = NULL;
 480                         if (rw == WRITE &&
 481                             test_bit(R5_ReWrite, &sh->dev[i].flags))
 482                                 atomic_add(STRIPE_SECTORS,
 483                                         &rdev->corrected_errors);
 484                         generic_make_request(bi);
 485                 } else {
 486                         if (rw == WRITE)
 487                                 set_bit(STRIPE_DEGRADED, &sh->state);
 488                         pr_debug("skip op %ld on disc %d for sector %llu\n",
 489                                 bi->bi_rw, i, (unsigned long long)sh->sector);
 490                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 491                         set_bit(STRIPE_HANDLE, &sh->state);
 492                 }
 493         }
 494 }
 495
 496 static struct dma_async_tx_descriptor *
 497 async_copy_data(int frombio, struct bio *bio, struct page *page,
 498         sector_t sector, struct dma_async_tx_descriptor *tx)
 499 {
 500         struct bio_vec *bvl;
 501         struct page *bio_page;
 502         int i;
 503         int page_offset;
 504         struct async_submit_ctl submit;
 505         enum async_tx_flags flags = 0;
 506
 507         if (bio->bi_sector >= sector)
 508                 page_offset = (signed)(bio->bi_sector - sector) * 512;
 509         else
 510                 page_offset = (signed)(sector - bio->bi_sector) * -512;
 511
 512         if (frombio)
 513                 flags |= ASYNC_TX_FENCE;
 514         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
 515
 516         bio_for_each_segment(bvl, bio, i) {
 517                 int len = bio_iovec_idx(bio, i)->bv_len;
 518                 int clen;
 519                 int b_offset = 0;
 520
 521                 if (page_offset < 0) {
 522                         b_offset = -page_offset;
 523                         page_offset += b_offset;
 524                         len -= b_offset;
 525                 }
 526
 527                 if (len > 0 && page_offset + len > STRIPE_SIZE)
 528                         clen = STRIPE_SIZE - page_offset;
 529                 else
 530                         clen = len;
 531
 532                 if (clen > 0) {
 533                         b_offset += bio_iovec_idx(bio, i)->bv_offset;
 534                         bio_page = bio_iovec_idx(bio, i)->bv_page;
 535                         if (frombio)
 536                                 tx = async_memcpy(page, bio_page, page_offset,
 537                                                   b_offset, clen, &submit);
 538                         else
 539                                 tx = async_memcpy(bio_page, page, b_offset,
 540                                                   page_offset, clen, &submit);
 541                 }
 542                 /* chain the operations */
 543                 submit.depend_tx = tx;
 544
 545                 if (clen < len) /* hit end of page */
 546                         break;
 547                 page_offset +=  len;
 548         }
 549
 550         return tx;
 551 }
 552
 553 static void ops_complete_biofill(void *stripe_head_ref)
 554 {
 555         struct stripe_head *sh = stripe_head_ref;
 556         struct bio *return_bi = NULL;
 557         raid5_conf_t *conf = sh->raid_conf;
 558         int i;
 559
 560         pr_debug("%s: stripe %llu\n", __func__,
 561                 (unsigned long long)sh->sector);
 562
 563         /* clear completed biofills */
 564         spin_lock_irq(&conf->device_lock);
 565         for (i = sh->disks; i--; ) {
 566                 struct r5dev *dev = &sh->dev[i];
 567
 568                 /* acknowledge completion of a biofill operation */
 569                 /* and check if we need to reply to a read request,
 570                  * new R5_Wantfill requests are held off until
 571                  * !STRIPE_BIOFILL_RUN
 572                  */
 573                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
 574                         struct bio *rbi, *rbi2;
 575
 576                         BUG_ON(!dev->read);
 577                         rbi = dev->read;
 578                         dev->read = NULL;
 579                         while (rbi && rbi->bi_sector <
 580                                 dev->sector + STRIPE_SECTORS) {
 581                                 rbi2 = r5_next_bio(rbi, dev->sector);
 582                                 if (!raid5_dec_bi_phys_segments(rbi)) {
 583                                         rbi->bi_next = return_bi;
 584                                         return_bi = rbi;
 585                                 }
 586                                 rbi = rbi2;
 587                         }
 588                 }
 589         }
 590         spin_unlock_irq(&conf->device_lock);
 591         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
 592
 593         return_io(return_bi);
 594
 595         set_bit(STRIPE_HANDLE, &sh->state);
 596         release_stripe(sh);
 597 }
 598
 599 static void ops_run_biofill(struct stripe_head *sh)
 600 {
 601         struct dma_async_tx_descriptor *tx = NULL;
 602         raid5_conf_t *conf = sh->raid_conf;
 603         struct async_submit_ctl submit;
 604         int i;
 605
 606         pr_debug("%s: stripe %llu\n", __func__,
 607                 (unsigned long long)sh->sector);
 608
 609         for (i = sh->disks; i--; ) {
 610                 struct r5dev *dev = &sh->dev[i];
 611                 if (test_bit(R5_Wantfill, &dev->flags)) {
 612                         struct bio *rbi;
 613                         spin_lock_irq(&conf->device_lock);
 614                         dev->read = rbi = dev->toread;
 615                         dev->toread = NULL;
 616                         spin_unlock_irq(&conf->device_lock);
 617                         while (rbi && rbi->bi_sector <
 618                                 dev->sector + STRIPE_SECTORS) {
 619                                 tx = async_copy_data(0, rbi, dev->page,
 620                                         dev->sector, tx);
 621                                 rbi = r5_next_bio(rbi, dev->sector);
 622                         }
 623                 }
 624         }
 625
 626         atomic_inc(&sh->count);
 627         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
 628         async_trigger_callback(&submit);
 629 }
 630
 631 static void mark_target_uptodate(struct stripe_head *sh, int target)
 632 {
 633         struct r5dev *tgt;
 634
 635         if (target < 0)
 636                 return;
 637
 638         tgt = &sh->dev[target];
 639         set_bit(R5_UPTODATE, &tgt->flags);
 640         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
 641         clear_bit(R5_Wantcompute, &tgt->flags);
 642 }
 643
 644 static void ops_complete_compute(void *stripe_head_ref)
 645 {
 646         struct stripe_head *sh = stripe_head_ref;
 647
 648         pr_debug("%s: stripe %llu\n", __func__,
 649                 (unsigned long long)sh->sector);
 650
 651         /* mark the computed target(s) as uptodate */
 652         mark_target_uptodate(sh, sh->ops.target);
 653         mark_target_uptodate(sh, sh->ops.target2);
 654
 655         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
 656         if (sh->check_state == check_state_compute_run)
 657                 sh->check_state = check_state_compute_result;
 658         set_bit(STRIPE_HANDLE, &sh->state);
 659         release_stripe(sh);
 660 }
 661
 662 /* return a pointer to the address conversion region of the scribble buffer */
 663 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
 664                                  struct raid5_percpu *percpu)
 665 {
 666         return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
 667 }
 668
 669 static struct dma_async_tx_descriptor *
 670 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
 671 {
 672         int disks = sh->disks;
 673         struct page **xor_srcs = percpu->scribble;
 674         int target = sh->ops.target;
 675         struct r5dev *tgt = &sh->dev[target];
 676         struct page *xor_dest = tgt->page;
 677         int count = 0;
 678         struct dma_async_tx_descriptor *tx;
 679         struct async_submit_ctl submit;
 680         int i;
 681
 682         pr_debug("%s: stripe %llu block: %d\n",
 683                 __func__, (unsigned long long)sh->sector, target);
 684         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
 685
 686         for (i = disks; i--; )
 687                 if (i != target)
 688                         xor_srcs[count++] = sh->dev[i].page;
 689
 690         atomic_inc(&sh->count);
 691
 692         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
 693                           ops_complete_compute, sh, to_addr_conv(sh, percpu));
 694         if (unlikely(count == 1))
 695                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
 696         else
 697                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
 698
 699         return tx;
 700 }
 701
 702 /* set_syndrome_sources - populate source buffers for gen_syndrome
 703  * @srcs - (struct page *) array of size sh->disks
 704  * @sh - stripe_head to parse
 705  *
 706  * Populates srcs in proper layout order for the stripe and returns the
 707  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
 708  * destination buffer is recorded in srcs[count] and the Q destination
 709  * is recorded in srcs[count+1]].
 710  */
 711 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
 712 {
 713         int disks = sh->disks;
 714         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
 715         int d0_idx = raid6_d0(sh);
 716         int count;
 717         int i;
 718
 719         for (i = 0; i < disks; i++)
 720                 srcs[i] = (void *)raid6_empty_zero_page;
 721
 722         count = 0;
 723         i = d0_idx;
 724         do {
 725                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
 726
 727                 srcs[slot] = sh->dev[i].page;
 728                 i = raid6_next_disk(i, disks);
 729         } while (i != d0_idx);
 730         BUG_ON(count != syndrome_disks);
 731
 732         return count;
 733 }
 734
 735 static struct dma_async_tx_descriptor *
 736 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
 737 {
 738         int disks = sh->disks;
 739         struct page **blocks = percpu->scribble;
 740         int target;
 741         int qd_idx = sh->qd_idx;
 742         struct dma_async_tx_descriptor *tx;
 743         struct async_submit_ctl submit;
 744         struct r5dev *tgt;
 745         struct page *dest;
 746         int i;
 747         int count;
 748
 749         if (sh->ops.target < 0)
 750                 target = sh->ops.target2;
 751         else if (sh->ops.target2 < 0)
 752                 target = sh->ops.target;
 753         else
 754                 /* we should only have one valid target */
 755                 BUG();
 756         BUG_ON(target < 0);
 757         pr_debug("%s: stripe %llu block: %d\n",
 758                 __func__, (unsigned long long)sh->sector, target);
 759
 760         tgt = &sh->dev[target];
 761         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
 762         dest = tgt->page;
 763
 764         atomic_inc(&sh->count);
 765
 766         if (target == qd_idx) {
 767                 count = set_syndrome_sources(blocks, sh);
 768                 blocks[count] = NULL; /* regenerating p is not necessary */
 769                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
 770                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
 771                                   ops_complete_compute, sh,
 772                                   to_addr_conv(sh, percpu));
 773                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
 774         } else {
 775                 /* Compute any data- or p-drive using XOR */
 776                 count = 0;
 777                 for (i = disks; i-- ; ) {
 778                         if (i == target || i == qd_idx)
 779                                 continue;
 780                         blocks[count++] = sh->dev[i].page;
 781                 }
 782
 783                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
 784                                   NULL, ops_complete_compute, sh,
 785                                   to_addr_conv(sh, percpu));
 786                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
 787         }
 788
 789         return tx;
 790 }
 791
 792 static struct dma_async_tx_descriptor *
 793 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
 794 {
 795         int i, count, disks = sh->disks;
 796         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
 797         int d0_idx = raid6_d0(sh);
 798         int faila = -1, failb = -1;
 799         int target = sh->ops.target;
 800         int target2 = sh->ops.target2;
 801         struct r5dev *tgt = &sh->dev[target];
 802         struct r5dev *tgt2 = &sh->dev[target2];
 803         struct dma_async_tx_descriptor *tx;
 804         struct page **blocks = percpu->scribble;
 805         struct async_submit_ctl submit;
 806
 807         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
 808                  __func__, (unsigned long long)sh->sector, target, target2);
 809         BUG_ON(target < 0 || target2 < 0);
 810         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
 811         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
 812
 813         /* we need to open-code set_syndrome_sources to handle the
 814          * slot number conversion for 'faila' and 'failb'
 815          */
 816         for (i = 0; i < disks ; i++)
 817                 blocks[i] = (void *)raid6_empty_zero_page;
 818         count = 0;
 819         i = d0_idx;
 820         do {
 821                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
 822
 823                 blocks[slot] = sh->dev[i].page;
 824
 825                 if (i == target)
 826                         faila = slot;
 827                 if (i == target2)
 828                         failb = slot;
 829                 i = raid6_next_disk(i, disks);
 830         } while (i != d0_idx);
 831         BUG_ON(count != syndrome_disks);
 832
 833         BUG_ON(faila == failb);
 834         if (failb < faila)
 835                 swap(faila, failb);
 836         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
 837                  __func__, (unsigned long long)sh->sector, faila, failb);
 838
 839         atomic_inc(&sh->count);
 840
 841         if (failb == syndrome_disks+1) {
 842                 /* Q disk is one of the missing disks */
 843                 if (faila == syndrome_disks) {
 844                         /* Missing P+Q, just recompute */
 845                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
 846                                           ops_complete_compute, sh,
 847                                           to_addr_conv(sh, percpu));
 848                         return async_gen_syndrome(blocks, 0, count+2,
 849                                                   STRIPE_SIZE, &submit);
 850                 } else {
 851                         struct page *dest;
 852                         int data_target;
 853                         int qd_idx = sh->qd_idx;
 854
 855                         /* Missing D+Q: recompute D from P, then recompute Q */
 856                         if (target == qd_idx)
 857                                 data_target = target2;
 858                         else
 859                                 data_target = target;
 860
 861                         count = 0;
 862                         for (i = disks; i-- ; ) {
 863                                 if (i == data_target || i == qd_idx)
 864                                         continue;
 865                                 blocks[count++] = sh->dev[i].page;
 866                         }
 867                         dest = sh->dev[data_target].page;
 868                         init_async_submit(&submit,
 869                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
 870                                           NULL, NULL, NULL,
 871                                           to_addr_conv(sh, percpu));
 872                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
 873                                        &submit);
 874
 875                         count = set_syndrome_sources(blocks, sh);
 876                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
 877                                           ops_complete_compute, sh,
 878                                           to_addr_conv(sh, percpu));
 879                         return async_gen_syndrome(blocks, 0, count+2,
 880                                                   STRIPE_SIZE, &submit);
 881                 }
 882         } else {
 883                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
 884                                   ops_complete_compute, sh,
 885                                   to_addr_conv(sh, percpu));
 886                 if (failb == syndrome_disks) {
 887                         /* We're missing D+P. */
 888                         return async_raid6_datap_recov(syndrome_disks+2,
 889                                                        STRIPE_SIZE, faila,
 890                                                        blocks, &submit);
 891                 } else {
 892                         /* We're missing D+D. */
 893                         return async_raid6_2data_recov(syndrome_disks+2,
 894                                                        STRIPE_SIZE, faila, failb,
 895                                                        blocks, &submit);
 896                 }
 897         }
 898 }
 899
 900
 901 static void ops_complete_prexor(void *stripe_head_ref)
 902 {
 903         struct stripe_head *sh = stripe_head_ref;
 904
 905         pr_debug("%s: stripe %llu\n", __func__,
 906                 (unsigned long long)sh->sector);
 907 }
 908
 909 static struct dma_async_tx_descriptor *
 910 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
 911                struct dma_async_tx_descriptor *tx)
 912 {
 913         int disks = sh->disks;
 914         struct page **xor_srcs = percpu->scribble;
 915         int count = 0, pd_idx = sh->pd_idx, i;
 916         struct async_submit_ctl submit;
 917
 918         /* existing parity data subtracted */
 919         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
 920
 921         pr_debug("%s: stripe %llu\n", __func__,
 922                 (unsigned long long)sh->sector);
 923
 924         for (i = disks; i--; ) {
 925                 struct r5dev *dev = &sh->dev[i];
 926                 /* Only process blocks that are known to be uptodate */
 927                 if (test_bit(R5_Wantdrain, &dev->flags))
 928                         xor_srcs[count++] = dev->page;
 929         }
 930
 931         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
 932                           ops_complete_prexor, sh, to_addr_conv(sh, percpu));
 933         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
 934
 935         return tx;
 936 }
 937
 938 static struct dma_async_tx_descriptor *
 939 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
 940 {
 941         int disks = sh->disks;
 942         int i;
 943
 944         pr_debug("%s: stripe %llu\n", __func__,
 945                 (unsigned long long)sh->sector);
 946
 947         for (i = disks; i--; ) {
 948                 struct r5dev *dev = &sh->dev[i];
 949                 struct bio *chosen;
 950
 951                 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
 952                         struct bio *wbi;
 953
 954                         spin_lock(&sh->lock);
 955                         chosen = dev->towrite;
 956                         dev->towrite = NULL;
 957                         BUG_ON(dev->written);
 958                         wbi = dev->written = chosen;
 959                         spin_unlock(&sh->lock);
 960
 961                         while (wbi && wbi->bi_sector <
 962                                 dev->sector + STRIPE_SECTORS) {
 963                                 tx = async_copy_data(1, wbi, dev->page,
 964                                         dev->sector, tx);
 965                                 wbi = r5_next_bio(wbi, dev->sector);
 966                         }
 967                 }
 968         }
 969
 970         return tx;
 971 }
 972
 973 static void ops_complete_reconstruct(void *stripe_head_ref)
 974 {
 975         struct stripe_head *sh = stripe_head_ref;
 976         int disks = sh->disks;
 977         int pd_idx = sh->pd_idx;
 978         int qd_idx = sh->qd_idx;
 979         int i;
 980
 981         pr_debug("%s: stripe %llu\n", __func__,
 982                 (unsigned long long)sh->sector);
 983
 984         for (i = disks; i--; ) {
 985                 struct r5dev *dev = &sh->dev[i];
 986
 987                 if (dev->written || i == pd_idx || i == qd_idx)
 988                         set_bit(R5_UPTODATE, &dev->flags);
 989         }
 990
 991         if (sh->reconstruct_state == reconstruct_state_drain_run)
 992                 sh->reconstruct_state = reconstruct_state_drain_result;
 993         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
 994                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
 995         else {
 996                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
 997                 sh->reconstruct_state = reconstruct_state_result;
 998         }
 999
1000         set_bit(STRIPE_HANDLE, &sh->state);
1001         release_stripe(sh);
1002 }
1003
1004 static void
1005 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1006                      struct dma_async_tx_descriptor *tx)
1007 {
1008         int disks = sh->disks;
1009         struct page **xor_srcs = percpu->scribble;
1010         struct async_submit_ctl submit;
1011         int count = 0, pd_idx = sh->pd_idx, i;
1012         struct page *xor_dest;
1013         int prexor = 0;
1014         unsigned long flags;
1015
1016         pr_debug("%s: stripe %llu\n", __func__,
1017                 (unsigned long long)sh->sector);
1018
1019         /* check if prexor is active which means only process blocks
1020          * that are part of a read-modify-write (written)
1021          */
1022         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1023                 prexor = 1;
1024                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1025                 for (i = disks; i--; ) {
1026                         struct r5dev *dev = &sh->dev[i];
1027                         if (dev->written)
1028                                 xor_srcs[count++] = dev->page;
1029                 }
1030         } else {
1031                 xor_dest = sh->dev[pd_idx].page;
1032                 for (i = disks; i--; ) {
1033                         struct r5dev *dev = &sh->dev[i];
1034                         if (i != pd_idx)
1035                                 xor_srcs[count++] = dev->page;
1036                 }
1037         }
1038
1039         /* 1/ if we prexor'd then the dest is reused as a source
1040          * 2/ if we did not prexor then we are redoing the parity
1041          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1042          * for the synchronous xor case
1043          */
1044         flags = ASYNC_TX_ACK |
1045                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1046
1047         atomic_inc(&sh->count);
1048
1049         init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1050                           to_addr_conv(sh, percpu));
1051         if (unlikely(count == 1))
1052                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1053         else
1054                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1055 }
1056
1057 static void
1058 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1059                      struct dma_async_tx_descriptor *tx)
1060 {
1061         struct async_submit_ctl submit;
1062         struct page **blocks = percpu->scribble;
1063         int count;
1064
1065         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1066
1067         count = set_syndrome_sources(blocks, sh);
1068
1069         atomic_inc(&sh->count);
1070
1071         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1072                           sh, to_addr_conv(sh, percpu));
1073         async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1074 }
1075
1076 static void ops_complete_check(void *stripe_head_ref)
1077 {
1078         struct stripe_head *sh = stripe_head_ref;
1079
1080         pr_debug("%s: stripe %llu\n", __func__,
1081                 (unsigned long long)sh->sector);
1082
1083         sh->check_state = check_state_check_result;
1084         set_bit(STRIPE_HANDLE, &sh->state);
1085         release_stripe(sh);
1086 }
1087
1088 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1089 {
1090         int disks = sh->disks;
1091         int pd_idx = sh->pd_idx;
1092         int qd_idx = sh->qd_idx;
1093         struct page *xor_dest;
1094         struct page **xor_srcs = percpu->scribble;
1095         struct dma_async_tx_descriptor *tx;
1096         struct async_submit_ctl submit;
1097         int count;
1098         int i;
1099
1100         pr_debug("%s: stripe %llu\n", __func__,
1101                 (unsigned long long)sh->sector);
1102
1103         count = 0;
1104         xor_dest = sh->dev[pd_idx].page;
1105         xor_srcs[count++] = xor_dest;
1106         for (i = disks; i--; ) {
1107                 if (i == pd_idx || i == qd_idx)
1108                         continue;
1109                 xor_srcs[count++] = sh->dev[i].page;
1110         }
1111
1112         init_async_submit(&submit, 0, NULL, NULL, NULL,
1113                           to_addr_conv(sh, percpu));
1114         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1115                            &sh->ops.zero_sum_result, &submit);
1116
1117         atomic_inc(&sh->count);
1118         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1119         tx = async_trigger_callback(&submit);
1120 }
1121
1122 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1123 {
1124         struct page **srcs = percpu->scribble;
1125         struct async_submit_ctl submit;
1126         int count;
1127
1128         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1129                 (unsigned long long)sh->sector, checkp);
1130
1131         count = set_syndrome_sources(srcs, sh);
1132         if (!checkp)
1133                 srcs[count] = NULL;
1134
1135         atomic_inc(&sh->count);
1136         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1137                           sh, to_addr_conv(sh, percpu));
1138         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1139                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1140 }
1141
1142 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1143 {
1144         int overlap_clear = 0, i, disks = sh->disks;
1145         struct dma_async_tx_descriptor *tx = NULL;
1146         raid5_conf_t *conf = sh->raid_conf;
1147         int level = conf->level;
1148         struct raid5_percpu *percpu;
1149         unsigned long cpu;
1150
1151         cpu = get_cpu();
1152         percpu = per_cpu_ptr(conf->percpu, cpu);
1153         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1154                 ops_run_biofill(sh);
1155                 overlap_clear++;
1156         }
1157
1158         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1159                 if (level < 6)
1160                         tx = ops_run_compute5(sh, percpu);
1161                 else {
1162                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
1163                                 tx = ops_run_compute6_1(sh, percpu);
1164                         else
1165                                 tx = ops_run_compute6_2(sh, percpu);
1166                 }
1167                 /* terminate the chain if reconstruct is not set to be run */
1168                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1169                         async_tx_ack(tx);
1170         }
1171
1172         if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1173                 tx = ops_run_prexor(sh, percpu, tx);
1174
1175         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1176                 tx = ops_run_biodrain(sh, tx);
1177                 overlap_clear++;
1178         }
1179
1180         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1181                 if (level < 6)
1182                         ops_run_reconstruct5(sh, percpu, tx);
1183                 else
1184                         ops_run_reconstruct6(sh, percpu, tx);
1185         }
1186
1187         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1188                 if (sh->check_state == check_state_run)
1189                         ops_run_check_p(sh, percpu);
1190                 else if (sh->check_state == check_state_run_q)
1191                         ops_run_check_pq(sh, percpu, 0);
1192                 else if (sh->check_state == check_state_run_pq)
1193                         ops_run_check_pq(sh, percpu, 1);
1194                 else
1195                         BUG();
1196         }
1197
1198         if (overlap_clear)
1199                 for (i = disks; i--; ) {
1200                         struct r5dev *dev = &sh->dev[i];
1201                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1202                                 wake_up(&sh->raid_conf->wait_for_overlap);
1203                 }
1204         put_cpu();
1205 }
1206
1207 static int grow_one_stripe(raid5_conf_t *conf)
1208 {
1209         struct stripe_head *sh;
1210         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
1211         if (!sh)
1212                 return 0;
1213         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
1214         sh->raid_conf = conf;
1215         spin_lock_init(&sh->lock);
1216
1217         if (grow_buffers(sh, conf->raid_disks)) {
1218                 shrink_buffers(sh, conf->raid_disks);
1219                 kmem_cache_free(conf->slab_cache, sh);
1220                 return 0;
1221         }
1222         sh->disks = conf->raid_disks;
1223         /* we just created an active stripe so... */
1224         atomic_set(&sh->count, 1);
1225         atomic_inc(&conf->active_stripes);
1226         INIT_LIST_HEAD(&sh->lru);
1227         release_stripe(sh);
1228         return 1;
1229 }
1230
1231 static int grow_stripes(raid5_conf_t *conf, int num)
1232 {
1233         struct kmem_cache *sc;
1234         int devs = conf->raid_disks;
1235
1236         sprintf(conf->cache_name[0],
1237                 "raid%d-%s", conf->level, mdname(conf->mddev));
1238         sprintf(conf->cache_name[1],
1239                 "raid%d-%s-alt", conf->level, mdname(conf->mddev));
1240         conf->active_name = 0;
1241         sc = kmem_cache_create(conf->cache_name[conf->active_name],
1242                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1243                                0, 0, NULL);
1244         if (!sc)
1245                 return 1;
1246         conf->slab_cache = sc;
1247         conf->pool_size = devs;
1248         while (num--)
1249                 if (!grow_one_stripe(conf))
1250                         return 1;
1251         return 0;
1252 }
1253
1254 /**
1255  * scribble_len - return the required size of the scribble region
1256  * @num - total number of disks in the array
1257  *
1258  * The size must be enough to contain:
1259  * 1/ a struct page pointer for each device in the array +2
1260  * 2/ room to convert each entry in (1) to its corresponding dma
1261  *    (dma_map_page()) or page (page_address()) address.
1262  *
1263  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1264  * calculate over all devices (not just the data blocks), using zeros in place
1265  * of the P and Q blocks.
1266  */
1267 static size_t scribble_len(int num)
1268 {
1269         size_t len;
1270
1271         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1272
1273         return len;
1274 }
1275
1276 static int resize_stripes(raid5_conf_t *conf, int newsize)
1277 {
1278         /* Make all the stripes able to hold 'newsize' devices.
1279          * New slots in each stripe get 'page' set to a new page.
1280          *
1281          * This happens in stages:
1282          * 1/ create a new kmem_cache and allocate the required number of
1283          *    stripe_heads.
1284          * 2/ gather all the old stripe_heads and tranfer the pages across
1285          *    to the new stripe_heads.  This will have the side effect of
1286          *    freezing the array as once all stripe_heads have been collected,
1287          *    no IO will be possible.  Old stripe heads are freed once their
1288          *    pages have been transferred over, and the old kmem_cache is
1289          *    freed when all stripes are done.
1290          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
1291          *    we simple return a failre status - no need to clean anything up.
1292          * 4/ allocate new pages for the new slots in the new stripe_heads.
1293          *    If this fails, we don't bother trying the shrink the
1294          *    stripe_heads down again, we just leave them as they are.
1295          *    As each stripe_head is processed the new one is released into
1296          *    active service.
1297          *
1298          * Once step2 is started, we cannot afford to wait for a write,
1299          * so we use GFP_NOIO allocations.
1300          */
1301         struct stripe_head *osh, *nsh;
1302         LIST_HEAD(newstripes);
1303         struct disk_info *ndisks;
1304         unsigned long cpu;
1305         int err;
1306         struct kmem_cache *sc;
1307         int i;
1308
1309         if (newsize <= conf->pool_size)
1310                 return 0; /* never bother to shrink */
1311
1312         err = md_allow_write(conf->mddev);
1313         if (err)
1314                 return err;
1315
1316         /* Step 1 */
1317         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1318                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1319                                0, 0, NULL);
1320         if (!sc)
1321                 return -ENOMEM;
1322
1323         for (i = conf->max_nr_stripes; i; i--) {
1324                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1325                 if (!nsh)
1326                         break;
1327
1328                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1329
1330                 nsh->raid_conf = conf;
1331                 spin_lock_init(&nsh->lock);
1332
1333                 list_add(&nsh->lru, &newstripes);
1334         }
1335         if (i) {
1336                 /* didn't get enough, give up */
1337                 while (!list_empty(&newstripes)) {
1338                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1339                         list_del(&nsh->lru);
1340                         kmem_cache_free(sc, nsh);
1341                 }
1342                 kmem_cache_destroy(sc);
1343                 return -ENOMEM;
1344         }
1345         /* Step 2 - Must use GFP_NOIO now.
1346          * OK, we have enough stripes, start collecting inactive
1347          * stripes and copying them over
1348          */
1349         list_for_each_entry(nsh, &newstripes, lru) {
1350                 spin_lock_irq(&conf->device_lock);
1351                 wait_event_lock_irq(conf->wait_for_stripe,
1352                                     !list_empty(&conf->inactive_list),
1353                                     conf->device_lock,
1354                                     unplug_slaves(conf->mddev)
1355                         );
1356                 osh = get_free_stripe(conf);
1357                 spin_unlock_irq(&conf->device_lock);
1358                 atomic_set(&nsh->count, 1);
1359                 for(i=0; i<conf->pool_size; i++)
1360                         nsh->dev[i].page = osh->dev[i].page;
1361                 for( ; i<newsize; i++)
1362                         nsh->dev[i].page = NULL;
1363                 kmem_cache_free(conf->slab_cache, osh);
1364         }
1365         kmem_cache_destroy(conf->slab_cache);
1366
1367         /* Step 3.
1368          * At this point, we are holding all the stripes so the array
1369          * is completely stalled, so now is a good time to resize
1370          * conf->disks and the scribble region
1371          */
1372         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1373         if (ndisks) {
1374                 for (i=0; i<conf->raid_disks; i++)
1375                         ndisks[i] = conf->disks[i];
1376                 kfree(conf->disks);
1377                 conf->disks = ndisks;
1378         } else
1379                 err = -ENOMEM;
1380
1381         get_online_cpus();
1382         conf->scribble_len = scribble_len(newsize);
1383         for_each_present_cpu(cpu) {
1384                 struct raid5_percpu *percpu;
1385                 void *scribble;
1386
1387                 percpu = per_cpu_ptr(conf->percpu, cpu);
1388                 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1389
1390                 if (scribble) {
1391                         kfree(percpu->scribble);
1392                         percpu->scribble = scribble;
1393                 } else {
1394                         err = -ENOMEM;
1395                         break;
1396                 }
1397         }
1398         put_online_cpus();
1399
1400         /* Step 4, return new stripes to service */
1401         while(!list_empty(&newstripes)) {
1402                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1403                 list_del_init(&nsh->lru);
1404
1405                 for (i=conf->raid_disks; i < newsize; i++)
1406                         if (nsh->dev[i].page == NULL) {
1407                                 struct page *p = alloc_page(GFP_NOIO);
1408                                 nsh->dev[i].page = p;
1409                                 if (!p)
1410                                         err = -ENOMEM;
1411                         }
1412                 release_stripe(nsh);
1413         }
1414         /* critical section pass, GFP_NOIO no longer needed */
1415
1416         conf->slab_cache = sc;
1417         conf->active_name = 1-conf->active_name;
1418         conf->pool_size = newsize;
1419         return err;
1420 }
1421
1422 static int drop_one_stripe(raid5_conf_t *conf)
1423 {
1424         struct stripe_head *sh;
1425
1426         spin_lock_irq(&conf->device_lock);
1427         sh = get_free_stripe(conf);
1428         spin_unlock_irq(&conf->device_lock);
1429         if (!sh)
1430                 return 0;
1431         BUG_ON(atomic_read(&sh->count));
1432         shrink_buffers(sh, conf->pool_size);
1433         kmem_cache_free(conf->slab_cache, sh);
1434         atomic_dec(&conf->active_stripes);
1435         return 1;
1436 }
1437
1438 static void shrink_stripes(raid5_conf_t *conf)
1439 {
1440         while (drop_one_stripe(conf))
1441                 ;
1442
1443         if (conf->slab_cache)
1444                 kmem_cache_destroy(conf->slab_cache);
1445         conf->slab_cache = NULL;
1446 }
1447
1448 static void raid5_end_read_request(struct bio * bi, int error)
1449 {
1450         struct stripe_head *sh = bi->bi_private;
1451         raid5_conf_t *conf = sh->raid_conf;
1452         int disks = sh->disks, i;
1453         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1454         char b[BDEVNAME_SIZE];
1455         mdk_rdev_t *rdev;
1456
1457
1458         for (i=0 ; i<disks; i++)
1459                 if (bi == &sh->dev[i].req)
1460                         break;
1461
1462         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1463                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1464                 uptodate);
1465         if (i == disks) {
1466                 BUG();
1467                 return;
1468         }
1469
1470         if (uptodate) {
1471                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1472                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1473                         rdev = conf->disks[i].rdev;
1474                         printk_rl(KERN_INFO "raid5:%s: read error corrected"
1475                                   " (%lu sectors at %llu on %s)\n",
1476                                   mdname(conf->mddev), STRIPE_SECTORS,
1477                                   (unsigned long long)(sh->sector
1478                                                        + rdev->data_offset),
1479                                   bdevname(rdev->bdev, b));
1480                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1481                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1482                 }
1483                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1484                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1485         } else {
1486                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1487                 int retry = 0;
1488                 rdev = conf->disks[i].rdev;
1489
1490                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1491                 atomic_inc(&rdev->read_errors);
1492                 if (conf->mddev->degraded)
1493                         printk_rl(KERN_WARNING
1494                                   "raid5:%s: read error not correctable "
1495                                   "(sector %llu on %s).\n",
1496                                   mdname(conf->mddev),
1497                                   (unsigned long long)(sh->sector
1498                                                        + rdev->data_offset),
1499                                   bdn);
1500                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1501                         /* Oh, no!!! */
1502                         printk_rl(KERN_WARNING
1503                                   "raid5:%s: read error NOT corrected!! "
1504                                   "(sector %llu on %s).\n",
1505                                   mdname(conf->mddev),
1506                                   (unsigned long long)(sh->sector
1507                                                        + rdev->data_offset),
1508                                   bdn);
1509                 else if (atomic_read(&rdev->read_errors)
1510                          > conf->max_nr_stripes)
1511                         printk(KERN_WARNING
1512                                "raid5:%s: Too many read errors, failing device %s.\n",
1513                                mdname(conf->mddev), bdn);
1514                 else
1515                         retry = 1;
1516                 if (retry)
1517                         set_bit(R5_ReadError, &sh->dev[i].flags);
1518                 else {
1519                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1520                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1521                         md_error(conf->mddev, rdev);
1522                 }
1523         }
1524         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1525         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1526         set_bit(STRIPE_HANDLE, &sh->state);
1527         release_stripe(sh);
1528 }
1529
1530 static void raid5_end_write_request(struct bio *bi, int error)
1531 {
1532         struct stripe_head *sh = bi->bi_private;
1533         raid5_conf_t *conf = sh->raid_conf;
1534         int disks = sh->disks, i;
1535         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1536
1537         for (i=0 ; i<disks; i++)
1538                 if (bi == &sh->dev[i].req)
1539                         break;
1540
1541         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1542                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1543                 uptodate);
1544         if (i == disks) {
1545                 BUG();
1546                 return;
1547         }
1548
1549         if (!uptodate)
1550                 md_error(conf->mddev, conf->disks[i].rdev);
1551
1552         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1553
1554         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1555         set_bit(STRIPE_HANDLE, &sh->state);
1556         release_stripe(sh);
1557 }
1558
1559
1560 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1561
1562 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1563 {
1564         struct r5dev *dev = &sh->dev[i];
1565
1566         bio_init(&dev->req);
1567         dev->req.bi_io_vec = &dev->vec;
1568         dev->req.bi_vcnt++;
1569         dev->req.bi_max_vecs++;
1570         dev->vec.bv_page = dev->page;
1571         dev->vec.bv_len = STRIPE_SIZE;
1572         dev->vec.bv_offset = 0;
1573
1574         dev->req.bi_sector = sh->sector;
1575         dev->req.bi_private = sh;
1576
1577         dev->flags = 0;
1578         dev->sector = compute_blocknr(sh, i, previous);
1579 }
1580
1581 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1582 {
1583         char b[BDEVNAME_SIZE];
1584         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1585         pr_debug("raid5: error called\n");
1586
1587         if (!test_bit(Faulty, &rdev->flags)) {
1588                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1589                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1590                         unsigned long flags;
1591                         spin_lock_irqsave(&conf->device_lock, flags);
1592                         mddev->degraded++;
1593                         spin_unlock_irqrestore(&conf->device_lock, flags);
1594                         /*
1595                          * if recovery was running, make sure it aborts.
1596                          */
1597                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1598                 }
1599                 set_bit(Faulty, &rdev->flags);
1600                 printk(KERN_ALERT
1601                        "raid5: Disk failure on %s, disabling device.\n"
1602                        "raid5: Operation continuing on %d devices.\n",
1603                        bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1604         }
1605 }
1606
1607 /*
1608  * Input: a 'big' sector number,
1609  * Output: index of the data and parity disk, and the sector # in them.
1610  */
1611 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1612                                      int previous, int *dd_idx,
1613                                      struct stripe_head *sh)
1614 {
1615         long stripe;
1616         unsigned long chunk_number;
1617         unsigned int chunk_offset;
1618         int pd_idx, qd_idx;
1619         int ddf_layout = 0;
1620         sector_t new_sector;
1621         int algorithm = previous ? conf->prev_algo
1622                                  : conf->algorithm;
1623         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1624                                          : conf->chunk_sectors;
1625         int raid_disks = previous ? conf->previous_raid_disks
1626                                   : conf->raid_disks;
1627         int data_disks = raid_disks - conf->max_degraded;
1628
1629         /* First compute the information on this sector */
1630
1631         /*
1632          * Compute the chunk number and the sector offset inside the chunk
1633          */
1634         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1635         chunk_number = r_sector;
1636         BUG_ON(r_sector != chunk_number);
1637
1638         /*
1639          * Compute the stripe number
1640          */
1641         stripe = chunk_number / data_disks;
1642
1643         /*
1644          * Compute the data disk and parity disk indexes inside the stripe
1645          */
1646         *dd_idx = chunk_number % data_disks;
1647
1648         /*
1649          * Select the parity disk based on the user selected algorithm.
1650          */
1651         pd_idx = qd_idx = ~0;
1652         switch(conf->level) {
1653         case 4:
1654                 pd_idx = data_disks;
1655                 break;
1656         case 5:
1657                 switch (algorithm) {
1658                 case ALGORITHM_LEFT_ASYMMETRIC:
1659                         pd_idx = data_disks - stripe % raid_disks;
1660                         if (*dd_idx >= pd_idx)
1661                                 (*dd_idx)++;
1662                         break;
1663                 case ALGORITHM_RIGHT_ASYMMETRIC:
1664                         pd_idx = stripe % raid_disks;
1665                         if (*dd_idx >= pd_idx)
1666                                 (*dd_idx)++;
1667                         break;
1668                 case ALGORITHM_LEFT_SYMMETRIC:
1669                         pd_idx = data_disks - stripe % raid_disks;
1670                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1671                         break;
1672                 case ALGORITHM_RIGHT_SYMMETRIC:
1673                         pd_idx = stripe % raid_disks;
1674                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1675                         break;
1676                 case ALGORITHM_PARITY_0:
1677                         pd_idx = 0;
1678                         (*dd_idx)++;
1679                         break;
1680                 case ALGORITHM_PARITY_N:
1681                         pd_idx = data_disks;
1682                         break;
1683                 default:
1684                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1685                                 algorithm);
1686                         BUG();
1687                 }
1688                 break;
1689         case 6:
1690
1691                 switch (algorithm) {
1692                 case ALGORITHM_LEFT_ASYMMETRIC:
1693                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1694                         qd_idx = pd_idx + 1;
1695                         if (pd_idx == raid_disks-1) {
1696                                 (*dd_idx)++;    /* Q D D D P */
1697                                 qd_idx = 0;
1698                         } else if (*dd_idx >= pd_idx)
1699                                 (*dd_idx) += 2; /* D D P Q D */
1700                         break;
1701                 case ALGORITHM_RIGHT_ASYMMETRIC:
1702                         pd_idx = stripe % raid_disks;
1703                         qd_idx = pd_idx + 1;
1704                         if (pd_idx == raid_disks-1) {
1705                                 (*dd_idx)++;    /* Q D D D P */
1706                                 qd_idx = 0;
1707                         } else if (*dd_idx >= pd_idx)
1708                                 (*dd_idx) += 2; /* D D P Q D */
1709                         break;
1710                 case ALGORITHM_LEFT_SYMMETRIC:
1711                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1712                         qd_idx = (pd_idx + 1) % raid_disks;
1713                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1714                         break;
1715                 case ALGORITHM_RIGHT_SYMMETRIC:
1716                         pd_idx = stripe % raid_disks;
1717                         qd_idx = (pd_idx + 1) % raid_disks;
1718                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1719                         break;
1720
1721                 case ALGORITHM_PARITY_0:
1722                         pd_idx = 0;
1723                         qd_idx = 1;
1724                         (*dd_idx) += 2;
1725                         break;
1726                 case ALGORITHM_PARITY_N:
1727                         pd_idx = data_disks;
1728                         qd_idx = data_disks + 1;
1729                         break;
1730
1731                 case ALGORITHM_ROTATING_ZERO_RESTART:
1732                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
1733                          * of blocks for computing Q is different.
1734                          */
1735                         pd_idx = stripe % raid_disks;
1736                         qd_idx = pd_idx + 1;
1737                         if (pd_idx == raid_disks-1) {
1738                                 (*dd_idx)++;    /* Q D D D P */
1739                                 qd_idx = 0;
1740                         } else if (*dd_idx >= pd_idx)
1741                                 (*dd_idx) += 2; /* D D P Q D */
1742                         ddf_layout = 1;
1743                         break;
1744
1745                 case ALGORITHM_ROTATING_N_RESTART:
1746                         /* Same a left_asymmetric, by first stripe is
1747                          * D D D P Q  rather than
1748                          * Q D D D P
1749                          */
1750                         pd_idx = raid_disks - 1 - ((stripe + 1) % raid_disks);
1751                         qd_idx = pd_idx + 1;
1752                         if (pd_idx == raid_disks-1) {
1753                                 (*dd_idx)++;    /* Q D D D P */
1754                                 qd_idx = 0;
1755                         } else if (*dd_idx >= pd_idx)
1756                                 (*dd_idx) += 2; /* D D P Q D */
1757                         ddf_layout = 1;
1758                         break;
1759
1760                 case ALGORITHM_ROTATING_N_CONTINUE:
1761                         /* Same as left_symmetric but Q is before P */
1762                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1763                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1764                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1765                         ddf_layout = 1;
1766                         break;
1767
1768                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1769                         /* RAID5 left_asymmetric, with Q on last device */
1770                         pd_idx = data_disks - stripe % (raid_disks-1);
1771                         if (*dd_idx >= pd_idx)
1772                                 (*dd_idx)++;
1773                         qd_idx = raid_disks - 1;
1774                         break;
1775
1776                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1777                         pd_idx = stripe % (raid_disks-1);
1778                         if (*dd_idx >= pd_idx)
1779                                 (*dd_idx)++;
1780                         qd_idx = raid_disks - 1;
1781                         break;
1782
1783                 case ALGORITHM_LEFT_SYMMETRIC_6:
1784                         pd_idx = data_disks - stripe % (raid_disks-1);
1785                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1786                         qd_idx = raid_disks - 1;
1787                         break;
1788
1789                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1790                         pd_idx = stripe % (raid_disks-1);
1791                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1792                         qd_idx = raid_disks - 1;
1793                         break;
1794
1795                 case ALGORITHM_PARITY_0_6:
1796                         pd_idx = 0;
1797                         (*dd_idx)++;
1798                         qd_idx = raid_disks - 1;
1799                         break;
1800
1801
1802                 default:
1803                         printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1804                                algorithm);
1805                         BUG();
1806                 }
1807                 break;
1808         }
1809
1810         if (sh) {
1811                 sh->pd_idx = pd_idx;
1812                 sh->qd_idx = qd_idx;
1813                 sh->ddf_layout = ddf_layout;
1814         }
1815         /*
1816          * Finally, compute the new sector number
1817          */
1818         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1819         return new_sector;
1820 }
1821
1822
1823 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1824 {
1825         raid5_conf_t *conf = sh->raid_conf;
1826         int raid_disks = sh->disks;
1827         int data_disks = raid_disks - conf->max_degraded;
1828         sector_t new_sector = sh->sector, check;
1829         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1830                                          : conf->chunk_sectors;
1831         int algorithm = previous ? conf->prev_algo
1832                                  : conf->algorithm;
1833         sector_t stripe;
1834         int chunk_offset;
1835         int chunk_number, dummy1, dd_idx = i;
1836         sector_t r_sector;
1837         struct stripe_head sh2;
1838
1839
1840         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1841         stripe = new_sector;
1842         BUG_ON(new_sector != stripe);
1843
1844         if (i == sh->pd_idx)
1845                 return 0;
1846         switch(conf->level) {
1847         case 4: break;
1848         case 5:
1849                 switch (algorithm) {
1850                 case ALGORITHM_LEFT_ASYMMETRIC:
1851                 case ALGORITHM_RIGHT_ASYMMETRIC:
1852                         if (i > sh->pd_idx)
1853                                 i--;
1854                         break;
1855                 case ALGORITHM_LEFT_SYMMETRIC:
1856                 case ALGORITHM_RIGHT_SYMMETRIC:
1857                         if (i < sh->pd_idx)
1858                                 i += raid_disks;
1859                         i -= (sh->pd_idx + 1);
1860                         break;
1861                 case ALGORITHM_PARITY_0:
1862                         i -= 1;
1863                         break;
1864                 case ALGORITHM_PARITY_N:
1865                         break;
1866                 default:
1867                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1868                                algorithm);
1869                         BUG();
1870                 }
1871                 break;
1872         case 6:
1873                 if (i == sh->qd_idx)
1874                         return 0; /* It is the Q disk */
1875                 switch (algorithm) {
1876                 case ALGORITHM_LEFT_ASYMMETRIC:
1877                 case ALGORITHM_RIGHT_ASYMMETRIC:
1878                 case ALGORITHM_ROTATING_ZERO_RESTART:
1879                 case ALGORITHM_ROTATING_N_RESTART:
1880                         if (sh->pd_idx == raid_disks-1)
1881                                 i--;    /* Q D D D P */
1882                         else if (i > sh->pd_idx)
1883                                 i -= 2; /* D D P Q D */
1884                         break;
1885                 case ALGORITHM_LEFT_SYMMETRIC:
1886                 case ALGORITHM_RIGHT_SYMMETRIC:
1887                         if (sh->pd_idx == raid_disks-1)
1888                                 i--; /* Q D D D P */
1889                         else {
1890                                 /* D D P Q D */
1891                                 if (i < sh->pd_idx)
1892                                         i += raid_disks;
1893                                 i -= (sh->pd_idx + 2);
1894                         }
1895                         break;
1896                 case ALGORITHM_PARITY_0:
1897                         i -= 2;
1898                         break;
1899                 case ALGORITHM_PARITY_N:
1900                         break;
1901                 case ALGORITHM_ROTATING_N_CONTINUE:
1902                         if (sh->pd_idx == 0)
1903                                 i--;    /* P D D D Q */
1904                         else if (i > sh->pd_idx)
1905                                 i -= 2; /* D D Q P D */
1906                         break;
1907                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1908                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1909                         if (i > sh->pd_idx)
1910                                 i--;
1911                         break;
1912                 case ALGORITHM_LEFT_SYMMETRIC_6:
1913                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1914                         if (i < sh->pd_idx)
1915                                 i += data_disks + 1;
1916                         i -= (sh->pd_idx + 1);
1917                         break;
1918                 case ALGORITHM_PARITY_0_6:
1919                         i -= 1;
1920                         break;
1921                 default:
1922                         printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1923                                algorithm);
1924                         BUG();
1925                 }
1926                 break;
1927         }
1928
1929         chunk_number = stripe * data_disks + i;
1930         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1931
1932         check = raid5_compute_sector(conf, r_sector,
1933                                      previous, &dummy1, &sh2);
1934         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
1935                 || sh2.qd_idx != sh->qd_idx) {
1936                 printk(KERN_ERR "compute_blocknr: map not correct\n");
1937                 return 0;
1938         }
1939         return r_sector;
1940 }
1941
1942
1943 static void
1944 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
1945                          int rcw, int expand)
1946 {
1947         int i, pd_idx = sh->pd_idx, disks = sh->disks;
1948         raid5_conf_t *conf = sh->raid_conf;
1949         int level = conf->level;
1950
1951         if (rcw) {
1952                 /* if we are not expanding this is a proper write request, and
1953                  * there will be bios with new data to be drained into the
1954                  * stripe cache
1955                  */
1956                 if (!expand) {
1957                         sh->reconstruct_state = reconstruct_state_drain_run;
1958                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1959                 } else
1960                         sh->reconstruct_state = reconstruct_state_run;
1961
1962                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
1963
1964                 for (i = disks; i--; ) {
1965                         struct r5dev *dev = &sh->dev[i];
1966
1967                         if (dev->towrite) {
1968                                 set_bit(R5_LOCKED, &dev->flags);
1969                                 set_bit(R5_Wantdrain, &dev->flags);
1970                                 if (!expand)
1971                                         clear_bit(R5_UPTODATE, &dev->flags);
1972                                 s->locked++;
1973                         }
1974                 }
1975                 if (s->locked + conf->max_degraded == disks)
1976                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1977                                 atomic_inc(&conf->pending_full_writes);
1978         } else {
1979                 BUG_ON(level == 6);
1980                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1981                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1982
1983                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
1984                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
1985                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1986                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
1987
1988                 for (i = disks; i--; ) {
1989                         struct r5dev *dev = &sh->dev[i];
1990                         if (i == pd_idx)
1991                                 continue;
1992
1993                         if (dev->towrite &&
1994                             (test_bit(R5_UPTODATE, &dev->flags) ||
1995                              test_bit(R5_Wantcompute, &dev->flags))) {
1996                                 set_bit(R5_Wantdrain, &dev->flags);
1997                                 set_bit(R5_LOCKED, &dev->flags);
1998                                 clear_bit(R5_UPTODATE, &dev->flags);
1999                                 s->locked++;
2000                         }
2001                 }
2002         }
2003
2004         /* keep the parity disk(s) locked while asynchronous operations
2005          * are in flight
2006          */
2007         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2008         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2009         s->locked++;
2010
2011         if (level == 6) {
2012                 int qd_idx = sh->qd_idx;
2013                 struct r5dev *dev = &sh->dev[qd_idx];
2014
2015                 set_bit(R5_LOCKED, &dev->flags);
2016                 clear_bit(R5_UPTODATE, &dev->flags);
2017                 s->locked++;
2018         }
2019
2020         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2021                 __func__, (unsigned long long)sh->sector,
2022                 s->locked, s->ops_request);
2023 }
2024
2025 /*
2026  * Each stripe/dev can have one or more bion attached.
2027  * toread/towrite point to the first in a chain.
2028  * The bi_next chain must be in order.
2029  */
2030 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2031 {
2032         struct bio **bip;
2033         raid5_conf_t *conf = sh->raid_conf;
2034         int firstwrite=0;
2035
2036         pr_debug("adding bh b#%llu to stripe s#%llu\n",
2037                 (unsigned long long)bi->bi_sector,
2038                 (unsigned long long)sh->sector);
2039
2040
2041         spin_lock(&sh->lock);
2042         spin_lock_irq(&conf->device_lock);
2043         if (forwrite) {
2044                 bip = &sh->dev[dd_idx].towrite;
2045                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2046                         firstwrite = 1;
2047         } else
2048                 bip = &sh->dev[dd_idx].toread;
2049         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2050                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2051                         goto overlap;
2052                 bip = & (*bip)->bi_next;
2053         }
2054         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2055                 goto overlap;
2056
2057         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2058         if (*bip)
2059                 bi->bi_next = *bip;
2060         *bip = bi;
2061         bi->bi_phys_segments++;
2062         spin_unlock_irq(&conf->device_lock);
2063         spin_unlock(&sh->lock);
2064
2065         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2066                 (unsigned long long)bi->bi_sector,
2067                 (unsigned long long)sh->sector, dd_idx);
2068
2069         if (conf->mddev->bitmap && firstwrite) {
2070                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2071                                   STRIPE_SECTORS, 0);
2072                 sh->bm_seq = conf->seq_flush+1;
2073                 set_bit(STRIPE_BIT_DELAY, &sh->state);
2074         }
2075
2076         if (forwrite) {
2077                 /* check if page is covered */
2078                 sector_t sector = sh->dev[dd_idx].sector;
2079                 for (bi=sh->dev[dd_idx].towrite;
2080                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2081                              bi && bi->bi_sector <= sector;
2082                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2083                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2084                                 sector = bi->bi_sector + (bi->bi_size>>9);
2085                 }
2086                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2087                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2088         }
2089         return 1;
2090
2091  overlap:
2092         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2093         spin_unlock_irq(&conf->device_lock);
2094         spin_unlock(&sh->lock);
2095         return 0;
2096 }
2097
2098 static void end_reshape(raid5_conf_t *conf);
2099
2100 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2101                             struct stripe_head *sh)
2102 {
2103         int sectors_per_chunk =
2104                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2105         int dd_idx;
2106         int chunk_offset = sector_div(stripe, sectors_per_chunk);
2107         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2108
2109         raid5_compute_sector(conf,
2110                              stripe * (disks - conf->max_degraded)
2111                              *sectors_per_chunk + chunk_offset,
2112                              previous,
2113                              &dd_idx, sh);
2114 }
2115
2116 static void
2117 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2118                                 struct stripe_head_state *s, int disks,
2119                                 struct bio **return_bi)
2120 {
2121         int i;
2122         for (i = disks; i--; ) {
2123                 struct bio *bi;
2124                 int bitmap_end = 0;
2125
2126                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2127                         mdk_rdev_t *rdev;
2128                         rcu_read_lock();
2129                         rdev = rcu_dereference(conf->disks[i].rdev);
2130                         if (rdev && test_bit(In_sync, &rdev->flags))
2131                                 /* multiple read failures in one stripe */
2132                                 md_error(conf->mddev, rdev);
2133                         rcu_read_unlock();
2134                 }
2135                 spin_lock_irq(&conf->device_lock);
2136                 /* fail all writes first */
2137                 bi = sh->dev[i].towrite;
2138                 sh->dev[i].towrite = NULL;
2139                 if (bi) {
2140                         s->to_write--;
2141                         bitmap_end = 1;
2142                 }
2143
2144                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2145                         wake_up(&conf->wait_for_overlap);
2146
2147                 while (bi && bi->bi_sector <
2148                         sh->dev[i].sector + STRIPE_SECTORS) {
2149                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2150                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2151                         if (!raid5_dec_bi_phys_segments(bi)) {
2152                                 md_write_end(conf->mddev);
2153                                 bi->bi_next = *return_bi;
2154                                 *return_bi = bi;
2155                         }
2156                         bi = nextbi;
2157                 }
2158                 /* and fail all 'written' */
2159                 bi = sh->dev[i].written;
2160                 sh->dev[i].written = NULL;
2161                 if (bi) bitmap_end = 1;
2162                 while (bi && bi->bi_sector <
2163                        sh->dev[i].sector + STRIPE_SECTORS) {
2164                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2165                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2166                         if (!raid5_dec_bi_phys_segments(bi)) {
2167                                 md_write_end(conf->mddev);
2168                                 bi->bi_next = *return_bi;
2169                                 *return_bi = bi;
2170                         }
2171                         bi = bi2;
2172                 }
2173
2174                 /* fail any reads if this device is non-operational and
2175                  * the data has not reached the cache yet.
2176                  */
2177                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2178                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2179                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
2180                         bi = sh->dev[i].toread;
2181                         sh->dev[i].toread = NULL;
2182                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2183                                 wake_up(&conf->wait_for_overlap);
2184                         if (bi) s->to_read--;
2185                         while (bi && bi->bi_sector <
2186                                sh->dev[i].sector + STRIPE_SECTORS) {
2187                                 struct bio *nextbi =
2188                                         r5_next_bio(bi, sh->dev[i].sector);
2189                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2190                                 if (!raid5_dec_bi_phys_segments(bi)) {
2191                                         bi->bi_next = *return_bi;
2192                                         *return_bi = bi;
2193                                 }
2194                                 bi = nextbi;
2195                         }
2196                 }
2197                 spin_unlock_irq(&conf->device_lock);
2198                 if (bitmap_end)
2199                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2200                                         STRIPE_SECTORS, 0, 0);
2201         }
2202
2203         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2204                 if (atomic_dec_and_test(&conf->pending_full_writes))
2205                         md_wakeup_thread(conf->mddev->thread);
2206 }
2207
2208 /* fetch_block5 - checks the given member device to see if its data needs
2209  * to be read or computed to satisfy a request.
2210  *
2211  * Returns 1 when no more member devices need to be checked, otherwise returns
2212  * 0 to tell the loop in handle_stripe_fill5 to continue
2213  */
2214 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2215                         int disk_idx, int disks)
2216 {
2217         struct r5dev *dev = &sh->dev[disk_idx];
2218         struct r5dev *failed_dev = &sh->dev[s->failed_num];
2219
2220         /* is the data in this block needed, and can we get it? */
2221         if (!test_bit(R5_LOCKED, &dev->flags) &&
2222             !test_bit(R5_UPTODATE, &dev->flags) &&
2223             (dev->toread ||
2224              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2225              s->syncing || s->expanding ||
2226              (s->failed &&
2227               (failed_dev->toread ||
2228                (failed_dev->towrite &&
2229                 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2230                 /* We would like to get this block, possibly by computing it,
2231                  * otherwise read it if the backing disk is insync
2232                  */
2233                 if ((s->uptodate == disks - 1) &&
2234                     (s->failed && disk_idx == s->failed_num)) {
2235                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2236                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2237                         set_bit(R5_Wantcompute, &dev->flags);
2238                         sh->ops.target = disk_idx;
2239                         sh->ops.target2 = -1;
2240                         s->req_compute = 1;
2241                         /* Careful: from this point on 'uptodate' is in the eye
2242                          * of raid_run_ops which services 'compute' operations
2243                          * before writes. R5_Wantcompute flags a block that will
2244                          * be R5_UPTODATE by the time it is needed for a
2245                          * subsequent operation.
2246                          */
2247                         s->uptodate++;
2248                         return 1; /* uptodate + compute == disks */
2249                 } else if (test_bit(R5_Insync, &dev->flags)) {
2250                         set_bit(R5_LOCKED, &dev->flags);
2251                         set_bit(R5_Wantread, &dev->flags);
2252                         s->locked++;
2253                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2254                                 s->syncing);
2255                 }
2256         }
2257
2258         return 0;
2259 }
2260
2261 /**
2262  * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2263  */
2264 static void handle_stripe_fill5(struct stripe_head *sh,
2265                         struct stripe_head_state *s, int disks)
2266 {
2267         int i;
2268
2269         /* look for blocks to read/compute, skip this if a compute
2270          * is already in flight, or if the stripe contents are in the
2271          * midst of changing due to a write
2272          */
2273         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2274             !sh->reconstruct_state)
2275                 for (i = disks; i--; )
2276                         if (fetch_block5(sh, s, i, disks))
2277                                 break;
2278         set_bit(STRIPE_HANDLE, &sh->state);
2279 }
2280
2281 /* fetch_block6 - checks the given member device to see if its data needs
2282  * to be read or computed to satisfy a request.
2283  *
2284  * Returns 1 when no more member devices need to be checked, otherwise returns
2285  * 0 to tell the loop in handle_stripe_fill6 to continue
2286  */
2287 static int fetch_block6(struct stripe_head *sh, struct stripe_head_state *s,
2288                          struct r6_state *r6s, int disk_idx, int disks)
2289 {
2290         struct r5dev *dev = &sh->dev[disk_idx];
2291         struct r5dev *fdev[2] = { &sh->dev[r6s->failed_num[0]],
2292                                   &sh->dev[r6s->failed_num[1]] };
2293
2294         if (!test_bit(R5_LOCKED, &dev->flags) &&
2295             !test_bit(R5_UPTODATE, &dev->flags) &&
2296             (dev->toread ||
2297              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2298              s->syncing || s->expanding ||
2299              (s->failed >= 1 &&
2300               (fdev[0]->toread || s->to_write)) ||
2301              (s->failed >= 2 &&
2302               (fdev[1]->toread || s->to_write)))) {
2303                 /* we would like to get this block, possibly by computing it,
2304                  * otherwise read it if the backing disk is insync
2305                  */
2306                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2307                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2308                 if ((s->uptodate == disks - 1) &&
2309                     (s->failed && (disk_idx == r6s->failed_num[0] ||
2310                                    disk_idx == r6s->failed_num[1]))) {
2311                         /* have disk failed, and we're requested to fetch it;
2312                          * do compute it
2313                          */
2314                         pr_debug("Computing stripe %llu block %d\n",
2315                                (unsigned long long)sh->sector, disk_idx);
2316                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2317                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2318                         set_bit(R5_Wantcompute, &dev->flags);
2319                         sh->ops.target = disk_idx;
2320                         sh->ops.target2 = -1; /* no 2nd target */
2321                         s->req_compute = 1;
2322                         s->uptodate++;
2323                         return 1;
2324                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2325                         /* Computing 2-failure is *very* expensive; only
2326                          * do it if failed >= 2
2327                          */
2328                         int other;
2329                         for (other = disks; other--; ) {
2330                                 if (other == disk_idx)
2331                                         continue;
2332                                 if (!test_bit(R5_UPTODATE,
2333                                       &sh->dev[other].flags))
2334                                         break;
2335                         }
2336                         BUG_ON(other < 0);
2337                         pr_debug("Computing stripe %llu blocks %d,%d\n",
2338                                (unsigned long long)sh->sector,
2339                                disk_idx, other);
2340                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2341                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2342                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2343                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
2344                         sh->ops.target = disk_idx;
2345                         sh->ops.target2 = other;
2346                         s->uptodate += 2;
2347                         s->req_compute = 1;
2348                         return 1;
2349                 } else if (test_bit(R5_Insync, &dev->flags)) {
2350                         set_bit(R5_LOCKED, &dev->flags);
2351                         set_bit(R5_Wantread, &dev->flags);
2352                         s->locked++;
2353                         pr_debug("Reading block %d (sync=%d)\n",
2354                                 disk_idx, s->syncing);
2355                 }
2356         }
2357
2358         return 0;
2359 }
2360
2361 /**
2362  * handle_stripe_fill6 - read or compute data to satisfy pending requests.
2363  */
2364 static void handle_stripe_fill6(struct stripe_head *sh,
2365                         struct stripe_head_state *s, struct r6_state *r6s,
2366                         int disks)
2367 {
2368         int i;
2369
2370         /* look for blocks to read/compute, skip this if a compute
2371          * is already in flight, or if the stripe contents are in the
2372          * midst of changing due to a write
2373          */
2374         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2375             !sh->reconstruct_state)
2376                 for (i = disks; i--; )
2377                         if (fetch_block6(sh, s, r6s, i, disks))
2378                                 break;
2379         set_bit(STRIPE_HANDLE, &sh->state);
2380 }
2381
2382
2383 /* handle_stripe_clean_event
2384  * any written block on an uptodate or failed drive can be returned.
2385  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2386  * never LOCKED, so we don't need to test 'failed' directly.
2387  */
2388 static void handle_stripe_clean_event(raid5_conf_t *conf,
2389         struct stripe_head *sh, int disks, struct bio **return_bi)
2390 {
2391         int i;
2392         struct r5dev *dev;
2393
2394         for (i = disks; i--; )
2395                 if (sh->dev[i].written) {
2396                         dev = &sh->dev[i];
2397                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2398                                 test_bit(R5_UPTODATE, &dev->flags)) {
2399                                 /* We can return any write requests */
2400                                 struct bio *wbi, *wbi2;
2401                                 int bitmap_end = 0;
2402                                 pr_debug("Return write for disc %d\n", i);
2403                                 spin_lock_irq(&conf->device_lock);
2404                                 wbi = dev->written;
2405                                 dev->written = NULL;
2406                                 while (wbi && wbi->bi_sector <
2407                                         dev->sector + STRIPE_SECTORS) {
2408                                         wbi2 = r5_next_bio(wbi, dev->sector);
2409                                         if (!raid5_dec_bi_phys_segments(wbi)) {
2410                                                 md_write_end(conf->mddev);
2411                                                 wbi->bi_next = *return_bi;
2412                                                 *return_bi = wbi;
2413                                         }
2414                                         wbi = wbi2;
2415                                 }
2416                                 if (dev->towrite == NULL)
2417                                         bitmap_end = 1;
2418                                 spin_unlock_irq(&conf->device_lock);
2419                                 if (bitmap_end)
2420                                         bitmap_endwrite(conf->mddev->bitmap,
2421                                                         sh->sector,
2422                                                         STRIPE_SECTORS,
2423                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2424                                                         0);
2425                         }
2426                 }
2427
2428         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2429                 if (atomic_dec_and_test(&conf->pending_full_writes))
2430                         md_wakeup_thread(conf->mddev->thread);
2431 }
2432
2433 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2434                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2435 {
2436         int rmw = 0, rcw = 0, i;
2437         for (i = disks; i--; ) {
2438                 /* would I have to read this buffer for read_modify_write */
2439                 struct r5dev *dev = &sh->dev[i];
2440                 if ((dev->towrite || i == sh->pd_idx) &&
2441                     !test_bit(R5_LOCKED, &dev->flags) &&
2442                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2443                       test_bit(R5_Wantcompute, &dev->flags))) {
2444                         if (test_bit(R5_Insync, &dev->flags))
2445                                 rmw++;
2446                         else
2447                                 rmw += 2*disks;  /* cannot read it */
2448                 }
2449                 /* Would I have to read this buffer for reconstruct_write */
2450                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2451                     !test_bit(R5_LOCKED, &dev->flags) &&
2452                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2453                     test_bit(R5_Wantcompute, &dev->flags))) {
2454                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2455                         else
2456                                 rcw += 2*disks;
2457                 }
2458         }
2459         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2460                 (unsigned long long)sh->sector, rmw, rcw);
2461         set_bit(STRIPE_HANDLE, &sh->state);
2462         if (rmw < rcw && rmw > 0)
2463                 /* prefer read-modify-write, but need to get some data */
2464                 for (i = disks; i--; ) {
2465                         struct r5dev *dev = &sh->dev[i];
2466                         if ((dev->towrite || i == sh->pd_idx) &&
2467                             !test_bit(R5_LOCKED, &dev->flags) &&
2468                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2469                             test_bit(R5_Wantcompute, &dev->flags)) &&
2470                             test_bit(R5_Insync, &dev->flags)) {
2471                                 if (
2472                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2473                                         pr_debug("Read_old block "
2474                                                 "%d for r-m-w\n", i);
2475                                         set_bit(R5_LOCKED, &dev->flags);
2476                                         set_bit(R5_Wantread, &dev->flags);
2477                                         s->locked++;
2478                                 } else {
2479                                         set_bit(STRIPE_DELAYED, &sh->state);
2480                                         set_bit(STRIPE_HANDLE, &sh->state);
2481                                 }
2482                         }
2483                 }
2484         if (rcw <= rmw && rcw > 0)
2485                 /* want reconstruct write, but need to get some data */
2486                 for (i = disks; i--; ) {
2487                         struct r5dev *dev = &sh->dev[i];
2488                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2489                             i != sh->pd_idx &&
2490                             !test_bit(R5_LOCKED, &dev->flags) &&
2491                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2492                             test_bit(R5_Wantcompute, &dev->flags)) &&
2493                             test_bit(R5_Insync, &dev->flags)) {
2494                                 if (
2495                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2496                                         pr_debug("Read_old block "
2497                                                 "%d for Reconstruct\n", i);
2498                                         set_bit(R5_LOCKED, &dev->flags);
2499                                         set_bit(R5_Wantread, &dev->flags);
2500                                         s->locked++;
2501                                 } else {
2502                                         set_bit(STRIPE_DELAYED, &sh->state);
2503                                         set_bit(STRIPE_HANDLE, &sh->state);
2504                                 }
2505                         }
2506                 }
2507         /* now if nothing is locked, and if we have enough data,
2508          * we can start a write request
2509          */
2510         /* since handle_stripe can be called at any time we need to handle the
2511          * case where a compute block operation has been submitted and then a
2512          * subsequent call wants to start a write request.  raid_run_ops only
2513          * handles the case where compute block and reconstruct are requested
2514          * simultaneously.  If this is not the case then new writes need to be
2515          * held off until the compute completes.
2516          */
2517         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2518             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2519             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2520                 schedule_reconstruction(sh, s, rcw == 0, 0);
2521 }
2522
2523 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2524                 struct stripe_head *sh, struct stripe_head_state *s,
2525                 struct r6_state *r6s, int disks)
2526 {
2527         int rcw = 0, pd_idx = sh->pd_idx, i;
2528         int qd_idx = sh->qd_idx;
2529
2530         set_bit(STRIPE_HANDLE, &sh->state);
2531         for (i = disks; i--; ) {
2532                 struct r5dev *dev = &sh->dev[i];
2533                 /* check if we haven't enough data */
2534                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2535                     i != pd_idx && i != qd_idx &&
2536                     !test_bit(R5_LOCKED, &dev->flags) &&
2537                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2538                       test_bit(R5_Wantcompute, &dev->flags))) {
2539                         rcw++;
2540                         if (!test_bit(R5_Insync, &dev->flags))
2541                                 continue; /* it's a failed drive */
2542
2543                         if (
2544                           test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2545                                 pr_debug("Read_old stripe %llu "
2546                                         "block %d for Reconstruct\n",
2547                                      (unsigned long long)sh->sector, i);
2548                                 set_bit(R5_LOCKED, &dev->flags);
2549                                 set_bit(R5_Wantread, &dev->flags);
2550                                 s->locked++;
2551                         } else {
2552                                 pr_debug("Request delayed stripe %llu "
2553                                         "block %d for Reconstruct\n",
2554                                      (unsigned long long)sh->sector, i);
2555                                 set_bit(STRIPE_DELAYED, &sh->state);
2556                                 set_bit(STRIPE_HANDLE, &sh->state);
2557                         }
2558                 }
2559         }
2560         /* now if nothing is locked, and if we have enough data, we can start a
2561          * write request
2562          */
2563         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2564             s->locked == 0 && rcw == 0 &&
2565             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2566                 schedule_reconstruction(sh, s, 1, 0);
2567         }
2568 }
2569
2570 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2571                                 struct stripe_head_state *s, int disks)
2572 {
2573         struct r5dev *dev = NULL;
2574
2575         set_bit(STRIPE_HANDLE, &sh->state);
2576
2577         switch (sh->check_state) {
2578         case check_state_idle:
2579                 /* start a new check operation if there are no failures */
2580                 if (s->failed == 0) {
2581                         BUG_ON(s->uptodate != disks);
2582                         sh->check_state = check_state_run;
2583                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2584                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2585                         s->uptodate--;
2586                         break;
2587                 }
2588                 dev = &sh->dev[s->failed_num];
2589                 /* fall through */
2590         case check_state_compute_result:
2591                 sh->check_state = check_state_idle;
2592                 if (!dev)
2593                         dev = &sh->dev[sh->pd_idx];
2594
2595                 /* check that a write has not made the stripe insync */
2596                 if (test_bit(STRIPE_INSYNC, &sh->state))
2597                         break;
2598
2599                 /* either failed parity check, or recovery is happening */
2600                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2601                 BUG_ON(s->uptodate != disks);
2602
2603                 set_bit(R5_LOCKED, &dev->flags);
2604                 s->locked++;
2605                 set_bit(R5_Wantwrite, &dev->flags);
2606
2607                 clear_bit(STRIPE_DEGRADED, &sh->state);
2608                 set_bit(STRIPE_INSYNC, &sh->state);
2609                 break;
2610         case check_state_run:
2611                 break; /* we will be called again upon completion */
2612         case check_state_check_result:
2613                 sh->check_state = check_state_idle;
2614
2615                 /* if a failure occurred during the check operation, leave
2616                  * STRIPE_INSYNC not set and let the stripe be handled again
2617                  */
2618                 if (s->failed)
2619                         break;
2620
2621                 /* handle a successful check operation, if parity is correct
2622                  * we are done.  Otherwise update the mismatch count and repair
2623                  * parity if !MD_RECOVERY_CHECK
2624                  */
2625                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2626                         /* parity is correct (on disc,
2627                          * not in buffer any more)
2628                          */
2629                         set_bit(STRIPE_INSYNC, &sh->state);
2630                 else {
2631                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2632                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2633                                 /* don't try to repair!! */
2634                                 set_bit(STRIPE_INSYNC, &sh->state);
2635                         else {
2636                                 sh->check_state = check_state_compute_run;
2637                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2638                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2639                                 set_bit(R5_Wantcompute,
2640                                         &sh->dev[sh->pd_idx].flags);
2641                                 sh->ops.target = sh->pd_idx;
2642                                 sh->ops.target2 = -1;
2643                                 s->uptodate++;
2644                         }
2645                 }
2646                 break;
2647         case check_state_compute_run:
2648                 break;
2649         default:
2650                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2651                        __func__, sh->check_state,
2652                        (unsigned long long) sh->sector);
2653                 BUG();
2654         }
2655 }
2656
2657
2658 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2659                                   struct stripe_head_state *s,
2660                                   struct r6_state *r6s, int disks)
2661 {
2662         int pd_idx = sh->pd_idx;
2663         int qd_idx = sh->qd_idx;
2664         struct r5dev *dev;
2665
2666         set_bit(STRIPE_HANDLE, &sh->state);
2667
2668         BUG_ON(s->failed > 2);
2669
2670         /* Want to check and possibly repair P and Q.
2671          * However there could be one 'failed' device, in which
2672          * case we can only check one of them, possibly using the
2673          * other to generate missing data
2674          */
2675
2676         switch (sh->check_state) {
2677         case check_state_idle:
2678                 /* start a new check operation if there are < 2 failures */
2679                 if (s->failed == r6s->q_failed) {
2680                         /* The only possible failed device holds Q, so it
2681                          * makes sense to check P (If anything else were failed,
2682                          * we would have used P to recreate it).
2683                          */
2684                         sh->check_state = check_state_run;
2685                 }
2686                 if (!r6s->q_failed && s->failed < 2) {
2687                         /* Q is not failed, and we didn't use it to generate
2688                          * anything, so it makes sense to check it
2689                          */
2690                         if (sh->check_state == check_state_run)
2691                                 sh->check_state = check_state_run_pq;
2692                         else
2693                                 sh->check_state = check_state_run_q;
2694                 }
2695
2696                 /* discard potentially stale zero_sum_result */
2697                 sh->ops.zero_sum_result = 0;
2698
2699                 if (sh->check_state == check_state_run) {
2700                         /* async_xor_zero_sum destroys the contents of P */
2701                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2702                         s->uptodate--;
2703                 }
2704                 if (sh->check_state >= check_state_run &&
2705                     sh->check_state <= check_state_run_pq) {
2706                         /* async_syndrome_zero_sum preserves P and Q, so
2707                          * no need to mark them !uptodate here
2708                          */
2709                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2710                         break;
2711                 }
2712
2713                 /* we have 2-disk failure */
2714                 BUG_ON(s->failed != 2);
2715                 /* fall through */
2716         case check_state_compute_result:
2717                 sh->check_state = check_state_idle;
2718
2719                 /* check that a write has not made the stripe insync */
2720                 if (test_bit(STRIPE_INSYNC, &sh->state))
2721                         break;
2722
2723                 /* now write out any block on a failed drive,
2724                  * or P or Q if they were recomputed
2725                  */
2726                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2727                 if (s->failed == 2) {
2728                         dev = &sh->dev[r6s->failed_num[1]];
2729                         s->locked++;
2730                         set_bit(R5_LOCKED, &dev->flags);
2731                         set_bit(R5_Wantwrite, &dev->flags);
2732                 }
2733                 if (s->failed >= 1) {
2734                         dev = &sh->dev[r6s->failed_num[0]];
2735                         s->locked++;
2736                         set_bit(R5_LOCKED, &dev->flags);
2737                         set_bit(R5_Wantwrite, &dev->flags);
2738                 }
2739                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2740                         dev = &sh->dev[pd_idx];
2741                         s->locked++;
2742                         set_bit(R5_LOCKED, &dev->flags);
2743                         set_bit(R5_Wantwrite, &dev->flags);
2744                 }
2745                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2746                         dev = &sh->dev[qd_idx];
2747                         s->locked++;
2748                         set_bit(R5_LOCKED, &dev->flags);
2749                         set_bit(R5_Wantwrite, &dev->flags);
2750                 }
2751                 clear_bit(STRIPE_DEGRADED, &sh->state);
2752
2753                 set_bit(STRIPE_INSYNC, &sh->state);
2754                 break;
2755         case check_state_run:
2756         case check_state_run_q:
2757         case check_state_run_pq:
2758                 break; /* we will be called again upon completion */
2759         case check_state_check_result:
2760                 sh->check_state = check_state_idle;
2761
2762                 /* handle a successful check operation, if parity is correct
2763                  * we are done.  Otherwise update the mismatch count and repair
2764                  * parity if !MD_RECOVERY_CHECK
2765                  */
2766                 if (sh->ops.zero_sum_result == 0) {
2767                         /* both parities are correct */
2768                         if (!s->failed)
2769                                 set_bit(STRIPE_INSYNC, &sh->state);
2770                         else {
2771                                 /* in contrast to the raid5 case we can validate
2772                                  * parity, but still have a failure to write
2773                                  * back
2774                                  */
2775                                 sh->check_state = check_state_compute_result;
2776                                 /* Returning at this point means that we may go
2777                                  * off and bring p and/or q uptodate again so
2778                                  * we make sure to check zero_sum_result again
2779                                  * to verify if p or q need writeback
2780                                  */
2781                         }
2782                 } else {
2783                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2784                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2785                                 /* don't try to repair!! */
2786                                 set_bit(STRIPE_INSYNC, &sh->state);
2787                         else {
2788                                 int *target = &sh->ops.target;
2789
2790                                 sh->ops.target = -1;
2791                                 sh->ops.target2 = -1;
2792                                 sh->check_state = check_state_compute_run;
2793                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2794                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2795                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2796                                         set_bit(R5_Wantcompute,
2797                                                 &sh->dev[pd_idx].flags);
2798                                         *target = pd_idx;
2799                                         target = &sh->ops.target2;
2800                                         s->uptodate++;
2801                                 }
2802                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2803                                         set_bit(R5_Wantcompute,
2804                                                 &sh->dev[qd_idx].flags);
2805                                         *target = qd_idx;
2806                                         s->uptodate++;
2807                                 }
2808                         }
2809                 }
2810                 break;
2811         case check_state_compute_run:
2812                 break;
2813         default:
2814                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2815                        __func__, sh->check_state,
2816                        (unsigned long long) sh->sector);
2817                 BUG();
2818         }
2819 }
2820
2821 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2822                                 struct r6_state *r6s)
2823 {
2824         int i;
2825
2826         /* We have read all the blocks in this stripe and now we need to
2827          * copy some of them into a target stripe for expand.
2828          */
2829         struct dma_async_tx_descriptor *tx = NULL;
2830         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2831         for (i = 0; i < sh->disks; i++)
2832                 if (i != sh->pd_idx && i != sh->qd_idx) {
2833                         int dd_idx, j;
2834                         struct stripe_head *sh2;
2835                         struct async_submit_ctl submit;
2836
2837                         sector_t bn = compute_blocknr(sh, i, 1);
2838                         sector_t s = raid5_compute_sector(conf, bn, 0,
2839                                                           &dd_idx, NULL);
2840                         sh2 = get_active_stripe(conf, s, 0, 1, 1);
2841                         if (sh2 == NULL)
2842                                 /* so far only the early blocks of this stripe
2843                                  * have been requested.  When later blocks
2844                                  * get requested, we will try again
2845                                  */
2846                                 continue;
2847                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2848                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2849                                 /* must have already done this block */
2850                                 release_stripe(sh2);
2851                                 continue;
2852                         }
2853
2854                         /* place all the copies on one channel */
2855                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2856                         tx = async_memcpy(sh2->dev[dd_idx].page,
2857                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
2858                                           &submit);
2859
2860                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2861                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2862                         for (j = 0; j < conf->raid_disks; j++)
2863                                 if (j != sh2->pd_idx &&
2864                                     (!r6s || j != sh2->qd_idx) &&
2865                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2866                                         break;
2867                         if (j == conf->raid_disks) {
2868                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2869                                 set_bit(STRIPE_HANDLE, &sh2->state);
2870                         }
2871                         release_stripe(sh2);
2872
2873                 }
2874         /* done submitting copies, wait for them to complete */
2875         if (tx) {
2876                 async_tx_ack(tx);
2877                 dma_wait_for_async_tx(tx);
2878         }
2879 }
2880
2881
2882 /*
2883  * handle_stripe - do things to a stripe.
2884  *
2885  * We lock the stripe and then examine the state of various bits
2886  * to see what needs to be done.
2887  * Possible results:
2888  *    return some read request which now have data
2889  *    return some write requests which are safely on disc
2890  *    schedule a read on some buffers
2891  *    schedule a write of some buffers
2892  *    return confirmation of parity correctness
2893  *
2894  * buffers are taken off read_list or write_list, and bh_cache buffers
2895  * get BH_Lock set before the stripe lock is released.
2896  *
2897  */
2898
2899 static bool handle_stripe5(struct stripe_head *sh)
2900 {
2901         raid5_conf_t *conf = sh->raid_conf;
2902         int disks = sh->disks, i;
2903         struct bio *return_bi = NULL;
2904         struct stripe_head_state s;
2905         struct r5dev *dev;
2906         mdk_rdev_t *blocked_rdev = NULL;
2907         int prexor;
2908
2909         memset(&s, 0, sizeof(s));
2910         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2911                  "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2912                  atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2913                  sh->reconstruct_state);
2914
2915         spin_lock(&sh->lock);
2916         clear_bit(STRIPE_HANDLE, &sh->state);
2917         clear_bit(STRIPE_DELAYED, &sh->state);
2918
2919         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2920         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2921         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2922
2923         /* Now to look around and see what can be done */
2924         rcu_read_lock();
2925         for (i=disks; i--; ) {
2926                 mdk_rdev_t *rdev;
2927                 struct r5dev *dev = &sh->dev[i];
2928                 clear_bit(R5_Insync, &dev->flags);
2929
2930                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2931                         "written %p\n", i, dev->flags, dev->toread, dev->read,
2932                         dev->towrite, dev->written);
2933
2934                 /* maybe we can request a biofill operation
2935                  *
2936                  * new wantfill requests are only permitted while
2937                  * ops_complete_biofill is guaranteed to be inactive
2938                  */
2939                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2940                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2941                         set_bit(R5_Wantfill, &dev->flags);
2942
2943                 /* now count some things */
2944                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2945                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2946                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2947
2948                 if (test_bit(R5_Wantfill, &dev->flags))
2949                         s.to_fill++;
2950                 else if (dev->toread)
2951                         s.to_read++;
2952                 if (dev->towrite) {
2953                         s.to_write++;
2954                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2955                                 s.non_overwrite++;
2956                 }
2957                 if (dev->written)
2958                         s.written++;
2959                 rdev = rcu_dereference(conf->disks[i].rdev);
2960                 if (blocked_rdev == NULL &&
2961                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2962                         blocked_rdev = rdev;
2963                         atomic_inc(&rdev->nr_pending);
2964                 }
2965                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2966                         /* The ReadError flag will just be confusing now */
2967                         clear_bit(R5_ReadError, &dev->flags);
2968                         clear_bit(R5_ReWrite, &dev->flags);
2969                 }
2970                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2971                     || test_bit(R5_ReadError, &dev->flags)) {
2972                         s.failed++;
2973                         s.failed_num = i;
2974                 } else
2975                         set_bit(R5_Insync, &dev->flags);
2976         }
2977         rcu_read_unlock();
2978
2979         if (unlikely(blocked_rdev)) {
2980                 if (s.syncing || s.expanding || s.expanded ||
2981                     s.to_write || s.written) {
2982                         set_bit(STRIPE_HANDLE, &sh->state);
2983                         goto unlock;
2984                 }
2985                 /* There is nothing for the blocked_rdev to block */
2986                 rdev_dec_pending(blocked_rdev, conf->mddev);
2987                 blocked_rdev = NULL;
2988         }
2989
2990         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
2991                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
2992                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
2993         }
2994
2995         pr_debug("locked=%d uptodate=%d to_read=%d"
2996                 " to_write=%d failed=%d failed_num=%d\n",
2997                 s.locked, s.uptodate, s.to_read, s.to_write,
2998                 s.failed, s.failed_num);
2999         /* check if the array has lost two devices and, if so, some requests might
3000          * need to be failed
3001          */
3002         if (s.failed > 1 && s.to_read+s.to_write+s.written)
3003                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3004         if (s.failed > 1 && s.syncing) {
3005                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3006                 clear_bit(STRIPE_SYNCING, &sh->state);
3007                 s.syncing = 0;
3008         }
3009
3010         /* might be able to return some write requests if the parity block
3011          * is safe, or on a failed drive
3012          */
3013         dev = &sh->dev[sh->pd_idx];
3014         if ( s.written &&
3015              ((test_bit(R5_Insync, &dev->flags) &&
3016                !test_bit(R5_LOCKED, &dev->flags) &&
3017                test_bit(R5_UPTODATE, &dev->flags)) ||
3018                (s.failed == 1 && s.failed_num == sh->pd_idx)))
3019                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3020
3021         /* Now we might consider reading some blocks, either to check/generate
3022          * parity, or to satisfy requests
3023          * or to load a block that is being partially written.
3024          */
3025         if (s.to_read || s.non_overwrite ||
3026             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3027                 handle_stripe_fill5(sh, &s, disks);
3028
3029         /* Now we check to see if any write operations have recently
3030          * completed
3031          */
3032         prexor = 0;
3033         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3034                 prexor = 1;
3035         if (sh->reconstruct_state == reconstruct_state_drain_result ||
3036             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3037                 sh->reconstruct_state = reconstruct_state_idle;
3038
3039                 /* All the 'written' buffers and the parity block are ready to
3040                  * be written back to disk
3041                  */
3042                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3043                 for (i = disks; i--; ) {
3044                         dev = &sh->dev[i];
3045                         if (test_bit(R5_LOCKED, &dev->flags) &&
3046                                 (i == sh->pd_idx || dev->written)) {
3047                                 pr_debug("Writing block %d\n", i);
3048                                 set_bit(R5_Wantwrite, &dev->flags);
3049                                 if (prexor)
3050                                         continue;
3051                                 if (!test_bit(R5_Insync, &dev->flags) ||
3052                                     (i == sh->pd_idx && s.failed == 0))
3053                                         set_bit(STRIPE_INSYNC, &sh->state);
3054                         }
3055                 }
3056                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
3057                         atomic_dec(&conf->preread_active_stripes);
3058                         if (atomic_read(&conf->preread_active_stripes) <
3059                                 IO_THRESHOLD)
3060                                 md_wakeup_thread(conf->mddev->thread);
3061                 }
3062         }
3063
3064         /* Now to consider new write requests and what else, if anything
3065          * should be read.  We do not handle new writes when:
3066          * 1/ A 'write' operation (copy+xor) is already in flight.
3067          * 2/ A 'check' operation is in flight, as it may clobber the parity
3068          *    block.
3069          */
3070         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3071                 handle_stripe_dirtying5(conf, sh, &s, disks);
3072
3073         /* maybe we need to check and possibly fix the parity for this stripe
3074          * Any reads will already have been scheduled, so we just see if enough
3075          * data is available.  The parity check is held off while parity
3076          * dependent operations are in flight.
3077          */
3078         if (sh->check_state ||
3079             (s.syncing && s.locked == 0 &&
3080              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3081              !test_bit(STRIPE_INSYNC, &sh->state)))
3082                 handle_parity_checks5(conf, sh, &s, disks);
3083
3084         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3085                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3086                 clear_bit(STRIPE_SYNCING, &sh->state);
3087         }
3088
3089         /* If the failed drive is just a ReadError, then we might need to progress
3090          * the repair/check process
3091          */
3092         if (s.failed == 1 && !conf->mddev->ro &&
3093             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
3094             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
3095             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
3096                 ) {
3097                 dev = &sh->dev[s.failed_num];
3098                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3099                         set_bit(R5_Wantwrite, &dev->flags);
3100                         set_bit(R5_ReWrite, &dev->flags);
3101                         set_bit(R5_LOCKED, &dev->flags);
3102                         s.locked++;
3103                 } else {
3104                         /* let's read it back */
3105                         set_bit(R5_Wantread, &dev->flags);
3106                         set_bit(R5_LOCKED, &dev->flags);
3107                         s.locked++;
3108                 }
3109         }
3110
3111         /* Finish reconstruct operations initiated by the expansion process */
3112         if (sh->reconstruct_state == reconstruct_state_result) {
3113                 struct stripe_head *sh2
3114                         = get_active_stripe(conf, sh->sector, 1, 1, 1);
3115                 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3116                         /* sh cannot be written until sh2 has been read.
3117                          * so arrange for sh to be delayed a little
3118                          */
3119                         set_bit(STRIPE_DELAYED, &sh->state);
3120                         set_bit(STRIPE_HANDLE, &sh->state);
3121                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3122                                               &sh2->state))
3123                                 atomic_inc(&conf->preread_active_stripes);
3124                         release_stripe(sh2);
3125                         goto unlock;
3126                 }
3127                 if (sh2)
3128                         release_stripe(sh2);
3129
3130                 sh->reconstruct_state = reconstruct_state_idle;
3131                 clear_bit(STRIPE_EXPANDING, &sh->state);
3132                 for (i = conf->raid_disks; i--; ) {
3133                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3134                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3135                         s.locked++;
3136                 }
3137         }
3138
3139         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3140             !sh->reconstruct_state) {
3141                 /* Need to write out all blocks after computing parity */
3142                 sh->disks = conf->raid_disks;
3143                 stripe_set_idx(sh->sector, conf, 0, sh);
3144                 schedule_reconstruction(sh, &s, 1, 1);
3145         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3146                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3147                 atomic_dec(&conf->reshape_stripes);
3148                 wake_up(&conf->wait_for_overlap);
3149                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3150         }
3151
3152         if (s.expanding && s.locked == 0 &&
3153             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3154                 handle_stripe_expansion(conf, sh, NULL);
3155
3156  unlock:
3157         spin_unlock(&sh->lock);
3158
3159         /* wait for this device to become unblocked */
3160         if (unlikely(blocked_rdev))
3161                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3162
3163         if (s.ops_request)
3164                 raid_run_ops(sh, s.ops_request);
3165
3166         ops_run_io(sh, &s);
3167
3168         return_io(return_bi);
3169
3170         return blocked_rdev == NULL;
3171 }
3172
3173 static bool handle_stripe6(struct stripe_head *sh)
3174 {
3175         raid5_conf_t *conf = sh->raid_conf;
3176         int disks = sh->disks;
3177         struct bio *return_bi = NULL;
3178         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3179         struct stripe_head_state s;
3180         struct r6_state r6s;
3181         struct r5dev *dev, *pdev, *qdev;
3182         mdk_rdev_t *blocked_rdev = NULL;
3183
3184         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3185                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3186                (unsigned long long)sh->sector, sh->state,
3187                atomic_read(&sh->count), pd_idx, qd_idx,
3188                sh->check_state, sh->reconstruct_state);
3189         memset(&s, 0, sizeof(s));
3190
3191         spin_lock(&sh->lock);
3192         clear_bit(STRIPE_HANDLE, &sh->state);
3193         clear_bit(STRIPE_DELAYED, &sh->state);
3194
3195         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3196         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3197         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3198         /* Now to look around and see what can be done */
3199
3200         rcu_read_lock();
3201         for (i=disks; i--; ) {
3202                 mdk_rdev_t *rdev;
3203                 dev = &sh->dev[i];
3204                 clear_bit(R5_Insync, &dev->flags);
3205
3206                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3207                         i, dev->flags, dev->toread, dev->towrite, dev->written);
3208                 /* maybe we can reply to a read
3209                  *
3210                  * new wantfill requests are only permitted while
3211                  * ops_complete_biofill is guaranteed to be inactive
3212                  */
3213                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3214                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3215                         set_bit(R5_Wantfill, &dev->flags);
3216
3217                 /* now count some things */
3218                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3219                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3220                 if (test_bit(R5_Wantcompute, &dev->flags)) {
3221                         s.compute++;
3222                         BUG_ON(s.compute > 2);
3223                 }
3224
3225                 if (test_bit(R5_Wantfill, &dev->flags)) {
3226                         s.to_fill++;
3227                 } else if (dev->toread)
3228                         s.to_read++;
3229                 if (dev->towrite) {
3230                         s.to_write++;
3231                         if (!test_bit(R5_OVERWRITE, &dev->flags))
3232                                 s.non_overwrite++;
3233                 }
3234                 if (dev->written)
3235                         s.written++;
3236                 rdev = rcu_dereference(conf->disks[i].rdev);
3237                 if (blocked_rdev == NULL &&
3238                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3239                         blocked_rdev = rdev;
3240                         atomic_inc(&rdev->nr_pending);
3241                 }
3242                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3243                         /* The ReadError flag will just be confusing now */
3244                         clear_bit(R5_ReadError, &dev->flags);
3245                         clear_bit(R5_ReWrite, &dev->flags);
3246                 }
3247                 if (!rdev || !test_bit(In_sync, &rdev->flags)
3248                     || test_bit(R5_ReadError, &dev->flags)) {
3249                         if (s.failed < 2)
3250                                 r6s.failed_num[s.failed] = i;
3251                         s.failed++;
3252                 } else
3253                         set_bit(R5_Insync, &dev->flags);
3254         }
3255         rcu_read_unlock();
3256
3257         if (unlikely(blocked_rdev)) {
3258                 if (s.syncing || s.expanding || s.expanded ||
3259                     s.to_write || s.written) {
3260                         set_bit(STRIPE_HANDLE, &sh->state);
3261                         goto unlock;
3262                 }
3263                 /* There is nothing for the blocked_rdev to block */
3264                 rdev_dec_pending(blocked_rdev, conf->mddev);
3265                 blocked_rdev = NULL;
3266         }
3267
3268         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3269                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3270                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3271         }
3272
3273         pr_debug("locked=%d uptodate=%d to_read=%d"
3274                " to_write=%d failed=%d failed_num=%d,%d\n",
3275                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3276                r6s.failed_num[0], r6s.failed_num[1]);
3277         /* check if the array has lost >2 devices and, if so, some requests
3278          * might need to be failed
3279          */
3280         if (s.failed > 2 && s.to_read+s.to_write+s.written)
3281                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3282         if (s.failed > 2 && s.syncing) {
3283                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3284                 clear_bit(STRIPE_SYNCING, &sh->state);
3285                 s.syncing = 0;
3286         }
3287
3288         /*
3289          * might be able to return some write requests if the parity blocks
3290          * are safe, or on a failed drive
3291          */
3292         pdev = &sh->dev[pd_idx];
3293         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3294                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3295         qdev = &sh->dev[qd_idx];
3296         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == qd_idx)
3297                 || (s.failed >= 2 && r6s.failed_num[1] == qd_idx);
3298
3299         if ( s.written &&
3300              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3301                              && !test_bit(R5_LOCKED, &pdev->flags)
3302                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3303              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3304                              && !test_bit(R5_LOCKED, &qdev->flags)
3305                              && test_bit(R5_UPTODATE, &qdev->flags)))))
3306                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3307
3308         /* Now we might consider reading some blocks, either to check/generate
3309          * parity, or to satisfy requests
3310          * or to load a block that is being partially written.
3311          */
3312         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3313             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3314                 handle_stripe_fill6(sh, &s, &r6s, disks);
3315
3316         /* Now we check to see if any write operations have recently
3317          * completed
3318          */
3319         if (sh->reconstruct_state == reconstruct_state_drain_result) {
3320                 int qd_idx = sh->qd_idx;
3321
3322                 sh->reconstruct_state = reconstruct_state_idle;
3323                 /* All the 'written' buffers and the parity blocks are ready to
3324                  * be written back to disk
3325                  */
3326                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3327                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags));
3328                 for (i = disks; i--; ) {
3329                         dev = &sh->dev[i];
3330                         if (test_bit(R5_LOCKED, &dev->flags) &&
3331                             (i == sh->pd_idx || i == qd_idx ||
3332                              dev->written)) {
3333                                 pr_debug("Writing block %d\n", i);
3334                                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3335                                 set_bit(R5_Wantwrite, &dev->flags);
3336                                 if (!test_bit(R5_Insync, &dev->flags) ||
3337                                     ((i == sh->pd_idx || i == qd_idx) &&
3338                                       s.failed == 0))
3339                                         set_bit(STRIPE_INSYNC, &sh->state);
3340                         }
3341                 }
3342                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
3343                         atomic_dec(&conf->preread_active_stripes);
3344                         if (atomic_read(&conf->preread_active_stripes) <
3345                                 IO_THRESHOLD)
3346                                 md_wakeup_thread(conf->mddev->thread);
3347                 }
3348         }
3349
3350         /* Now to consider new write requests and what else, if anything
3351          * should be read.  We do not handle new writes when:
3352          * 1/ A 'write' operation (copy+gen_syndrome) is already in flight.
3353          * 2/ A 'check' operation is in flight, as it may clobber the parity
3354          *    block.
3355          */
3356         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3357                 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3358
3359         /* maybe we need to check and possibly fix the parity for this stripe
3360          * Any reads will already have been scheduled, so we just see if enough
3361          * data is available.  The parity check is held off while parity
3362          * dependent operations are in flight.
3363          */
3364         if (sh->check_state ||
3365             (s.syncing && s.locked == 0 &&
3366              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3367              !test_bit(STRIPE_INSYNC, &sh->state)))
3368                 handle_parity_checks6(conf, sh, &s, &r6s, disks);
3369
3370         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3371                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3372                 clear_bit(STRIPE_SYNCING, &sh->state);
3373         }
3374
3375         /* If the failed drives are just a ReadError, then we might need
3376          * to progress the repair/check process
3377          */
3378         if (s.failed <= 2 && !conf->mddev->ro)
3379                 for (i = 0; i < s.failed; i++) {
3380                         dev = &sh->dev[r6s.failed_num[i]];
3381                         if (test_bit(R5_ReadError, &dev->flags)
3382                             && !test_bit(R5_LOCKED, &dev->flags)
3383                             && test_bit(R5_UPTODATE, &dev->flags)
3384                                 ) {
3385                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3386                                         set_bit(R5_Wantwrite, &dev->flags);
3387                                         set_bit(R5_ReWrite, &dev->flags);
3388                                         set_bit(R5_LOCKED, &dev->flags);
3389                                         s.locked++;
3390                                 } else {
3391                                         /* let's read it back */
3392                                         set_bit(R5_Wantread, &dev->flags);
3393                                         set_bit(R5_LOCKED, &dev->flags);
3394                                         s.locked++;
3395                                 }
3396                         }
3397                 }
3398
3399         /* Finish reconstruct operations initiated by the expansion process */
3400         if (sh->reconstruct_state == reconstruct_state_result) {
3401                 sh->reconstruct_state = reconstruct_state_idle;
3402                 clear_bit(STRIPE_EXPANDING, &sh->state);
3403                 for (i = conf->raid_disks; i--; ) {
3404                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3405                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3406                         s.locked++;
3407                 }
3408         }
3409
3410         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3411             !sh->reconstruct_state) {
3412                 struct stripe_head *sh2
3413                         = get_active_stripe(conf, sh->sector, 1, 1, 1);
3414                 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3415                         /* sh cannot be written until sh2 has been read.
3416                          * so arrange for sh to be delayed a little
3417                          */
3418                         set_bit(STRIPE_DELAYED, &sh->state);
3419                         set_bit(STRIPE_HANDLE, &sh->state);
3420                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3421                                               &sh2->state))
3422                                 atomic_inc(&conf->preread_active_stripes);
3423                         release_stripe(sh2);
3424                         goto unlock;
3425                 }
3426                 if (sh2)
3427                         release_stripe(sh2);
3428
3429                 /* Need to write out all blocks after computing P&Q */
3430                 sh->disks = conf->raid_disks;
3431                 stripe_set_idx(sh->sector, conf, 0, sh);
3432                 schedule_reconstruction(sh, &s, 1, 1);
3433         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3434                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3435                 atomic_dec(&conf->reshape_stripes);
3436                 wake_up(&conf->wait_for_overlap);
3437                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3438         }
3439
3440         if (s.expanding && s.locked == 0 &&
3441             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3442                 handle_stripe_expansion(conf, sh, &r6s);
3443
3444  unlock:
3445         spin_unlock(&sh->lock);
3446
3447         /* wait for this device to become unblocked */
3448         if (unlikely(blocked_rdev))
3449                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3450
3451         if (s.ops_request)
3452                 raid_run_ops(sh, s.ops_request);
3453
3454         ops_run_io(sh, &s);
3455
3456         return_io(return_bi);
3457
3458         return blocked_rdev == NULL;
3459 }
3460
3461 /* returns true if the stripe was handled */
3462 static bool handle_stripe(struct stripe_head *sh)
3463 {
3464         if (sh->raid_conf->level == 6)
3465                 return handle_stripe6(sh);
3466         else
3467                 return handle_stripe5(sh);
3468 }
3469
3470 static void raid5_activate_delayed(raid5_conf_t *conf)
3471 {
3472         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3473                 while (!list_empty(&conf->delayed_list)) {
3474                         struct list_head *l = conf->delayed_list.next;
3475                         struct stripe_head *sh;
3476                         sh = list_entry(l, struct stripe_head, lru);
3477                         list_del_init(l);
3478                         clear_bit(STRIPE_DELAYED, &sh->state);
3479                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3480                                 atomic_inc(&conf->preread_active_stripes);
3481                         list_add_tail(&sh->lru, &conf->hold_list);
3482                 }
3483         } else
3484                 blk_plug_device(conf->mddev->queue);
3485 }
3486
3487 static void activate_bit_delay(raid5_conf_t *conf)
3488 {
3489         /* device_lock is held */
3490         struct list_head head;
3491         list_add(&head, &conf->bitmap_list);
3492         list_del_init(&conf->bitmap_list);
3493         while (!list_empty(&head)) {
3494                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3495                 list_del_init(&sh->lru);
3496                 atomic_inc(&sh->count);
3497                 __release_stripe(conf, sh);
3498         }
3499 }
3500
3501 static void unplug_slaves(mddev_t *mddev)
3502 {
3503         raid5_conf_t *conf = mddev->private;
3504         int i;
3505
3506         rcu_read_lock();
3507         for (i = 0; i < conf->raid_disks; i++) {
3508                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3509                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3510                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3511
3512                         atomic_inc(&rdev->nr_pending);
3513                         rcu_read_unlock();
3514
3515                         blk_unplug(r_queue);
3516
3517                         rdev_dec_pending(rdev, mddev);
3518                         rcu_read_lock();
3519                 }
3520         }
3521         rcu_read_unlock();
3522 }
3523
3524 static void raid5_unplug_device(struct request_queue *q)
3525 {
3526         mddev_t *mddev = q->queuedata;
3527         raid5_conf_t *conf = mddev->private;
3528         unsigned long flags;
3529
3530         spin_lock_irqsave(&conf->device_lock, flags);
3531
3532         if (blk_remove_plug(q)) {
3533                 conf->seq_flush++;
3534                 raid5_activate_delayed(conf);
3535         }
3536         md_wakeup_thread(mddev->thread);
3537
3538         spin_unlock_irqrestore(&conf->device_lock, flags);
3539
3540         unplug_slaves(mddev);
3541 }
3542
3543 static int raid5_congested(void *data, int bits)
3544 {
3545         mddev_t *mddev = data;
3546         raid5_conf_t *conf = mddev->private;
3547
3548         /* No difference between reads and writes.  Just check
3549          * how busy the stripe_cache is
3550          */
3551         if (conf->inactive_blocked)
3552                 return 1;
3553         if (conf->quiesce)
3554                 return 1;
3555         if (list_empty_careful(&conf->inactive_list))
3556                 return 1;
3557
3558         return 0;
3559 }
3560
3561 /* We want read requests to align with chunks where possible,
3562  * but write requests don't need to.
3563  */
3564 static int raid5_mergeable_bvec(struct request_queue *q,
3565                                 struct bvec_merge_data *bvm,
3566                                 struct bio_vec *biovec)
3567 {
3568         mddev_t *mddev = q->queuedata;
3569         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3570         int max;
3571         unsigned int chunk_sectors = mddev->chunk_sectors;
3572         unsigned int bio_sectors = bvm->bi_size >> 9;
3573
3574         if ((bvm->bi_rw & 1) == WRITE)
3575                 return biovec->bv_len; /* always allow writes to be mergeable */
3576
3577         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3578                 chunk_sectors = mddev->new_chunk_sectors;
3579         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3580         if (max < 0) max = 0;
3581         if (max <= biovec->bv_len && bio_sectors == 0)
3582                 return biovec->bv_len;
3583         else
3584                 return max;
3585 }
3586
3587
3588 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3589 {
3590         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3591         unsigned int chunk_sectors = mddev->chunk_sectors;
3592         unsigned int bio_sectors = bio->bi_size >> 9;
3593
3594         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3595                 chunk_sectors = mddev->new_chunk_sectors;
3596         return  chunk_sectors >=
3597                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3598 }
3599
3600 /*
3601  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3602  *  later sampled by raid5d.
3603  */
3604 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3605 {
3606         unsigned long flags;
3607
3608         spin_lock_irqsave(&conf->device_lock, flags);
3609
3610         bi->bi_next = conf->retry_read_aligned_list;
3611         conf->retry_read_aligned_list = bi;
3612
3613         spin_unlock_irqrestore(&conf->device_lock, flags);
3614         md_wakeup_thread(conf->mddev->thread);
3615 }
3616
3617
3618 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3619 {
3620         struct bio *bi;
3621
3622         bi = conf->retry_read_aligned;
3623         if (bi) {
3624                 conf->retry_read_aligned = NULL;
3625                 return bi;
3626         }
3627         bi = conf->retry_read_aligned_list;
3628         if(bi) {
3629                 conf->retry_read_aligned_list = bi->bi_next;
3630                 bi->bi_next = NULL;
3631                 /*
3632                  * this sets the active strip count to 1 and the processed
3633                  * strip count to zero (upper 8 bits)
3634                  */
3635                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3636         }
3637
3638         return bi;
3639 }
3640
3641
3642 /*
3643  *  The "raid5_align_endio" should check if the read succeeded and if it
3644  *  did, call bio_endio on the original bio (having bio_put the new bio
3645  *  first).
3646  *  If the read failed..
3647  */
3648 static void raid5_align_endio(struct bio *bi, int error)
3649 {
3650         struct bio* raid_bi  = bi->bi_private;
3651         mddev_t *mddev;
3652         raid5_conf_t *conf;
3653         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3654         mdk_rdev_t *rdev;
3655
3656         bio_put(bi);
3657
3658         mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3659         conf = mddev->private;
3660         rdev = (void*)raid_bi->bi_next;
3661         raid_bi->bi_next = NULL;
3662
3663         rdev_dec_pending(rdev, conf->mddev);
3664
3665         if (!error && uptodate) {
3666                 bio_endio(raid_bi, 0);
3667                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3668                         wake_up(&conf->wait_for_stripe);
3669                 return;
3670         }
3671
3672
3673         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3674
3675         add_bio_to_retry(raid_bi, conf);
3676 }
3677
3678 static int bio_fits_rdev(struct bio *bi)
3679 {
3680         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3681
3682         if ((bi->bi_size>>9) > queue_max_sectors(q))
3683                 return 0;
3684         blk_recount_segments(q, bi);
3685         if (bi->bi_phys_segments > queue_max_phys_segments(q))
3686                 return 0;
3687
3688         if (q->merge_bvec_fn)
3689                 /* it's too hard to apply the merge_bvec_fn at this stage,
3690                  * just just give up
3691                  */
3692                 return 0;
3693
3694         return 1;
3695 }
3696
3697
3698 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3699 {
3700         mddev_t *mddev = q->queuedata;
3701         raid5_conf_t *conf = mddev->private;
3702         unsigned int dd_idx;
3703         struct bio* align_bi;
3704         mdk_rdev_t *rdev;
3705
3706         if (!in_chunk_boundary(mddev, raid_bio)) {
3707                 pr_debug("chunk_aligned_read : non aligned\n");
3708                 return 0;
3709         }
3710         /*
3711          * use bio_clone to make a copy of the bio
3712          */
3713         align_bi = bio_clone(raid_bio, GFP_NOIO);
3714         if (!align_bi)
3715                 return 0;
3716         /*
3717          *   set bi_end_io to a new function, and set bi_private to the
3718          *     original bio.
3719          */
3720         align_bi->bi_end_io  = raid5_align_endio;
3721         align_bi->bi_private = raid_bio;
3722         /*
3723          *      compute position
3724          */
3725         align_bi->bi_sector =  raid5_compute_sector(conf, raid_bio->bi_sector,
3726                                                     0,
3727                                                     &dd_idx, NULL);
3728
3729         rcu_read_lock();
3730         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3731         if (rdev && test_bit(In_sync, &rdev->flags)) {
3732                 atomic_inc(&rdev->nr_pending);
3733                 rcu_read_unlock();
3734                 raid_bio->bi_next = (void*)rdev;
3735                 align_bi->bi_bdev =  rdev->bdev;
3736                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3737                 align_bi->bi_sector += rdev->data_offset;
3738
3739                 if (!bio_fits_rdev(align_bi)) {
3740                         /* too big in some way */
3741                         bio_put(align_bi);
3742                         rdev_dec_pending(rdev, mddev);
3743                         return 0;
3744                 }
3745
3746                 spin_lock_irq(&conf->device_lock);
3747                 wait_event_lock_irq(conf->wait_for_stripe,
3748                                     conf->quiesce == 0,
3749                                     conf->device_lock, /* nothing */);
3750                 atomic_inc(&conf->active_aligned_reads);
3751                 spin_unlock_irq(&conf->device_lock);
3752
3753                 generic_make_request(align_bi);
3754                 return 1;
3755         } else {
3756                 rcu_read_unlock();
3757                 bio_put(align_bi);
3758                 return 0;
3759         }
3760 }
3761
3762 /* __get_priority_stripe - get the next stripe to process
3763  *
3764  * Full stripe writes are allowed to pass preread active stripes up until
3765  * the bypass_threshold is exceeded.  In general the bypass_count
3766  * increments when the handle_list is handled before the hold_list; however, it
3767  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3768  * stripe with in flight i/o.  The bypass_count will be reset when the
3769  * head of the hold_list has changed, i.e. the head was promoted to the
3770  * handle_list.
3771  */
3772 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3773 {
3774         struct stripe_head *sh;
3775
3776         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3777                   __func__,
3778                   list_empty(&conf->handle_list) ? "empty" : "busy",
3779                   list_empty(&conf->hold_list) ? "empty" : "busy",
3780                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
3781
3782         if (!list_empty(&conf->handle_list)) {
3783                 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3784
3785                 if (list_empty(&conf->hold_list))
3786                         conf->bypass_count = 0;
3787                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3788                         if (conf->hold_list.next == conf->last_hold)
3789                                 conf->bypass_count++;
3790                         else {
3791                                 conf->last_hold = conf->hold_list.next;
3792                                 conf->bypass_count -= conf->bypass_threshold;
3793                                 if (conf->bypass_count < 0)
3794                                         conf->bypass_count = 0;
3795                         }
3796                 }
3797         } else if (!list_empty(&conf->hold_list) &&
3798                    ((conf->bypass_threshold &&
3799                      conf->bypass_count > conf->bypass_threshold) ||
3800                     atomic_read(&conf->pending_full_writes) == 0)) {
3801                 sh = list_entry(conf->hold_list.next,
3802                                 typeof(*sh), lru);
3803                 conf->bypass_count -= conf->bypass_threshold;
3804                 if (conf->bypass_count < 0)
3805                         conf->bypass_count = 0;
3806         } else
3807                 return NULL;
3808
3809         list_del_init(&sh->lru);
3810         atomic_inc(&sh->count);
3811         BUG_ON(atomic_read(&sh->count) != 1);
3812         return sh;
3813 }
3814
3815 static int make_request(struct request_queue *q, struct bio * bi)
3816 {
3817         mddev_t *mddev = q->queuedata;
3818         raid5_conf_t *conf = mddev->private;
3819         int dd_idx;
3820         sector_t new_sector;
3821         sector_t logical_sector, last_sector;
3822         struct stripe_head *sh;
3823         const int rw = bio_data_dir(bi);
3824         int cpu, remaining;
3825
3826         if (unlikely(bio_barrier(bi))) {
3827                 bio_endio(bi, -EOPNOTSUPP);
3828                 return 0;
3829         }
3830
3831         md_write_start(mddev, bi);
3832
3833         cpu = part_stat_lock();
3834         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
3835         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
3836                       bio_sectors(bi));
3837         part_stat_unlock();
3838
3839         if (rw == READ &&
3840              mddev->reshape_position == MaxSector &&
3841              chunk_aligned_read(q,bi))
3842                 return 0;
3843
3844         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3845         last_sector = bi->bi_sector + (bi->bi_size>>9);
3846         bi->bi_next = NULL;
3847         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
3848
3849         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3850                 DEFINE_WAIT(w);
3851                 int disks, data_disks;
3852                 int previous;
3853
3854         retry:
3855                 previous = 0;
3856                 disks = conf->raid_disks;
3857                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3858                 if (unlikely(conf->reshape_progress != MaxSector)) {
3859                         /* spinlock is needed as reshape_progress may be
3860                          * 64bit on a 32bit platform, and so it might be
3861                          * possible to see a half-updated value
3862                          * Ofcourse reshape_progress could change after
3863                          * the lock is dropped, so once we get a reference
3864                          * to the stripe that we think it is, we will have
3865                          * to check again.
3866                          */
3867                         spin_lock_irq(&conf->device_lock);
3868                         if (mddev->delta_disks < 0
3869                             ? logical_sector < conf->reshape_progress
3870                             : logical_sector >= conf->reshape_progress) {
3871                                 disks = conf->previous_raid_disks;
3872                                 previous = 1;
3873                         } else {
3874                                 if (mddev->delta_disks < 0
3875                                     ? logical_sector < conf->reshape_safe
3876                                     : logical_sector >= conf->reshape_safe) {
3877                                         spin_unlock_irq(&conf->device_lock);
3878                                         schedule();
3879                                         goto retry;
3880                                 }
3881                         }
3882                         spin_unlock_irq(&conf->device_lock);
3883                 }
3884                 data_disks = disks - conf->max_degraded;
3885
3886                 new_sector = raid5_compute_sector(conf, logical_sector,
3887                                                   previous,
3888                                                   &dd_idx, NULL);
3889                 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3890                         (unsigned long long)new_sector,
3891                         (unsigned long long)logical_sector);
3892
3893                 sh = get_active_stripe(conf, new_sector, previous,
3894                                        (bi->bi_rw&RWA_MASK), 0);
3895                 if (sh) {
3896                         if (unlikely(previous)) {
3897                                 /* expansion might have moved on while waiting for a
3898                                  * stripe, so we must do the range check again.
3899                                  * Expansion could still move past after this
3900                                  * test, but as we are holding a reference to
3901                                  * 'sh', we know that if that happens,
3902                                  *  STRIPE_EXPANDING will get set and the expansion
3903                                  * won't proceed until we finish with the stripe.
3904                                  */
3905                                 int must_retry = 0;
3906                                 spin_lock_irq(&conf->device_lock);
3907                                 if (mddev->delta_disks < 0
3908                                     ? logical_sector >= conf->reshape_progress
3909                                     : logical_sector < conf->reshape_progress)
3910                                         /* mismatch, need to try again */
3911                                         must_retry = 1;
3912                                 spin_unlock_irq(&conf->device_lock);
3913                                 if (must_retry) {
3914                                         release_stripe(sh);
3915                                         schedule();
3916                                         goto retry;
3917                                 }
3918                         }
3919
3920                         if (bio_data_dir(bi) == WRITE &&
3921                             logical_sector >= mddev->suspend_lo &&
3922                             logical_sector < mddev->suspend_hi) {
3923                                 release_stripe(sh);
3924                                 /* As the suspend_* range is controlled by
3925                                  * userspace, we want an interruptible
3926                                  * wait.
3927                                  */
3928                                 flush_signals(current);
3929                                 prepare_to_wait(&conf->wait_for_overlap,
3930                                                 &w, TASK_INTERRUPTIBLE);
3931                                 if (logical_sector >= mddev->suspend_lo &&
3932                                     logical_sector < mddev->suspend_hi)
3933                                         schedule();
3934                                 goto retry;
3935                         }
3936
3937                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3938                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3939                                 /* Stripe is busy expanding or
3940                                  * add failed due to overlap.  Flush everything
3941                                  * and wait a while
3942                                  */
3943                                 raid5_unplug_device(mddev->queue);
3944                                 release_stripe(sh);
3945                                 schedule();
3946                                 goto retry;
3947                         }
3948                         finish_wait(&conf->wait_for_overlap, &w);
3949                         set_bit(STRIPE_HANDLE, &sh->state);
3950                         clear_bit(STRIPE_DELAYED, &sh->state);
3951                         release_stripe(sh);
3952                 } else {
3953                         /* cannot get stripe for read-ahead, just give-up */
3954                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
3955                         finish_wait(&conf->wait_for_overlap, &w);
3956                         break;
3957                 }
3958
3959         }
3960         spin_lock_irq(&conf->device_lock);
3961         remaining = raid5_dec_bi_phys_segments(bi);
3962         spin_unlock_irq(&conf->device_lock);
3963         if (remaining == 0) {
3964
3965                 if ( rw == WRITE )
3966                         md_write_end(mddev);
3967
3968                 bio_endio(bi, 0);
3969         }
3970         return 0;
3971 }
3972
3973 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
3974
3975 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3976 {
3977         /* reshaping is quite different to recovery/resync so it is
3978          * handled quite separately ... here.
3979          *
3980          * On each call to sync_request, we gather one chunk worth of
3981          * destination stripes and flag them as expanding.
3982          * Then we find all the source stripes and request reads.
3983          * As the reads complete, handle_stripe will copy the data
3984          * into the destination stripe and release that stripe.
3985          */
3986         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3987         struct stripe_head *sh;
3988         sector_t first_sector, last_sector;
3989         int raid_disks = conf->previous_raid_disks;
3990         int data_disks = raid_disks - conf->max_degraded;
3991         int new_data_disks = conf->raid_disks - conf->max_degraded;
3992         int i;
3993         int dd_idx;
3994         sector_t writepos, readpos, safepos;
3995         sector_t stripe_addr;
3996         int reshape_sectors;
3997         struct list_head stripes;
3998
3999         if (sector_nr == 0) {
4000                 /* If restarting in the middle, skip the initial sectors */
4001                 if (mddev->delta_disks < 0 &&
4002                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4003                         sector_nr = raid5_size(mddev, 0, 0)
4004                                 - conf->reshape_progress;
4005                 } else if (mddev->delta_disks >= 0 &&
4006                            conf->reshape_progress > 0)
4007                         sector_nr = conf->reshape_progress;
4008                 sector_div(sector_nr, new_data_disks);
4009                 if (sector_nr) {
4010                         *skipped = 1;
4011                         return sector_nr;
4012                 }
4013         }
4014
4015         /* We need to process a full chunk at a time.
4016          * If old and new chunk sizes differ, we need to process the
4017          * largest of these
4018          */
4019         if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4020                 reshape_sectors = mddev->new_chunk_sectors;
4021         else
4022                 reshape_sectors = mddev->chunk_sectors;
4023
4024         /* we update the metadata when there is more than 3Meg
4025          * in the block range (that is rather arbitrary, should
4026          * probably be time based) or when the data about to be
4027          * copied would over-write the source of the data at
4028          * the front of the range.
4029          * i.e. one new_stripe along from reshape_progress new_maps
4030          * to after where reshape_safe old_maps to
4031          */
4032         writepos = conf->reshape_progress;
4033         sector_div(writepos, new_data_disks);
4034         readpos = conf->reshape_progress;
4035         sector_div(readpos, data_disks);
4036         safepos = conf->reshape_safe;
4037         sector_div(safepos, data_disks);
4038         if (mddev->delta_disks < 0) {
4039                 writepos -= min_t(sector_t, reshape_sectors, writepos);
4040                 readpos += reshape_sectors;
4041                 safepos += reshape_sectors;
4042         } else {
4043                 writepos += reshape_sectors;
4044                 readpos -= min_t(sector_t, reshape_sectors, readpos);
4045                 safepos -= min_t(sector_t, reshape_sectors, safepos);
4046         }
4047
4048         /* 'writepos' is the most advanced device address we might write.
4049          * 'readpos' is the least advanced device address we might read.
4050          * 'safepos' is the least address recorded in the metadata as having
4051          *     been reshaped.
4052          * If 'readpos' is behind 'writepos', then there is no way that we can
4053          * ensure safety in the face of a crash - that must be done by userspace
4054          * making a backup of the data.  So in that case there is no particular
4055          * rush to update metadata.
4056          * Otherwise if 'safepos' is behind 'writepos', then we really need to
4057          * update the metadata to advance 'safepos' to match 'readpos' so that
4058          * we can be safe in the event of a crash.
4059          * So we insist on updating metadata if safepos is behind writepos and
4060          * readpos is beyond writepos.
4061          * In any case, update the metadata every 10 seconds.
4062          * Maybe that number should be configurable, but I'm not sure it is
4063          * worth it.... maybe it could be a multiple of safemode_delay???
4064          */
4065         if ((mddev->delta_disks < 0
4066              ? (safepos > writepos && readpos < writepos)
4067              : (safepos < writepos && readpos > writepos)) ||
4068             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4069                 /* Cannot proceed until we've updated the superblock... */
4070                 wait_event(conf->wait_for_overlap,
4071                            atomic_read(&conf->reshape_stripes)==0);
4072                 mddev->reshape_position = conf->reshape_progress;
4073                 mddev->curr_resync_completed = mddev->curr_resync;
4074                 conf->reshape_checkpoint = jiffies;
4075                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4076                 md_wakeup_thread(mddev->thread);
4077                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4078                            kthread_should_stop());
4079                 spin_lock_irq(&conf->device_lock);
4080                 conf->reshape_safe = mddev->reshape_position;
4081                 spin_unlock_irq(&conf->device_lock);
4082                 wake_up(&conf->wait_for_overlap);
4083                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4084         }
4085
4086         if (mddev->delta_disks < 0) {
4087                 BUG_ON(conf->reshape_progress == 0);
4088                 stripe_addr = writepos;
4089                 BUG_ON((mddev->dev_sectors &
4090                         ~((sector_t)reshape_sectors - 1))
4091                        - reshape_sectors - stripe_addr
4092                        != sector_nr);
4093         } else {
4094                 BUG_ON(writepos != sector_nr + reshape_sectors);
4095                 stripe_addr = sector_nr;
4096         }
4097         INIT_LIST_HEAD(&stripes);
4098         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4099                 int j;
4100                 int skipped = 0;
4101                 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4102                 set_bit(STRIPE_EXPANDING, &sh->state);
4103                 atomic_inc(&conf->reshape_stripes);
4104                 /* If any of this stripe is beyond the end of the old
4105                  * array, then we need to zero those blocks
4106                  */
4107                 for (j=sh->disks; j--;) {
4108                         sector_t s;
4109                         if (j == sh->pd_idx)
4110                                 continue;
4111                         if (conf->level == 6 &&
4112                             j == sh->qd_idx)
4113                                 continue;
4114                         s = compute_blocknr(sh, j, 0);
4115                         if (s < raid5_size(mddev, 0, 0)) {
4116                                 skipped = 1;
4117                                 continue;
4118                         }
4119                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4120                         set_bit(R5_Expanded, &sh->dev[j].flags);
4121                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
4122                 }
4123                 if (!skipped) {
4124                         set_bit(STRIPE_EXPAND_READY, &sh->state);
4125                         set_bit(STRIPE_HANDLE, &sh->state);
4126                 }
4127                 list_add(&sh->lru, &stripes);
4128         }
4129         spin_lock_irq(&conf->device_lock);
4130         if (mddev->delta_disks < 0)
4131                 conf->reshape_progress -= reshape_sectors * new_data_disks;
4132         else
4133                 conf->reshape_progress += reshape_sectors * new_data_disks;
4134         spin_unlock_irq(&conf->device_lock);
4135         /* Ok, those stripe are ready. We can start scheduling
4136          * reads on the source stripes.
4137          * The source stripes are determined by mapping the first and last
4138          * block on the destination stripes.
4139          */
4140         first_sector =
4141                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4142                                      1, &dd_idx, NULL);
4143         last_sector =
4144                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4145                                             * new_data_disks - 1),
4146                                      1, &dd_idx, NULL);
4147         if (last_sector >= mddev->dev_sectors)
4148                 last_sector = mddev->dev_sectors - 1;
4149         while (first_sector <= last_sector) {
4150                 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4151                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4152                 set_bit(STRIPE_HANDLE, &sh->state);
4153                 release_stripe(sh);
4154                 first_sector += STRIPE_SECTORS;
4155         }
4156         /* Now that the sources are clearly marked, we can release
4157          * the destination stripes
4158          */
4159         while (!list_empty(&stripes)) {
4160                 sh = list_entry(stripes.next, struct stripe_head, lru);
4161                 list_del_init(&sh->lru);
4162                 release_stripe(sh);
4163         }
4164         /* If this takes us to the resync_max point where we have to pause,
4165          * then we need to write out the superblock.
4166          */
4167         sector_nr += reshape_sectors;
4168         if ((sector_nr - mddev->curr_resync_completed) * 2
4169             >= mddev->resync_max - mddev->curr_resync_completed) {
4170                 /* Cannot proceed until we've updated the superblock... */
4171                 wait_event(conf->wait_for_overlap,
4172                            atomic_read(&conf->reshape_stripes) == 0);
4173                 mddev->reshape_position = conf->reshape_progress;
4174                 mddev->curr_resync_completed = mddev->curr_resync + reshape_sectors;
4175                 conf->reshape_checkpoint = jiffies;
4176                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4177                 md_wakeup_thread(mddev->thread);
4178                 wait_event(mddev->sb_wait,
4179                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4180                            || kthread_should_stop());
4181                 spin_lock_irq(&conf->device_lock);
4182                 conf->reshape_safe = mddev->reshape_position;
4183                 spin_unlock_irq(&conf->device_lock);
4184                 wake_up(&conf->wait_for_overlap);
4185                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4186         }
4187         return reshape_sectors;
4188 }
4189
4190 /* FIXME go_faster isn't used */
4191 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
4192 {
4193         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4194         struct stripe_head *sh;
4195         sector_t max_sector = mddev->dev_sectors;
4196         int sync_blocks;
4197         int still_degraded = 0;
4198         int i;
4199
4200         if (sector_nr >= max_sector) {
4201                 /* just being told to finish up .. nothing much to do */
4202                 unplug_slaves(mddev);
4203
4204                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4205                         end_reshape(conf);
4206                         return 0;
4207                 }
4208
4209                 if (mddev->curr_resync < max_sector) /* aborted */
4210                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4211                                         &sync_blocks, 1);
4212                 else /* completed sync */
4213                         conf->fullsync = 0;
4214                 bitmap_close_sync(mddev->bitmap);
4215
4216                 return 0;
4217         }
4218
4219         /* Allow raid5_quiesce to complete */
4220         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4221
4222         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4223                 return reshape_request(mddev, sector_nr, skipped);
4224
4225         /* No need to check resync_max as we never do more than one
4226          * stripe, and as resync_max will always be on a chunk boundary,
4227          * if the check in md_do_sync didn't fire, there is no chance
4228          * of overstepping resync_max here
4229          */
4230
4231         /* if there is too many failed drives and we are trying
4232          * to resync, then assert that we are finished, because there is
4233          * nothing we can do.
4234          */
4235         if (mddev->degraded >= conf->max_degraded &&
4236             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4237                 sector_t rv = mddev->dev_sectors - sector_nr;
4238                 *skipped = 1;
4239                 return rv;
4240         }
4241         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4242             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4243             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4244                 /* we can skip this block, and probably more */
4245                 sync_blocks /= STRIPE_SECTORS;
4246                 *skipped = 1;
4247                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4248         }
4249
4250
4251         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4252
4253         sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4254         if (sh == NULL) {
4255                 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4256                 /* make sure we don't swamp the stripe cache if someone else
4257                  * is trying to get access
4258                  */
4259                 schedule_timeout_uninterruptible(1);
4260         }
4261         /* Need to check if array will still be degraded after recovery/resync
4262          * We don't need to check the 'failed' flag as when that gets set,
4263          * recovery aborts.
4264          */
4265         for (i = 0; i < conf->raid_disks; i++)
4266                 if (conf->disks[i].rdev == NULL)
4267                         still_degraded = 1;
4268
4269         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4270
4271         spin_lock(&sh->lock);
4272         set_bit(STRIPE_SYNCING, &sh->state);
4273         clear_bit(STRIPE_INSYNC, &sh->state);
4274         spin_unlock(&sh->lock);
4275
4276         /* wait for any blocked device to be handled */
4277         while (unlikely(!handle_stripe(sh)))
4278                 ;
4279         release_stripe(sh);
4280
4281         return STRIPE_SECTORS;
4282 }
4283
4284 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4285 {
4286         /* We may not be able to submit a whole bio at once as there
4287          * may not be enough stripe_heads available.
4288          * We cannot pre-allocate enough stripe_heads as we may need
4289          * more than exist in the cache (if we allow ever large chunks).
4290          * So we do one stripe head at a time and record in
4291          * ->bi_hw_segments how many have been done.
4292          *
4293          * We *know* that this entire raid_bio is in one chunk, so
4294          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4295          */
4296         struct stripe_head *sh;
4297         int dd_idx;
4298         sector_t sector, logical_sector, last_sector;
4299         int scnt = 0;
4300         int remaining;
4301         int handled = 0;
4302
4303         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4304         sector = raid5_compute_sector(conf, logical_sector,
4305                                       0, &dd_idx, NULL);
4306         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4307
4308         for (; logical_sector < last_sector;
4309              logical_sector += STRIPE_SECTORS,
4310                      sector += STRIPE_SECTORS,
4311                      scnt++) {
4312
4313                 if (scnt < raid5_bi_hw_segments(raid_bio))
4314                         /* already done this stripe */
4315                         continue;
4316
4317                 sh = get_active_stripe(conf, sector, 0, 1, 0);
4318
4319                 if (!sh) {
4320                         /* failed to get a stripe - must wait */
4321                         raid5_set_bi_hw_segments(raid_bio, scnt);
4322                         conf->retry_read_aligned = raid_bio;
4323                         return handled;
4324                 }
4325
4326                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4327                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4328                         release_stripe(sh);
4329                         raid5_set_bi_hw_segments(raid_bio, scnt);
4330                         conf->retry_read_aligned = raid_bio;
4331                         return handled;
4332                 }
4333
4334                 handle_stripe(sh);
4335                 release_stripe(sh);
4336                 handled++;
4337         }
4338         spin_lock_irq(&conf->device_lock);
4339         remaining = raid5_dec_bi_phys_segments(raid_bio);
4340         spin_unlock_irq(&conf->device_lock);
4341         if (remaining == 0)
4342                 bio_endio(raid_bio, 0);
4343         if (atomic_dec_and_test(&conf->active_aligned_reads))
4344                 wake_up(&conf->wait_for_stripe);
4345         return handled;
4346 }
4347
4348 #ifdef CONFIG_MULTICORE_RAID456
4349 static void __process_stripe(void *param, async_cookie_t cookie)
4350 {
4351         struct stripe_head *sh = param;
4352
4353         handle_stripe(sh);
4354         release_stripe(sh);
4355 }
4356
4357 static void process_stripe(struct stripe_head *sh, struct list_head *domain)
4358 {
4359         async_schedule_domain(__process_stripe, sh, domain);
4360 }
4361
4362 static void synchronize_stripe_processing(struct list_head *domain)
4363 {
4364         async_synchronize_full_domain(domain);
4365 }
4366 #else
4367 static void process_stripe(struct stripe_head *sh, struct list_head *domain)
4368 {
4369         handle_stripe(sh);
4370         release_stripe(sh);
4371         cond_resched();
4372 }
4373
4374 static void synchronize_stripe_processing(struct list_head *domain)
4375 {
4376 }
4377 #endif
4378
4379
4380 /*
4381  * This is our raid5 kernel thread.
4382  *
4383  * We scan the hash table for stripes which can be handled now.
4384  * During the scan, completed stripes are saved for us by the interrupt
4385  * handler, so that they will not have to wait for our next wakeup.
4386  */
4387 static void raid5d(mddev_t *mddev)
4388 {
4389         struct stripe_head *sh;
4390         raid5_conf_t *conf = mddev->private;
4391         int handled;
4392         LIST_HEAD(raid_domain);
4393
4394         pr_debug("+++ raid5d active\n");
4395
4396         md_check_recovery(mddev);
4397
4398         handled = 0;
4399         spin_lock_irq(&conf->device_lock);
4400         while (1) {
4401                 struct bio *bio;
4402
4403                 if (conf->seq_flush != conf->seq_write) {
4404                         int seq = conf->seq_flush;
4405                         spin_unlock_irq(&conf->device_lock);
4406                         bitmap_unplug(mddev->bitmap);
4407                         spin_lock_irq(&conf->device_lock);
4408                         conf->seq_write = seq;
4409                         activate_bit_delay(conf);
4410                 }
4411
4412                 while ((bio = remove_bio_from_retry(conf))) {
4413                         int ok;
4414                         spin_unlock_irq(&conf->device_lock);
4415                         ok = retry_aligned_read(conf, bio);
4416                         spin_lock_irq(&conf->device_lock);
4417                         if (!ok)
4418                                 break;
4419                         handled++;
4420                 }
4421
4422                 sh = __get_priority_stripe(conf);
4423
4424                 if (!sh)
4425                         break;
4426                 spin_unlock_irq(&conf->device_lock);
4427
4428                 handled++;
4429                 process_stripe(sh, &raid_domain);
4430
4431                 spin_lock_irq(&conf->device_lock);
4432         }
4433         pr_debug("%d stripes handled\n", handled);
4434
4435         spin_unlock_irq(&conf->device_lock);
4436
4437         synchronize_stripe_processing(&raid_domain);
4438         async_tx_issue_pending_all();
4439         unplug_slaves(mddev);
4440
4441         pr_debug("--- raid5d inactive\n");
4442 }
4443
4444 static ssize_t
4445 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4446 {
4447         raid5_conf_t *conf = mddev->private;
4448         if (conf)
4449                 return sprintf(page, "%d\n", conf->max_nr_stripes);
4450         else
4451                 return 0;
4452 }
4453
4454 static ssize_t
4455 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4456 {
4457         raid5_conf_t *conf = mddev->private;
4458         unsigned long new;
4459         int err;
4460
4461         if (len >= PAGE_SIZE)
4462                 return -EINVAL;
4463         if (!conf)
4464                 return -ENODEV;
4465
4466         if (strict_strtoul(page, 10, &new))
4467                 return -EINVAL;
4468         if (new <= 16 || new > 32768)
4469                 return -EINVAL;
4470         while (new < conf->max_nr_stripes) {
4471                 if (drop_one_stripe(conf))
4472                         conf->max_nr_stripes--;
4473                 else
4474                         break;
4475         }
4476         err = md_allow_write(mddev);
4477         if (err)
4478                 return err;
4479         while (new > conf->max_nr_stripes) {
4480                 if (grow_one_stripe(conf))
4481                         conf->max_nr_stripes++;
4482                 else break;
4483         }
4484         return len;
4485 }
4486
4487 static struct md_sysfs_entry
4488 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4489                                 raid5_show_stripe_cache_size,
4490                                 raid5_store_stripe_cache_size);
4491
4492 static ssize_t
4493 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4494 {
4495         raid5_conf_t *conf = mddev->private;
4496         if (conf)
4497                 return sprintf(page, "%d\n", conf->bypass_threshold);
4498         else
4499                 return 0;
4500 }
4501
4502 static ssize_t
4503 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4504 {
4505         raid5_conf_t *conf = mddev->private;
4506         unsigned long new;
4507         if (len >= PAGE_SIZE)
4508                 return -EINVAL;
4509         if (!conf)
4510                 return -ENODEV;
4511
4512         if (strict_strtoul(page, 10, &new))
4513                 return -EINVAL;
4514         if (new > conf->max_nr_stripes)
4515                 return -EINVAL;
4516         conf->bypass_threshold = new;
4517         return len;
4518 }
4519
4520 static struct md_sysfs_entry
4521 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4522                                         S_IRUGO | S_IWUSR,
4523                                         raid5_show_preread_threshold,
4524                                         raid5_store_preread_threshold);
4525
4526 static ssize_t
4527 stripe_cache_active_show(mddev_t *mddev, char *page)
4528 {
4529         raid5_conf_t *conf = mddev->private;
4530         if (conf)
4531                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4532         else
4533                 return 0;
4534 }
4535
4536 static struct md_sysfs_entry
4537 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4538
4539 static struct attribute *raid5_attrs[] =  {
4540         &raid5_stripecache_size.attr,
4541         &raid5_stripecache_active.attr,
4542         &raid5_preread_bypass_threshold.attr,
4543         NULL,
4544 };
4545 static struct attribute_group raid5_attrs_group = {
4546         .name = NULL,
4547         .attrs = raid5_attrs,
4548 };
4549
4550 static sector_t
4551 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4552 {
4553         raid5_conf_t *conf = mddev->private;
4554
4555         if (!sectors)
4556                 sectors = mddev->dev_sectors;
4557         if (!raid_disks) {
4558                 /* size is defined by the smallest of previous and new size */
4559                 if (conf->raid_disks < conf->previous_raid_disks)
4560                         raid_disks = conf->raid_disks;
4561                 else
4562                         raid_disks = conf->previous_raid_disks;
4563         }
4564
4565         sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4566         sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4567         return sectors * (raid_disks - conf->max_degraded);
4568 }
4569
4570 static void raid5_free_percpu(raid5_conf_t *conf)
4571 {
4572         struct raid5_percpu *percpu;
4573         unsigned long cpu;
4574
4575         if (!conf->percpu)
4576                 return;
4577
4578         get_online_cpus();
4579         for_each_possible_cpu(cpu) {
4580                 percpu = per_cpu_ptr(conf->percpu, cpu);
4581                 safe_put_page(percpu->spare_page);
4582                 kfree(percpu->scribble);
4583         }
4584 #ifdef CONFIG_HOTPLUG_CPU
4585         unregister_cpu_notifier(&conf->cpu_notify);
4586 #endif
4587         put_online_cpus();
4588
4589         free_percpu(conf->percpu);
4590 }
4591
4592 static void free_conf(raid5_conf_t *conf)
4593 {
4594         shrink_stripes(conf);
4595         raid5_free_percpu(conf);
4596         kfree(conf->disks);
4597         kfree(conf->stripe_hashtbl);
4598         kfree(conf);
4599 }
4600
4601 #ifdef CONFIG_HOTPLUG_CPU
4602 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4603                               void *hcpu)
4604 {
4605         raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
4606         long cpu = (long)hcpu;
4607         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4608
4609         switch (action) {
4610         case CPU_UP_PREPARE:
4611         case CPU_UP_PREPARE_FROZEN:
4612                 if (conf->level == 6 && !percpu->spare_page)
4613                         percpu->spare_page = alloc_page(GFP_KERNEL);
4614                 if (!percpu->scribble)
4615                         percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4616
4617                 if (!percpu->scribble ||
4618                     (conf->level == 6 && !percpu->spare_page)) {
4619                         safe_put_page(percpu->spare_page);
4620                         kfree(percpu->scribble);
4621                         pr_err("%s: failed memory allocation for cpu%ld\n",
4622                                __func__, cpu);
4623                         return NOTIFY_BAD;
4624                 }
4625                 break;
4626         case CPU_DEAD:
4627         case CPU_DEAD_FROZEN:
4628                 safe_put_page(percpu->spare_page);
4629                 kfree(percpu->scribble);
4630                 percpu->spare_page = NULL;
4631                 percpu->scribble = NULL;
4632                 break;
4633         default:
4634                 break;
4635         }
4636         return NOTIFY_OK;
4637 }
4638 #endif
4639
4640 static int raid5_alloc_percpu(raid5_conf_t *conf)
4641 {
4642         unsigned long cpu;
4643         struct page *spare_page;
4644         struct raid5_percpu *allcpus;
4645         void *scribble;
4646         int err;
4647
4648         allcpus = alloc_percpu(struct raid5_percpu);
4649         if (!allcpus)
4650                 return -ENOMEM;
4651         conf->percpu = allcpus;
4652
4653         get_online_cpus();
4654         err = 0;
4655         for_each_present_cpu(cpu) {
4656                 if (conf->level == 6) {
4657                         spare_page = alloc_page(GFP_KERNEL);
4658                         if (!spare_page) {
4659                                 err = -ENOMEM;
4660                                 break;
4661                         }
4662                         per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4663                 }
4664                 scribble = kmalloc(scribble_len(conf->raid_disks), GFP_KERNEL);
4665                 if (!scribble) {
4666                         err = -ENOMEM;
4667                         break;
4668                 }
4669                 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4670         }
4671 #ifdef CONFIG_HOTPLUG_CPU
4672         conf->cpu_notify.notifier_call = raid456_cpu_notify;
4673         conf->cpu_notify.priority = 0;
4674         if (err == 0)
4675                 err = register_cpu_notifier(&conf->cpu_notify);
4676 #endif
4677         put_online_cpus();
4678
4679         return err;
4680 }
4681
4682 static raid5_conf_t *setup_conf(mddev_t *mddev)
4683 {
4684         raid5_conf_t *conf;
4685         int raid_disk, memory;
4686         mdk_rdev_t *rdev;
4687         struct disk_info *disk;
4688
4689         if (mddev->new_level != 5
4690             && mddev->new_level != 4
4691             && mddev->new_level != 6) {
4692                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4693                        mdname(mddev), mddev->new_level);
4694                 return ERR_PTR(-EIO);
4695         }
4696         if ((mddev->new_level == 5
4697              && !algorithm_valid_raid5(mddev->new_layout)) ||
4698             (mddev->new_level == 6
4699              && !algorithm_valid_raid6(mddev->new_layout))) {
4700                 printk(KERN_ERR "raid5: %s: layout %d not supported\n",
4701                        mdname(mddev), mddev->new_layout);
4702                 return ERR_PTR(-EIO);
4703         }
4704         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4705                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4706                        mdname(mddev), mddev->raid_disks);
4707                 return ERR_PTR(-EINVAL);
4708         }
4709
4710         if (!mddev->new_chunk_sectors ||
4711             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4712             !is_power_of_2(mddev->new_chunk_sectors)) {
4713                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4714                        mddev->new_chunk_sectors << 9, mdname(mddev));
4715                 return ERR_PTR(-EINVAL);
4716         }
4717
4718         conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4719         if (conf == NULL)
4720                 goto abort;
4721
4722         conf->raid_disks = mddev->raid_disks;
4723         conf->scribble_len = scribble_len(conf->raid_disks);
4724         if (mddev->reshape_position == MaxSector)
4725                 conf->previous_raid_disks = mddev->raid_disks;
4726         else
4727                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4728
4729         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4730                               GFP_KERNEL);
4731         if (!conf->disks)
4732                 goto abort;
4733
4734         conf->mddev = mddev;
4735
4736         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4737                 goto abort;
4738
4739         conf->level = mddev->new_level;
4740         if (raid5_alloc_percpu(conf) != 0)
4741                 goto abort;
4742
4743         spin_lock_init(&conf->device_lock);
4744         init_waitqueue_head(&conf->wait_for_stripe);
4745         init_waitqueue_head(&conf->wait_for_overlap);
4746         INIT_LIST_HEAD(&conf->handle_list);
4747         INIT_LIST_HEAD(&conf->hold_list);
4748         INIT_LIST_HEAD(&conf->delayed_list);
4749         INIT_LIST_HEAD(&conf->bitmap_list);
4750         INIT_LIST_HEAD(&conf->inactive_list);
4751         atomic_set(&conf->active_stripes, 0);
4752         atomic_set(&conf->preread_active_stripes, 0);
4753         atomic_set(&conf->active_aligned_reads, 0);
4754         conf->bypass_threshold = BYPASS_THRESHOLD;
4755
4756         pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4757
4758         list_for_each_entry(rdev, &mddev->disks, same_set) {
4759                 raid_disk = rdev->raid_disk;
4760                 if (raid_disk >= conf->raid_disks
4761                     || raid_disk < 0)
4762                         continue;
4763                 disk = conf->disks + raid_disk;
4764
4765                 disk->rdev = rdev;
4766
4767                 if (test_bit(In_sync, &rdev->flags)) {
4768                         char b[BDEVNAME_SIZE];
4769                         printk(KERN_INFO "raid5: device %s operational as raid"
4770                                 " disk %d\n", bdevname(rdev->bdev,b),
4771                                 raid_disk);
4772                 } else
4773                         /* Cannot rely on bitmap to complete recovery */
4774                         conf->fullsync = 1;
4775         }
4776
4777         conf->chunk_sectors = mddev->new_chunk_sectors;
4778         conf->level = mddev->new_level;
4779         if (conf->level == 6)
4780                 conf->max_degraded = 2;
4781         else
4782                 conf->max_degraded = 1;
4783         conf->algorithm = mddev->new_layout;
4784         conf->max_nr_stripes = NR_STRIPES;
4785         conf->reshape_progress = mddev->reshape_position;
4786         if (conf->reshape_progress != MaxSector) {
4787                 conf->prev_chunk_sectors = mddev->chunk_sectors;
4788                 conf->prev_algo = mddev->layout;
4789         }
4790
4791         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4792                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4793         if (grow_stripes(conf, conf->max_nr_stripes)) {
4794                 printk(KERN_ERR
4795                         "raid5: couldn't allocate %dkB for buffers\n", memory);
4796                 goto abort;
4797         } else
4798                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4799                         memory, mdname(mddev));
4800
4801         conf->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4802         if (!conf->thread) {
4803                 printk(KERN_ERR
4804                        "raid5: couldn't allocate thread for %s\n",
4805                        mdname(mddev));
4806                 goto abort;
4807         }
4808
4809         return conf;
4810
4811  abort:
4812         if (conf) {
4813                 free_conf(conf);
4814                 return ERR_PTR(-EIO);
4815         } else
4816                 return ERR_PTR(-ENOMEM);
4817 }
4818
4819 static int run(mddev_t *mddev)
4820 {
4821         raid5_conf_t *conf;
4822         int working_disks = 0, chunk_size;
4823         mdk_rdev_t *rdev;
4824
4825         if (mddev->recovery_cp != MaxSector)
4826                 printk(KERN_NOTICE "raid5: %s is not clean"
4827                        " -- starting background reconstruction\n",
4828                        mdname(mddev));
4829         if (mddev->reshape_position != MaxSector) {
4830                 /* Check that we can continue the reshape.
4831                  * Currently only disks can change, it must
4832                  * increase, and we must be past the point where
4833                  * a stripe over-writes itself
4834                  */
4835                 sector_t here_new, here_old;
4836                 int old_disks;
4837                 int max_degraded = (mddev->level == 6 ? 2 : 1);
4838
4839                 if (mddev->new_level != mddev->level) {
4840                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4841                                "required - aborting.\n",
4842                                mdname(mddev));
4843                         return -EINVAL;
4844                 }
4845                 old_disks = mddev->raid_disks - mddev->delta_disks;
4846                 /* reshape_position must be on a new-stripe boundary, and one
4847                  * further up in new geometry must map after here in old
4848                  * geometry.
4849                  */
4850                 here_new = mddev->reshape_position;
4851                 if (sector_div(here_new, mddev->new_chunk_sectors *
4852                                (mddev->raid_disks - max_degraded))) {
4853                         printk(KERN_ERR "raid5: reshape_position not "
4854                                "on a stripe boundary\n");
4855                         return -EINVAL;
4856                 }
4857                 /* here_new is the stripe we will write to */
4858                 here_old = mddev->reshape_position;
4859                 sector_div(here_old, mddev->chunk_sectors *
4860                            (old_disks-max_degraded));
4861                 /* here_old is the first stripe that we might need to read
4862                  * from */
4863                 if (mddev->delta_disks == 0) {
4864                         /* We cannot be sure it is safe to start an in-place
4865                          * reshape.  It is only safe if user-space if monitoring
4866                          * and taking constant backups.
4867                          * mdadm always starts a situation like this in
4868                          * readonly mode so it can take control before
4869                          * allowing any writes.  So just check for that.
4870                          */
4871                         if ((here_new * mddev->new_chunk_sectors !=
4872                              here_old * mddev->chunk_sectors) ||
4873                             mddev->ro == 0) {
4874                                 printk(KERN_ERR "raid5: in-place reshape must be started"
4875                                        " in read-only mode - aborting\n");
4876                                 return -EINVAL;
4877                         }
4878                 } else if (mddev->delta_disks < 0
4879                     ? (here_new * mddev->new_chunk_sectors <=
4880                        here_old * mddev->chunk_sectors)
4881                     : (here_new * mddev->new_chunk_sectors >=
4882                        here_old * mddev->chunk_sectors)) {
4883                         /* Reading from the same stripe as writing to - bad */
4884                         printk(KERN_ERR "raid5: reshape_position too early for "
4885                                "auto-recovery - aborting.\n");
4886                         return -EINVAL;
4887                 }
4888                 printk(KERN_INFO "raid5: reshape will continue\n");
4889                 /* OK, we should be able to continue; */
4890         } else {
4891                 BUG_ON(mddev->level != mddev->new_level);
4892                 BUG_ON(mddev->layout != mddev->new_layout);
4893                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4894                 BUG_ON(mddev->delta_disks != 0);
4895         }
4896
4897         if (mddev->private == NULL)
4898                 conf = setup_conf(mddev);
4899         else
4900                 conf = mddev->private;
4901
4902         if (IS_ERR(conf))
4903                 return PTR_ERR(conf);
4904
4905         mddev->thread = conf->thread;
4906         conf->thread = NULL;
4907         mddev->private = conf;
4908
4909         /*
4910          * 0 for a fully functional array, 1 or 2 for a degraded array.
4911          */
4912         list_for_each_entry(rdev, &mddev->disks, same_set)
4913                 if (rdev->raid_disk >= 0 &&
4914                     test_bit(In_sync, &rdev->flags))
4915                         working_disks++;
4916
4917         mddev->degraded = conf->raid_disks - working_disks;
4918
4919         if (mddev->degraded > conf->max_degraded) {
4920                 printk(KERN_ERR "raid5: not enough operational devices for %s"
4921                         " (%d/%d failed)\n",
4922                         mdname(mddev), mddev->degraded, conf->raid_disks);
4923                 goto abort;
4924         }
4925
4926         /* device size must be a multiple of chunk size */
4927         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
4928         mddev->resync_max_sectors = mddev->dev_sectors;
4929
4930         if (mddev->degraded > 0 &&
4931             mddev->recovery_cp != MaxSector) {
4932                 if (mddev->ok_start_degraded)
4933                         printk(KERN_WARNING
4934                                "raid5: starting dirty degraded array: %s"
4935                                "- data corruption possible.\n",
4936                                mdname(mddev));
4937                 else {
4938                         printk(KERN_ERR
4939                                "raid5: cannot start dirty degraded array for %s\n",
4940                                mdname(mddev));
4941                         goto abort;
4942                 }
4943         }
4944
4945         if (mddev->degraded == 0)
4946                 printk("raid5: raid level %d set %s active with %d out of %d"
4947                        " devices, algorithm %d\n", conf->level, mdname(mddev),
4948                        mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4949                        mddev->new_layout);
4950         else
4951                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4952                         " out of %d devices, algorithm %d\n", conf->level,
4953                         mdname(mddev), mddev->raid_disks - mddev->degraded,
4954                         mddev->raid_disks, mddev->new_layout);
4955
4956         print_raid5_conf(conf);
4957
4958         if (conf->reshape_progress != MaxSector) {
4959                 printk("...ok start reshape thread\n");
4960                 conf->reshape_safe = conf->reshape_progress;
4961                 atomic_set(&conf->reshape_stripes, 0);
4962                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4963                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4964                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4965                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4966                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4967                                                         "%s_reshape");
4968         }
4969
4970         /* read-ahead size must cover two whole stripes, which is
4971          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4972          */
4973         {
4974                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4975                 int stripe = data_disks *
4976                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
4977                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4978                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4979         }
4980
4981         /* Ok, everything is just fine now */
4982         if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4983                 printk(KERN_WARNING
4984                        "raid5: failed to create sysfs attributes for %s\n",
4985                        mdname(mddev));
4986
4987         mddev->queue->queue_lock = &conf->device_lock;
4988
4989         mddev->queue->unplug_fn = raid5_unplug_device;
4990         mddev->queue->backing_dev_info.congested_data = mddev;
4991         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4992
4993         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4994
4995         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4996         chunk_size = mddev->chunk_sectors << 9;
4997         blk_queue_io_min(mddev->queue, chunk_size);
4998         blk_queue_io_opt(mddev->queue, chunk_size *
4999                          (conf->raid_disks - conf->max_degraded));
5000
5001         list_for_each_entry(rdev, &mddev->disks, same_set)
5002                 disk_stack_limits(mddev->gendisk, rdev->bdev,
5003                                   rdev->data_offset << 9);
5004
5005         return 0;
5006 abort:
5007         md_unregister_thread(mddev->thread);
5008         mddev->thread = NULL;
5009         if (conf) {
5010                 print_raid5_conf(conf);
5011                 free_conf(conf);
5012         }
5013         mddev->private = NULL;
5014         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
5015         return -EIO;
5016 }
5017
5018
5019
5020 static int stop(mddev_t *mddev)
5021 {
5022         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
5023
5024         md_unregister_thread(mddev->thread);
5025         mddev->thread = NULL;
5026         mddev->queue->backing_dev_info.congested_fn = NULL;
5027         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
5028         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
5029         free_conf(conf);
5030         mddev->private = NULL;
5031         return 0;
5032 }
5033
5034 #ifdef DEBUG
5035 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
5036 {
5037         int i;
5038
5039         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
5040                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
5041         seq_printf(seq, "sh %llu,  count %d.\n",
5042                    (unsigned long long)sh->sector, atomic_read(&sh->count));
5043         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
5044         for (i = 0; i < sh->disks; i++) {
5045                 seq_printf(seq, "(cache%d: %p %ld) ",
5046                            i, sh->dev[i].page, sh->dev[i].flags);
5047         }
5048         seq_printf(seq, "\n");
5049 }
5050
5051 static void printall(struct seq_file *seq, raid5_conf_t *conf)
5052 {
5053         struct stripe_head *sh;
5054         struct hlist_node *hn;
5055         int i;
5056
5057         spin_lock_irq(&conf->device_lock);
5058         for (i = 0; i < NR_HASH; i++) {
5059                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
5060                         if (sh->raid_conf != conf)
5061                                 continue;
5062                         print_sh(seq, sh);
5063                 }
5064         }
5065         spin_unlock_irq(&conf->device_lock);
5066 }
5067 #endif
5068
5069 static void status(struct seq_file *seq, mddev_t *mddev)
5070 {
5071         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
5072         int i;
5073
5074         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5075                 mddev->chunk_sectors / 2, mddev->layout);
5076         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5077         for (i = 0; i < conf->raid_disks; i++)
5078                 seq_printf (seq, "%s",
5079                                conf->disks[i].rdev &&
5080                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5081         seq_printf (seq, "]");
5082 #ifdef DEBUG
5083         seq_printf (seq, "\n");
5084         printall(seq, conf);
5085 #endif
5086 }
5087
5088 static void print_raid5_conf (raid5_conf_t *conf)
5089 {
5090         int i;
5091         struct disk_info *tmp;
5092
5093         printk("RAID5 conf printout:\n");
5094         if (!conf) {
5095                 printk("(conf==NULL)\n");
5096                 return;
5097         }
5098         printk(" --- rd:%d wd:%d\n", conf->raid_disks,
5099                  conf->raid_disks - conf->mddev->degraded);
5100
5101         for (i = 0; i < conf->raid_disks; i++) {
5102                 char b[BDEVNAME_SIZE];
5103                 tmp = conf->disks + i;
5104                 if (tmp->rdev)
5105                 printk(" disk %d, o:%d, dev:%s\n",
5106                         i, !test_bit(Faulty, &tmp->rdev->flags),
5107                         bdevname(tmp->rdev->bdev,b));
5108         }
5109 }
5110
5111 static int raid5_spare_active(mddev_t *mddev)
5112 {
5113         int i;
5114         raid5_conf_t *conf = mddev->private;
5115         struct disk_info *tmp;
5116
5117         for (i = 0; i < conf->raid_disks; i++) {
5118                 tmp = conf->disks + i;
5119                 if (tmp->rdev
5120                     && !test_bit(Faulty, &tmp->rdev->flags)
5121                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5122                         unsigned long flags;
5123                         spin_lock_irqsave(&conf->device_lock, flags);
5124                         mddev->degraded--;
5125                         spin_unlock_irqrestore(&conf->device_lock, flags);
5126                 }
5127         }
5128         print_raid5_conf(conf);
5129         return 0;
5130 }
5131
5132 static int raid5_remove_disk(mddev_t *mddev, int number)
5133 {
5134         raid5_conf_t *conf = mddev->private;
5135         int err = 0;
5136         mdk_rdev_t *rdev;
5137         struct disk_info *p = conf->disks + number;
5138
5139         print_raid5_conf(conf);
5140         rdev = p->rdev;
5141         if (rdev) {
5142                 if (number >= conf->raid_disks &&
5143                     conf->reshape_progress == MaxSector)
5144                         clear_bit(In_sync, &rdev->flags);
5145
5146                 if (test_bit(In_sync, &rdev->flags) ||
5147                     atomic_read(&rdev->nr_pending)) {
5148                         err = -EBUSY;
5149                         goto abort;
5150                 }
5151                 /* Only remove non-faulty devices if recovery
5152                  * isn't possible.
5153                  */
5154                 if (!test_bit(Faulty, &rdev->flags) &&
5155                     mddev->degraded <= conf->max_degraded &&
5156                     number < conf->raid_disks) {
5157                         err = -EBUSY;
5158                         goto abort;
5159                 }
5160                 p->rdev = NULL;
5161                 synchronize_rcu();
5162                 if (atomic_read(&rdev->nr_pending)) {
5163                         /* lost the race, try later */
5164                         err = -EBUSY;
5165                         p->rdev = rdev;
5166                 }
5167         }
5168 abort:
5169
5170         print_raid5_conf(conf);
5171         return err;
5172 }
5173
5174 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
5175 {
5176         raid5_conf_t *conf = mddev->private;
5177         int err = -EEXIST;
5178         int disk;
5179         struct disk_info *p;
5180         int first = 0;
5181         int last = conf->raid_disks - 1;
5182
5183         if (mddev->degraded > conf->max_degraded)
5184                 /* no point adding a device */
5185                 return -EINVAL;
5186
5187         if (rdev->raid_disk >= 0)
5188                 first = last = rdev->raid_disk;
5189
5190         /*
5191          * find the disk ... but prefer rdev->saved_raid_disk
5192          * if possible.
5193          */
5194         if (rdev->saved_raid_disk >= 0 &&
5195             rdev->saved_raid_disk >= first &&
5196             conf->disks[rdev->saved_raid_disk].rdev == NULL)
5197                 disk = rdev->saved_raid_disk;
5198         else
5199                 disk = first;
5200         for ( ; disk <= last ; disk++)
5201                 if ((p=conf->disks + disk)->rdev == NULL) {
5202                         clear_bit(In_sync, &rdev->flags);
5203                         rdev->raid_disk = disk;
5204                         err = 0;
5205                         if (rdev->saved_raid_disk != disk)
5206                                 conf->fullsync = 1;
5207                         rcu_assign_pointer(p->rdev, rdev);
5208                         break;
5209                 }
5210         print_raid5_conf(conf);
5211         return err;
5212 }
5213
5214 static int raid5_resize(mddev_t *mddev, sector_t sectors)
5215 {
5216         /* no resync is happening, and there is enough space
5217          * on all devices, so we can resize.
5218          * We need to make sure resync covers any new space.
5219          * If the array is shrinking we should possibly wait until
5220          * any io in the removed space completes, but it hardly seems
5221          * worth it.
5222          */
5223         sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5224         md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5225                                                mddev->raid_disks));
5226         if (mddev->array_sectors >
5227             raid5_size(mddev, sectors, mddev->raid_disks))
5228                 return -EINVAL;
5229         set_capacity(mddev->gendisk, mddev->array_sectors);
5230         mddev->changed = 1;
5231         revalidate_disk(mddev->gendisk);
5232         if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) {
5233                 mddev->recovery_cp = mddev->dev_sectors;
5234                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5235         }
5236         mddev->dev_sectors = sectors;
5237         mddev->resync_max_sectors = sectors;
5238         return 0;
5239 }
5240
5241 static int check_stripe_cache(mddev_t *mddev)
5242 {
5243         /* Can only proceed if there are plenty of stripe_heads.
5244          * We need a minimum of one full stripe,, and for sensible progress
5245          * it is best to have about 4 times that.
5246          * If we require 4 times, then the default 256 4K stripe_heads will
5247          * allow for chunk sizes up to 256K, which is probably OK.
5248          * If the chunk size is greater, user-space should request more
5249          * stripe_heads first.
5250          */
5251         raid5_conf_t *conf = mddev->private;
5252         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5253             > conf->max_nr_stripes ||
5254             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5255             > conf->max_nr_stripes) {
5256                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
5257                        ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5258                         / STRIPE_SIZE)*4);
5259                 return 0;
5260         }
5261         return 1;
5262 }
5263
5264 static int check_reshape(mddev_t *mddev)
5265 {
5266         raid5_conf_t *conf = mddev->private;
5267
5268         if (mddev->delta_disks == 0 &&
5269             mddev->new_layout == mddev->layout &&
5270             mddev->new_chunk_sectors == mddev->chunk_sectors)
5271                 return 0; /* nothing to do */
5272         if (mddev->bitmap)
5273                 /* Cannot grow a bitmap yet */
5274                 return -EBUSY;
5275         if (mddev->degraded > conf->max_degraded)
5276                 return -EINVAL;
5277         if (mddev->delta_disks < 0) {
5278                 /* We might be able to shrink, but the devices must
5279                  * be made bigger first.
5280                  * For raid6, 4 is the minimum size.
5281                  * Otherwise 2 is the minimum
5282                  */
5283                 int min = 2;
5284                 if (mddev->level == 6)
5285                         min = 4;
5286                 if (mddev->raid_disks + mddev->delta_disks < min)
5287                         return -EINVAL;
5288         }
5289
5290         if (!check_stripe_cache(mddev))
5291                 return -ENOSPC;
5292
5293         return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5294 }
5295
5296 static int raid5_start_reshape(mddev_t *mddev)
5297 {
5298         raid5_conf_t *conf = mddev->private;
5299         mdk_rdev_t *rdev;
5300         int spares = 0;
5301         int added_devices = 0;
5302         unsigned long flags;
5303
5304         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5305                 return -EBUSY;
5306
5307         if (!check_stripe_cache(mddev))
5308                 return -ENOSPC;
5309
5310         list_for_each_entry(rdev, &mddev->disks, same_set)
5311                 if (rdev->raid_disk < 0 &&
5312                     !test_bit(Faulty, &rdev->flags))
5313                         spares++;
5314
5315         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5316                 /* Not enough devices even to make a degraded array
5317                  * of that size
5318                  */
5319                 return -EINVAL;
5320
5321         /* Refuse to reduce size of the array.  Any reductions in
5322          * array size must be through explicit setting of array_size
5323          * attribute.
5324          */
5325         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5326             < mddev->array_sectors) {
5327                 printk(KERN_ERR "md: %s: array size must be reduced "
5328                        "before number of disks\n", mdname(mddev));
5329                 return -EINVAL;
5330         }
5331
5332         atomic_set(&conf->reshape_stripes, 0);
5333         spin_lock_irq(&conf->device_lock);
5334         conf->previous_raid_disks = conf->raid_disks;
5335         conf->raid_disks += mddev->delta_disks;
5336         conf->prev_chunk_sectors = conf->chunk_sectors;
5337         conf->chunk_sectors = mddev->new_chunk_sectors;
5338         conf->prev_algo = conf->algorithm;
5339         conf->algorithm = mddev->new_layout;
5340         if (mddev->delta_disks < 0)
5341                 conf->reshape_progress = raid5_size(mddev, 0, 0);
5342         else
5343                 conf->reshape_progress = 0;
5344         conf->reshape_safe = conf->reshape_progress;
5345         conf->generation++;
5346         spin_unlock_irq(&conf->device_lock);
5347
5348         /* Add some new drives, as many as will fit.
5349          * We know there are enough to make the newly sized array work.
5350          */
5351         list_for_each_entry(rdev, &mddev->disks, same_set)
5352                 if (rdev->raid_disk < 0 &&
5353                     !test_bit(Faulty, &rdev->flags)) {
5354                         if (raid5_add_disk(mddev, rdev) == 0) {
5355                                 char nm[20];
5356                                 set_bit(In_sync, &rdev->flags);
5357                                 added_devices++;
5358                                 rdev->recovery_offset = 0;
5359                                 sprintf(nm, "rd%d", rdev->raid_disk);
5360                                 if (sysfs_create_link(&mddev->kobj,
5361                                                       &rdev->kobj, nm))
5362                                         printk(KERN_WARNING
5363                                                "raid5: failed to create "
5364                                                " link %s for %s\n",
5365                                                nm, mdname(mddev));
5366                         } else
5367                                 break;
5368                 }
5369
5370         if (mddev->delta_disks > 0) {
5371                 spin_lock_irqsave(&conf->device_lock, flags);
5372                 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks)
5373                         - added_devices;
5374                 spin_unlock_irqrestore(&conf->device_lock, flags);
5375         }
5376         mddev->raid_disks = conf->raid_disks;
5377         mddev->reshape_position = conf->reshape_progress;
5378         set_bit(MD_CHANGE_DEVS, &mddev->flags);
5379
5380         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5381         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5382         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5383         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5384         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5385                                                 "%s_reshape");
5386         if (!mddev->sync_thread) {
5387                 mddev->recovery = 0;
5388                 spin_lock_irq(&conf->device_lock);
5389                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5390                 conf->reshape_progress = MaxSector;
5391                 spin_unlock_irq(&conf->device_lock);
5392                 return -EAGAIN;
5393         }
5394         conf->reshape_checkpoint = jiffies;
5395         md_wakeup_thread(mddev->sync_thread);
5396         md_new_event(mddev);
5397         return 0;
5398 }
5399
5400 /* This is called from the reshape thread and should make any
5401  * changes needed in 'conf'
5402  */
5403 static void end_reshape(raid5_conf_t *conf)
5404 {
5405
5406         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5407
5408                 spin_lock_irq(&conf->device_lock);
5409                 conf->previous_raid_disks = conf->raid_disks;
5410                 conf->reshape_progress = MaxSector;
5411                 spin_unlock_irq(&conf->device_lock);
5412                 wake_up(&conf->wait_for_overlap);
5413
5414                 /* read-ahead size must cover two whole stripes, which is
5415                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5416                  */
5417                 {
5418                         int data_disks = conf->raid_disks - conf->max_degraded;
5419                         int stripe = data_disks * ((conf->chunk_sectors << 9)
5420                                                    / PAGE_SIZE);
5421                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5422                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5423                 }
5424         }
5425 }
5426
5427 /* This is called from the raid5d thread with mddev_lock held.
5428  * It makes config changes to the device.
5429  */
5430 static void raid5_finish_reshape(mddev_t *mddev)
5431 {
5432         raid5_conf_t *conf = mddev->private;
5433
5434         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5435
5436                 if (mddev->delta_disks > 0) {
5437                         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5438                         set_capacity(mddev->gendisk, mddev->array_sectors);
5439                         mddev->changed = 1;
5440                         revalidate_disk(mddev->gendisk);
5441                 } else {
5442                         int d;
5443                         mddev->degraded = conf->raid_disks;
5444                         for (d = 0; d < conf->raid_disks ; d++)
5445                                 if (conf->disks[d].rdev &&
5446                                     test_bit(In_sync,
5447                                              &conf->disks[d].rdev->flags))
5448                                         mddev->degraded--;
5449                         for (d = conf->raid_disks ;
5450                              d < conf->raid_disks - mddev->delta_disks;
5451                              d++) {
5452                                 mdk_rdev_t *rdev = conf->disks[d].rdev;
5453                                 if (rdev && raid5_remove_disk(mddev, d) == 0) {
5454                                         char nm[20];
5455                                         sprintf(nm, "rd%d", rdev->raid_disk);
5456                                         sysfs_remove_link(&mddev->kobj, nm);
5457                                         rdev->raid_disk = -1;
5458                                 }
5459                         }
5460                 }
5461                 mddev->layout = conf->algorithm;
5462                 mddev->chunk_sectors = conf->chunk_sectors;
5463                 mddev->reshape_position = MaxSector;
5464                 mddev->delta_disks = 0;
5465         }
5466 }
5467
5468 static void raid5_quiesce(mddev_t *mddev, int state)
5469 {
5470         raid5_conf_t *conf = mddev->private;
5471
5472         switch(state) {
5473         case 2: /* resume for a suspend */
5474                 wake_up(&conf->wait_for_overlap);
5475                 break;
5476
5477         case 1: /* stop all writes */
5478                 spin_lock_irq(&conf->device_lock);
5479                 /* '2' tells resync/reshape to pause so that all
5480                  * active stripes can drain
5481                  */
5482                 conf->quiesce = 2;
5483                 wait_event_lock_irq(conf->wait_for_stripe,
5484                                     atomic_read(&conf->active_stripes) == 0 &&
5485                                     atomic_read(&conf->active_aligned_reads) == 0,
5486                                     conf->device_lock, /* nothing */);
5487                 conf->quiesce = 1;
5488                 spin_unlock_irq(&conf->device_lock);
5489                 /* allow reshape to continue */
5490                 wake_up(&conf->wait_for_overlap);
5491                 break;
5492
5493         case 0: /* re-enable writes */
5494                 spin_lock_irq(&conf->device_lock);
5495                 conf->quiesce = 0;
5496                 wake_up(&conf->wait_for_stripe);
5497                 wake_up(&conf->wait_for_overlap);
5498                 spin_unlock_irq(&conf->device_lock);
5499                 break;
5500         }
5501 }
5502
5503
5504 static void *raid5_takeover_raid1(mddev_t *mddev)
5505 {
5506         int chunksect;
5507
5508         if (mddev->raid_disks != 2 ||
5509             mddev->degraded > 1)
5510                 return ERR_PTR(-EINVAL);
5511
5512         /* Should check if there are write-behind devices? */
5513
5514         chunksect = 64*2; /* 64K by default */
5515
5516         /* The array must be an exact multiple of chunksize */
5517         while (chunksect && (mddev->array_sectors & (chunksect-1)))
5518                 chunksect >>= 1;
5519
5520         if ((chunksect<<9) < STRIPE_SIZE)
5521                 /* array size does not allow a suitable chunk size */
5522                 return ERR_PTR(-EINVAL);
5523
5524         mddev->new_level = 5;
5525         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5526         mddev->new_chunk_sectors = chunksect;
5527
5528         return setup_conf(mddev);
5529 }
5530
5531 static void *raid5_takeover_raid6(mddev_t *mddev)
5532 {
5533         int new_layout;
5534
5535         switch (mddev->layout) {
5536         case ALGORITHM_LEFT_ASYMMETRIC_6:
5537                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5538                 break;
5539         case ALGORITHM_RIGHT_ASYMMETRIC_6:
5540                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5541                 break;
5542         case ALGORITHM_LEFT_SYMMETRIC_6:
5543                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5544                 break;
5545         case ALGORITHM_RIGHT_SYMMETRIC_6:
5546                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5547                 break;
5548         case ALGORITHM_PARITY_0_6:
5549                 new_layout = ALGORITHM_PARITY_0;
5550                 break;
5551         case ALGORITHM_PARITY_N:
5552                 new_layout = ALGORITHM_PARITY_N;
5553                 break;
5554         default:
5555                 return ERR_PTR(-EINVAL);
5556         }
5557         mddev->new_level = 5;
5558         mddev->new_layout = new_layout;
5559         mddev->delta_disks = -1;
5560         mddev->raid_disks -= 1;
5561         return setup_conf(mddev);
5562 }
5563
5564
5565 static int raid5_check_reshape(mddev_t *mddev)
5566 {
5567         /* For a 2-drive array, the layout and chunk size can be changed
5568          * immediately as not restriping is needed.
5569          * For larger arrays we record the new value - after validation
5570          * to be used by a reshape pass.
5571          */
5572         raid5_conf_t *conf = mddev->private;
5573         int new_chunk = mddev->new_chunk_sectors;
5574
5575         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5576                 return -EINVAL;
5577         if (new_chunk > 0) {
5578                 if (!is_power_of_2(new_chunk))
5579                         return -EINVAL;
5580                 if (new_chunk < (PAGE_SIZE>>9))
5581                         return -EINVAL;
5582                 if (mddev->array_sectors & (new_chunk-1))
5583                         /* not factor of array size */
5584                         return -EINVAL;
5585         }
5586
5587         /* They look valid */
5588
5589         if (mddev->raid_disks == 2) {
5590                 /* can make the change immediately */
5591                 if (mddev->new_layout >= 0) {
5592                         conf->algorithm = mddev->new_layout;
5593                         mddev->layout = mddev->new_layout;
5594                 }
5595                 if (new_chunk > 0) {
5596                         conf->chunk_sectors = new_chunk ;
5597                         mddev->chunk_sectors = new_chunk;
5598                 }
5599                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5600                 md_wakeup_thread(mddev->thread);
5601         }
5602         return check_reshape(mddev);
5603 }
5604
5605 static int raid6_check_reshape(mddev_t *mddev)
5606 {
5607         int new_chunk = mddev->new_chunk_sectors;
5608
5609         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5610                 return -EINVAL;
5611         if (new_chunk > 0) {
5612                 if (!is_power_of_2(new_chunk))
5613                         return -EINVAL;
5614                 if (new_chunk < (PAGE_SIZE >> 9))
5615                         return -EINVAL;
5616                 if (mddev->array_sectors & (new_chunk-1))
5617                         /* not factor of array size */
5618                         return -EINVAL;
5619         }
5620
5621         /* They look valid */
5622         return check_reshape(mddev);
5623 }
5624
5625 static void *raid5_takeover(mddev_t *mddev)
5626 {
5627         /* raid5 can take over:
5628          *  raid0 - if all devices are the same - make it a raid4 layout
5629          *  raid1 - if there are two drives.  We need to know the chunk size
5630          *  raid4 - trivial - just use a raid4 layout.
5631          *  raid6 - Providing it is a *_6 layout
5632          */
5633
5634         if (mddev->level == 1)
5635                 return raid5_takeover_raid1(mddev);
5636         if (mddev->level == 4) {
5637                 mddev->new_layout = ALGORITHM_PARITY_N;
5638                 mddev->new_level = 5;
5639                 return setup_conf(mddev);
5640         }
5641         if (mddev->level == 6)
5642                 return raid5_takeover_raid6(mddev);
5643
5644         return ERR_PTR(-EINVAL);
5645 }
5646
5647
5648 static struct mdk_personality raid5_personality;
5649
5650 static void *raid6_takeover(mddev_t *mddev)
5651 {
5652         /* Currently can only take over a raid5.  We map the
5653          * personality to an equivalent raid6 personality
5654          * with the Q block at the end.
5655          */
5656         int new_layout;
5657
5658         if (mddev->pers != &raid5_personality)
5659                 return ERR_PTR(-EINVAL);
5660         if (mddev->degraded > 1)
5661                 return ERR_PTR(-EINVAL);
5662         if (mddev->raid_disks > 253)
5663                 return ERR_PTR(-EINVAL);
5664         if (mddev->raid_disks < 3)
5665                 return ERR_PTR(-EINVAL);
5666
5667         switch (mddev->layout) {
5668         case ALGORITHM_LEFT_ASYMMETRIC:
5669                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5670                 break;
5671         case ALGORITHM_RIGHT_ASYMMETRIC:
5672                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5673                 break;
5674         case ALGORITHM_LEFT_SYMMETRIC:
5675                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5676                 break;
5677         case ALGORITHM_RIGHT_SYMMETRIC:
5678                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5679                 break;
5680         case ALGORITHM_PARITY_0:
5681                 new_layout = ALGORITHM_PARITY_0_6;
5682                 break;
5683         case ALGORITHM_PARITY_N:
5684                 new_layout = ALGORITHM_PARITY_N;
5685                 break;
5686         default:
5687                 return ERR_PTR(-EINVAL);
5688         }
5689         mddev->new_level = 6;
5690         mddev->new_layout = new_layout;
5691         mddev->delta_disks = 1;
5692         mddev->raid_disks += 1;
5693         return setup_conf(mddev);
5694 }
5695
5696
5697 static struct mdk_personality raid6_personality =
5698 {
5699         .name           = "raid6",
5700         .level          = 6,
5701         .owner          = THIS_MODULE,
5702         .make_request   = make_request,
5703         .run            = run,
5704         .stop           = stop,
5705         .status         = status,
5706         .error_handler  = error,
5707         .hot_add_disk   = raid5_add_disk,
5708         .hot_remove_disk= raid5_remove_disk,
5709         .spare_active   = raid5_spare_active,
5710         .sync_request   = sync_request,
5711         .resize         = raid5_resize,
5712         .size           = raid5_size,
5713         .check_reshape  = raid6_check_reshape,
5714         .start_reshape  = raid5_start_reshape,
5715         .finish_reshape = raid5_finish_reshape,
5716         .quiesce        = raid5_quiesce,
5717         .takeover       = raid6_takeover,
5718 };
5719 static struct mdk_personality raid5_personality =
5720 {
5721         .name           = "raid5",
5722         .level          = 5,
5723         .owner          = THIS_MODULE,
5724         .make_request   = make_request,
5725         .run            = run,
5726         .stop           = stop,
5727         .status         = status,
5728         .error_handler  = error,
5729         .hot_add_disk   = raid5_add_disk,
5730         .hot_remove_disk= raid5_remove_disk,
5731         .spare_active   = raid5_spare_active,
5732         .sync_request   = sync_request,
5733         .resize         = raid5_resize,
5734         .size           = raid5_size,
5735         .check_reshape  = raid5_check_reshape,
5736         .start_reshape  = raid5_start_reshape,
5737         .finish_reshape = raid5_finish_reshape,
5738         .quiesce        = raid5_quiesce,
5739         .takeover       = raid5_takeover,
5740 };
5741
5742 static struct mdk_personality raid4_personality =
5743 {
5744         .name           = "raid4",
5745         .level          = 4,
5746         .owner          = THIS_MODULE,
5747         .make_request   = make_request,
5748         .run            = run,
5749         .stop           = stop,
5750         .status         = status,
5751         .error_handler  = error,
5752         .hot_add_disk   = raid5_add_disk,
5753         .hot_remove_disk= raid5_remove_disk,
5754         .spare_active   = raid5_spare_active,
5755         .sync_request   = sync_request,
5756         .resize         = raid5_resize,
5757         .size           = raid5_size,
5758         .check_reshape  = raid5_check_reshape,
5759         .start_reshape  = raid5_start_reshape,
5760         .finish_reshape = raid5_finish_reshape,
5761         .quiesce        = raid5_quiesce,
5762 };
5763
5764 static int __init raid5_init(void)
5765 {
5766         register_md_personality(&raid6_personality);
5767         register_md_personality(&raid5_personality);
5768         register_md_personality(&raid4_personality);
5769         return 0;
5770 }
5771
5772 static void raid5_exit(void)
5773 {
5774         unregister_md_personality(&raid6_personality);
5775         unregister_md_personality(&raid5_personality);
5776         unregister_md_personality(&raid4_personality);
5777 }
5778
5779 module_init(raid5_init);
5780 module_exit(raid5_exit);
5781 MODULE_LICENSE("GPL");
5782 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5783 MODULE_ALIAS("md-raid5");
5784 MODULE_ALIAS("md-raid4");
5785 MODULE_ALIAS("md-level-5");
5786 MODULE_ALIAS("md-level-4");
5787 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5788 MODULE_ALIAS("md-raid6");
5789 MODULE_ALIAS("md-level-6");
5790
5791 /* This used to be two separate modules, they were: */
5792 MODULE_ALIAS("raid5");
5793 MODULE_ALIAS("raid6");