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 #include <linux/raid/bitmap.h>
  33
  34 /*
  35  * Stripe cache
  36  */
  37
  38 #define NR_STRIPES              256
  39 #define STRIPE_SIZE             PAGE_SIZE
  40 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
  41 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
  42 #define IO_THRESHOLD            1
  43 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_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 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 if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 102                                  conf->seq_write == sh->bm_seq)
 103                                 list_add_tail(&sh->lru, &conf->bitmap_list);
 104                         else {
 105                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
 106                                 list_add_tail(&sh->lru, &conf->handle_list);
 107                         }
 108                         md_wakeup_thread(conf->mddev->thread);
 109                 } else {
 110                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
 111                                 atomic_dec(&conf->preread_active_stripes);
 112                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
 113                                         md_wakeup_thread(conf->mddev->thread);
 114                         }
 115                         list_add_tail(&sh->lru, &conf->inactive_list);
 116                         atomic_dec(&conf->active_stripes);
 117                         if (!conf->inactive_blocked ||
 118                             atomic_read(&conf->active_stripes) < (conf->max_nr_stripes*3/4))
 119                                 wake_up(&conf->wait_for_stripe);
 120                 }
 121         }
 122 }
 123 static void release_stripe(struct stripe_head *sh)
 124 {
 125         raid6_conf_t *conf = sh->raid_conf;
 126         unsigned long flags;
 127
 128         spin_lock_irqsave(&conf->device_lock, flags);
 129         __release_stripe(conf, sh);
 130         spin_unlock_irqrestore(&conf->device_lock, flags);
 131 }
 132
 133 static inline void remove_hash(struct stripe_head *sh)
 134 {
 135         PRINTK("remove_hash(), stripe %llu\n", (unsigned long long)sh->sector);
 136
 137         hlist_del_init(&sh->hash);
 138 }
 139
 140 static inline void insert_hash(raid6_conf_t *conf, struct stripe_head *sh)
 141 {
 142         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 143
 144         PRINTK("insert_hash(), stripe %llu\n", (unsigned long long)sh->sector);
 145
 146         CHECK_DEVLOCK();
 147         hlist_add_head(&sh->hash, hp);
 148 }
 149
 150
 151 /* find an idle stripe, make sure it is unhashed, and return it. */
 152 static struct stripe_head *get_free_stripe(raid6_conf_t *conf)
 153 {
 154         struct stripe_head *sh = NULL;
 155         struct list_head *first;
 156
 157         CHECK_DEVLOCK();
 158         if (list_empty(&conf->inactive_list))
 159                 goto out;
 160         first = conf->inactive_list.next;
 161         sh = list_entry(first, struct stripe_head, lru);
 162         list_del_init(first);
 163         remove_hash(sh);
 164         atomic_inc(&conf->active_stripes);
 165 out:
 166         return sh;
 167 }
 168
 169 static void shrink_buffers(struct stripe_head *sh, int num)
 170 {
 171         struct page *p;
 172         int i;
 173
 174         for (i=0; i<num ; i++) {
 175                 p = sh->dev[i].page;
 176                 if (!p)
 177                         continue;
 178                 sh->dev[i].page = NULL;
 179                 put_page(p);
 180         }
 181 }
 182
 183 static int grow_buffers(struct stripe_head *sh, int num)
 184 {
 185         int i;
 186
 187         for (i=0; i<num; i++) {
 188                 struct page *page;
 189
 190                 if (!(page = alloc_page(GFP_KERNEL))) {
 191                         return 1;
 192                 }
 193                 sh->dev[i].page = page;
 194         }
 195         return 0;
 196 }
 197
 198 static void raid6_build_block (struct stripe_head *sh, int i);
 199
 200 static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx)
 201 {
 202         raid6_conf_t *conf = sh->raid_conf;
 203         int disks = conf->raid_disks, i;
 204
 205         if (atomic_read(&sh->count) != 0)
 206                 BUG();
 207         if (test_bit(STRIPE_HANDLE, &sh->state))
 208                 BUG();
 209
 210         CHECK_DEVLOCK();
 211         PRINTK("init_stripe called, stripe %llu\n",
 212                 (unsigned long long)sh->sector);
 213
 214         remove_hash(sh);
 215
 216         sh->sector = sector;
 217         sh->pd_idx = pd_idx;
 218         sh->state = 0;
 219
 220         for (i=disks; i--; ) {
 221                 struct r5dev *dev = &sh->dev[i];
 222
 223                 if (dev->toread || dev->towrite || dev->written ||
 224                     test_bit(R5_LOCKED, &dev->flags)) {
 225                         PRINTK("sector=%llx i=%d %p %p %p %d\n",
 226                                (unsigned long long)sh->sector, i, dev->toread,
 227                                dev->towrite, dev->written,
 228                                test_bit(R5_LOCKED, &dev->flags));
 229                         BUG();
 230                 }
 231                 dev->flags = 0;
 232                 raid6_build_block(sh, i);
 233         }
 234         insert_hash(conf, sh);
 235 }
 236
 237 static struct stripe_head *__find_stripe(raid6_conf_t *conf, sector_t sector)
 238 {
 239         struct stripe_head *sh;
 240         struct hlist_node *hn;
 241
 242         CHECK_DEVLOCK();
 243         PRINTK("__find_stripe, sector %llu\n", (unsigned long long)sector);
 244         hlist_for_each_entry (sh, hn,  stripe_hash(conf, sector), hash)
 245                 if (sh->sector == sector)
 246                         return sh;
 247         PRINTK("__stripe %llu not in cache\n", (unsigned long long)sector);
 248         return NULL;
 249 }
 250
 251 static void unplug_slaves(mddev_t *mddev);
 252
 253 static struct stripe_head *get_active_stripe(raid6_conf_t *conf, sector_t sector,
 254                                              int pd_idx, int noblock)
 255 {
 256         struct stripe_head *sh;
 257
 258         PRINTK("get_stripe, sector %llu\n", (unsigned long long)sector);
 259
 260         spin_lock_irq(&conf->device_lock);
 261
 262         do {
 263                 wait_event_lock_irq(conf->wait_for_stripe,
 264                                     conf->quiesce == 0,
 265                                     conf->device_lock, /* nothing */);
 266                 sh = __find_stripe(conf, sector);
 267                 if (!sh) {
 268                         if (!conf->inactive_blocked)
 269                                 sh = get_free_stripe(conf);
 270                         if (noblock && sh == NULL)
 271                                 break;
 272                         if (!sh) {
 273                                 conf->inactive_blocked = 1;
 274                                 wait_event_lock_irq(conf->wait_for_stripe,
 275                                                     !list_empty(&conf->inactive_list) &&
 276                                                     (atomic_read(&conf->active_stripes)
 277                                                      < (conf->max_nr_stripes *3/4)
 278                                                      || !conf->inactive_blocked),
 279                                                     conf->device_lock,
 280                                                     unplug_slaves(conf->mddev);
 281                                         );
 282                                 conf->inactive_blocked = 0;
 283                         } else
 284                                 init_stripe(sh, sector, pd_idx);
 285                 } else {
 286                         if (atomic_read(&sh->count)) {
 287                                 if (!list_empty(&sh->lru))
 288                                         BUG();
 289                         } else {
 290                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 291                                         atomic_inc(&conf->active_stripes);
 292                                 if (list_empty(&sh->lru))
 293                                         BUG();
 294                                 list_del_init(&sh->lru);
 295                         }
 296                 }
 297         } while (sh == NULL);
 298
 299         if (sh)
 300                 atomic_inc(&sh->count);
 301
 302         spin_unlock_irq(&conf->device_lock);
 303         return sh;
 304 }
 305
 306 static int grow_one_stripe(raid6_conf_t *conf)
 307 {
 308         struct stripe_head *sh;
 309         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
 310         if (!sh)
 311                 return 0;
 312         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
 313         sh->raid_conf = conf;
 314         spin_lock_init(&sh->lock);
 315
 316         if (grow_buffers(sh, conf->raid_disks)) {
 317                 shrink_buffers(sh, conf->raid_disks);
 318                 kmem_cache_free(conf->slab_cache, sh);
 319                 return 0;
 320         }
 321         /* we just created an active stripe so... */
 322         atomic_set(&sh->count, 1);
 323         atomic_inc(&conf->active_stripes);
 324         INIT_LIST_HEAD(&sh->lru);
 325         release_stripe(sh);
 326         return 1;
 327 }
 328
 329 static int grow_stripes(raid6_conf_t *conf, int num)
 330 {
 331         kmem_cache_t *sc;
 332         int devs = conf->raid_disks;
 333
 334         sprintf(conf->cache_name, "raid6/%s", mdname(conf->mddev));
 335
 336         sc = kmem_cache_create(conf->cache_name,
 337                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
 338                                0, 0, NULL, NULL);
 339         if (!sc)
 340                 return 1;
 341         conf->slab_cache = sc;
 342         while (num--)
 343                 if (!grow_one_stripe(conf))
 344                         return 1;
 345         return 0;
 346 }
 347
 348 static int drop_one_stripe(raid6_conf_t *conf)
 349 {
 350         struct stripe_head *sh;
 351         spin_lock_irq(&conf->device_lock);
 352         sh = get_free_stripe(conf);
 353         spin_unlock_irq(&conf->device_lock);
 354         if (!sh)
 355                 return 0;
 356         if (atomic_read(&sh->count))
 357                 BUG();
 358         shrink_buffers(sh, conf->raid_disks);
 359         kmem_cache_free(conf->slab_cache, sh);
 360         atomic_dec(&conf->active_stripes);
 361         return 1;
 362 }
 363
 364 static void shrink_stripes(raid6_conf_t *conf)
 365 {
 366         while (drop_one_stripe(conf))
 367                 ;
 368
 369         if (conf->slab_cache)
 370                 kmem_cache_destroy(conf->slab_cache);
 371         conf->slab_cache = NULL;
 372 }
 373
 374 static int raid6_end_read_request(struct bio * bi, unsigned int bytes_done,
 375                                   int error)
 376 {
 377         struct stripe_head *sh = bi->bi_private;
 378         raid6_conf_t *conf = sh->raid_conf;
 379         int disks = conf->raid_disks, i;
 380         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
 381
 382         if (bi->bi_size)
 383                 return 1;
 384
 385         for (i=0 ; i<disks; i++)
 386                 if (bi == &sh->dev[i].req)
 387                         break;
 388
 389         PRINTK("end_read_request %llu/%d, count: %d, uptodate %d.\n",
 390                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
 391                 uptodate);
 392         if (i == disks) {
 393                 BUG();
 394                 return 0;
 395         }
 396
 397         if (uptodate) {
 398 #if 0
 399                 struct bio *bio;
 400                 unsigned long flags;
 401                 spin_lock_irqsave(&conf->device_lock, flags);
 402                 /* we can return a buffer if we bypassed the cache or
 403                  * if the top buffer is not in highmem.  If there are
 404                  * multiple buffers, leave the extra work to
 405                  * handle_stripe
 406                  */
 407                 buffer = sh->bh_read[i];
 408                 if (buffer &&
 409                     (!PageHighMem(buffer->b_page)
 410                      || buffer->b_page == bh->b_page )
 411                         ) {
 412                         sh->bh_read[i] = buffer->b_reqnext;
 413                         buffer->b_reqnext = NULL;
 414                 } else
 415                         buffer = NULL;
 416                 spin_unlock_irqrestore(&conf->device_lock, flags);
 417                 if (sh->bh_page[i]==bh->b_page)
 418                         set_buffer_uptodate(bh);
 419                 if (buffer) {
 420                         if (buffer->b_page != bh->b_page)
 421                                 memcpy(buffer->b_data, bh->b_data, bh->b_size);
 422                         buffer->b_end_io(buffer, 1);
 423                 }
 424 #else
 425                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
 426 #endif
 427                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
 428                         printk(KERN_INFO "raid6: read error corrected!!\n");
 429                         clear_bit(R5_ReadError, &sh->dev[i].flags);
 430                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
 431                 }
 432                 if (atomic_read(&conf->disks[i].rdev->read_errors))
 433                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
 434         } else {
 435                 int retry = 0;
 436                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
 437                 atomic_inc(&conf->disks[i].rdev->read_errors);
 438                 if (conf->mddev->degraded)
 439                         printk(KERN_WARNING "raid6: read error not correctable.\n");
 440                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
 441                         /* Oh, no!!! */
 442                         printk(KERN_WARNING "raid6: read error NOT corrected!!\n");
 443                 else if (atomic_read(&conf->disks[i].rdev->read_errors)
 444                          > conf->max_nr_stripes)
 445                         printk(KERN_WARNING
 446                                "raid6: Too many read errors, failing device.\n");
 447                 else
 448                         retry = 1;
 449                 if (retry)
 450                         set_bit(R5_ReadError, &sh->dev[i].flags);
 451                 else {
 452                         clear_bit(R5_ReadError, &sh->dev[i].flags);
 453                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
 454                         md_error(conf->mddev, conf->disks[i].rdev);
 455                 }
 456         }
 457         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
 458 #if 0
 459         /* must restore b_page before unlocking buffer... */
 460         if (sh->bh_page[i] != bh->b_page) {
 461                 bh->b_page = sh->bh_page[i];
 462                 bh->b_data = page_address(bh->b_page);
 463                 clear_buffer_uptodate(bh);
 464         }
 465 #endif
 466         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 467         set_bit(STRIPE_HANDLE, &sh->state);
 468         release_stripe(sh);
 469         return 0;
 470 }
 471
 472 static int raid6_end_write_request (struct bio *bi, unsigned int bytes_done,
 473                                     int error)
 474 {
 475         struct stripe_head *sh = bi->bi_private;
 476         raid6_conf_t *conf = sh->raid_conf;
 477         int disks = conf->raid_disks, i;
 478         unsigned long flags;
 479         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
 480
 481         if (bi->bi_size)
 482                 return 1;
 483
 484         for (i=0 ; i<disks; i++)
 485                 if (bi == &sh->dev[i].req)
 486                         break;
 487
 488         PRINTK("end_write_request %llu/%d, count %d, uptodate: %d.\n",
 489                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
 490                 uptodate);
 491         if (i == disks) {
 492                 BUG();
 493                 return 0;
 494         }
 495
 496         spin_lock_irqsave(&conf->device_lock, flags);
 497         if (!uptodate)
 498                 md_error(conf->mddev, conf->disks[i].rdev);
 499
 500         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
 501
 502         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 503         set_bit(STRIPE_HANDLE, &sh->state);
 504         __release_stripe(conf, sh);
 505         spin_unlock_irqrestore(&conf->device_lock, flags);
 506         return 0;
 507 }
 508
 509
 510 static sector_t compute_blocknr(struct stripe_head *sh, int i);
 511
 512 static void raid6_build_block (struct stripe_head *sh, int i)
 513 {
 514         struct r5dev *dev = &sh->dev[i];
 515         int pd_idx = sh->pd_idx;
 516         int qd_idx = raid6_next_disk(pd_idx, sh->raid_conf->raid_disks);
 517
 518         bio_init(&dev->req);
 519         dev->req.bi_io_vec = &dev->vec;
 520         dev->req.bi_vcnt++;
 521         dev->req.bi_max_vecs++;
 522         dev->vec.bv_page = dev->page;
 523         dev->vec.bv_len = STRIPE_SIZE;
 524         dev->vec.bv_offset = 0;
 525
 526         dev->req.bi_sector = sh->sector;
 527         dev->req.bi_private = sh;
 528
 529         dev->flags = 0;
 530         if (i != pd_idx && i != qd_idx)
 531                 dev->sector = compute_blocknr(sh, i);
 532 }
 533
 534 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 535 {
 536         char b[BDEVNAME_SIZE];
 537         raid6_conf_t *conf = (raid6_conf_t *) mddev->private;
 538         PRINTK("raid6: error called\n");
 539
 540         if (!test_bit(Faulty, &rdev->flags)) {
 541                 mddev->sb_dirty = 1;
 542                 if (test_bit(In_sync, &rdev->flags)) {
 543                         conf->working_disks--;
 544                         mddev->degraded++;
 545                         conf->failed_disks++;
 546                         clear_bit(In_sync, &rdev->flags);
 547                         /*
 548                          * if recovery was running, make sure it aborts.
 549                          */
 550                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 551                 }
 552                 set_bit(Faulty, &rdev->flags);
 553                 printk (KERN_ALERT
 554                         "raid6: Disk failure on %s, disabling device."
 555                         " Operation continuing on %d devices\n",
 556                         bdevname(rdev->bdev,b), conf->working_disks);
 557         }
 558 }
 559
 560 /*
 561  * Input: a 'big' sector number,
 562  * Output: index of the data and parity disk, and the sector # in them.
 563  */
 564 static sector_t raid6_compute_sector(sector_t r_sector, unsigned int raid_disks,
 565                         unsigned int data_disks, unsigned int * dd_idx,
 566                         unsigned int * pd_idx, raid6_conf_t *conf)
 567 {
 568         long stripe;
 569         unsigned long chunk_number;
 570         unsigned int chunk_offset;
 571         sector_t new_sector;
 572         int sectors_per_chunk = conf->chunk_size >> 9;
 573
 574         /* First compute the information on this sector */
 575
 576         /*
 577          * Compute the chunk number and the sector offset inside the chunk
 578          */
 579         chunk_offset = sector_div(r_sector, sectors_per_chunk);
 580         chunk_number = r_sector;
 581         if ( r_sector != chunk_number ) {
 582                 printk(KERN_CRIT "raid6: ERROR: r_sector = %llu, chunk_number = %lu\n",
 583                        (unsigned long long)r_sector, (unsigned long)chunk_number);
 584                 BUG();
 585         }
 586
 587         /*
 588          * Compute the stripe number
 589          */
 590         stripe = chunk_number / data_disks;
 591
 592         /*
 593          * Compute the data disk and parity disk indexes inside the stripe
 594          */
 595         *dd_idx = chunk_number % data_disks;
 596
 597         /*
 598          * Select the parity disk based on the user selected algorithm.
 599          */
 600
 601         /**** FIX THIS ****/
 602         switch (conf->algorithm) {
 603         case ALGORITHM_LEFT_ASYMMETRIC:
 604                 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
 605                 if (*pd_idx == raid_disks-1)
 606                         (*dd_idx)++;    /* Q D D D P */
 607                 else if (*dd_idx >= *pd_idx)
 608                         (*dd_idx) += 2; /* D D P Q D */
 609                 break;
 610         case ALGORITHM_RIGHT_ASYMMETRIC:
 611                 *pd_idx = stripe % raid_disks;
 612                 if (*pd_idx == raid_disks-1)
 613                         (*dd_idx)++;    /* Q D D D P */
 614                 else if (*dd_idx >= *pd_idx)
 615                         (*dd_idx) += 2; /* D D P Q D */
 616                 break;
 617         case ALGORITHM_LEFT_SYMMETRIC:
 618                 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
 619                 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
 620                 break;
 621         case ALGORITHM_RIGHT_SYMMETRIC:
 622                 *pd_idx = stripe % raid_disks;
 623                 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
 624                 break;
 625         default:
 626                 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
 627                         conf->algorithm);
 628         }
 629
 630         PRINTK("raid6: chunk_number = %lu, pd_idx = %u, dd_idx = %u\n",
 631                chunk_number, *pd_idx, *dd_idx);
 632
 633         /*
 634          * Finally, compute the new sector number
 635          */
 636         new_sector = (sector_t) stripe * sectors_per_chunk + chunk_offset;
 637         return new_sector;
 638 }
 639
 640
 641 static sector_t compute_blocknr(struct stripe_head *sh, int i)
 642 {
 643         raid6_conf_t *conf = sh->raid_conf;
 644         int raid_disks = conf->raid_disks, data_disks = raid_disks - 2;
 645         sector_t new_sector = sh->sector, check;
 646         int sectors_per_chunk = conf->chunk_size >> 9;
 647         sector_t stripe;
 648         int chunk_offset;
 649         int chunk_number, dummy1, dummy2, dd_idx = i;
 650         sector_t r_sector;
 651         int i0 = i;
 652
 653         chunk_offset = sector_div(new_sector, sectors_per_chunk);
 654         stripe = new_sector;
 655         if ( new_sector != stripe ) {
 656                 printk(KERN_CRIT "raid6: ERROR: new_sector = %llu, stripe = %lu\n",
 657                        (unsigned long long)new_sector, (unsigned long)stripe);
 658                 BUG();
 659         }
 660
 661         switch (conf->algorithm) {
 662                 case ALGORITHM_LEFT_ASYMMETRIC:
 663                 case ALGORITHM_RIGHT_ASYMMETRIC:
 664                         if (sh->pd_idx == raid_disks-1)
 665                                 i--;    /* Q D D D P */
 666                         else if (i > sh->pd_idx)
 667                                 i -= 2; /* D D P Q D */
 668                         break;
 669                 case ALGORITHM_LEFT_SYMMETRIC:
 670                 case ALGORITHM_RIGHT_SYMMETRIC:
 671                         if (sh->pd_idx == raid_disks-1)
 672                                 i--; /* Q D D D P */
 673                         else {
 674                                 /* D D P Q D */
 675                                 if (i < sh->pd_idx)
 676                                         i += raid_disks;
 677                                 i -= (sh->pd_idx + 2);
 678                         }
 679                         break;
 680                 default:
 681                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
 682                                 conf->algorithm);
 683         }
 684
 685         PRINTK("raid6: compute_blocknr: pd_idx = %u, i0 = %u, i = %u\n", sh->pd_idx, i0, i);
 686
 687         chunk_number = stripe * data_disks + i;
 688         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
 689
 690         check = raid6_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
 691         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
 692                 printk(KERN_CRIT "raid6: compute_blocknr: map not correct\n");
 693                 return 0;
 694         }
 695         return r_sector;
 696 }
 697
 698
 699
 700 /*
 701  * Copy data between a page in the stripe cache, and one or more bion
 702  * The page could align with the middle of the bio, or there could be
 703  * several bion, each with several bio_vecs, which cover part of the page
 704  * Multiple bion are linked together on bi_next.  There may be extras
 705  * at the end of this list.  We ignore them.
 706  */
 707 static void copy_data(int frombio, struct bio *bio,
 708                      struct page *page,
 709                      sector_t sector)
 710 {
 711         char *pa = page_address(page);
 712         struct bio_vec *bvl;
 713         int i;
 714         int page_offset;
 715
 716         if (bio->bi_sector >= sector)
 717                 page_offset = (signed)(bio->bi_sector - sector) * 512;
 718         else
 719                 page_offset = (signed)(sector - bio->bi_sector) * -512;
 720         bio_for_each_segment(bvl, bio, i) {
 721                 int len = bio_iovec_idx(bio,i)->bv_len;
 722                 int clen;
 723                 int b_offset = 0;
 724
 725                 if (page_offset < 0) {
 726                         b_offset = -page_offset;
 727                         page_offset += b_offset;
 728                         len -= b_offset;
 729                 }
 730
 731                 if (len > 0 && page_offset + len > STRIPE_SIZE)
 732                         clen = STRIPE_SIZE - page_offset;
 733                 else clen = len;
 734
 735                 if (clen > 0) {
 736                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
 737                         if (frombio)
 738                                 memcpy(pa+page_offset, ba+b_offset, clen);
 739                         else
 740                                 memcpy(ba+b_offset, pa+page_offset, clen);
 741                         __bio_kunmap_atomic(ba, KM_USER0);
 742                 }
 743                 if (clen < len) /* hit end of page */
 744                         break;
 745                 page_offset +=  len;
 746         }
 747 }
 748
 749 #define check_xor()     do {                                            \
 750                            if (count == MAX_XOR_BLOCKS) {               \
 751                                 xor_block(count, STRIPE_SIZE, ptr);     \
 752                                 count = 1;                              \
 753                            }                                            \
 754                         } while(0)
 755
 756 /* Compute P and Q syndromes */
 757 static void compute_parity(struct stripe_head *sh, int method)
 758 {
 759         raid6_conf_t *conf = sh->raid_conf;
 760         int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = conf->raid_disks, count;
 761         struct bio *chosen;
 762         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
 763         void *ptrs[disks];
 764
 765         qd_idx = raid6_next_disk(pd_idx, disks);
 766         d0_idx = raid6_next_disk(qd_idx, disks);
 767
 768         PRINTK("compute_parity, stripe %llu, method %d\n",
 769                 (unsigned long long)sh->sector, method);
 770
 771         switch(method) {
 772         case READ_MODIFY_WRITE:
 773                 BUG();          /* READ_MODIFY_WRITE N/A for RAID-6 */
 774         case RECONSTRUCT_WRITE:
 775                 for (i= disks; i-- ;)
 776                         if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
 777                                 chosen = sh->dev[i].towrite;
 778                                 sh->dev[i].towrite = NULL;
 779
 780                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
 781                                         wake_up(&conf->wait_for_overlap);
 782
 783                                 if (sh->dev[i].written) BUG();
 784                                 sh->dev[i].written = chosen;
 785                         }
 786                 break;
 787         case CHECK_PARITY:
 788                 BUG();          /* Not implemented yet */
 789         }
 790
 791         for (i = disks; i--;)
 792                 if (sh->dev[i].written) {
 793                         sector_t sector = sh->dev[i].sector;
 794                         struct bio *wbi = sh->dev[i].written;
 795                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
 796                                 copy_data(1, wbi, sh->dev[i].page, sector);
 797                                 wbi = r5_next_bio(wbi, sector);
 798                         }
 799
 800                         set_bit(R5_LOCKED, &sh->dev[i].flags);
 801                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
 802                 }
 803
 804 //      switch(method) {
 805 //      case RECONSTRUCT_WRITE:
 806 //      case CHECK_PARITY:
 807 //      case UPDATE_PARITY:
 808                 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
 809                 /* FIX: Is this ordering of drives even remotely optimal? */
 810                 count = 0;
 811                 i = d0_idx;
 812                 do {
 813                         ptrs[count++] = page_address(sh->dev[i].page);
 814                         if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
 815                                 printk("block %d/%d not uptodate on parity calc\n", i,count);
 816                         i = raid6_next_disk(i, disks);
 817                 } while ( i != d0_idx );
 818 //              break;
 819 //      }
 820
 821         raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
 822
 823         switch(method) {
 824         case RECONSTRUCT_WRITE:
 825                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
 826                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
 827                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
 828                 set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
 829                 break;
 830         case UPDATE_PARITY:
 831                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
 832                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
 833                 break;
 834         }
 835 }
 836
 837 /* Compute one missing block */
 838 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
 839 {
 840         raid6_conf_t *conf = sh->raid_conf;
 841         int i, count, disks = conf->raid_disks;
 842         void *ptr[MAX_XOR_BLOCKS], *p;
 843         int pd_idx = sh->pd_idx;
 844         int qd_idx = raid6_next_disk(pd_idx, disks);
 845
 846         PRINTK("compute_block_1, stripe %llu, idx %d\n",
 847                 (unsigned long long)sh->sector, dd_idx);
 848
 849         if ( dd_idx == qd_idx ) {
 850                 /* We're actually computing the Q drive */
 851                 compute_parity(sh, UPDATE_PARITY);
 852         } else {
 853                 ptr[0] = page_address(sh->dev[dd_idx].page);
 854                 if (!nozero) memset(ptr[0], 0, STRIPE_SIZE);
 855                 count = 1;
 856                 for (i = disks ; i--; ) {
 857                         if (i == dd_idx || i == qd_idx)
 858                                 continue;
 859                         p = page_address(sh->dev[i].page);
 860                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
 861                                 ptr[count++] = p;
 862                         else
 863                                 printk("compute_block() %d, stripe %llu, %d"
 864                                        " not present\n", dd_idx,
 865                                        (unsigned long long)sh->sector, i);
 866
 867                         check_xor();
 868                 }
 869                 if (count != 1)
 870                         xor_block(count, STRIPE_SIZE, ptr);
 871                 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
 872                 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
 873         }
 874 }
 875
 876 /* Compute two missing blocks */
 877 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
 878 {
 879         raid6_conf_t *conf = sh->raid_conf;
 880         int i, count, disks = conf->raid_disks;
 881         int pd_idx = sh->pd_idx;
 882         int qd_idx = raid6_next_disk(pd_idx, disks);
 883         int d0_idx = raid6_next_disk(qd_idx, disks);
 884         int faila, failb;
 885
 886         /* faila and failb are disk numbers relative to d0_idx */
 887         /* pd_idx become disks-2 and qd_idx become disks-1 */
 888         faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
 889         failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
 890
 891         BUG_ON(faila == failb);
 892         if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
 893
 894         PRINTK("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
 895                (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
 896
 897         if ( failb == disks-1 ) {
 898                 /* Q disk is one of the missing disks */
 899                 if ( faila == disks-2 ) {
 900                         /* Missing P+Q, just recompute */
 901                         compute_parity(sh, UPDATE_PARITY);
 902                         return;
 903                 } else {
 904                         /* We're missing D+Q; recompute D from P */
 905                         compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
 906                         compute_parity(sh, UPDATE_PARITY); /* Is this necessary? */
 907                         return;
 908                 }
 909         }
 910
 911         /* We're missing D+P or D+D; build pointer table */
 912         {
 913                 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
 914                 void *ptrs[disks];
 915
 916                 count = 0;
 917                 i = d0_idx;
 918                 do {
 919                         ptrs[count++] = page_address(sh->dev[i].page);
 920                         i = raid6_next_disk(i, disks);
 921                         if (i != dd_idx1 && i != dd_idx2 &&
 922                             !test_bit(R5_UPTODATE, &sh->dev[i].flags))
 923                                 printk("compute_2 with missing block %d/%d\n", count, i);
 924                 } while ( i != d0_idx );
 925
 926                 if ( failb == disks-2 ) {
 927                         /* We're missing D+P. */
 928                         raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
 929                 } else {
 930                         /* We're missing D+D. */
 931                         raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
 932                 }
 933
 934                 /* Both the above update both missing blocks */
 935                 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
 936                 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
 937         }
 938 }
 939
 940
 941 /*
 942  * Each stripe/dev can have one or more bion attached.
 943  * toread/towrite point to the first in a chain.
 944  * The bi_next chain must be in order.
 945  */
 946 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
 947 {
 948         struct bio **bip;
 949         raid6_conf_t *conf = sh->raid_conf;
 950         int firstwrite=0;
 951
 952         PRINTK("adding bh b#%llu to stripe s#%llu\n",
 953                 (unsigned long long)bi->bi_sector,
 954                 (unsigned long long)sh->sector);
 955
 956
 957         spin_lock(&sh->lock);
 958         spin_lock_irq(&conf->device_lock);
 959         if (forwrite) {
 960                 bip = &sh->dev[dd_idx].towrite;
 961                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
 962                         firstwrite = 1;
 963         } else
 964                 bip = &sh->dev[dd_idx].toread;
 965         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
 966                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
 967                         goto overlap;
 968                 bip = &(*bip)->bi_next;
 969         }
 970         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
 971                 goto overlap;
 972
 973         if (*bip && bi->bi_next && (*bip) != bi->bi_next)
 974                 BUG();
 975         if (*bip)
 976                 bi->bi_next = *bip;
 977         *bip = bi;
 978         bi->bi_phys_segments ++;
 979         spin_unlock_irq(&conf->device_lock);
 980         spin_unlock(&sh->lock);
 981
 982         PRINTK("added bi b#%llu to stripe s#%llu, disk %d.\n",
 983                 (unsigned long long)bi->bi_sector,
 984                 (unsigned long long)sh->sector, dd_idx);
 985
 986         if (conf->mddev->bitmap && firstwrite) {
 987                 sh->bm_seq = conf->seq_write;
 988                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
 989                                   STRIPE_SECTORS, 0);
 990                 set_bit(STRIPE_BIT_DELAY, &sh->state);
 991         }
 992
 993         if (forwrite) {
 994                 /* check if page is covered */
 995                 sector_t sector = sh->dev[dd_idx].sector;
 996                 for (bi=sh->dev[dd_idx].towrite;
 997                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
 998                              bi && bi->bi_sector <= sector;
 999                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1000                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1001                                 sector = bi->bi_sector + (bi->bi_size>>9);
1002                 }
1003                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1004                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1005         }
1006         return 1;
1007
1008  overlap:
1009         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1010         spin_unlock_irq(&conf->device_lock);
1011         spin_unlock(&sh->lock);
1012         return 0;
1013 }
1014
1015
1016 static int page_is_zero(struct page *p)
1017 {
1018         char *a = page_address(p);
1019         return ((*(u32*)a) == 0 &&
1020                 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1021 }
1022 /*
1023  * handle_stripe - do things to a stripe.
1024  *
1025  * We lock the stripe and then examine the state of various bits
1026  * to see what needs to be done.
1027  * Possible results:
1028  *    return some read request which now have data
1029  *    return some write requests which are safely on disc
1030  *    schedule a read on some buffers
1031  *    schedule a write of some buffers
1032  *    return confirmation of parity correctness
1033  *
1034  * Parity calculations are done inside the stripe lock
1035  * buffers are taken off read_list or write_list, and bh_cache buffers
1036  * get BH_Lock set before the stripe lock is released.
1037  *
1038  */
1039
1040 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
1041 {
1042         raid6_conf_t *conf = sh->raid_conf;
1043         int disks = conf->raid_disks;
1044         struct bio *return_bi= NULL;
1045         struct bio *bi;
1046         int i;
1047         int syncing;
1048         int locked=0, uptodate=0, to_read=0, to_write=0, failed=0, written=0;
1049         int non_overwrite = 0;
1050         int failed_num[2] = {0, 0};
1051         struct r5dev *dev, *pdev, *qdev;
1052         int pd_idx = sh->pd_idx;
1053         int qd_idx = raid6_next_disk(pd_idx, disks);
1054         int p_failed, q_failed;
1055
1056         PRINTK("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d, qd_idx=%d\n",
1057                (unsigned long long)sh->sector, sh->state, atomic_read(&sh->count),
1058                pd_idx, qd_idx);
1059
1060         spin_lock(&sh->lock);
1061         clear_bit(STRIPE_HANDLE, &sh->state);
1062         clear_bit(STRIPE_DELAYED, &sh->state);
1063
1064         syncing = test_bit(STRIPE_SYNCING, &sh->state);
1065         /* Now to look around and see what can be done */
1066
1067         rcu_read_lock();
1068         for (i=disks; i--; ) {
1069                 mdk_rdev_t *rdev;
1070                 dev = &sh->dev[i];
1071                 clear_bit(R5_Insync, &dev->flags);
1072
1073                 PRINTK("check %d: state 0x%lx read %p write %p written %p\n",
1074                         i, dev->flags, dev->toread, dev->towrite, dev->written);
1075                 /* maybe we can reply to a read */
1076                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
1077                         struct bio *rbi, *rbi2;
1078                         PRINTK("Return read for disc %d\n", i);
1079                         spin_lock_irq(&conf->device_lock);
1080                         rbi = dev->toread;
1081                         dev->toread = NULL;
1082                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1083                                 wake_up(&conf->wait_for_overlap);
1084                         spin_unlock_irq(&conf->device_lock);
1085                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1086                                 copy_data(0, rbi, dev->page, dev->sector);
1087                                 rbi2 = r5_next_bio(rbi, dev->sector);
1088                                 spin_lock_irq(&conf->device_lock);
1089                                 if (--rbi->bi_phys_segments == 0) {
1090                                         rbi->bi_next = return_bi;
1091                                         return_bi = rbi;
1092                                 }
1093                                 spin_unlock_irq(&conf->device_lock);
1094                                 rbi = rbi2;
1095                         }
1096                 }
1097
1098                 /* now count some things */
1099                 if (test_bit(R5_LOCKED, &dev->flags)) locked++;
1100                 if (test_bit(R5_UPTODATE, &dev->flags)) uptodate++;
1101
1102
1103                 if (dev->toread) to_read++;
1104                 if (dev->towrite) {
1105                         to_write++;
1106                         if (!test_bit(R5_OVERWRITE, &dev->flags))
1107                                 non_overwrite++;
1108                 }
1109                 if (dev->written) written++;
1110                 rdev = rcu_dereference(conf->disks[i].rdev);
1111                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
1112                         /* The ReadError flag will just be confusing now */
1113                         clear_bit(R5_ReadError, &dev->flags);
1114                         clear_bit(R5_ReWrite, &dev->flags);
1115                 }
1116                 if (!rdev || !test_bit(In_sync, &rdev->flags)
1117                     || test_bit(R5_ReadError, &dev->flags)) {
1118                         if ( failed < 2 )
1119                                 failed_num[failed] = i;
1120                         failed++;
1121                 } else
1122                         set_bit(R5_Insync, &dev->flags);
1123         }
1124         rcu_read_unlock();
1125         PRINTK("locked=%d uptodate=%d to_read=%d"
1126                " to_write=%d failed=%d failed_num=%d,%d\n",
1127                locked, uptodate, to_read, to_write, failed,
1128                failed_num[0], failed_num[1]);
1129         /* check if the array has lost >2 devices and, if so, some requests might
1130          * need to be failed
1131          */
1132         if (failed > 2 && to_read+to_write+written) {
1133                 for (i=disks; i--; ) {
1134                         int bitmap_end = 0;
1135
1136                         if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1137                                 mdk_rdev_t *rdev;
1138                                 rcu_read_lock();
1139                                 rdev = rcu_dereference(conf->disks[i].rdev);
1140                                 if (rdev && test_bit(In_sync, &rdev->flags))
1141                                         /* multiple read failures in one stripe */
1142                                         md_error(conf->mddev, rdev);
1143                                 rcu_read_unlock();
1144                         }
1145
1146                         spin_lock_irq(&conf->device_lock);
1147                         /* fail all writes first */
1148                         bi = sh->dev[i].towrite;
1149                         sh->dev[i].towrite = NULL;
1150                         if (bi) { to_write--; bitmap_end = 1; }
1151
1152                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1153                                 wake_up(&conf->wait_for_overlap);
1154
1155                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1156                                 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1157                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1158                                 if (--bi->bi_phys_segments == 0) {
1159                                         md_write_end(conf->mddev);
1160                                         bi->bi_next = return_bi;
1161                                         return_bi = bi;
1162                                 }
1163                                 bi = nextbi;
1164                         }
1165                         /* and fail all 'written' */
1166                         bi = sh->dev[i].written;
1167                         sh->dev[i].written = NULL;
1168                         if (bi) bitmap_end = 1;
1169                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS) {
1170                                 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1171                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1172                                 if (--bi->bi_phys_segments == 0) {
1173                                         md_write_end(conf->mddev);
1174                                         bi->bi_next = return_bi;
1175                                         return_bi = bi;
1176                                 }
1177                                 bi = bi2;
1178                         }
1179
1180                         /* fail any reads if this device is non-operational */
1181                         if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1182                             test_bit(R5_ReadError, &sh->dev[i].flags)) {
1183                                 bi = sh->dev[i].toread;
1184                                 sh->dev[i].toread = NULL;
1185                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1186                                         wake_up(&conf->wait_for_overlap);
1187                                 if (bi) to_read--;
1188                                 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1189                                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1190                                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1191                                         if (--bi->bi_phys_segments == 0) {
1192                                                 bi->bi_next = return_bi;
1193                                                 return_bi = bi;
1194                                         }
1195                                         bi = nextbi;
1196                                 }
1197                         }
1198                         spin_unlock_irq(&conf->device_lock);
1199                         if (bitmap_end)
1200                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1201                                                 STRIPE_SECTORS, 0, 0);
1202                 }
1203         }
1204         if (failed > 2 && syncing) {
1205                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
1206                 clear_bit(STRIPE_SYNCING, &sh->state);
1207                 syncing = 0;
1208         }
1209
1210         /*
1211          * might be able to return some write requests if the parity blocks
1212          * are safe, or on a failed drive
1213          */
1214         pdev = &sh->dev[pd_idx];
1215         p_failed = (failed >= 1 && failed_num[0] == pd_idx)
1216                 || (failed >= 2 && failed_num[1] == pd_idx);
1217         qdev = &sh->dev[qd_idx];
1218         q_failed = (failed >= 1 && failed_num[0] == qd_idx)
1219                 || (failed >= 2 && failed_num[1] == qd_idx);
1220
1221         if ( written &&
1222              ( p_failed || ((test_bit(R5_Insync, &pdev->flags)
1223                              && !test_bit(R5_LOCKED, &pdev->flags)
1224                              && test_bit(R5_UPTODATE, &pdev->flags))) ) &&
1225              ( q_failed || ((test_bit(R5_Insync, &qdev->flags)
1226                              && !test_bit(R5_LOCKED, &qdev->flags)
1227                              && test_bit(R5_UPTODATE, &qdev->flags))) ) ) {
1228                 /* any written block on an uptodate or failed drive can be
1229                  * returned.  Note that if we 'wrote' to a failed drive,
1230                  * it will be UPTODATE, but never LOCKED, so we don't need
1231                  * to test 'failed' directly.
1232                  */
1233                 for (i=disks; i--; )
1234                         if (sh->dev[i].written) {
1235                                 dev = &sh->dev[i];
1236                                 if (!test_bit(R5_LOCKED, &dev->flags) &&
1237                                     test_bit(R5_UPTODATE, &dev->flags) ) {
1238                                         /* We can return any write requests */
1239                                         int bitmap_end = 0;
1240                                         struct bio *wbi, *wbi2;
1241                                         PRINTK("Return write for stripe %llu disc %d\n",
1242                                                (unsigned long long)sh->sector, i);
1243                                         spin_lock_irq(&conf->device_lock);
1244                                         wbi = dev->written;
1245                                         dev->written = NULL;
1246                                         while (wbi && wbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1247                                                 wbi2 = r5_next_bio(wbi, dev->sector);
1248                                                 if (--wbi->bi_phys_segments == 0) {
1249                                                         md_write_end(conf->mddev);
1250                                                         wbi->bi_next = return_bi;
1251                                                         return_bi = wbi;
1252                                                 }
1253                                                 wbi = wbi2;
1254                                         }
1255                                         if (dev->towrite == NULL)
1256                                                 bitmap_end = 1;
1257                                         spin_unlock_irq(&conf->device_lock);
1258                                         if (bitmap_end)
1259                                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1260                                                                 STRIPE_SECTORS,
1261                                                                 !test_bit(STRIPE_DEGRADED, &sh->state), 0);
1262                                 }
1263                         }
1264         }
1265
1266         /* Now we might consider reading some blocks, either to check/generate
1267          * parity, or to satisfy requests
1268          * or to load a block that is being partially written.
1269          */
1270         if (to_read || non_overwrite || (to_write && failed) || (syncing && (uptodate < disks))) {
1271                 for (i=disks; i--;) {
1272                         dev = &sh->dev[i];
1273                         if (!test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1274                             (dev->toread ||
1275                              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1276                              syncing ||
1277                              (failed >= 1 && (sh->dev[failed_num[0]].toread || to_write)) ||
1278                              (failed >= 2 && (sh->dev[failed_num[1]].toread || to_write))
1279                                     )
1280                                 ) {
1281                                 /* we would like to get this block, possibly
1282                                  * by computing it, but we might not be able to
1283                                  */
1284                                 if (uptodate == disks-1) {
1285                                         PRINTK("Computing stripe %llu block %d\n",
1286                                                (unsigned long long)sh->sector, i);
1287                                         compute_block_1(sh, i, 0);
1288                                         uptodate++;
1289                                 } else if ( uptodate == disks-2 && failed >= 2 ) {
1290                                         /* Computing 2-failure is *very* expensive; only do it if failed >= 2 */
1291                                         int other;
1292                                         for (other=disks; other--;) {
1293                                                 if ( other == i )
1294                                                         continue;
1295                                                 if ( !test_bit(R5_UPTODATE, &sh->dev[other].flags) )
1296                                                         break;
1297                                         }
1298                                         BUG_ON(other < 0);
1299                                         PRINTK("Computing stripe %llu blocks %d,%d\n",
1300                                                (unsigned long long)sh->sector, i, other);
1301                                         compute_block_2(sh, i, other);
1302                                         uptodate += 2;
1303                                 } else if (test_bit(R5_Insync, &dev->flags)) {
1304                                         set_bit(R5_LOCKED, &dev->flags);
1305                                         set_bit(R5_Wantread, &dev->flags);
1306 #if 0
1307                                         /* if I am just reading this block and we don't have
1308                                            a failed drive, or any pending writes then sidestep the cache */
1309                                         if (sh->bh_read[i] && !sh->bh_read[i]->b_reqnext &&
1310                                             ! syncing && !failed && !to_write) {
1311                                                 sh->bh_cache[i]->b_page =  sh->bh_read[i]->b_page;
1312                                                 sh->bh_cache[i]->b_data =  sh->bh_read[i]->b_data;
1313                                         }
1314 #endif
1315                                         locked++;
1316                                         PRINTK("Reading block %d (sync=%d)\n",
1317                                                 i, syncing);
1318                                 }
1319                         }
1320                 }
1321                 set_bit(STRIPE_HANDLE, &sh->state);
1322         }
1323
1324         /* now to consider writing and what else, if anything should be read */
1325         if (to_write) {
1326                 int rcw=0, must_compute=0;
1327                 for (i=disks ; i--;) {
1328                         dev = &sh->dev[i];
1329                         /* Would I have to read this buffer for reconstruct_write */
1330                         if (!test_bit(R5_OVERWRITE, &dev->flags)
1331                             && i != pd_idx && i != qd_idx
1332                             && (!test_bit(R5_LOCKED, &dev->flags)
1333 #if 0
1334                                 || sh->bh_page[i] != bh->b_page
1335 #endif
1336                                     ) &&
1337                             !test_bit(R5_UPTODATE, &dev->flags)) {
1338                                 if (test_bit(R5_Insync, &dev->flags)) rcw++;
1339                                 else {
1340                                         PRINTK("raid6: must_compute: disk %d flags=%#lx\n", i, dev->flags);
1341                                         must_compute++;
1342                                 }
1343                         }
1344                 }
1345                 PRINTK("for sector %llu, rcw=%d, must_compute=%d\n",
1346                        (unsigned long long)sh->sector, rcw, must_compute);
1347                 set_bit(STRIPE_HANDLE, &sh->state);
1348
1349                 if (rcw > 0)
1350                         /* want reconstruct write, but need to get some data */
1351                         for (i=disks; i--;) {
1352                                 dev = &sh->dev[i];
1353                                 if (!test_bit(R5_OVERWRITE, &dev->flags)
1354                                     && !(failed == 0 && (i == pd_idx || i == qd_idx))
1355                                     && !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1356                                     test_bit(R5_Insync, &dev->flags)) {
1357                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1358                                         {
1359                                                 PRINTK("Read_old stripe %llu block %d for Reconstruct\n",
1360                                                        (unsigned long long)sh->sector, i);
1361                                                 set_bit(R5_LOCKED, &dev->flags);
1362                                                 set_bit(R5_Wantread, &dev->flags);
1363                                                 locked++;
1364                                         } else {
1365                                                 PRINTK("Request delayed stripe %llu block %d for Reconstruct\n",
1366                                                        (unsigned long long)sh->sector, i);
1367                                                 set_bit(STRIPE_DELAYED, &sh->state);
1368                                                 set_bit(STRIPE_HANDLE, &sh->state);
1369                                         }
1370                                 }
1371                         }
1372                 /* now if nothing is locked, and if we have enough data, we can start a write request */
1373                 if (locked == 0 && rcw == 0 &&
1374                     !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
1375                         if ( must_compute > 0 ) {
1376                                 /* We have failed blocks and need to compute them */
1377                                 switch ( failed ) {
1378                                 case 0: BUG();
1379                                 case 1: compute_block_1(sh, failed_num[0], 0); break;
1380                                 case 2: compute_block_2(sh, failed_num[0], failed_num[1]); break;
1381                                 default: BUG(); /* This request should have been failed? */
1382                                 }
1383                         }
1384
1385                         PRINTK("Computing parity for stripe %llu\n", (unsigned long long)sh->sector);
1386                         compute_parity(sh, RECONSTRUCT_WRITE);
1387                         /* now every locked buffer is ready to be written */
1388                         for (i=disks; i--;)
1389                                 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
1390                                         PRINTK("Writing stripe %llu block %d\n",
1391                                                (unsigned long long)sh->sector, i);
1392                                         locked++;
1393                                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
1394                                 }
1395                         /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
1396                         set_bit(STRIPE_INSYNC, &sh->state);
1397
1398                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
1399                                 atomic_dec(&conf->preread_active_stripes);
1400                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
1401                                         md_wakeup_thread(conf->mddev->thread);
1402                         }
1403                 }
1404         }
1405
1406         /* maybe we need to check and possibly fix the parity for this stripe
1407          * Any reads will already have been scheduled, so we just see if enough data
1408          * is available
1409          */
1410         if (syncing && locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state)) {
1411                 int update_p = 0, update_q = 0;
1412                 struct r5dev *dev;
1413
1414                 set_bit(STRIPE_HANDLE, &sh->state);
1415
1416                 BUG_ON(failed>2);
1417                 BUG_ON(uptodate < disks);
1418                 /* Want to check and possibly repair P and Q.
1419                  * However there could be one 'failed' device, in which
1420                  * case we can only check one of them, possibly using the
1421                  * other to generate missing data
1422                  */
1423
1424                 /* If !tmp_page, we cannot do the calculations,
1425                  * but as we have set STRIPE_HANDLE, we will soon be called
1426                  * by stripe_handle with a tmp_page - just wait until then.
1427                  */
1428                 if (tmp_page) {
1429                         if (failed == q_failed) {
1430                                 /* The only possible failed device holds 'Q', so it makes
1431                                  * sense to check P (If anything else were failed, we would
1432                                  * have used P to recreate it).
1433                                  */
1434                                 compute_block_1(sh, pd_idx, 1);
1435                                 if (!page_is_zero(sh->dev[pd_idx].page)) {
1436                                         compute_block_1(sh,pd_idx,0);
1437                                         update_p = 1;
1438                                 }
1439                         }
1440                         if (!q_failed && failed < 2) {
1441                                 /* q is not failed, and we didn't use it to generate
1442                                  * anything, so it makes sense to check it
1443                                  */
1444                                 memcpy(page_address(tmp_page),
1445                                        page_address(sh->dev[qd_idx].page),
1446                                        STRIPE_SIZE);
1447                                 compute_parity(sh, UPDATE_PARITY);
1448                                 if (memcmp(page_address(tmp_page),
1449                                            page_address(sh->dev[qd_idx].page),
1450                                            STRIPE_SIZE)!= 0) {
1451                                         clear_bit(STRIPE_INSYNC, &sh->state);
1452                                         update_q = 1;
1453                                 }
1454                         }
1455                         if (update_p || update_q) {
1456                                 conf->mddev->resync_mismatches += STRIPE_SECTORS;
1457                                 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
1458                                         /* don't try to repair!! */
1459                                         update_p = update_q = 0;
1460                         }
1461
1462                         /* now write out any block on a failed drive,
1463                          * or P or Q if they need it
1464                          */
1465
1466                         if (failed == 2) {
1467                                 dev = &sh->dev[failed_num[1]];
1468                                 locked++;
1469                                 set_bit(R5_LOCKED, &dev->flags);
1470                                 set_bit(R5_Wantwrite, &dev->flags);
1471                         }
1472                         if (failed >= 1) {
1473                                 dev = &sh->dev[failed_num[0]];
1474                                 locked++;
1475                                 set_bit(R5_LOCKED, &dev->flags);
1476                                 set_bit(R5_Wantwrite, &dev->flags);
1477                         }
1478
1479                         if (update_p) {
1480                                 dev = &sh->dev[pd_idx];
1481                                 locked ++;
1482                                 set_bit(R5_LOCKED, &dev->flags);
1483                                 set_bit(R5_Wantwrite, &dev->flags);
1484                         }
1485                         if (update_q) {
1486                                 dev = &sh->dev[qd_idx];
1487                                 locked++;
1488                                 set_bit(R5_LOCKED, &dev->flags);
1489                                 set_bit(R5_Wantwrite, &dev->flags);
1490                         }
1491                         clear_bit(STRIPE_DEGRADED, &sh->state);
1492
1493                         set_bit(STRIPE_INSYNC, &sh->state);
1494                 }
1495         }
1496
1497         if (syncing && locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
1498                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
1499                 clear_bit(STRIPE_SYNCING, &sh->state);
1500         }
1501
1502         /* If the failed drives are just a ReadError, then we might need
1503          * to progress the repair/check process
1504          */
1505         if (failed <= 2 && ! conf->mddev->ro)
1506                 for (i=0; i<failed;i++) {
1507                         dev = &sh->dev[failed_num[i]];
1508                         if (test_bit(R5_ReadError, &dev->flags)
1509                             && !test_bit(R5_LOCKED, &dev->flags)
1510                             && test_bit(R5_UPTODATE, &dev->flags)
1511                                 ) {
1512                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
1513                                         set_bit(R5_Wantwrite, &dev->flags);
1514                                         set_bit(R5_ReWrite, &dev->flags);
1515                                         set_bit(R5_LOCKED, &dev->flags);
1516                                 } else {
1517                                         /* let's read it back */
1518                                         set_bit(R5_Wantread, &dev->flags);
1519                                         set_bit(R5_LOCKED, &dev->flags);
1520                                 }
1521                         }
1522                 }
1523         spin_unlock(&sh->lock);
1524
1525         while ((bi=return_bi)) {
1526                 int bytes = bi->bi_size;
1527
1528                 return_bi = bi->bi_next;
1529                 bi->bi_next = NULL;
1530                 bi->bi_size = 0;
1531                 bi->bi_end_io(bi, bytes, 0);
1532         }
1533         for (i=disks; i-- ;) {
1534                 int rw;
1535                 struct bio *bi;
1536                 mdk_rdev_t *rdev;
1537                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
1538                         rw = 1;
1539                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1540                         rw = 0;
1541                 else
1542                         continue;
1543
1544                 bi = &sh->dev[i].req;
1545
1546                 bi->bi_rw = rw;
1547                 if (rw)
1548                         bi->bi_end_io = raid6_end_write_request;
1549                 else
1550                         bi->bi_end_io = raid6_end_read_request;
1551
1552                 rcu_read_lock();
1553                 rdev = rcu_dereference(conf->disks[i].rdev);
1554                 if (rdev && test_bit(Faulty, &rdev->flags))
1555                         rdev = NULL;
1556                 if (rdev)
1557                         atomic_inc(&rdev->nr_pending);
1558                 rcu_read_unlock();
1559
1560                 if (rdev) {
1561                         if (syncing)
1562                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1563
1564                         bi->bi_bdev = rdev->bdev;
1565                         PRINTK("for %llu schedule op %ld on disc %d\n",
1566                                 (unsigned long long)sh->sector, bi->bi_rw, i);
1567                         atomic_inc(&sh->count);
1568                         bi->bi_sector = sh->sector + rdev->data_offset;
1569                         bi->bi_flags = 1 << BIO_UPTODATE;
1570                         bi->bi_vcnt = 1;
1571                         bi->bi_max_vecs = 1;
1572                         bi->bi_idx = 0;
1573                         bi->bi_io_vec = &sh->dev[i].vec;
1574                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1575                         bi->bi_io_vec[0].bv_offset = 0;
1576                         bi->bi_size = STRIPE_SIZE;
1577                         bi->bi_next = NULL;
1578                         if (rw == WRITE &&
1579                             test_bit(R5_ReWrite, &sh->dev[i].flags))
1580                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1581                         generic_make_request(bi);
1582                 } else {
1583                         if (rw == 1)
1584                                 set_bit(STRIPE_DEGRADED, &sh->state);
1585                         PRINTK("skip op %ld on disc %d for sector %llu\n",
1586                                 bi->bi_rw, i, (unsigned long long)sh->sector);
1587                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1588                         set_bit(STRIPE_HANDLE, &sh->state);
1589                 }
1590         }
1591 }
1592
1593 static void raid6_activate_delayed(raid6_conf_t *conf)
1594 {
1595         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
1596                 while (!list_empty(&conf->delayed_list)) {
1597                         struct list_head *l = conf->delayed_list.next;
1598                         struct stripe_head *sh;
1599                         sh = list_entry(l, struct stripe_head, lru);
1600                         list_del_init(l);
1601                         clear_bit(STRIPE_DELAYED, &sh->state);
1602                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1603                                 atomic_inc(&conf->preread_active_stripes);
1604                         list_add_tail(&sh->lru, &conf->handle_list);
1605                 }
1606         }
1607 }
1608
1609 static void activate_bit_delay(raid6_conf_t *conf)
1610 {
1611         /* device_lock is held */
1612         struct list_head head;
1613         list_add(&head, &conf->bitmap_list);
1614         list_del_init(&conf->bitmap_list);
1615         while (!list_empty(&head)) {
1616                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
1617                 list_del_init(&sh->lru);
1618                 atomic_inc(&sh->count);
1619                 __release_stripe(conf, sh);
1620         }
1621 }
1622
1623 static void unplug_slaves(mddev_t *mddev)
1624 {
1625         raid6_conf_t *conf = mddev_to_conf(mddev);
1626         int i;
1627
1628         rcu_read_lock();
1629         for (i=0; i<mddev->raid_disks; i++) {
1630                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
1631                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
1632                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
1633
1634                         atomic_inc(&rdev->nr_pending);
1635                         rcu_read_unlock();
1636
1637                         if (r_queue->unplug_fn)
1638                                 r_queue->unplug_fn(r_queue);
1639
1640                         rdev_dec_pending(rdev, mddev);
1641                         rcu_read_lock();
1642                 }
1643         }
1644         rcu_read_unlock();
1645 }
1646
1647 static void raid6_unplug_device(request_queue_t *q)
1648 {
1649         mddev_t *mddev = q->queuedata;
1650         raid6_conf_t *conf = mddev_to_conf(mddev);
1651         unsigned long flags;
1652
1653         spin_lock_irqsave(&conf->device_lock, flags);
1654
1655         if (blk_remove_plug(q)) {
1656                 conf->seq_flush++;
1657                 raid6_activate_delayed(conf);
1658         }
1659         md_wakeup_thread(mddev->thread);
1660
1661         spin_unlock_irqrestore(&conf->device_lock, flags);
1662
1663         unplug_slaves(mddev);
1664 }
1665
1666 static int raid6_issue_flush(request_queue_t *q, struct gendisk *disk,
1667                              sector_t *error_sector)
1668 {
1669         mddev_t *mddev = q->queuedata;
1670         raid6_conf_t *conf = mddev_to_conf(mddev);
1671         int i, ret = 0;
1672
1673         rcu_read_lock();
1674         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
1675                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
1676                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
1677                         struct block_device *bdev = rdev->bdev;
1678                         request_queue_t *r_queue = bdev_get_queue(bdev);
1679
1680                         if (!r_queue->issue_flush_fn)
1681                                 ret = -EOPNOTSUPP;
1682                         else {
1683                                 atomic_inc(&rdev->nr_pending);
1684                                 rcu_read_unlock();
1685                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
1686                                                               error_sector);
1687                                 rdev_dec_pending(rdev, mddev);
1688                                 rcu_read_lock();
1689                         }
1690                 }
1691         }
1692         rcu_read_unlock();
1693         return ret;
1694 }
1695
1696 static inline void raid6_plug_device(raid6_conf_t *conf)
1697 {
1698         spin_lock_irq(&conf->device_lock);
1699         blk_plug_device(conf->mddev->queue);
1700         spin_unlock_irq(&conf->device_lock);
1701 }
1702
1703 static int make_request (request_queue_t *q, struct bio * bi)
1704 {
1705         mddev_t *mddev = q->queuedata;
1706         raid6_conf_t *conf = mddev_to_conf(mddev);
1707         const unsigned int raid_disks = conf->raid_disks;
1708         const unsigned int data_disks = raid_disks - 2;
1709         unsigned int dd_idx, pd_idx;
1710         sector_t new_sector;
1711         sector_t logical_sector, last_sector;
1712         struct stripe_head *sh;
1713         const int rw = bio_data_dir(bi);
1714
1715         if (unlikely(bio_barrier(bi))) {
1716                 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
1717                 return 0;
1718         }
1719
1720         md_write_start(mddev, bi);
1721
1722         disk_stat_inc(mddev->gendisk, ios[rw]);
1723         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
1724
1725         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
1726         last_sector = bi->bi_sector + (bi->bi_size>>9);
1727
1728         bi->bi_next = NULL;
1729         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
1730
1731         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
1732                 DEFINE_WAIT(w);
1733
1734                 new_sector = raid6_compute_sector(logical_sector,
1735                                                   raid_disks, data_disks, &dd_idx, &pd_idx, conf);
1736
1737                 PRINTK("raid6: make_request, sector %llu logical %llu\n",
1738                        (unsigned long long)new_sector,
1739                        (unsigned long long)logical_sector);
1740
1741         retry:
1742                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
1743                 sh = get_active_stripe(conf, new_sector, pd_idx, (bi->bi_rw&RWA_MASK));
1744                 if (sh) {
1745                         if (!add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
1746                                 /* Add failed due to overlap.  Flush everything
1747                                  * and wait a while
1748                                  */
1749                                 raid6_unplug_device(mddev->queue);
1750                                 release_stripe(sh);
1751                                 schedule();
1752                                 goto retry;
1753                         }
1754                         finish_wait(&conf->wait_for_overlap, &w);
1755                         raid6_plug_device(conf);
1756                         handle_stripe(sh, NULL);
1757                         release_stripe(sh);
1758                 } else {
1759                         /* cannot get stripe for read-ahead, just give-up */
1760                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1761                         finish_wait(&conf->wait_for_overlap, &w);
1762                         break;
1763                 }
1764
1765         }
1766         spin_lock_irq(&conf->device_lock);
1767         if (--bi->bi_phys_segments == 0) {
1768                 int bytes = bi->bi_size;
1769
1770                 if (rw == WRITE )
1771                         md_write_end(mddev);
1772                 bi->bi_size = 0;
1773                 bi->bi_end_io(bi, bytes, 0);
1774         }
1775         spin_unlock_irq(&conf->device_lock);
1776         return 0;
1777 }
1778
1779 /* FIXME go_faster isn't used */
1780 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1781 {
1782         raid6_conf_t *conf = (raid6_conf_t *) mddev->private;
1783         struct stripe_head *sh;
1784         int sectors_per_chunk = conf->chunk_size >> 9;
1785         sector_t x;
1786         unsigned long stripe;
1787         int chunk_offset;
1788         int dd_idx, pd_idx;
1789         sector_t first_sector;
1790         int raid_disks = conf->raid_disks;
1791         int data_disks = raid_disks - 2;
1792         sector_t max_sector = mddev->size << 1;
1793         int sync_blocks;
1794         int still_degraded = 0;
1795         int i;
1796
1797         if (sector_nr >= max_sector) {
1798                 /* just being told to finish up .. nothing much to do */
1799                 unplug_slaves(mddev);
1800
1801                 if (mddev->curr_resync < max_sector) /* aborted */
1802                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1803                                         &sync_blocks, 1);
1804                 else /* completed sync */
1805                         conf->fullsync = 0;
1806                 bitmap_close_sync(mddev->bitmap);
1807
1808                 return 0;
1809         }
1810         /* if there are 2 or more failed drives and we are trying
1811          * to resync, then assert that we are finished, because there is
1812          * nothing we can do.
1813          */
1814         if (mddev->degraded >= 2 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1815                 sector_t rv = (mddev->size << 1) - sector_nr;
1816                 *skipped = 1;
1817                 return rv;
1818         }
1819         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1820             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1821             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
1822                 /* we can skip this block, and probably more */
1823                 sync_blocks /= STRIPE_SECTORS;
1824                 *skipped = 1;
1825                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
1826         }
1827
1828         x = sector_nr;
1829         chunk_offset = sector_div(x, sectors_per_chunk);
1830         stripe = x;
1831         BUG_ON(x != stripe);
1832
1833         first_sector = raid6_compute_sector((sector_t)stripe*data_disks*sectors_per_chunk
1834                 + chunk_offset, raid_disks, data_disks, &dd_idx, &pd_idx, conf);
1835         sh = get_active_stripe(conf, sector_nr, pd_idx, 1);
1836         if (sh == NULL) {
1837                 sh = get_active_stripe(conf, sector_nr, pd_idx, 0);
1838                 /* make sure we don't swamp the stripe cache if someone else
1839                  * is trying to get access
1840                  */
1841                 schedule_timeout_uninterruptible(1);
1842         }
1843         /* Need to check if array will still be degraded after recovery/resync
1844          * We don't need to check the 'failed' flag as when that gets set,
1845          * recovery aborts.
1846          */
1847         for (i=0; i<mddev->raid_disks; i++)
1848                 if (conf->disks[i].rdev == NULL)
1849                         still_degraded = 1;
1850
1851         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
1852
1853         spin_lock(&sh->lock);
1854         set_bit(STRIPE_SYNCING, &sh->state);
1855         clear_bit(STRIPE_INSYNC, &sh->state);
1856         spin_unlock(&sh->lock);
1857
1858         handle_stripe(sh, NULL);
1859         release_stripe(sh);
1860
1861         return STRIPE_SECTORS;
1862 }
1863
1864 /*
1865  * This is our raid6 kernel thread.
1866  *
1867  * We scan the hash table for stripes which can be handled now.
1868  * During the scan, completed stripes are saved for us by the interrupt
1869  * handler, so that they will not have to wait for our next wakeup.
1870  */
1871 static void raid6d (mddev_t *mddev)
1872 {
1873         struct stripe_head *sh;
1874         raid6_conf_t *conf = mddev_to_conf(mddev);
1875         int handled;
1876
1877         PRINTK("+++ raid6d active\n");
1878
1879         md_check_recovery(mddev);
1880
1881         handled = 0;
1882         spin_lock_irq(&conf->device_lock);
1883         while (1) {
1884                 struct list_head *first;
1885
1886                 if (conf->seq_flush - conf->seq_write > 0) {
1887                         int seq = conf->seq_flush;
1888                         spin_unlock_irq(&conf->device_lock);
1889                         bitmap_unplug(mddev->bitmap);
1890                         spin_lock_irq(&conf->device_lock);
1891                         conf->seq_write = seq;
1892                         activate_bit_delay(conf);
1893                 }
1894
1895                 if (list_empty(&conf->handle_list) &&
1896                     atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
1897                     !blk_queue_plugged(mddev->queue) &&
1898                     !list_empty(&conf->delayed_list))
1899                         raid6_activate_delayed(conf);
1900
1901                 if (list_empty(&conf->handle_list))
1902                         break;
1903
1904                 first = conf->handle_list.next;
1905                 sh = list_entry(first, struct stripe_head, lru);
1906
1907                 list_del_init(first);
1908                 atomic_inc(&sh->count);
1909                 if (atomic_read(&sh->count)!= 1)
1910                         BUG();
1911                 spin_unlock_irq(&conf->device_lock);
1912
1913                 handled++;
1914                 handle_stripe(sh, conf->spare_page);
1915                 release_stripe(sh);
1916
1917                 spin_lock_irq(&conf->device_lock);
1918         }
1919         PRINTK("%d stripes handled\n", handled);
1920
1921         spin_unlock_irq(&conf->device_lock);
1922
1923         unplug_slaves(mddev);
1924
1925         PRINTK("--- raid6d inactive\n");
1926 }
1927
1928 static ssize_t
1929 raid6_show_stripe_cache_size(mddev_t *mddev, char *page)
1930 {
1931         raid6_conf_t *conf = mddev_to_conf(mddev);
1932         if (conf)
1933                 return sprintf(page, "%d\n", conf->max_nr_stripes);
1934         else
1935                 return 0;
1936 }
1937
1938 static ssize_t
1939 raid6_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
1940 {
1941         raid6_conf_t *conf = mddev_to_conf(mddev);
1942         char *end;
1943         int new;
1944         if (len >= PAGE_SIZE)
1945                 return -EINVAL;
1946         if (!conf)
1947                 return -ENODEV;
1948
1949         new = simple_strtoul(page, &end, 10);
1950         if (!*page || (*end && *end != '\n') )
1951                 return -EINVAL;
1952         if (new <= 16 || new > 32768)
1953                 return -EINVAL;
1954         while (new < conf->max_nr_stripes) {
1955                 if (drop_one_stripe(conf))
1956                         conf->max_nr_stripes--;
1957                 else
1958                         break;
1959         }
1960         while (new > conf->max_nr_stripes) {
1961                 if (grow_one_stripe(conf))
1962                         conf->max_nr_stripes++;
1963                 else break;
1964         }
1965         return len;
1966 }
1967
1968 static struct md_sysfs_entry
1969 raid6_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
1970                                 raid6_show_stripe_cache_size,
1971                                 raid6_store_stripe_cache_size);
1972
1973 static ssize_t
1974 stripe_cache_active_show(mddev_t *mddev, char *page)
1975 {
1976         raid6_conf_t *conf = mddev_to_conf(mddev);
1977         if (conf)
1978                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
1979         else
1980                 return 0;
1981 }
1982
1983 static struct md_sysfs_entry
1984 raid6_stripecache_active = __ATTR_RO(stripe_cache_active);
1985
1986 static struct attribute *raid6_attrs[] =  {
1987         &raid6_stripecache_size.attr,
1988         &raid6_stripecache_active.attr,
1989         NULL,
1990 };
1991 static struct attribute_group raid6_attrs_group = {
1992         .name = NULL,
1993         .attrs = raid6_attrs,
1994 };
1995
1996 static int run(mddev_t *mddev)
1997 {
1998         raid6_conf_t *conf;
1999         int raid_disk, memory;
2000         mdk_rdev_t *rdev;
2001         struct disk_info *disk;
2002         struct list_head *tmp;
2003
2004         if (mddev->level != 6) {
2005                 PRINTK("raid6: %s: raid level not set to 6 (%d)\n", mdname(mddev), mddev->level);
2006                 return -EIO;
2007         }
2008
2009         mddev->private = kzalloc(sizeof (raid6_conf_t)
2010                                  + mddev->raid_disks * sizeof(struct disk_info),
2011                                  GFP_KERNEL);
2012         if ((conf = mddev->private) == NULL)
2013                 goto abort;
2014         conf->mddev = mddev;
2015
2016         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
2017                 goto abort;
2018
2019         conf->spare_page = alloc_page(GFP_KERNEL);
2020         if (!conf->spare_page)
2021                 goto abort;
2022
2023         spin_lock_init(&conf->device_lock);
2024         init_waitqueue_head(&conf->wait_for_stripe);
2025         init_waitqueue_head(&conf->wait_for_overlap);
2026         INIT_LIST_HEAD(&conf->handle_list);
2027         INIT_LIST_HEAD(&conf->delayed_list);
2028         INIT_LIST_HEAD(&conf->bitmap_list);
2029         INIT_LIST_HEAD(&conf->inactive_list);
2030         atomic_set(&conf->active_stripes, 0);
2031         atomic_set(&conf->preread_active_stripes, 0);
2032
2033         PRINTK("raid6: run(%s) called.\n", mdname(mddev));
2034
2035         ITERATE_RDEV(mddev,rdev,tmp) {
2036                 raid_disk = rdev->raid_disk;
2037                 if (raid_disk >= mddev->raid_disks
2038                     || raid_disk < 0)
2039                         continue;
2040                 disk = conf->disks + raid_disk;
2041
2042                 disk->rdev = rdev;
2043
2044                 if (test_bit(In_sync, &rdev->flags)) {
2045                         char b[BDEVNAME_SIZE];
2046                         printk(KERN_INFO "raid6: device %s operational as raid"
2047                                " disk %d\n", bdevname(rdev->bdev,b),
2048                                raid_disk);
2049                         conf->working_disks++;
2050                 }
2051         }
2052
2053         conf->raid_disks = mddev->raid_disks;
2054
2055         /*
2056          * 0 for a fully functional array, 1 or 2 for a degraded array.
2057          */
2058         mddev->degraded = conf->failed_disks = conf->raid_disks - conf->working_disks;
2059         conf->mddev = mddev;
2060         conf->chunk_size = mddev->chunk_size;
2061         conf->level = mddev->level;
2062         conf->algorithm = mddev->layout;
2063         conf->max_nr_stripes = NR_STRIPES;
2064
2065         /* device size must be a multiple of chunk size */
2066         mddev->size &= ~(mddev->chunk_size/1024 -1);
2067         mddev->resync_max_sectors = mddev->size << 1;
2068
2069         if (conf->raid_disks < 4) {
2070                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
2071                        mdname(mddev), conf->raid_disks);
2072                 goto abort;
2073         }
2074         if (!conf->chunk_size || conf->chunk_size % 4) {
2075                 printk(KERN_ERR "raid6: invalid chunk size %d for %s\n",
2076                        conf->chunk_size, mdname(mddev));
2077                 goto abort;
2078         }
2079         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
2080                 printk(KERN_ERR
2081                        "raid6: unsupported parity algorithm %d for %s\n",
2082                        conf->algorithm, mdname(mddev));
2083                 goto abort;
2084         }
2085         if (mddev->degraded > 2) {
2086                 printk(KERN_ERR "raid6: not enough operational devices for %s"
2087                        " (%d/%d failed)\n",
2088                        mdname(mddev), conf->failed_disks, conf->raid_disks);
2089                 goto abort;
2090         }
2091
2092         if (mddev->degraded > 0 &&
2093             mddev->recovery_cp != MaxSector) {
2094                 if (mddev->ok_start_degraded)
2095                         printk(KERN_WARNING "raid6: starting dirty degraded array:%s"
2096                                "- data corruption possible.\n",
2097                                mdname(mddev));
2098                 else {
2099                         printk(KERN_ERR "raid6: cannot start dirty degraded array"
2100                                " for %s\n", mdname(mddev));
2101                         goto abort;
2102                 }
2103         }
2104
2105         {
2106                 mddev->thread = md_register_thread(raid6d, mddev, "%s_raid6");
2107                 if (!mddev->thread) {
2108                         printk(KERN_ERR
2109                                "raid6: couldn't allocate thread for %s\n",
2110                                mdname(mddev));
2111                         goto abort;
2112                 }
2113         }
2114
2115         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
2116                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
2117         if (grow_stripes(conf, conf->max_nr_stripes)) {
2118                 printk(KERN_ERR
2119                        "raid6: couldn't allocate %dkB for buffers\n", memory);
2120                 shrink_stripes(conf);
2121                 md_unregister_thread(mddev->thread);
2122                 goto abort;
2123         } else
2124                 printk(KERN_INFO "raid6: allocated %dkB for %s\n",
2125                        memory, mdname(mddev));
2126
2127         if (mddev->degraded == 0)
2128                 printk(KERN_INFO "raid6: raid level %d set %s active with %d out of %d"
2129                        " devices, algorithm %d\n", conf->level, mdname(mddev),
2130                        mddev->raid_disks-mddev->degraded, mddev->raid_disks,
2131                        conf->algorithm);
2132         else
2133                 printk(KERN_ALERT "raid6: raid level %d set %s active with %d"
2134                        " out of %d devices, algorithm %d\n", conf->level,
2135                        mdname(mddev), mddev->raid_disks - mddev->degraded,
2136                        mddev->raid_disks, conf->algorithm);
2137
2138         print_raid6_conf(conf);
2139
2140         /* read-ahead size must cover two whole stripes, which is
2141          * 2 * (n-2) * chunksize where 'n' is the number of raid devices
2142          */
2143         {
2144                 int stripe = (mddev->raid_disks-2) * mddev->chunk_size
2145                         / PAGE_SIZE;
2146                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
2147                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
2148         }
2149
2150         /* Ok, everything is just fine now */
2151         mddev->array_size =  mddev->size * (mddev->raid_disks - 2);
2152
2153         mddev->queue->unplug_fn = raid6_unplug_device;
2154         mddev->queue->issue_flush_fn = raid6_issue_flush;
2155         return 0;
2156 abort:
2157         if (conf) {
2158                 print_raid6_conf(conf);
2159                 safe_put_page(conf->spare_page);
2160                 kfree(conf->stripe_hashtbl);
2161                 kfree(conf);
2162         }
2163         mddev->private = NULL;
2164         printk(KERN_ALERT "raid6: failed to run raid set %s\n", mdname(mddev));
2165         return -EIO;
2166 }
2167
2168
2169
2170 static int stop (mddev_t *mddev)
2171 {
2172         raid6_conf_t *conf = (raid6_conf_t *) mddev->private;
2173
2174         md_unregister_thread(mddev->thread);
2175         mddev->thread = NULL;
2176         shrink_stripes(conf);
2177         kfree(conf->stripe_hashtbl);
2178         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2179         sysfs_remove_group(&mddev->kobj, &raid6_attrs_group);
2180         kfree(conf);
2181         mddev->private = NULL;
2182         return 0;
2183 }
2184
2185 #if RAID6_DUMPSTATE
2186 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
2187 {
2188         int i;
2189
2190         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
2191                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
2192         seq_printf(seq, "sh %llu,  count %d.\n",
2193                    (unsigned long long)sh->sector, atomic_read(&sh->count));
2194         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
2195         for (i = 0; i < sh->raid_conf->raid_disks; i++) {
2196                 seq_printf(seq, "(cache%d: %p %ld) ",
2197                            i, sh->dev[i].page, sh->dev[i].flags);
2198         }
2199         seq_printf(seq, "\n");
2200 }
2201
2202 static void printall (struct seq_file *seq, raid6_conf_t *conf)
2203 {
2204         struct stripe_head *sh;
2205         struct hlist_node *hn;
2206         int i;
2207
2208         spin_lock_irq(&conf->device_lock);
2209         for (i = 0; i < NR_HASH; i++) {
2210                 sh = conf->stripe_hashtbl[i];
2211                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
2212                         if (sh->raid_conf != conf)
2213                                 continue;
2214                         print_sh(seq, sh);
2215                 }
2216         }
2217         spin_unlock_irq(&conf->device_lock);
2218 }
2219 #endif
2220
2221 static void status (struct seq_file *seq, mddev_t *mddev)
2222 {
2223         raid6_conf_t *conf = (raid6_conf_t *) mddev->private;
2224         int i;
2225
2226         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
2227         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->working_disks);
2228         for (i = 0; i < conf->raid_disks; i++)
2229                 seq_printf (seq, "%s",
2230                             conf->disks[i].rdev &&
2231                             test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
2232         seq_printf (seq, "]");
2233 #if RAID6_DUMPSTATE
2234         seq_printf (seq, "\n");
2235         printall(seq, conf);
2236 #endif
2237 }
2238
2239 static void print_raid6_conf (raid6_conf_t *conf)
2240 {
2241         int i;
2242         struct disk_info *tmp;
2243
2244         printk("RAID6 conf printout:\n");
2245         if (!conf) {
2246                 printk("(conf==NULL)\n");
2247                 return;
2248         }
2249         printk(" --- rd:%d wd:%d fd:%d\n", conf->raid_disks,
2250                  conf->working_disks, conf->failed_disks);
2251
2252         for (i = 0; i < conf->raid_disks; i++) {
2253                 char b[BDEVNAME_SIZE];
2254                 tmp = conf->disks + i;
2255                 if (tmp->rdev)
2256                 printk(" disk %d, o:%d, dev:%s\n",
2257                         i, !test_bit(Faulty, &tmp->rdev->flags),
2258                         bdevname(tmp->rdev->bdev,b));
2259         }
2260 }
2261
2262 static int raid6_spare_active(mddev_t *mddev)
2263 {
2264         int i;
2265         raid6_conf_t *conf = mddev->private;
2266         struct disk_info *tmp;
2267
2268         for (i = 0; i < conf->raid_disks; i++) {
2269                 tmp = conf->disks + i;
2270                 if (tmp->rdev
2271                     && !test_bit(Faulty, &tmp->rdev->flags)
2272                     && !test_bit(In_sync, &tmp->rdev->flags)) {
2273                         mddev->degraded--;
2274                         conf->failed_disks--;
2275                         conf->working_disks++;
2276                         set_bit(In_sync, &tmp->rdev->flags);
2277                 }
2278         }
2279         print_raid6_conf(conf);
2280         return 0;
2281 }
2282
2283 static int raid6_remove_disk(mddev_t *mddev, int number)
2284 {
2285         raid6_conf_t *conf = mddev->private;
2286         int err = 0;
2287         mdk_rdev_t *rdev;
2288         struct disk_info *p = conf->disks + number;
2289
2290         print_raid6_conf(conf);
2291         rdev = p->rdev;
2292         if (rdev) {
2293                 if (test_bit(In_sync, &rdev->flags) ||
2294                     atomic_read(&rdev->nr_pending)) {
2295                         err = -EBUSY;
2296                         goto abort;
2297                 }
2298                 p->rdev = NULL;
2299                 synchronize_rcu();
2300                 if (atomic_read(&rdev->nr_pending)) {
2301                         /* lost the race, try later */
2302                         err = -EBUSY;
2303                         p->rdev = rdev;
2304                 }
2305         }
2306
2307 abort:
2308
2309         print_raid6_conf(conf);
2310         return err;
2311 }
2312
2313 static int raid6_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
2314 {
2315         raid6_conf_t *conf = mddev->private;
2316         int found = 0;
2317         int disk;
2318         struct disk_info *p;
2319
2320         if (mddev->degraded > 2)
2321                 /* no point adding a device */
2322                 return 0;
2323         /*
2324          * find the disk ... but prefer rdev->saved_raid_disk
2325          * if possible.
2326          */
2327         if (rdev->saved_raid_disk >= 0 &&
2328             conf->disks[rdev->saved_raid_disk].rdev == NULL)
2329                 disk = rdev->saved_raid_disk;
2330         else
2331                 disk = 0;
2332         for ( ; disk < mddev->raid_disks; disk++)
2333                 if ((p=conf->disks + disk)->rdev == NULL) {
2334                         clear_bit(In_sync, &rdev->flags);
2335                         rdev->raid_disk = disk;
2336                         found = 1;
2337                         if (rdev->saved_raid_disk != disk)
2338                                 conf->fullsync = 1;
2339                         rcu_assign_pointer(p->rdev, rdev);
2340                         break;
2341                 }
2342         print_raid6_conf(conf);
2343         return found;
2344 }
2345
2346 static int raid6_resize(mddev_t *mddev, sector_t sectors)
2347 {
2348         /* no resync is happening, and there is enough space
2349          * on all devices, so we can resize.
2350          * We need to make sure resync covers any new space.
2351          * If the array is shrinking we should possibly wait until
2352          * any io in the removed space completes, but it hardly seems
2353          * worth it.
2354          */
2355         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
2356         mddev->array_size = (sectors * (mddev->raid_disks-2))>>1;
2357         set_capacity(mddev->gendisk, mddev->array_size << 1);
2358         mddev->changed = 1;
2359         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
2360                 mddev->recovery_cp = mddev->size << 1;
2361                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2362         }
2363         mddev->size = sectors /2;
2364         mddev->resync_max_sectors = sectors;
2365         return 0;
2366 }
2367
2368 static void raid6_quiesce(mddev_t *mddev, int state)
2369 {
2370         raid6_conf_t *conf = mddev_to_conf(mddev);
2371
2372         switch(state) {
2373         case 1: /* stop all writes */
2374                 spin_lock_irq(&conf->device_lock);
2375                 conf->quiesce = 1;
2376                 wait_event_lock_irq(conf->wait_for_stripe,
2377                                     atomic_read(&conf->active_stripes) == 0,
2378                                     conf->device_lock, /* nothing */);
2379                 spin_unlock_irq(&conf->device_lock);
2380                 break;
2381
2382         case 0: /* re-enable writes */
2383                 spin_lock_irq(&conf->device_lock);
2384                 conf->quiesce = 0;
2385                 wake_up(&conf->wait_for_stripe);
2386                 spin_unlock_irq(&conf->device_lock);
2387                 break;
2388         }
2389 }
2390
2391 static struct mdk_personality raid6_personality =
2392 {
2393         .name           = "raid6",
2394         .level          = 6,
2395         .owner          = THIS_MODULE,
2396         .make_request   = make_request,
2397         .run            = run,
2398         .stop           = stop,
2399         .status         = status,
2400         .error_handler  = error,
2401         .hot_add_disk   = raid6_add_disk,
2402         .hot_remove_disk= raid6_remove_disk,
2403         .spare_active   = raid6_spare_active,
2404         .sync_request   = sync_request,
2405         .resize         = raid6_resize,
2406         .quiesce        = raid6_quiesce,
2407 };
2408
2409 static int __init raid6_init(void)
2410 {
2411         int e;
2412
2413         e = raid6_select_algo();
2414         if ( e )
2415                 return e;
2416
2417         return register_md_personality(&raid6_personality);
2418 }
2419
2420 static void raid6_exit (void)
2421 {
2422         unregister_md_personality(&raid6_personality);
2423 }
2424
2425 module_init(raid6_init);
2426 module_exit(raid6_exit);
2427 MODULE_LICENSE("GPL");
2428 MODULE_ALIAS("md-personality-8"); /* RAID6 */
2429 MODULE_ALIAS("md-raid6");
2430 MODULE_ALIAS("md-level-6");