drivers/md/raid6main.c

   1 /*
   2  * raid6main.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-6 management functions.  This code is derived from raid5.c.
   8  * Last merge from raid5.c bkcvs version 1.79 (kernel 2.6.1).
   9  *
  10  * Thanks to Penguin Computing for making the RAID-6 development possible
  11  * by donating a test server!
  12  *
  13  * This program is free software; you can redistribute it and/or modify
  14  * it under the terms of the GNU General Public License as published by
  15  * the Free Software Foundation; either version 2, or (at your option)
  16  * any later version.
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * (for example /usr/src/linux/COPYING); if not, write to the Free
  20  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  21  */
  22
  23
  24 #include <linux/config.h>
  25 #include <linux/module.h>
  26 #include <linux/slab.h>
  27 #include <linux/highmem.h>
  28 #include <linux/bitops.h>
  29 #include <asm/atomic.h>
  30 #include "raid6.h"
  31
  32 /*
  33  * Stripe cache
  34  */
  35
  36 #define NR_STRIPES              256
  37 #define STRIPE_SIZE             PAGE_SIZE
  38 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
  39 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
  40 #define IO_THRESHOLD            1
  41 #define HASH_PAGES              1
  42 #define HASH_PAGES_ORDER        0
  43 #define NR_HASH                 (HASH_PAGES * PAGE_SIZE / sizeof(struct stripe_head *))
  44 #define HASH_MASK               (NR_HASH - 1)
  45
  46 #define stripe_hash(conf, sect) ((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK])
  47
  48 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
  49  * order without overlap.  There may be several bio's per stripe+device, and
  50  * a bio could span several devices.
  51  * When walking this list for a particular stripe+device, we must never proceed
  52  * beyond a bio that extends past this device, as the next bio might no longer
  53  * be valid.
  54  * This macro is used to determine the 'next' bio in the list, given the sector
  55  * of the current stripe+device
  56  */
  57 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
  58 /*
  59  * The following can be used to debug the driver
  60  */
  61 #define RAID6_DEBUG     0       /* Extremely verbose printk */
  62 #define RAID6_PARANOIA  1       /* Check spinlocks */
  63 #define RAID6_DUMPSTATE 0       /* Include stripe cache state in /proc/mdstat */
  64 #if RAID6_PARANOIA && defined(CONFIG_SMP)
  65 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
  66 #else
  67 # define CHECK_DEVLOCK()
  68 #endif
  69
  70 #define PRINTK(x...) ((void)(RAID6_DEBUG && printk(KERN_DEBUG x)))
  71 #if RAID6_DEBUG
  72 #undef inline
  73 #undef __inline__
  74 #define inline
  75 #define __inline__
  76 #endif
  77
  78 #if !RAID6_USE_EMPTY_ZERO_PAGE
  79 /* In .bss so it's zeroed */
  80 const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
  81 #endif
  82
  83 static inline int raid6_next_disk(int disk, int raid_disks)
  84 {
  85         disk++;
  86         return (disk < raid_disks) ? disk : 0;
  87 }
  88
  89 static void print_raid6_conf (raid6_conf_t *conf);
  90
  91 static inline void __release_stripe(raid6_conf_t *conf, struct stripe_head *sh)
  92 {
  93         if (atomic_dec_and_test(&sh->count)) {
  94                 if (!list_empty(&sh->lru))
  95                         BUG();
  96                 if (atomic_read(&conf->active_stripes)==0)
  97                         BUG();
  98                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
  99                         if (test_bit(STRIPE_DELAYED, &sh->state))
 100                                 list_add_tail(&sh->lru, &conf->delayed_list);
 101                         else
 102                                 list_add_tail(&sh->lru, &conf->handle_list);
 103                         md_wakeup_thread(conf->mddev->thread);
 104                 } else {
 105                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
 106                                 atomic_dec(&conf->preread_active_stripes);
 107                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
 108                                         md_wakeup_thread(conf->mddev->thread);
 109                         }
 110                         list_add_tail(&sh->lru, &conf->inactive_list);
 111                         atomic_dec(&conf->active_stripes);
 112                         if (!conf->inactive_blocked ||
 113                             atomic_read(&conf->active_stripes) < (NR_STRIPES*3/4))
 114                                 wake_up(&conf->wait_for_stripe);
 115                 }
 116         }
 117 }
 118 static void release_stripe(struct stripe_head *sh)
 119 {
 120         raid6_conf_t *conf = sh->raid_conf;
 121         unsigned long flags;
 122
 123         spin_lock_irqsave(&conf->device_lock, flags);
 124         __release_stripe(conf, sh);
 125         spin_unlock_irqrestore(&conf->device_lock, flags);
 126 }
 127
 128 static void remove_hash(struct stripe_head *sh)
 129 {
 130         PRINTK("remove_hash(), stripe %llu\n", (unsigned long long)sh->sector);
 131
 132         if (sh->hash_pprev) {
 133                 if (sh->hash_next)
 134                         sh->hash_next->hash_pprev = sh->hash_pprev;
 135                 *sh->hash_pprev = sh->hash_next;
 136                 sh->hash_pprev = NULL;
 137         }
 138 }
 139
 140 static __inline__ void insert_hash(raid6_conf_t *conf, struct stripe_head *sh)
 141 {
 142         struct stripe_head **shp = &stripe_hash(conf, sh->sector);
 143
 144         PRINTK("insert_hash(), stripe %llu\n", (unsigned long long)sh->sector);
 145
 146         CHECK_DEVLOCK();
 147         if ((sh->hash_next = *shp) != NULL)
 148                 (*shp)->hash_pprev = &sh->hash_next;
 149         *shp = sh;
 150         sh->hash_pprev = shp;
 151 }
 152
 153
 154 /* find an idle stripe, make sure it is unhashed, and return it. */
 155 static struct stripe_head *get_free_stripe(raid6_conf_t *conf)
 156 {
 157         struct stripe_head *sh = NULL;
 158         struct list_head *first;
 159
 160         CHECK_DEVLOCK();
 161         if (list_empty(&conf->inactive_list))
 162                 goto out;
 163         first = conf->inactive_list.next;
 164         sh = list_entry(first, struct stripe_head, lru);
 165         list_del_init(first);
 166         remove_hash(sh);
 167         atomic_inc(&conf->active_stripes);
 168 out:
 169         return sh;
 170 }
 171
 172 static void shrink_buffers(struct stripe_head *sh, int num)
 173 {
 174         struct page *p;
 175         int i;
 176
 177         for (i=0; i<num ; i++) {
 178                 p = sh->dev[i].page;
 179                 if (!p)
 180                         continue;
 181                 sh->dev[i].page = NULL;
 182                 page_cache_release(p);
 183         }
 184 }
 185
 186 static int grow_buffers(struct stripe_head *sh, int num)
 187 {
 188         int i;
 189
 190         for (i=0; i<num; i++) {
 191                 struct page *page;
 192
 193                 if (!(page = alloc_page(GFP_KERNEL))) {
 194                         return 1;
 195                 }
 196                 sh->dev[i].page = page;
 197         }
 198         return 0;
 199 }
 200
 201 static void raid6_build_block (struct stripe_head *sh, int i);
 202
 203 static inline void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx)
 204 {
 205         raid6_conf_t *conf = sh->raid_conf;
 206         int disks = conf->raid_disks, i;
 207
 208         if (atomic_read(&sh->count) != 0)
 209                 BUG();
 210         if (test_bit(STRIPE_HANDLE, &sh->state))
 211                 BUG();
 212
 213         CHECK_DEVLOCK();
 214         PRINTK("init_stripe called, stripe %llu\n",
 215                 (unsigned long long)sh->sector);
 216
 217         remove_hash(sh);
 218
 219         sh->sector = sector;
 220         sh->pd_idx = pd_idx;
 221         sh->state = 0;
 222
 223         for (i=disks; i--; ) {
 224                 struct r5dev *dev = &sh->dev[i];
 225
 226                 if (dev->toread || dev->towrite || dev->written ||
 227                     test_bit(R5_LOCKED, &dev->flags)) {
 228                         PRINTK("sector=%llx i=%d %p %p %p %d\n",
 229                                (unsigned long long)sh->sector, i, dev->toread,
 230                                dev->towrite, dev->written,
 231                                test_bit(R5_LOCKED, &dev->flags));
 232                         BUG();
 233                 }
 234                 dev->flags = 0;
 235                 raid6_build_block(sh, i);
 236         }
 237         insert_hash(conf, sh);
 238 }
 239
 240 static struct stripe_head *__find_stripe(raid6_conf_t *conf, sector_t sector)
 241 {
 242         struct stripe_head *sh;
 243
 244         CHECK_DEVLOCK();
 245         PRINTK("__find_stripe, sector %llu\n", (unsigned long long)sector);
 246         for (sh = stripe_hash(conf, sector); sh; sh = sh->hash_next)
 247                 if (sh->sector == sector)
 248                         return sh;
 249         PRINTK("__stripe %llu not in cache\n", (unsigned long long)sector);
 250         return NULL;
 251 }
 252
 253 static void unplug_slaves(mddev_t *mddev);
 254
 255 static struct stripe_head *get_active_stripe(raid6_conf_t *conf, sector_t sector,
 256                                              int pd_idx, int noblock)
 257 {
 258         struct stripe_head *sh;
 259
 260         PRINTK("get_stripe, sector %llu\n", (unsigned long long)sector);
 261
 262         spin_lock_irq(&conf->device_lock);
 263
 264         do {
 265                 sh = __find_stripe(conf, sector);
 266                 if (!sh) {
 267                         if (!conf->inactive_blocked)
 268                                 sh = get_free_stripe(conf);
 269                         if (noblock && sh == NULL)
 270                                 break;
 271                         if (!sh) {
 272                                 conf->inactive_blocked = 1;
 273                                 wait_event_lock_irq(conf->wait_for_stripe,
 274                                                     !list_empty(&conf->inactive_list) &&
 275                                                     (atomic_read(&conf->active_stripes) < (NR_STRIPES *3/4)
 276                                                      || !conf->inactive_blocked),
 277                                                     conf->device_lock,
 278                                                     unplug_slaves(conf->mddev);
 279                                         );
 280                                 conf->inactive_blocked = 0;
 281                         } else
 282                                 init_stripe(sh, sector, pd_idx);
 283                 } else {
 284                         if (atomic_read(&sh->count)) {
 285                                 if (!list_empty(&sh->lru))
 286                                         BUG();
 287                         } else {
 288                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 289                                         atomic_inc(&conf->active_stripes);
 290                                 if (list_empty(&sh->lru))
 291                                         BUG();
 292                                 list_del_init(&sh->lru);
 293                         }
 294                 }
 295         } while (sh == NULL);
 296
 297         if (sh)
 298                 atomic_inc(&sh->count);
 299
 300         spin_unlock_irq(&conf->device_lock);
 301         return sh;
 302 }
 303
 304 static int grow_stripes(raid6_conf_t *conf, int num)
 305 {
 306         struct stripe_head *sh;
 307         kmem_cache_t *sc;
 308         int devs = conf->raid_disks;
 309
 310         sprintf(conf->cache_name, "raid6/%s", mdname(conf->mddev));
 311
 312         sc = kmem_cache_create(conf->cache_name,
 313                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
 314                                0, 0, NULL, NULL);
 315         if (!sc)
 316                 return 1;
 317         conf->slab_cache = sc;
 318         while (num--) {
 319                 sh = kmem_cache_alloc(sc, GFP_KERNEL);
 320                 if (!sh)
 321                         return 1;
 322                 memset(sh, 0, sizeof(*sh) + (devs-1)*sizeof(struct r5dev));
 323                 sh->raid_conf = conf;
 324                 spin_lock_init(&sh->lock);
 325
 326                 if (grow_buffers(sh, conf->raid_disks)) {
 327                         shrink_buffers(sh, conf->raid_disks);
 328                         kmem_cache_free(sc, sh);
 329                         return 1;
 330                 }
 331                 /* we just created an active stripe so... */
 332                 atomic_set(&sh->count, 1);
 333                 atomic_inc(&conf->active_stripes);
 334                 INIT_LIST_HEAD(&sh->lru);
 335                 release_stripe(sh);
 336         }
 337         return 0;
 338 }
 339
 340 static void shrink_stripes(raid6_conf_t *conf)
 341 {
 342         struct stripe_head *sh;
 343
 344         while (1) {
 345                 spin_lock_irq(&conf->device_lock);
 346                 sh = get_free_stripe(conf);
 347                 spin_unlock_irq(&conf->device_lock);
 348                 if (!sh)
 349                         break;
 350                 if (atomic_read(&sh->count))
 351                         BUG();
 352                 shrink_buffers(sh, conf->raid_disks);
 353                 kmem_cache_free(conf->slab_cache, sh);
 354                 atomic_dec(&conf->active_stripes);
 355         }
 356         kmem_cache_destroy(conf->slab_cache);
 357         conf->slab_cache = NULL;
 358 }
 359
 360 static int raid6_end_read_request (struct bio * bi, unsigned int bytes_done,
 361                                    int error)
 362 {
 363         struct stripe_head *sh = bi->bi_private;
 364         raid6_conf_t *conf = sh->raid_conf;
 365         int disks = conf->raid_disks, i;
 366         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
 367
 368         if (bi->bi_size)
 369                 return 1;
 370
 371         for (i=0 ; i<disks; i++)
 372                 if (bi == &sh->dev[i].req)
 373                         break;
 374
 375         PRINTK("end_read_request %llu/%d, count: %d, uptodate %d.\n",
 376                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
 377                 uptodate);
 378         if (i == disks) {
 379                 BUG();
 380                 return 0;
 381         }
 382
 383         if (uptodate) {
 384 #if 0
 385                 struct bio *bio;
 386                 unsigned long flags;
 387                 spin_lock_irqsave(&conf->device_lock, flags);
 388                 /* we can return a buffer if we bypassed the cache or
 389                  * if the top buffer is not in highmem.  If there are
 390                  * multiple buffers, leave the extra work to
 391                  * handle_stripe
 392                  */
 393                 buffer = sh->bh_read[i];
 394                 if (buffer &&
 395                     (!PageHighMem(buffer->b_page)
 396                      || buffer->b_page == bh->b_page )
 397                         ) {
 398                         sh->bh_read[i] = buffer->b_reqnext;
 399                         buffer->b_reqnext = NULL;
 400                 } else
 401                         buffer = NULL;
 402                 spin_unlock_irqrestore(&conf->device_lock, flags);
 403                 if (sh->bh_page[i]==bh->b_page)
 404                         set_buffer_uptodate(bh);
 405                 if (buffer) {
 406                         if (buffer->b_page != bh->b_page)
 407                                 memcpy(buffer->b_data, bh->b_data, bh->b_size);
 408                         buffer->b_end_io(buffer, 1);
 409                 }
 410 #else
 411                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
 412 #endif
 413         } else {
 414                 md_error(conf->mddev, conf->disks[i].rdev);
 415                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
 416         }
 417         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
 418 #if 0
 419         /* must restore b_page before unlocking buffer... */
 420         if (sh->bh_page[i] != bh->b_page) {
 421                 bh->b_page = sh->bh_page[i];
 422                 bh->b_data = page_address(bh->b_page);
 423                 clear_buffer_uptodate(bh);
 424         }
 425 #endif
 426         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 427         set_bit(STRIPE_HANDLE, &sh->state);
 428         release_stripe(sh);
 429         return 0;
 430 }
 431
 432 static int raid6_end_write_request (struct bio *bi, unsigned int bytes_done,
 433                                     int error)
 434 {
 435         struct stripe_head *sh = bi->bi_private;
 436         raid6_conf_t *conf = sh->raid_conf;
 437         int disks = conf->raid_disks, i;
 438         unsigned long flags;
 439         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
 440
 441         if (bi->bi_size)
 442                 return 1;
 443
 444         for (i=0 ; i<disks; i++)
 445                 if (bi == &sh->dev[i].req)
 446                         break;
 447
 448         PRINTK("end_write_request %llu/%d, count %d, uptodate: %d.\n",
 449                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
 450                 uptodate);
 451         if (i == disks) {
 452                 BUG();
 453                 return 0;
 454         }
 455
 456         spin_lock_irqsave(&conf->device_lock, flags);
 457         if (!uptodate)
 458                 md_error(conf->mddev, conf->disks[i].rdev);
 459
 460         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
 461
 462         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 463         set_bit(STRIPE_HANDLE, &sh->state);
 464         __release_stripe(conf, sh);
 465         spin_unlock_irqrestore(&conf->device_lock, flags);
 466         return 0;
 467 }
 468
 469
 470 static sector_t compute_blocknr(struct stripe_head *sh, int i);
 471
 472 static void raid6_build_block (struct stripe_head *sh, int i)
 473 {
 474         struct r5dev *dev = &sh->dev[i];
 475         int pd_idx = sh->pd_idx;
 476         int qd_idx = raid6_next_disk(pd_idx, sh->raid_conf->raid_disks);
 477
 478         bio_init(&dev->req);
 479         dev->req.bi_io_vec = &dev->vec;
 480         dev->req.bi_vcnt++;
 481         dev->req.bi_max_vecs++;
 482         dev->vec.bv_page = dev->page;
 483         dev->vec.bv_len = STRIPE_SIZE;
 484         dev->vec.bv_offset = 0;
 485
 486         dev->req.bi_sector = sh->sector;
 487         dev->req.bi_private = sh;
 488
 489         dev->flags = 0;
 490         if (i != pd_idx && i != qd_idx)
 491                 dev->sector = compute_blocknr(sh, i);
 492 }
 493
 494 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 495 {
 496         char b[BDEVNAME_SIZE];
 497         raid6_conf_t *conf = (raid6_conf_t *) mddev->private;
 498         PRINTK("raid6: error called\n");
 499
 500         if (!rdev->faulty) {
 501                 mddev->sb_dirty = 1;
 502                 if (rdev->in_sync) {
 503                         conf->working_disks--;
 504                         mddev->degraded++;
 505                         conf->failed_disks++;
 506                         rdev->in_sync = 0;
 507                         /*
 508                          * if recovery was running, make sure it aborts.
 509                          */
 510                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 511                 }
 512                 rdev->faulty = 1;
 513                 printk (KERN_ALERT
 514                         "raid6: Disk failure on %s, disabling device."
 515                         " Operation continuing on %d devices\n",
 516                         bdevname(rdev->bdev,b), conf->working_disks);
 517         }
 518 }
 519
 520 /*
 521  * Input: a 'big' sector number,
 522  * Output: index of the data and parity disk, and the sector # in them.
 523  */
 524 static sector_t raid6_compute_sector(sector_t r_sector, unsigned int raid_disks,
 525                         unsigned int data_disks, unsigned int * dd_idx,
 526                         unsigned int * pd_idx, raid6_conf_t *conf)
 527 {
 528         long stripe;
 529         unsigned long chunk_number;
 530         unsigned int chunk_offset;
 531         sector_t new_sector;
 532         int sectors_per_chunk = conf->chunk_size >> 9;
 533
 534         /* First compute the information on this sector */
 535
 536         /*
 537          * Compute the chunk number and the sector offset inside the chunk
 538          */
 539         chunk_offset = sector_div(r_sector, sectors_per_chunk);
 540         chunk_number = r_sector;
 541         if ( r_sector != chunk_number ) {
 542                 printk(KERN_CRIT "raid6: ERROR: r_sector = %llu, chunk_number = %lu\n",
 543                        (unsigned long long)r_sector, (unsigned long)chunk_number);
 544                 BUG();
 545         }
 546
 547         /*
 548          * Compute the stripe number
 549          */
 550         stripe = chunk_number / data_disks;
 551
 552         /*
 553          * Compute the data disk and parity disk indexes inside the stripe
 554          */
 555         *dd_idx = chunk_number % data_disks;
 556
 557         /*
 558          * Select the parity disk based on the user selected algorithm.
 559          */
 560
 561         /**** FIX THIS ****/
 562         switch (conf->algorithm) {
 563         case ALGORITHM_LEFT_ASYMMETRIC:
 564                 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
 565                 if (*pd_idx == raid_disks-1)
 566                         (*dd_idx)++;    /* Q D D D P */
 567                 else if (*dd_idx >= *pd_idx)
 568                         (*dd_idx) += 2; /* D D P Q D */
 569                 break;
 570         case ALGORITHM_RIGHT_ASYMMETRIC:
 571                 *pd_idx = stripe % raid_disks;
 572                 if (*pd_idx == raid_disks-1)
 573                         (*dd_idx)++;    /* Q D D D P */
 574                 else if (*dd_idx >= *pd_idx)
 575                         (*dd_idx) += 2; /* D D P Q D */
 576                 break;
 577         case ALGORITHM_LEFT_SYMMETRIC:
 578                 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
 579                 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
 580                 break;
 581         case ALGORITHM_RIGHT_SYMMETRIC:
 582                 *pd_idx = stripe % raid_disks;
 583                 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
 584                 break;
 585         default:
 586                 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
 587                         conf->algorithm);
 588         }
 589
 590         PRINTK("raid6: chunk_number = %lu, pd_idx = %u, dd_idx = %u\n",
 591                chunk_number, *pd_idx, *dd_idx);
 592
 593         /*
 594          * Finally, compute the new sector number
 595          */
 596         new_sector = (sector_t) stripe * sectors_per_chunk + chunk_offset;
 597         return new_sector;
 598 }
 599
 600
 601 static sector_t compute_blocknr(struct stripe_head *sh, int i)
 602 {
 603         raid6_conf_t *conf = sh->raid_conf;
 604         int raid_disks = conf->raid_disks, data_disks = raid_disks - 2;
 605         sector_t new_sector = sh->sector, check;
 606         int sectors_per_chunk = conf->chunk_size >> 9;
 607         sector_t stripe;
 608         int chunk_offset;
 609         int chunk_number, dummy1, dummy2, dd_idx = i;
 610         sector_t r_sector;
 611         int i0 = i;
 612
 613         chunk_offset = sector_div(new_sector, sectors_per_chunk);
 614         stripe = new_sector;
 615         if ( new_sector != stripe ) {
 616                 printk(KERN_CRIT "raid6: ERROR: new_sector = %llu, stripe = %lu\n",
 617                        (unsigned long long)new_sector, (unsigned long)stripe);
 618                 BUG();
 619         }
 620
 621         switch (conf->algorithm) {
 622                 case ALGORITHM_LEFT_ASYMMETRIC:
 623                 case ALGORITHM_RIGHT_ASYMMETRIC:
 624                         if (sh->pd_idx == raid_disks-1)
 625                                 i--;    /* Q D D D P */
 626                         else if (i > sh->pd_idx)
 627                                 i -= 2; /* D D P Q D */
 628                         break;
 629                 case ALGORITHM_LEFT_SYMMETRIC:
 630                 case ALGORITHM_RIGHT_SYMMETRIC:
 631                         if (sh->pd_idx == raid_disks-1)
 632                                 i--; /* Q D D D P */
 633                         else {
 634                                 /* D D P Q D */
 635                                 if (i < sh->pd_idx)
 636                                         i += raid_disks;
 637                                 i -= (sh->pd_idx + 2);
 638                         }
 639                         break;
 640                 default:
 641                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
 642                                 conf->algorithm);
 643         }
 644
 645         PRINTK("raid6: compute_blocknr: pd_idx = %u, i0 = %u, i = %u\n", sh->pd_idx, i0, i);
 646
 647         chunk_number = stripe * data_disks + i;
 648         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
 649
 650         check = raid6_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
 651         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
 652                 printk(KERN_CRIT "raid6: compute_blocknr: map not correct\n");
 653                 return 0;
 654         }
 655         return r_sector;
 656 }
 657
 658
 659
 660 /*
 661  * Copy data between a page in the stripe cache, and one or more bion
 662  * The page could align with the middle of the bio, or there could be
 663  * several bion, each with several bio_vecs, which cover part of the page
 664  * Multiple bion are linked together on bi_next.  There may be extras
 665  * at the end of this list.  We ignore them.
 666  */
 667 static void copy_data(int frombio, struct bio *bio,
 668                      struct page *page,
 669                      sector_t sector)
 670 {
 671         char *pa = page_address(page);
 672         struct bio_vec *bvl;
 673         int i;
 674         int page_offset;
 675
 676         if (bio->bi_sector >= sector)
 677                 page_offset = (signed)(bio->bi_sector - sector) * 512;
 678         else
 679                 page_offset = (signed)(sector - bio->bi_sector) * -512;
 680         bio_for_each_segment(bvl, bio, i) {
 681                 int len = bio_iovec_idx(bio,i)->bv_len;
 682                 int clen;
 683                 int b_offset = 0;
 684
 685                 if (page_offset < 0) {
 686                         b_offset = -page_offset;
 687                         page_offset += b_offset;
 688                         len -= b_offset;
 689                 }
 690
 691                 if (len > 0 && page_offset + len > STRIPE_SIZE)
 692                         clen = STRIPE_SIZE - page_offset;
 693                 else clen = len;
 694
 695                 if (clen > 0) {
 696                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
 697                         if (frombio)
 698                                 memcpy(pa+page_offset, ba+b_offset, clen);
 699                         else
 700                                 memcpy(ba+b_offset, pa+page_offset, clen);
 701                         __bio_kunmap_atomic(ba, KM_USER0);
 702                 }
 703                 if (clen < len) /* hit end of page */
 704                         break;
 705                 page_offset +=  len;
 706         }
 707 }
 708
 709 #define check_xor()     do {                                            \
 710                            if (count == MAX_XOR_BLOCKS) {               \
 711                                 xor_block(count, STRIPE_SIZE, ptr);     \
 712                                 count = 1;                              \
 713                            }                                            \
 714                         } while(0)
 715
 716 /* Compute P and Q syndromes */
 717 static void compute_parity(struct stripe_head *sh, int method)
 718 {
 719         raid6_conf_t *conf = sh->raid_conf;
 720         int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = conf->raid_disks, count;
 721         struct bio *chosen;
 722         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
 723         void *ptrs[disks];
 724
 725         qd_idx = raid6_next_disk(pd_idx, disks);
 726         d0_idx = raid6_next_disk(qd_idx, disks);
 727
 728         PRINTK("compute_parity, stripe %llu, method %d\n",
 729                 (unsigned long long)sh->sector, method);
 730
 731         switch(method) {
 732         case READ_MODIFY_WRITE:
 733                 BUG();          /* READ_MODIFY_WRITE N/A for RAID-6 */
 734         case RECONSTRUCT_WRITE:
 735                 for (i= disks; i-- ;)
 736                         if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
 737                                 chosen = sh->dev[i].towrite;
 738                                 sh->dev[i].towrite = NULL;
 739
 740                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
 741                                         wake_up(&conf->wait_for_overlap);
 742
 743                                 if (sh->dev[i].written) BUG();
 744                                 sh->dev[i].written = chosen;
 745                         }
 746                 break;
 747         case CHECK_PARITY:
 748                 BUG();          /* Not implemented yet */
 749         }
 750
 751         for (i = disks; i--;)
 752                 if (sh->dev[i].written) {
 753                         sector_t sector = sh->dev[i].sector;
 754                         struct bio *wbi = sh->dev[i].written;
 755                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
 756                                 copy_data(1, wbi, sh->dev[i].page, sector);
 757                                 wbi = r5_next_bio(wbi, sector);
 758                         }
 759
 760                         set_bit(R5_LOCKED, &sh->dev[i].flags);
 761                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
 762                 }
 763
 764 //      switch(method) {
 765 //      case RECONSTRUCT_WRITE:
 766 //      case CHECK_PARITY:
 767 //      case UPDATE_PARITY:
 768                 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
 769                 /* FIX: Is this ordering of drives even remotely optimal? */
 770                 count = 0;
 771                 i = d0_idx;
 772                 do {
 773                         ptrs[count++] = page_address(sh->dev[i].page);
 774                         if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
 775                                 printk("block %d/%d not uptodate on parity calc\n", i,count);
 776                         i = raid6_next_disk(i, disks);
 777                 } while ( i != d0_idx );
 778 //              break;
 779 //      }
 780
 781         raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
 782
 783         switch(method) {
 784         case RECONSTRUCT_WRITE:
 785                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
 786                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
 787                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
 788                 set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
 789                 break;
 790         case UPDATE_PARITY:
 791                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
 792                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
 793                 break;
 794         }
 795 }
 796
 797 /* Compute one missing block */
 798 static void compute_block_1(struct stripe_head *sh, int dd_idx)
 799 {
 800         raid6_conf_t *conf = sh->raid_conf;
 801         int i, count, disks = conf->raid_disks;
 802         void *ptr[MAX_XOR_BLOCKS], *p;
 803         int pd_idx = sh->pd_idx;
 804         int qd_idx = raid6_next_disk(pd_idx, disks);
 805
 806         PRINTK("compute_block_1, stripe %llu, idx %d\n",
 807                 (unsigned long long)sh->sector, dd_idx);
 808
 809         if ( dd_idx == qd_idx ) {
 810                 /* We're actually computing the Q drive */
 811                 compute_parity(sh, UPDATE_PARITY);
 812         } else {
 813                 ptr[0] = page_address(sh->dev[dd_idx].page);
 814                 memset(ptr[0], 0, STRIPE_SIZE);
 815                 count = 1;
 816                 for (i = disks ; i--; ) {
 817                         if (i == dd_idx || i == qd_idx)
 818                                 continue;
 819                         p = page_address(sh->dev[i].page);
 820                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
 821                                 ptr[count++] = p;
 822                         else
 823                                 printk("compute_block() %d, stripe %llu, %d"
 824                                        " not present\n", dd_idx,
 825                                        (unsigned long long)sh->sector, i);
 826
 827                         check_xor();
 828                 }
 829                 if (count != 1)
 830                         xor_block(count, STRIPE_SIZE, ptr);
 831                 set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
 832         }
 833 }
 834
 835 /* Compute two missing blocks */
 836 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
 837 {
 838         raid6_conf_t *conf = sh->raid_conf;
 839         int i, count, disks = conf->raid_disks;
 840         int pd_idx = sh->pd_idx;
 841         int qd_idx = raid6_next_disk(pd_idx, disks);
 842         int d0_idx = raid6_next_disk(qd_idx, disks);
 843         int faila, failb;
 844
 845         /* faila and failb are disk numbers relative to d0_idx */
 846         /* pd_idx become disks-2 and qd_idx become disks-1 */
 847         faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
 848         failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
 849
 850         BUG_ON(faila == failb);
 851         if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
 852
 853         PRINTK("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
 854                (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
 855
 856         if ( failb == disks-1 ) {
 857                 /* Q disk is one of the missing disks */
 858                 if ( faila == disks-2 ) {
 859                         /* Missing P+Q, just recompute */
 860                         compute_parity(sh, UPDATE_PARITY);
 861                         return;
 862                 } else {
 863                         /* We're missing D+Q; recompute D from P */
 864                         compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1);
 865                         compute_parity(sh, UPDATE_PARITY); /* Is this necessary? */
 866                         return;
 867                 }
 868         }
 869
 870         /* We're missing D+P or D+D; build pointer table */
 871         {
 872                 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
 873                 void *ptrs[disks];
 874
 875                 count = 0;
 876                 i = d0_idx;
 877                 do {
 878                         ptrs[count++] = page_address(sh->dev[i].page);
 879                         i = raid6_next_disk(i, disks);
 880                         if (i != dd_idx1 && i != dd_idx2 &&
 881                             !test_bit(R5_UPTODATE, &sh->dev[i].flags))
 882                                 printk("compute_2 with missing block %d/%d\n", count, i);
 883                 } while ( i != d0_idx );
 884
 885                 if ( failb == disks-2 ) {
 886                         /* We're missing D+P. */
 887                         raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
 888                 } else {
 889                         /* We're missing D+D. */
 890                         raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
 891                 }
 892
 893                 /* Both the above update both missing blocks */
 894                 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
 895                 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
 896         }
 897 }
 898
 899
 900 /*
 901  * Each stripe/dev can have one or more bion attached.
 902  * toread/towrite point to the first in a chain.
 903  * The bi_next chain must be in order.
 904  */
 905 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
 906 {
 907         struct bio **bip;
 908         raid6_conf_t *conf = sh->raid_conf;
 909
 910         PRINTK("adding bh b#%llu to stripe s#%llu\n",
 911                 (unsigned long long)bi->bi_sector,
 912                 (unsigned long long)sh->sector);
 913
 914
 915         spin_lock(&sh->lock);
 916         spin_lock_irq(&conf->device_lock);
 917         if (forwrite)
 918                 bip = &sh->dev[dd_idx].towrite;
 919         else
 920                 bip = &sh->dev[dd_idx].toread;
 921         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
 922                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
 923                         goto overlap;
 924                 bip = &(*bip)->bi_next;
 925         }
 926         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
 927                 goto overlap;
 928
 929         if (*bip && bi->bi_next && (*bip) != bi->bi_next)
 930                 BUG();
 931         if (*bip)
 932                 bi->bi_next = *bip;
 933         *bip = bi;
 934         bi->bi_phys_segments ++;
 935         spin_unlock_irq(&conf->device_lock);
 936         spin_unlock(&sh->lock);
 937
 938         PRINTK("added bi b#%llu to stripe s#%llu, disk %d.\n",
 939                 (unsigned long long)bi->bi_sector,
 940                 (unsigned long long)sh->sector, dd_idx);
 941
 942         if (forwrite) {
 943                 /* check if page is covered */
 944                 sector_t sector = sh->dev[dd_idx].sector;
 945                 for (bi=sh->dev[dd_idx].towrite;
 946                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
 947                              bi && bi->bi_sector <= sector;
 948                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
 949                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
 950                                 sector = bi->bi_sector + (bi->bi_size>>9);
 951                 }
 952                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
 953                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
 954         }
 955         return 1;
 956
 957  overlap:
 958         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
 959         spin_unlock_irq(&conf->device_lock);
 960         spin_unlock(&sh->lock);
 961         return 0;
 962 }
 963
 964
 965 /*
 966  * handle_stripe - do things to a stripe.
 967  *
 968  * We lock the stripe and then examine the state of various bits
 969  * to see what needs to be done.
 970  * Possible results:
 971  *    return some read request which now have data
 972  *    return some write requests which are safely on disc
 973  *    schedule a read on some buffers
 974  *    schedule a write of some buffers
 975  *    return confirmation of parity correctness
 976  *
 977  * Parity calculations are done inside the stripe lock
 978  * buffers are taken off read_list or write_list, and bh_cache buffers
 979  * get BH_Lock set before the stripe lock is released.
 980  *
 981  */
 982
 983 static void handle_stripe(struct stripe_head *sh)
 984 {
 985         raid6_conf_t *conf = sh->raid_conf;
 986         int disks = conf->raid_disks;
 987         struct bio *return_bi= NULL;
 988         struct bio *bi;
 989         int i;
 990         int syncing;
 991         int locked=0, uptodate=0, to_read=0, to_write=0, failed=0, written=0;
 992         int non_overwrite = 0;
 993         int failed_num[2] = {0, 0};
 994         struct r5dev *dev, *pdev, *qdev;
 995         int pd_idx = sh->pd_idx;
 996         int qd_idx = raid6_next_disk(pd_idx, disks);
 997         int p_failed, q_failed;
 998
 999         PRINTK("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d, qd_idx=%d\n",
1000                (unsigned long long)sh->sector, sh->state, atomic_read(&sh->count),
1001                pd_idx, qd_idx);
1002
1003         spin_lock(&sh->lock);
1004         clear_bit(STRIPE_HANDLE, &sh->state);
1005         clear_bit(STRIPE_DELAYED, &sh->state);
1006
1007         syncing = test_bit(STRIPE_SYNCING, &sh->state);
1008         /* Now to look around and see what can be done */
1009
1010         for (i=disks; i--; ) {
1011                 mdk_rdev_t *rdev;
1012                 dev = &sh->dev[i];
1013                 clear_bit(R5_Insync, &dev->flags);
1014                 clear_bit(R5_Syncio, &dev->flags);
1015
1016                 PRINTK("check %d: state 0x%lx read %p write %p written %p\n",
1017                         i, dev->flags, dev->toread, dev->towrite, dev->written);
1018                 /* maybe we can reply to a read */
1019                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
1020                         struct bio *rbi, *rbi2;
1021                         PRINTK("Return read for disc %d\n", i);
1022                         spin_lock_irq(&conf->device_lock);
1023                         rbi = dev->toread;
1024                         dev->toread = NULL;
1025                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1026                                 wake_up(&conf->wait_for_overlap);
1027                         spin_unlock_irq(&conf->device_lock);
1028                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1029                                 copy_data(0, rbi, dev->page, dev->sector);
1030                                 rbi2 = r5_next_bio(rbi, dev->sector);
1031                                 spin_lock_irq(&conf->device_lock);
1032                                 if (--rbi->bi_phys_segments == 0) {
1033                                         rbi->bi_next = return_bi;
1034                                         return_bi = rbi;
1035                                 }
1036                                 spin_unlock_irq(&conf->device_lock);
1037                                 rbi = rbi2;
1038                         }
1039                 }
1040
1041                 /* now count some things */
1042                 if (test_bit(R5_LOCKED, &dev->flags)) locked++;
1043                 if (test_bit(R5_UPTODATE, &dev->flags)) uptodate++;
1044
1045
1046                 if (dev->toread) to_read++;
1047                 if (dev->towrite) {
1048                         to_write++;
1049                         if (!test_bit(R5_OVERWRITE, &dev->flags))
1050                                 non_overwrite++;
1051                 }
1052                 if (dev->written) written++;
1053                 rdev = conf->disks[i].rdev; /* FIXME, should I be looking rdev */
1054                 if (!rdev || !rdev->in_sync) {
1055                         if ( failed < 2 )
1056                                 failed_num[failed] = i;
1057                         failed++;
1058                 } else
1059                         set_bit(R5_Insync, &dev->flags);
1060         }
1061         PRINTK("locked=%d uptodate=%d to_read=%d"
1062                " to_write=%d failed=%d failed_num=%d,%d\n",
1063                locked, uptodate, to_read, to_write, failed,
1064                failed_num[0], failed_num[1]);
1065         /* check if the array has lost >2 devices and, if so, some requests might
1066          * need to be failed
1067          */
1068         if (failed > 2 && to_read+to_write+written) {
1069                 spin_lock_irq(&conf->device_lock);
1070                 for (i=disks; i--; ) {
1071                         /* fail all writes first */
1072                         bi = sh->dev[i].towrite;
1073                         sh->dev[i].towrite = NULL;
1074                         if (bi) to_write--;
1075
1076                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1077                                 wake_up(&conf->wait_for_overlap);
1078
1079                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1080                                 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1081                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1082                                 if (--bi->bi_phys_segments == 0) {
1083                                         md_write_end(conf->mddev);
1084                                         bi->bi_next = return_bi;
1085                                         return_bi = bi;
1086                                 }
1087                                 bi = nextbi;
1088                         }
1089                         /* and fail all 'written' */
1090                         bi = sh->dev[i].written;
1091                         sh->dev[i].written = NULL;
1092                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS) {
1093                                 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1094                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1095                                 if (--bi->bi_phys_segments == 0) {
1096                                         md_write_end(conf->mddev);
1097                                         bi->bi_next = return_bi;
1098                                         return_bi = bi;
1099                                 }
1100                                 bi = bi2;
1101                         }
1102
1103                         /* fail any reads if this device is non-operational */
1104                         if (!test_bit(R5_Insync, &sh->dev[i].flags)) {
1105                                 bi = sh->dev[i].toread;
1106                                 sh->dev[i].toread = NULL;
1107                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1108                                         wake_up(&conf->wait_for_overlap);
1109                                 if (bi) to_read--;
1110                                 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1111                                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1112                                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1113                                         if (--bi->bi_phys_segments == 0) {
1114                                                 bi->bi_next = return_bi;
1115                                                 return_bi = bi;
1116                                         }
1117                                         bi = nextbi;
1118                                 }
1119                         }
1120                 }
1121                 spin_unlock_irq(&conf->device_lock);
1122         }
1123         if (failed > 2 && syncing) {
1124                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
1125                 clear_bit(STRIPE_SYNCING, &sh->state);
1126                 syncing = 0;
1127         }
1128
1129         /*
1130          * might be able to return some write requests if the parity blocks
1131          * are safe, or on a failed drive
1132          */
1133         pdev = &sh->dev[pd_idx];
1134         p_failed = (failed >= 1 && failed_num[0] == pd_idx)
1135                 || (failed >= 2 && failed_num[1] == pd_idx);
1136         qdev = &sh->dev[qd_idx];
1137         q_failed = (failed >= 1 && failed_num[0] == qd_idx)
1138                 || (failed >= 2 && failed_num[1] == qd_idx);
1139
1140         if ( written &&
1141              ( p_failed || ((test_bit(R5_Insync, &pdev->flags)
1142                              && !test_bit(R5_LOCKED, &pdev->flags)
1143                              && test_bit(R5_UPTODATE, &pdev->flags))) ) &&
1144              ( q_failed || ((test_bit(R5_Insync, &qdev->flags)
1145                              && !test_bit(R5_LOCKED, &qdev->flags)
1146                              && test_bit(R5_UPTODATE, &qdev->flags))) ) ) {
1147                 /* any written block on an uptodate or failed drive can be
1148                  * returned.  Note that if we 'wrote' to a failed drive,
1149                  * it will be UPTODATE, but never LOCKED, so we don't need
1150                  * to test 'failed' directly.
1151                  */
1152                 for (i=disks; i--; )
1153                         if (sh->dev[i].written) {
1154                                 dev = &sh->dev[i];
1155                                 if (!test_bit(R5_LOCKED, &dev->flags) &&
1156                                     test_bit(R5_UPTODATE, &dev->flags) ) {
1157                                         /* We can return any write requests */
1158                                         struct bio *wbi, *wbi2;
1159                                         PRINTK("Return write for stripe %llu disc %d\n",
1160                                                (unsigned long long)sh->sector, i);
1161                                         spin_lock_irq(&conf->device_lock);
1162                                         wbi = dev->written;
1163                                         dev->written = NULL;
1164                                         while (wbi && wbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1165                                                 wbi2 = r5_next_bio(wbi, dev->sector);
1166                                                 if (--wbi->bi_phys_segments == 0) {
1167                                                         md_write_end(conf->mddev);
1168                                                         wbi->bi_next = return_bi;
1169                                                         return_bi = wbi;
1170                                                 }
1171                                                 wbi = wbi2;
1172                                         }
1173                                         spin_unlock_irq(&conf->device_lock);
1174                                 }
1175                         }
1176         }
1177
1178         /* Now we might consider reading some blocks, either to check/generate
1179          * parity, or to satisfy requests
1180          * or to load a block that is being partially written.
1181          */
1182         if (to_read || non_overwrite || (to_write && failed) || (syncing && (uptodate < disks))) {
1183                 for (i=disks; i--;) {
1184                         dev = &sh->dev[i];
1185                         if (!test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1186                             (dev->toread ||
1187                              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1188                              syncing ||
1189                              (failed >= 1 && (sh->dev[failed_num[0]].toread || to_write)) ||
1190                              (failed >= 2 && (sh->dev[failed_num[1]].toread || to_write))
1191                                     )
1192                                 ) {
1193                                 /* we would like to get this block, possibly
1194                                  * by computing it, but we might not be able to
1195                                  */
1196                                 if (uptodate == disks-1) {
1197                                         PRINTK("Computing stripe %llu block %d\n",
1198                                                (unsigned long long)sh->sector, i);
1199                                         compute_block_1(sh, i);
1200                                         uptodate++;
1201                                 } else if ( uptodate == disks-2 && failed >= 2 ) {
1202                                         /* Computing 2-failure is *very* expensive; only do it if failed >= 2 */
1203                                         int other;
1204                                         for (other=disks; other--;) {
1205                                                 if ( other == i )
1206                                                         continue;
1207                                                 if ( !test_bit(R5_UPTODATE, &sh->dev[other].flags) )
1208                                                         break;
1209                                         }
1210                                         BUG_ON(other < 0);
1211                                         PRINTK("Computing stripe %llu blocks %d,%d\n",
1212                                                (unsigned long long)sh->sector, i, other);
1213                                         compute_block_2(sh, i, other);
1214                                         uptodate += 2;
1215                                 } else if (test_bit(R5_Insync, &dev->flags)) {
1216                                         set_bit(R5_LOCKED, &dev->flags);
1217                                         set_bit(R5_Wantread, &dev->flags);
1218 #if 0
1219                                         /* if I am just reading this block and we don't have
1220                                            a failed drive, or any pending writes then sidestep the cache */
1221                                         if (sh->bh_read[i] && !sh->bh_read[i]->b_reqnext &&
1222                                             ! syncing && !failed && !to_write) {
1223                                                 sh->bh_cache[i]->b_page =  sh->bh_read[i]->b_page;
1224                                                 sh->bh_cache[i]->b_data =  sh->bh_read[i]->b_data;
1225                                         }
1226 #endif
1227                                         locked++;
1228                                         PRINTK("Reading block %d (sync=%d)\n",
1229                                                 i, syncing);
1230                                         if (syncing)
1231                                                 md_sync_acct(conf->disks[i].rdev->bdev,
1232                                                              STRIPE_SECTORS);
1233                                 }
1234                         }
1235                 }
1236                 set_bit(STRIPE_HANDLE, &sh->state);
1237         }
1238
1239         /* now to consider writing and what else, if anything should be read */
1240         if (to_write) {
1241                 int rcw=0, must_compute=0;
1242                 for (i=disks ; i--;) {
1243                         dev = &sh->dev[i];
1244                         /* Would I have to read this buffer for reconstruct_write */
1245                         if (!test_bit(R5_OVERWRITE, &dev->flags)
1246                             && i != pd_idx && i != qd_idx
1247                             && (!test_bit(R5_LOCKED, &dev->flags)
1248 #if 0
1249                                 || sh->bh_page[i] != bh->b_page
1250 #endif
1251                                     ) &&
1252                             !test_bit(R5_UPTODATE, &dev->flags)) {
1253                                 if (test_bit(R5_Insync, &dev->flags)) rcw++;
1254                                 else {
1255                                         PRINTK("raid6: must_compute: disk %d flags=%#lx\n", i, dev->flags);
1256                                         must_compute++;
1257                                 }
1258                         }
1259                 }
1260                 PRINTK("for sector %llu, rcw=%d, must_compute=%d\n",
1261                        (unsigned long long)sh->sector, rcw, must_compute);
1262                 set_bit(STRIPE_HANDLE, &sh->state);
1263
1264                 if (rcw > 0)
1265                         /* want reconstruct write, but need to get some data */
1266                         for (i=disks; i--;) {
1267                                 dev = &sh->dev[i];
1268                                 if (!test_bit(R5_OVERWRITE, &dev->flags)
1269                                     && !(failed == 0 && (i == pd_idx || i == qd_idx))
1270                                     && !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1271                                     test_bit(R5_Insync, &dev->flags)) {
1272                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1273                                         {
1274                                                 PRINTK("Read_old stripe %llu block %d for Reconstruct\n",
1275                                                        (unsigned long long)sh->sector, i);
1276                                                 set_bit(R5_LOCKED, &dev->flags);
1277                                                 set_bit(R5_Wantread, &dev->flags);
1278                                                 locked++;
1279                                         } else {
1280                                                 PRINTK("Request delayed stripe %llu block %d for Reconstruct\n",
1281                                                        (unsigned long long)sh->sector, i);
1282                                                 set_bit(STRIPE_DELAYED, &sh->state);
1283                                                 set_bit(STRIPE_HANDLE, &sh->state);
1284                                         }
1285                                 }
1286                         }
1287                 /* now if nothing is locked, and if we have enough data, we can start a write request */
1288                 if (locked == 0 && rcw == 0) {
1289                         if ( must_compute > 0 ) {
1290                                 /* We have failed blocks and need to compute them */
1291                                 switch ( failed ) {
1292                                 case 0: BUG();
1293                                 case 1: compute_block_1(sh, failed_num[0]); break;
1294                                 case 2: compute_block_2(sh, failed_num[0], failed_num[1]); break;
1295                                 default: BUG(); /* This request should have been failed? */
1296                                 }
1297                         }
1298
1299                         PRINTK("Computing parity for stripe %llu\n", (unsigned long long)sh->sector);
1300                         compute_parity(sh, RECONSTRUCT_WRITE);
1301                         /* now every locked buffer is ready to be written */
1302                         for (i=disks; i--;)
1303                                 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
1304                                         PRINTK("Writing stripe %llu block %d\n",
1305                                                (unsigned long long)sh->sector, i);
1306                                         locked++;
1307                                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
1308 #if 0 /**** FIX: I don't understand the logic here... ****/
1309                                         if (!test_bit(R5_Insync, &sh->dev[i].flags)
1310                                             || ((i==pd_idx || i==qd_idx) && failed == 0)) /* FIX? */
1311                                                 set_bit(STRIPE_INSYNC, &sh->state);
1312 #endif
1313                                 }
1314                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
1315                                 atomic_dec(&conf->preread_active_stripes);
1316                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
1317                                         md_wakeup_thread(conf->mddev->thread);
1318                         }
1319                 }
1320         }
1321
1322         /* maybe we need to check and possibly fix the parity for this stripe
1323          * Any reads will already have been scheduled, so we just see if enough data
1324          * is available
1325          */
1326         if (syncing && locked == 0 &&
1327             !test_bit(STRIPE_INSYNC, &sh->state) && failed <= 2) {
1328                 set_bit(STRIPE_HANDLE, &sh->state);
1329 #if 0 /* RAID-6: Don't support CHECK PARITY yet */
1330                 if (failed == 0) {
1331                         char *pagea;
1332                         if (uptodate != disks)
1333                                 BUG();
1334                         compute_parity(sh, CHECK_PARITY);
1335                         uptodate--;
1336                         pagea = page_address(sh->dev[pd_idx].page);
1337                         if ((*(u32*)pagea) == 0 &&
1338                             !memcmp(pagea, pagea+4, STRIPE_SIZE-4)) {
1339                                 /* parity is correct (on disc, not in buffer any more) */
1340                                 set_bit(STRIPE_INSYNC, &sh->state);
1341                         }
1342                 }
1343 #endif
1344                 if (!test_bit(STRIPE_INSYNC, &sh->state)) {
1345                         int failed_needupdate[2];
1346                         struct r5dev *adev, *bdev;
1347
1348                         if ( failed < 1 )
1349                                 failed_num[0] = pd_idx;
1350                         if ( failed < 2 )
1351                                 failed_num[1] = (failed_num[0] == qd_idx) ? pd_idx : qd_idx;
1352
1353                         failed_needupdate[0] = !test_bit(R5_UPTODATE, &sh->dev[failed_num[0]].flags);
1354                         failed_needupdate[1] = !test_bit(R5_UPTODATE, &sh->dev[failed_num[1]].flags);
1355
1356                         PRINTK("sync: failed=%d num=%d,%d fnu=%u%u\n",
1357                                failed, failed_num[0], failed_num[1], failed_needupdate[0], failed_needupdate[1]);
1358
1359 #if 0  /* RAID-6: This code seems to require that CHECK_PARITY destroys the uptodateness of the parity */
1360                         /* should be able to compute the missing block(s) and write to spare */
1361                         if ( failed_needupdate[0] ^ failed_needupdate[1] ) {
1362                                 if (uptodate+1 != disks)
1363                                         BUG();
1364                                 compute_block_1(sh, failed_needupdate[0] ? failed_num[0] : failed_num[1]);
1365                                 uptodate++;
1366                         } else if ( failed_needupdate[0] & failed_needupdate[1] ) {
1367                                 if (uptodate+2 != disks)
1368                                         BUG();
1369                                 compute_block_2(sh, failed_num[0], failed_num[1]);
1370                                 uptodate += 2;
1371                         }
1372 #else
1373                         compute_block_2(sh, failed_num[0], failed_num[1]);
1374                         uptodate += failed_needupdate[0] + failed_needupdate[1];
1375 #endif
1376
1377                         if (uptodate != disks)
1378                                 BUG();
1379
1380                         PRINTK("Marking for sync stripe %llu blocks %d,%d\n",
1381                                (unsigned long long)sh->sector, failed_num[0], failed_num[1]);
1382
1383                         /**** FIX: Should we really do both of these unconditionally? ****/
1384                         adev = &sh->dev[failed_num[0]];
1385                         locked += !test_bit(R5_LOCKED, &adev->flags);
1386                         set_bit(R5_LOCKED, &adev->flags);
1387                         set_bit(R5_Wantwrite, &adev->flags);
1388                         bdev = &sh->dev[failed_num[1]];
1389                         locked += !test_bit(R5_LOCKED, &bdev->flags);
1390                         set_bit(R5_LOCKED, &bdev->flags);
1391                         set_bit(R5_Wantwrite, &bdev->flags);
1392
1393                         set_bit(STRIPE_INSYNC, &sh->state);
1394                         set_bit(R5_Syncio, &adev->flags);
1395                         set_bit(R5_Syncio, &bdev->flags);
1396                 }
1397         }
1398         if (syncing && locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
1399                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
1400                 clear_bit(STRIPE_SYNCING, &sh->state);
1401         }
1402
1403         spin_unlock(&sh->lock);
1404
1405         while ((bi=return_bi)) {
1406                 int bytes = bi->bi_size;
1407
1408                 return_bi = bi->bi_next;
1409                 bi->bi_next = NULL;
1410                 bi->bi_size = 0;
1411                 bi->bi_end_io(bi, bytes, 0);
1412         }
1413         for (i=disks; i-- ;) {
1414                 int rw;
1415                 struct bio *bi;
1416                 mdk_rdev_t *rdev;
1417                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
1418                         rw = 1;
1419                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1420                         rw = 0;
1421                 else
1422                         continue;
1423
1424                 bi = &sh->dev[i].req;
1425
1426                 bi->bi_rw = rw;
1427                 if (rw)
1428                         bi->bi_end_io = raid6_end_write_request;
1429                 else
1430                         bi->bi_end_io = raid6_end_read_request;
1431
1432                 rcu_read_lock();
1433                 rdev = conf->disks[i].rdev;
1434                 if (rdev && rdev->faulty)
1435                         rdev = NULL;
1436                 if (rdev)
1437                         atomic_inc(&rdev->nr_pending);
1438                 rcu_read_unlock();
1439
1440                 if (rdev) {
1441                         if (test_bit(R5_Syncio, &sh->dev[i].flags))
1442                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1443
1444                         bi->bi_bdev = rdev->bdev;
1445                         PRINTK("for %llu schedule op %ld on disc %d\n",
1446                                 (unsigned long long)sh->sector, bi->bi_rw, i);
1447                         atomic_inc(&sh->count);
1448                         bi->bi_sector = sh->sector + rdev->data_offset;
1449                         bi->bi_flags = 1 << BIO_UPTODATE;
1450                         bi->bi_vcnt = 1;
1451                         bi->bi_max_vecs = 1;
1452                         bi->bi_idx = 0;
1453                         bi->bi_io_vec = &sh->dev[i].vec;
1454                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1455                         bi->bi_io_vec[0].bv_offset = 0;
1456                         bi->bi_size = STRIPE_SIZE;
1457                         bi->bi_next = NULL;
1458                         generic_make_request(bi);
1459                 } else {
1460                         PRINTK("skip op %ld on disc %d for sector %llu\n",
1461                                 bi->bi_rw, i, (unsigned long long)sh->sector);
1462                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1463                         set_bit(STRIPE_HANDLE, &sh->state);
1464                 }
1465         }
1466 }
1467
1468 static inline void raid6_activate_delayed(raid6_conf_t *conf)
1469 {
1470         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
1471                 while (!list_empty(&conf->delayed_list)) {
1472                         struct list_head *l = conf->delayed_list.next;
1473                         struct stripe_head *sh;
1474                         sh = list_entry(l, struct stripe_head, lru);
1475                         list_del_init(l);
1476                         clear_bit(STRIPE_DELAYED, &sh->state);
1477                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1478                                 atomic_inc(&conf->preread_active_stripes);
1479                         list_add_tail(&sh->lru, &conf->handle_list);
1480                 }
1481         }
1482 }
1483
1484 static void unplug_slaves(mddev_t *mddev)
1485 {
1486         raid6_conf_t *conf = mddev_to_conf(mddev);
1487         int i;
1488
1489         rcu_read_lock();
1490         for (i=0; i<mddev->raid_disks; i++) {
1491                 mdk_rdev_t *rdev = conf->disks[i].rdev;
1492                 if (rdev && !rdev->faulty && atomic_read(&rdev->nr_pending)) {
1493                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
1494
1495                         atomic_inc(&rdev->nr_pending);
1496                         rcu_read_unlock();
1497
1498                         if (r_queue->unplug_fn)
1499                                 r_queue->unplug_fn(r_queue);
1500
1501                         rdev_dec_pending(rdev, mddev);
1502                         rcu_read_lock();
1503                 }
1504         }
1505         rcu_read_unlock();
1506 }
1507
1508 static void raid6_unplug_device(request_queue_t *q)
1509 {
1510         mddev_t *mddev = q->queuedata;
1511         raid6_conf_t *conf = mddev_to_conf(mddev);
1512         unsigned long flags;
1513
1514         spin_lock_irqsave(&conf->device_lock, flags);
1515
1516         if (blk_remove_plug(q))
1517                 raid6_activate_delayed(conf);
1518         md_wakeup_thread(mddev->thread);
1519
1520         spin_unlock_irqrestore(&conf->device_lock, flags);
1521
1522         unplug_slaves(mddev);
1523 }
1524
1525 static int raid6_issue_flush(request_queue_t *q, struct gendisk *disk,
1526                              sector_t *error_sector)
1527 {
1528         mddev_t *mddev = q->queuedata;
1529         raid6_conf_t *conf = mddev_to_conf(mddev);
1530         int i, ret = 0;
1531
1532         rcu_read_lock();
1533         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
1534                 mdk_rdev_t *rdev = conf->disks[i].rdev;
1535                 if (rdev && !rdev->faulty) {
1536                         struct block_device *bdev = rdev->bdev;
1537                         request_queue_t *r_queue = bdev_get_queue(bdev);
1538
1539                         if (!r_queue->issue_flush_fn)
1540                                 ret = -EOPNOTSUPP;
1541                         else {
1542                                 atomic_inc(&rdev->nr_pending);
1543                                 rcu_read_unlock();
1544                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
1545                                                               error_sector);
1546                                 rdev_dec_pending(rdev, mddev);
1547                                 rcu_read_lock();
1548                         }
1549                 }
1550         }
1551         rcu_read_unlock();
1552         return ret;
1553 }
1554
1555 static inline void raid6_plug_device(raid6_conf_t *conf)
1556 {
1557         spin_lock_irq(&conf->device_lock);
1558         blk_plug_device(conf->mddev->queue);
1559         spin_unlock_irq(&conf->device_lock);
1560 }
1561
1562 static int make_request (request_queue_t *q, struct bio * bi)
1563 {
1564         mddev_t *mddev = q->queuedata;
1565         raid6_conf_t *conf = mddev_to_conf(mddev);
1566         const unsigned int raid_disks = conf->raid_disks;
1567         const unsigned int data_disks = raid_disks - 2;
1568         unsigned int dd_idx, pd_idx;
1569         sector_t new_sector;
1570         sector_t logical_sector, last_sector;
1571         struct stripe_head *sh;
1572
1573         md_write_start(mddev, bi);
1574
1575         if (bio_data_dir(bi)==WRITE) {
1576                 disk_stat_inc(mddev->gendisk, writes);
1577                 disk_stat_add(mddev->gendisk, write_sectors, bio_sectors(bi));
1578         } else {
1579                 disk_stat_inc(mddev->gendisk, reads);
1580                 disk_stat_add(mddev->gendisk, read_sectors, bio_sectors(bi));
1581         }
1582
1583         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
1584         last_sector = bi->bi_sector + (bi->bi_size>>9);
1585
1586         bi->bi_next = NULL;
1587         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
1588
1589         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
1590                 DEFINE_WAIT(w);
1591
1592                 new_sector = raid6_compute_sector(logical_sector,
1593                                                   raid_disks, data_disks, &dd_idx, &pd_idx, conf);
1594
1595                 PRINTK("raid6: make_request, sector %llu logical %llu\n",
1596                        (unsigned long long)new_sector,
1597                        (unsigned long long)logical_sector);
1598
1599         retry:
1600                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
1601                 sh = get_active_stripe(conf, new_sector, pd_idx, (bi->bi_rw&RWA_MASK));
1602                 if (sh) {
1603                         if (!add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
1604                                 /* Add failed due to overlap.  Flush everything
1605                                  * and wait a while
1606                                  */
1607                                 raid6_unplug_device(mddev->queue);
1608                                 release_stripe(sh);
1609                                 schedule();
1610                                 goto retry;
1611                         }
1612                         finish_wait(&conf->wait_for_overlap, &w);
1613                         raid6_plug_device(conf);
1614                         handle_stripe(sh);
1615                         release_stripe(sh);
1616                 } else {
1617                         /* cannot get stripe for read-ahead, just give-up */
1618                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1619                         finish_wait(&conf->wait_for_overlap, &w);
1620                         break;
1621                 }
1622
1623         }
1624         spin_lock_irq(&conf->device_lock);
1625         if (--bi->bi_phys_segments == 0) {
1626                 int bytes = bi->bi_size;
1627
1628                 if ( bio_data_dir(bi) == WRITE )
1629                         md_write_end(mddev);
1630                 bi->bi_size = 0;
1631                 bi->bi_end_io(bi, bytes, 0);
1632         }
1633         spin_unlock_irq(&conf->device_lock);
1634         return 0;
1635 }
1636
1637 /* FIXME go_faster isn't used */
1638 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1639 {
1640         raid6_conf_t *conf = (raid6_conf_t *) mddev->private;
1641         struct stripe_head *sh;
1642         int sectors_per_chunk = conf->chunk_size >> 9;
1643         sector_t x;
1644         unsigned long stripe;
1645         int chunk_offset;
1646         int dd_idx, pd_idx;
1647         sector_t first_sector;
1648         int raid_disks = conf->raid_disks;
1649         int data_disks = raid_disks - 2;
1650
1651         if (sector_nr >= mddev->size <<1) {
1652                 /* just being told to finish up .. nothing much to do */
1653                 unplug_slaves(mddev);
1654                 return 0;
1655         }
1656         /* if there are 2 or more failed drives and we are trying
1657          * to resync, then assert that we are finished, because there is
1658          * nothing we can do.
1659          */
1660         if (mddev->degraded >= 2 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1661                 sector_t rv = (mddev->size << 1) - sector_nr;
1662                 *skipped = 1;
1663                 return rv;
1664         }
1665
1666         x = sector_nr;
1667         chunk_offset = sector_div(x, sectors_per_chunk);
1668         stripe = x;
1669         BUG_ON(x != stripe);
1670
1671         first_sector = raid6_compute_sector((sector_t)stripe*data_disks*sectors_per_chunk
1672                 + chunk_offset, raid_disks, data_disks, &dd_idx, &pd_idx, conf);
1673         sh = get_active_stripe(conf, sector_nr, pd_idx, 1);
1674         if (sh == NULL) {
1675                 sh = get_active_stripe(conf, sector_nr, pd_idx, 0);
1676                 /* make sure we don't swamp the stripe cache if someone else
1677                  * is trying to get access
1678                  */
1679                 set_current_state(TASK_UNINTERRUPTIBLE);
1680                 schedule_timeout(1);
1681         }
1682         spin_lock(&sh->lock);
1683         set_bit(STRIPE_SYNCING, &sh->state);
1684         clear_bit(STRIPE_INSYNC, &sh->state);
1685         spin_unlock(&sh->lock);
1686
1687         handle_stripe(sh);
1688         release_stripe(sh);
1689
1690         return STRIPE_SECTORS;
1691 }
1692
1693 /*
1694  * This is our raid6 kernel thread.
1695  *
1696  * We scan the hash table for stripes which can be handled now.
1697  * During the scan, completed stripes are saved for us by the interrupt
1698  * handler, so that they will not have to wait for our next wakeup.
1699  */
1700 static void raid6d (mddev_t *mddev)
1701 {
1702         struct stripe_head *sh;
1703         raid6_conf_t *conf = mddev_to_conf(mddev);
1704         int handled;
1705
1706         PRINTK("+++ raid6d active\n");
1707
1708         md_check_recovery(mddev);
1709
1710         handled = 0;
1711         spin_lock_irq(&conf->device_lock);
1712         while (1) {
1713                 struct list_head *first;
1714
1715                 if (list_empty(&conf->handle_list) &&
1716                     atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
1717                     !blk_queue_plugged(mddev->queue) &&
1718                     !list_empty(&conf->delayed_list))
1719                         raid6_activate_delayed(conf);
1720
1721                 if (list_empty(&conf->handle_list))
1722                         break;
1723
1724                 first = conf->handle_list.next;
1725                 sh = list_entry(first, struct stripe_head, lru);
1726
1727                 list_del_init(first);
1728                 atomic_inc(&sh->count);
1729                 if (atomic_read(&sh->count)!= 1)
1730                         BUG();
1731                 spin_unlock_irq(&conf->device_lock);
1732
1733                 handled++;
1734                 handle_stripe(sh);
1735                 release_stripe(sh);
1736
1737                 spin_lock_irq(&conf->device_lock);
1738         }
1739         PRINTK("%d stripes handled\n", handled);
1740
1741         spin_unlock_irq(&conf->device_lock);
1742
1743         unplug_slaves(mddev);
1744
1745         PRINTK("--- raid6d inactive\n");
1746 }
1747
1748 static int run (mddev_t *mddev)
1749 {
1750         raid6_conf_t *conf;
1751         int raid_disk, memory;
1752         mdk_rdev_t *rdev;
1753         struct disk_info *disk;
1754         struct list_head *tmp;
1755
1756         if (mddev->level != 6) {
1757                 PRINTK("raid6: %s: raid level not set to 6 (%d)\n", mdname(mddev), mddev->level);
1758                 return -EIO;
1759         }
1760
1761         mddev->private = kmalloc (sizeof (raid6_conf_t)
1762                                   + mddev->raid_disks * sizeof(struct disk_info),
1763                                   GFP_KERNEL);
1764         if ((conf = mddev->private) == NULL)
1765                 goto abort;
1766         memset (conf, 0, sizeof (*conf) + mddev->raid_disks * sizeof(struct disk_info) );
1767         conf->mddev = mddev;
1768
1769         if ((conf->stripe_hashtbl = (struct stripe_head **) __get_free_pages(GFP_ATOMIC, HASH_PAGES_ORDER)) == NULL)
1770                 goto abort;
1771         memset(conf->stripe_hashtbl, 0, HASH_PAGES * PAGE_SIZE);
1772
1773         spin_lock_init(&conf->device_lock);
1774         init_waitqueue_head(&conf->wait_for_stripe);
1775         init_waitqueue_head(&conf->wait_for_overlap);
1776         INIT_LIST_HEAD(&conf->handle_list);
1777         INIT_LIST_HEAD(&conf->delayed_list);
1778         INIT_LIST_HEAD(&conf->inactive_list);
1779         atomic_set(&conf->active_stripes, 0);
1780         atomic_set(&conf->preread_active_stripes, 0);
1781
1782         PRINTK("raid6: run(%s) called.\n", mdname(mddev));
1783
1784         ITERATE_RDEV(mddev,rdev,tmp) {
1785                 raid_disk = rdev->raid_disk;
1786                 if (raid_disk >= mddev->raid_disks
1787                     || raid_disk < 0)
1788                         continue;
1789                 disk = conf->disks + raid_disk;
1790
1791                 disk->rdev = rdev;
1792
1793                 if (rdev->in_sync) {
1794                         char b[BDEVNAME_SIZE];
1795                         printk(KERN_INFO "raid6: device %s operational as raid"
1796                                " disk %d\n", bdevname(rdev->bdev,b),
1797                                raid_disk);
1798                         conf->working_disks++;
1799                 }
1800         }
1801
1802         conf->raid_disks = mddev->raid_disks;
1803
1804         /*
1805          * 0 for a fully functional array, 1 or 2 for a degraded array.
1806          */
1807         mddev->degraded = conf->failed_disks = conf->raid_disks - conf->working_disks;
1808         conf->mddev = mddev;
1809         conf->chunk_size = mddev->chunk_size;
1810         conf->level = mddev->level;
1811         conf->algorithm = mddev->layout;
1812         conf->max_nr_stripes = NR_STRIPES;
1813
1814         /* device size must be a multiple of chunk size */
1815         mddev->size &= ~(mddev->chunk_size/1024 -1);
1816
1817         if (conf->raid_disks < 4) {
1818                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
1819                        mdname(mddev), conf->raid_disks);
1820                 goto abort;
1821         }
1822         if (!conf->chunk_size || conf->chunk_size % 4) {
1823                 printk(KERN_ERR "raid6: invalid chunk size %d for %s\n",
1824                        conf->chunk_size, mdname(mddev));
1825                 goto abort;
1826         }
1827         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
1828                 printk(KERN_ERR
1829                        "raid6: unsupported parity algorithm %d for %s\n",
1830                        conf->algorithm, mdname(mddev));
1831                 goto abort;
1832         }
1833         if (mddev->degraded > 2) {
1834                 printk(KERN_ERR "raid6: not enough operational devices for %s"
1835                        " (%d/%d failed)\n",
1836                        mdname(mddev), conf->failed_disks, conf->raid_disks);
1837                 goto abort;
1838         }
1839
1840 #if 0                           /* FIX: For now */
1841         if (mddev->degraded > 0 &&
1842             mddev->recovery_cp != MaxSector) {
1843                 printk(KERN_ERR "raid6: cannot start dirty degraded array for %s\n", mdname(mddev));
1844                 goto abort;
1845         }
1846 #endif
1847
1848         {
1849                 mddev->thread = md_register_thread(raid6d, mddev, "%s_raid6");
1850                 if (!mddev->thread) {
1851                         printk(KERN_ERR
1852                                "raid6: couldn't allocate thread for %s\n",
1853                                mdname(mddev));
1854                         goto abort;
1855                 }
1856         }
1857
1858         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
1859                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
1860         if (grow_stripes(conf, conf->max_nr_stripes)) {
1861                 printk(KERN_ERR
1862                        "raid6: couldn't allocate %dkB for buffers\n", memory);
1863                 shrink_stripes(conf);
1864                 md_unregister_thread(mddev->thread);
1865                 goto abort;
1866         } else
1867                 printk(KERN_INFO "raid6: allocated %dkB for %s\n",
1868                        memory, mdname(mddev));
1869
1870         if (mddev->degraded == 0)
1871                 printk(KERN_INFO "raid6: raid level %d set %s active with %d out of %d"
1872                        " devices, algorithm %d\n", conf->level, mdname(mddev),
1873                        mddev->raid_disks-mddev->degraded, mddev->raid_disks,
1874                        conf->algorithm);
1875         else
1876                 printk(KERN_ALERT "raid6: raid level %d set %s active with %d"
1877                        " out of %d devices, algorithm %d\n", conf->level,
1878                        mdname(mddev), mddev->raid_disks - mddev->degraded,
1879                        mddev->raid_disks, conf->algorithm);
1880
1881         print_raid6_conf(conf);
1882
1883         /* read-ahead size must cover two whole stripes, which is
1884          * 2 * (n-2) * chunksize where 'n' is the number of raid devices
1885          */
1886         {
1887                 int stripe = (mddev->raid_disks-2) * mddev->chunk_size
1888                         / PAGE_CACHE_SIZE;
1889                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
1890                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
1891         }
1892
1893         /* Ok, everything is just fine now */
1894         mddev->array_size =  mddev->size * (mddev->raid_disks - 2);
1895
1896         mddev->queue->unplug_fn = raid6_unplug_device;
1897         mddev->queue->issue_flush_fn = raid6_issue_flush;
1898         return 0;
1899 abort:
1900         if (conf) {
1901                 print_raid6_conf(conf);
1902                 if (conf->stripe_hashtbl)
1903                         free_pages((unsigned long) conf->stripe_hashtbl,
1904                                                         HASH_PAGES_ORDER);
1905                 kfree(conf);
1906         }
1907         mddev->private = NULL;
1908         printk(KERN_ALERT "raid6: failed to run raid set %s\n", mdname(mddev));
1909         return -EIO;
1910 }
1911
1912
1913
1914 static int stop (mddev_t *mddev)
1915 {
1916         raid6_conf_t *conf = (raid6_conf_t *) mddev->private;
1917
1918         md_unregister_thread(mddev->thread);
1919         mddev->thread = NULL;
1920         shrink_stripes(conf);
1921         free_pages((unsigned long) conf->stripe_hashtbl, HASH_PAGES_ORDER);
1922         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
1923         kfree(conf);
1924         mddev->private = NULL;
1925         return 0;
1926 }
1927
1928 #if RAID6_DUMPSTATE
1929 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
1930 {
1931         int i;
1932
1933         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
1934                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
1935         seq_printf(seq, "sh %llu,  count %d.\n",
1936                    (unsigned long long)sh->sector, atomic_read(&sh->count));
1937         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
1938         for (i = 0; i < sh->raid_conf->raid_disks; i++) {
1939                 seq_printf(seq, "(cache%d: %p %ld) ",
1940                            i, sh->dev[i].page, sh->dev[i].flags);
1941         }
1942         seq_printf(seq, "\n");
1943 }
1944
1945 static void printall (struct seq_file *seq, raid6_conf_t *conf)
1946 {
1947         struct stripe_head *sh;
1948         int i;
1949
1950         spin_lock_irq(&conf->device_lock);
1951         for (i = 0; i < NR_HASH; i++) {
1952                 sh = conf->stripe_hashtbl[i];
1953                 for (; sh; sh = sh->hash_next) {
1954                         if (sh->raid_conf != conf)
1955                                 continue;
1956                         print_sh(seq, sh);
1957                 }
1958         }
1959         spin_unlock_irq(&conf->device_lock);
1960 }
1961 #endif
1962
1963 static void status (struct seq_file *seq, mddev_t *mddev)
1964 {
1965         raid6_conf_t *conf = (raid6_conf_t *) mddev->private;
1966         int i;
1967
1968         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
1969         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->working_disks);
1970         for (i = 0; i < conf->raid_disks; i++)
1971                 seq_printf (seq, "%s",
1972                             conf->disks[i].rdev &&
1973                             conf->disks[i].rdev->in_sync ? "U" : "_");
1974         seq_printf (seq, "]");
1975 #if RAID6_DUMPSTATE
1976         seq_printf (seq, "\n");
1977         printall(seq, conf);
1978 #endif
1979 }
1980
1981 static void print_raid6_conf (raid6_conf_t *conf)
1982 {
1983         int i;
1984         struct disk_info *tmp;
1985
1986         printk("RAID6 conf printout:\n");
1987         if (!conf) {
1988                 printk("(conf==NULL)\n");
1989                 return;
1990         }
1991         printk(" --- rd:%d wd:%d fd:%d\n", conf->raid_disks,
1992                  conf->working_disks, conf->failed_disks);
1993
1994         for (i = 0; i < conf->raid_disks; i++) {
1995                 char b[BDEVNAME_SIZE];
1996                 tmp = conf->disks + i;
1997                 if (tmp->rdev)
1998                 printk(" disk %d, o:%d, dev:%s\n",
1999                         i, !tmp->rdev->faulty,
2000                         bdevname(tmp->rdev->bdev,b));
2001         }
2002 }
2003
2004 static int raid6_spare_active(mddev_t *mddev)
2005 {
2006         int i;
2007         raid6_conf_t *conf = mddev->private;
2008         struct disk_info *tmp;
2009
2010         for (i = 0; i < conf->raid_disks; i++) {
2011                 tmp = conf->disks + i;
2012                 if (tmp->rdev
2013                     && !tmp->rdev->faulty
2014                     && !tmp->rdev->in_sync) {
2015                         mddev->degraded--;
2016                         conf->failed_disks--;
2017                         conf->working_disks++;
2018                         tmp->rdev->in_sync = 1;
2019                 }
2020         }
2021         print_raid6_conf(conf);
2022         return 0;
2023 }
2024
2025 static int raid6_remove_disk(mddev_t *mddev, int number)
2026 {
2027         raid6_conf_t *conf = mddev->private;
2028         int err = 0;
2029         mdk_rdev_t *rdev;
2030         struct disk_info *p = conf->disks + number;
2031
2032         print_raid6_conf(conf);
2033         rdev = p->rdev;
2034         if (rdev) {
2035                 if (rdev->in_sync ||
2036                     atomic_read(&rdev->nr_pending)) {
2037                         err = -EBUSY;
2038                         goto abort;
2039                 }
2040                 p->rdev = NULL;
2041                 synchronize_rcu();
2042                 if (atomic_read(&rdev->nr_pending)) {
2043                         /* lost the race, try later */
2044                         err = -EBUSY;
2045                         p->rdev = rdev;
2046                 }
2047         }
2048
2049 abort:
2050
2051         print_raid6_conf(conf);
2052         return err;
2053 }
2054
2055 static int raid6_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
2056 {
2057         raid6_conf_t *conf = mddev->private;
2058         int found = 0;
2059         int disk;
2060         struct disk_info *p;
2061
2062         if (mddev->degraded > 2)
2063                 /* no point adding a device */
2064                 return 0;
2065         /*
2066          * find the disk ...
2067          */
2068         for (disk=0; disk < mddev->raid_disks; disk++)
2069                 if ((p=conf->disks + disk)->rdev == NULL) {
2070                         rdev->in_sync = 0;
2071                         rdev->raid_disk = disk;
2072                         found = 1;
2073                         p->rdev = rdev;
2074                         break;
2075                 }
2076         print_raid6_conf(conf);
2077         return found;
2078 }
2079
2080 static int raid6_resize(mddev_t *mddev, sector_t sectors)
2081 {
2082         /* no resync is happening, and there is enough space
2083          * on all devices, so we can resize.
2084          * We need to make sure resync covers any new space.
2085          * If the array is shrinking we should possibly wait until
2086          * any io in the removed space completes, but it hardly seems
2087          * worth it.
2088          */
2089         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
2090         mddev->array_size = (sectors * (mddev->raid_disks-2))>>1;
2091         set_capacity(mddev->gendisk, mddev->array_size << 1);
2092         mddev->changed = 1;
2093         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
2094                 mddev->recovery_cp = mddev->size << 1;
2095                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2096         }
2097         mddev->size = sectors /2;
2098         mddev->resync_max_sectors = sectors;
2099         return 0;
2100 }
2101
2102 static mdk_personality_t raid6_personality=
2103 {
2104         .name           = "raid6",
2105         .owner          = THIS_MODULE,
2106         .make_request   = make_request,
2107         .run            = run,
2108         .stop           = stop,
2109         .status         = status,
2110         .error_handler  = error,
2111         .hot_add_disk   = raid6_add_disk,
2112         .hot_remove_disk= raid6_remove_disk,
2113         .spare_active   = raid6_spare_active,
2114         .sync_request   = sync_request,
2115         .resize         = raid6_resize,
2116 };
2117
2118 static int __init raid6_init (void)
2119 {
2120         int e;
2121
2122         e = raid6_select_algo();
2123         if ( e )
2124                 return e;
2125
2126         return register_md_personality (RAID6, &raid6_personality);
2127 }
2128
2129 static void raid6_exit (void)
2130 {
2131         unregister_md_personality (RAID6);
2132 }
2133
2134 module_init(raid6_init);
2135 module_exit(raid6_exit);
2136 MODULE_LICENSE("GPL");
2137 MODULE_ALIAS("md-personality-8"); /* RAID6 */