drivers/md/raid1.c

   1 /*
   2  * raid1.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   5  *
   6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   7  *
   8  * RAID-1 management functions.
   9  *
  10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  11  *
  12  * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  13  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  14  *
  15  * This program is free software; you can redistribute it and/or modify
  16  * it under the terms of the GNU General Public License as published by
  17  * the Free Software Foundation; either version 2, or (at your option)
  18  * any later version.
  19  *
  20  * You should have received a copy of the GNU General Public License
  21  * (for example /usr/src/linux/COPYING); if not, write to the Free
  22  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  23  */
  24
  25 #include <linux/raid/raid1.h>
  26
  27 #define MAJOR_NR MD_MAJOR
  28 #define MD_DRIVER
  29 #define MD_PERSONALITY
  30
  31 /*
  32  * Number of guaranteed r1bios in case of extreme VM load:
  33  */
  34 #define NR_RAID1_BIOS 256
  35
  36 static mdk_personality_t raid1_personality;
  37 static spinlock_t retry_list_lock = SPIN_LOCK_UNLOCKED;
  38 static LIST_HEAD(retry_list_head);
  39
  40 static void * r1bio_pool_alloc(int gfp_flags, void *data)
  41 {
  42         mddev_t *mddev = data;
  43         r1bio_t *r1_bio;
  44
  45         /* allocate a r1bio with room for raid_disks entries in the bios array */
  46         r1_bio = kmalloc(sizeof(r1bio_t) + sizeof(struct bio*)*mddev->raid_disks,
  47                          gfp_flags);
  48         if (r1_bio)
  49                 memset(r1_bio, 0, sizeof(*r1_bio) + sizeof(struct bio*)*mddev->raid_disks);
  50
  51         return r1_bio;
  52 }
  53
  54 static void r1bio_pool_free(void *r1_bio, void *data)
  55 {
  56         kfree(r1_bio);
  57 }
  58
  59 #define RESYNC_BLOCK_SIZE (64*1024)
  60 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  61 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  62 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  63 #define RESYNC_WINDOW (2048*1024)
  64
  65 static void * r1buf_pool_alloc(int gfp_flags, void *data)
  66 {
  67         conf_t *conf = data;
  68         struct page *page;
  69         r1bio_t *r1_bio;
  70         struct bio *bio;
  71         int i, j;
  72
  73         r1_bio = r1bio_pool_alloc(gfp_flags, conf->mddev);
  74         if (!r1_bio)
  75                 return NULL;
  76
  77         /*
  78          * Allocate bios : 1 for reading, n-1 for writing
  79          */
  80         for (j = conf->raid_disks ; j-- ; ) {
  81                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
  82                 if (!bio)
  83                         goto out_free_bio;
  84                 r1_bio->bios[j] = bio;
  85         }
  86         /*
  87          * Allocate RESYNC_PAGES data pages and attach them to
  88          * the first bio;
  89          */
  90         bio = r1_bio->bios[0];
  91         for (i = 0; i < RESYNC_PAGES; i++) {
  92                 page = alloc_page(gfp_flags);
  93                 if (unlikely(!page))
  94                         goto out_free_pages;
  95
  96                 bio->bi_io_vec[i].bv_page = page;
  97         }
  98
  99         r1_bio->master_bio = bio;
 100
 101         return r1_bio;
 102
 103 out_free_pages:
 104         for ( ; i > 0 ; i--)
 105                 __free_page(bio->bi_io_vec[i-1].bv_page);
 106 out_free_bio:
 107         while ( j < conf->raid_disks )
 108                 bio_put(r1_bio->bios[++j]);
 109         r1bio_pool_free(r1_bio, conf->mddev);
 110         return NULL;
 111 }
 112
 113 static void r1buf_pool_free(void *__r1_bio, void *data)
 114 {
 115         int i;
 116         conf_t *conf = data;
 117         r1bio_t *r1bio = __r1_bio;
 118         struct bio *bio = r1bio->bios[0];
 119
 120         for (i = 0; i < RESYNC_PAGES; i++) {
 121                 __free_page(bio->bi_io_vec[i].bv_page);
 122                 bio->bi_io_vec[i].bv_page = NULL;
 123         }
 124         for (i=0 ; i < conf->raid_disks; i++)
 125                 bio_put(r1bio->bios[i]);
 126
 127         r1bio_pool_free(r1bio, conf->mddev);
 128 }
 129
 130 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
 131 {
 132         int i;
 133
 134         for (i = 0; i < conf->raid_disks; i++) {
 135                 struct bio **bio = r1_bio->bios + i;
 136                 if (*bio)
 137                         bio_put(*bio);
 138                 *bio = NULL;
 139         }
 140 }
 141
 142 static inline void free_r1bio(r1bio_t *r1_bio)
 143 {
 144         unsigned long flags;
 145
 146         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 147
 148         /*
 149          * Wake up any possible resync thread that waits for the device
 150          * to go idle.
 151          */
 152         spin_lock_irqsave(&conf->resync_lock, flags);
 153         if (!--conf->nr_pending) {
 154                 wake_up(&conf->wait_idle);
 155                 wake_up(&conf->wait_resume);
 156         }
 157         spin_unlock_irqrestore(&conf->resync_lock, flags);
 158
 159         put_all_bios(conf, r1_bio);
 160         mempool_free(r1_bio, conf->r1bio_pool);
 161 }
 162
 163 static inline void put_buf(r1bio_t *r1_bio)
 164 {
 165         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 166         unsigned long flags;
 167
 168         mempool_free(r1_bio, conf->r1buf_pool);
 169
 170         spin_lock_irqsave(&conf->resync_lock, flags);
 171         if (!conf->barrier)
 172                 BUG();
 173         --conf->barrier;
 174         wake_up(&conf->wait_resume);
 175         wake_up(&conf->wait_idle);
 176
 177         if (!--conf->nr_pending) {
 178                 wake_up(&conf->wait_idle);
 179                 wake_up(&conf->wait_resume);
 180         }
 181         spin_unlock_irqrestore(&conf->resync_lock, flags);
 182 }
 183
 184 static int map(mddev_t *mddev, mdk_rdev_t **rdevp)
 185 {
 186         conf_t *conf = mddev_to_conf(mddev);
 187         int i, disks = conf->raid_disks;
 188
 189         /*
 190          * Later we do read balancing on the read side
 191          * now we use the first available disk.
 192          */
 193
 194         spin_lock_irq(&conf->device_lock);
 195         for (i = 0; i < disks; i++) {
 196                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
 197                 if (rdev && rdev->in_sync) {
 198                         *rdevp = rdev;
 199                         atomic_inc(&rdev->nr_pending);
 200                         spin_unlock_irq(&conf->device_lock);
 201                         return 0;
 202                 }
 203         }
 204         spin_unlock_irq(&conf->device_lock);
 205
 206         printk(KERN_ERR "raid1_map(): huh, no more operational devices?\n");
 207         return -1;
 208 }
 209
 210 static void reschedule_retry(r1bio_t *r1_bio)
 211 {
 212         unsigned long flags;
 213         mddev_t *mddev = r1_bio->mddev;
 214
 215         spin_lock_irqsave(&retry_list_lock, flags);
 216         list_add(&r1_bio->retry_list, &retry_list_head);
 217         spin_unlock_irqrestore(&retry_list_lock, flags);
 218
 219         md_wakeup_thread(mddev->thread);
 220 }
 221
 222 /*
 223  * raid_end_bio_io() is called when we have finished servicing a mirrored
 224  * operation and are ready to return a success/failure code to the buffer
 225  * cache layer.
 226  */
 227 static void raid_end_bio_io(r1bio_t *r1_bio)
 228 {
 229         struct bio *bio = r1_bio->master_bio;
 230
 231         bio_endio(bio, bio->bi_size,
 232                 test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
 233         free_r1bio(r1_bio);
 234 }
 235
 236 /*
 237  * Update disk head position estimator based on IRQ completion info.
 238  */
 239 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
 240 {
 241         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 242
 243         conf->mirrors[disk].head_position =
 244                 r1_bio->sector + (r1_bio->sectors);
 245 }
 246
 247 static int raid1_end_read_request(struct bio *bio, unsigned int bytes_done, int error)
 248 {
 249         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 250         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 251         int mirror;
 252         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 253
 254         if (bio->bi_size)
 255                 return 1;
 256
 257         mirror = r1_bio->read_disk;
 258         /*
 259          * this branch is our 'one mirror IO has finished' event handler:
 260          */
 261         if (!uptodate)
 262                 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 263         else
 264                 /*
 265                  * Set R1BIO_Uptodate in our master bio, so that
 266                  * we will return a good error code for to the higher
 267                  * levels even if IO on some other mirrored buffer fails.
 268                  *
 269                  * The 'master' represents the composite IO operation to
 270                  * user-side. So if something waits for IO, then it will
 271                  * wait for the 'master' bio.
 272                  */
 273                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 274
 275         update_head_pos(mirror, r1_bio);
 276
 277         /*
 278          * we have only one bio on the read side
 279          */
 280         if (uptodate)
 281                 raid_end_bio_io(r1_bio);
 282         else {
 283                 /*
 284                  * oops, read error:
 285                  */
 286                 char b[BDEVNAME_SIZE];
 287                 printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
 288                        bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
 289                 reschedule_retry(r1_bio);
 290         }
 291
 292         atomic_dec(&conf->mirrors[mirror].rdev->nr_pending);
 293         return 0;
 294 }
 295
 296 static int raid1_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
 297 {
 298         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 299         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 300         int mirror;
 301         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 302
 303         if (bio->bi_size)
 304                 return 1;
 305
 306         for (mirror = 0; mirror < conf->raid_disks; mirror++)
 307                 if (r1_bio->bios[mirror] == bio)
 308                         break;
 309
 310         /*
 311          * this branch is our 'one mirror IO has finished' event handler:
 312          */
 313         if (!uptodate)
 314                 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 315         else
 316                 /*
 317                  * Set R1BIO_Uptodate in our master bio, so that
 318                  * we will return a good error code for to the higher
 319                  * levels even if IO on some other mirrored buffer fails.
 320                  *
 321                  * The 'master' represents the composite IO operation to
 322                  * user-side. So if something waits for IO, then it will
 323                  * wait for the 'master' bio.
 324                  */
 325                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 326
 327         update_head_pos(mirror, r1_bio);
 328
 329         /*
 330          *
 331          * Let's see if all mirrored write operations have finished
 332          * already.
 333          */
 334         if (atomic_dec_and_test(&r1_bio->remaining)) {
 335                 md_write_end(r1_bio->mddev);
 336                 raid_end_bio_io(r1_bio);
 337         }
 338
 339         atomic_dec(&conf->mirrors[mirror].rdev->nr_pending);
 340         return 0;
 341 }
 342
 343
 344 /*
 345  * This routine returns the disk from which the requested read should
 346  * be done. There is a per-array 'next expected sequential IO' sector
 347  * number - if this matches on the next IO then we use the last disk.
 348  * There is also a per-disk 'last know head position' sector that is
 349  * maintained from IRQ contexts, both the normal and the resync IO
 350  * completion handlers update this position correctly. If there is no
 351  * perfect sequential match then we pick the disk whose head is closest.
 352  *
 353  * If there are 2 mirrors in the same 2 devices, performance degrades
 354  * because position is mirror, not device based.
 355  *
 356  * The rdev for the device selected will have nr_pending incremented.
 357  */
 358 static int read_balance(conf_t *conf, struct bio *bio, r1bio_t *r1_bio)
 359 {
 360         const unsigned long this_sector = r1_bio->sector;
 361         int new_disk = conf->last_used, disk = new_disk;
 362         const int sectors = bio->bi_size >> 9;
 363         sector_t new_distance, current_distance;
 364
 365         spin_lock_irq(&conf->device_lock);
 366         /*
 367          * Check if it if we can balance. We can balance on the whole
 368          * device if no resync is going on, or below the resync window.
 369          * We take the first readable disk when above the resync window.
 370          */
 371         if (!conf->mddev->in_sync && (this_sector + sectors >= conf->next_resync)) {
 372                 /* make sure that disk is operational */
 373                 new_disk = 0;
 374
 375                 while (!conf->mirrors[new_disk].rdev ||
 376                        !conf->mirrors[new_disk].rdev->in_sync) {
 377                         new_disk++;
 378                         if (new_disk == conf->raid_disks) {
 379                                 new_disk = 0;
 380                                 break;
 381                         }
 382                 }
 383                 goto rb_out;
 384         }
 385
 386
 387         /* make sure the disk is operational */
 388         while (!conf->mirrors[new_disk].rdev ||
 389                !conf->mirrors[new_disk].rdev->in_sync) {
 390                 if (new_disk <= 0)
 391                         new_disk = conf->raid_disks;
 392                 new_disk--;
 393                 if (new_disk == disk) {
 394                         new_disk = conf->last_used;
 395                         goto rb_out;
 396                 }
 397         }
 398         disk = new_disk;
 399         /* now disk == new_disk == starting point for search */
 400
 401         /*
 402          * Don't change to another disk for sequential reads:
 403          */
 404         if (conf->next_seq_sect == this_sector)
 405                 goto rb_out;
 406         if (this_sector == conf->mirrors[new_disk].head_position)
 407                 goto rb_out;
 408
 409         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
 410
 411         /* Find the disk whose head is closest */
 412
 413         do {
 414                 if (disk <= 0)
 415                         disk = conf->raid_disks;
 416                 disk--;
 417
 418                 if (!conf->mirrors[disk].rdev ||
 419                     !conf->mirrors[disk].rdev->in_sync)
 420                         continue;
 421
 422                 if (!atomic_read(&conf->mirrors[disk].rdev->nr_pending)) {
 423                         new_disk = disk;
 424                         break;
 425                 }
 426                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
 427                 if (new_distance < current_distance) {
 428                         current_distance = new_distance;
 429                         new_disk = disk;
 430                 }
 431         } while (disk != conf->last_used);
 432
 433 rb_out:
 434         r1_bio->read_disk = new_disk;
 435         conf->next_seq_sect = this_sector + sectors;
 436
 437         conf->last_used = new_disk;
 438
 439         if (conf->mirrors[new_disk].rdev)
 440                 atomic_inc(&conf->mirrors[new_disk].rdev->nr_pending);
 441         spin_unlock_irq(&conf->device_lock);
 442
 443         return new_disk;
 444 }
 445
 446 /*
 447  * Throttle resync depth, so that we can both get proper overlapping of
 448  * requests, but are still able to handle normal requests quickly.
 449  */
 450 #define RESYNC_DEPTH 32
 451
 452 static void device_barrier(conf_t *conf, sector_t sect)
 453 {
 454         spin_lock_irq(&conf->resync_lock);
 455         wait_event_lock_irq(conf->wait_idle, !waitqueue_active(&conf->wait_resume), conf->resync_lock);
 456
 457         if (!conf->barrier++) {
 458                 wait_event_lock_irq(conf->wait_idle, !conf->nr_pending, conf->resync_lock);
 459                 if (conf->nr_pending)
 460                         BUG();
 461         }
 462         wait_event_lock_irq(conf->wait_resume, conf->barrier < RESYNC_DEPTH, conf->resync_lock);
 463         conf->next_resync = sect;
 464         spin_unlock_irq(&conf->resync_lock);
 465 }
 466
 467 static int make_request(request_queue_t *q, struct bio * bio)
 468 {
 469         mddev_t *mddev = q->queuedata;
 470         conf_t *conf = mddev_to_conf(mddev);
 471         mirror_info_t *mirror;
 472         r1bio_t *r1_bio;
 473         struct bio *read_bio;
 474         int i, disks = conf->raid_disks;
 475
 476         /*
 477          * Register the new request and wait if the reconstruction
 478          * thread has put up a bar for new requests.
 479          * Continue immediately if no resync is active currently.
 480          */
 481         spin_lock_irq(&conf->resync_lock);
 482         wait_event_lock_irq(conf->wait_resume, !conf->barrier, conf->resync_lock);
 483         conf->nr_pending++;
 484         spin_unlock_irq(&conf->resync_lock);
 485
 486         if (bio_data_dir(bio)==WRITE) {
 487                 disk_stat_inc(mddev->gendisk, writes);
 488                 disk_stat_add(mddev->gendisk, write_sectors, bio_sectors(bio));
 489         } else {
 490                 disk_stat_inc(mddev->gendisk, reads);
 491                 disk_stat_add(mddev->gendisk, read_sectors, bio_sectors(bio));
 492         }
 493
 494         /*
 495          * make_request() can abort the operation when READA is being
 496          * used and no empty request is available.
 497          *
 498          */
 499         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 500
 501         r1_bio->master_bio = bio;
 502         r1_bio->sectors = bio->bi_size >> 9;
 503
 504         r1_bio->mddev = mddev;
 505         r1_bio->sector = bio->bi_sector;
 506
 507         if (bio_data_dir(bio) == READ) {
 508                 /*
 509                  * read balancing logic:
 510                  */
 511                 mirror = conf->mirrors + read_balance(conf, bio, r1_bio);
 512
 513                 read_bio = bio_clone(bio, GFP_NOIO);
 514
 515                 r1_bio->bios[r1_bio->read_disk] = read_bio;
 516
 517                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 518                 read_bio->bi_bdev = mirror->rdev->bdev;
 519                 read_bio->bi_end_io = raid1_end_read_request;
 520                 read_bio->bi_rw = READ;
 521                 read_bio->bi_private = r1_bio;
 522
 523                 generic_make_request(read_bio);
 524                 return 0;
 525         }
 526
 527         /*
 528          * WRITE:
 529          */
 530         /* first select target devices under spinlock and
 531          * inc refcount on their rdev.  Record them by setting
 532          * bios[x] to bio
 533          */
 534         spin_lock_irq(&conf->device_lock);
 535         for (i = 0;  i < disks; i++) {
 536                 if (conf->mirrors[i].rdev &&
 537                     !conf->mirrors[i].rdev->faulty) {
 538                         atomic_inc(&conf->mirrors[i].rdev->nr_pending);
 539                         r1_bio->bios[i] = bio;
 540                 } else
 541                         r1_bio->bios[i] = NULL;
 542         }
 543         spin_unlock_irq(&conf->device_lock);
 544
 545         atomic_set(&r1_bio->remaining, 1);
 546         md_write_start(mddev);
 547         for (i = 0; i < disks; i++) {
 548                 struct bio *mbio;
 549                 if (!r1_bio->bios[i])
 550                         continue;
 551
 552                 mbio = bio_clone(bio, GFP_NOIO);
 553                 r1_bio->bios[i] = mbio;
 554
 555                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 556                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 557                 mbio->bi_end_io = raid1_end_write_request;
 558                 mbio->bi_rw = WRITE;
 559                 mbio->bi_private = r1_bio;
 560
 561                 atomic_inc(&r1_bio->remaining);
 562                 generic_make_request(mbio);
 563         }
 564
 565         if (atomic_dec_and_test(&r1_bio->remaining)) {
 566                 md_write_end(mddev);
 567                 raid_end_bio_io(r1_bio);
 568         }
 569
 570         return 0;
 571 }
 572
 573 static void status(struct seq_file *seq, mddev_t *mddev)
 574 {
 575         conf_t *conf = mddev_to_conf(mddev);
 576         int i;
 577
 578         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 579                                                 conf->working_disks);
 580         for (i = 0; i < conf->raid_disks; i++)
 581                 seq_printf(seq, "%s",
 582                               conf->mirrors[i].rdev &&
 583                               conf->mirrors[i].rdev->in_sync ? "U" : "_");
 584         seq_printf(seq, "]");
 585 }
 586
 587
 588 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 589 {
 590         char b[BDEVNAME_SIZE];
 591         conf_t *conf = mddev_to_conf(mddev);
 592
 593         /*
 594          * If it is not operational, then we have already marked it as dead
 595          * else if it is the last working disks, ignore the error, let the
 596          * next level up know.
 597          * else mark the drive as failed
 598          */
 599         if (rdev->in_sync
 600             && conf->working_disks == 1)
 601                 /*
 602                  * Don't fail the drive, act as though we were just a
 603                  * normal single drive
 604                  */
 605                 return;
 606         if (rdev->in_sync) {
 607                 mddev->degraded++;
 608                 conf->working_disks--;
 609                 /*
 610                  * if recovery is running, make sure it aborts.
 611                  */
 612                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 613         }
 614         rdev->in_sync = 0;
 615         rdev->faulty = 1;
 616         mddev->sb_dirty = 1;
 617         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device. \n"
 618                 "       Operation continuing on %d devices\n",
 619                 bdevname(rdev->bdev,b), conf->working_disks);
 620 }
 621
 622 static void print_conf(conf_t *conf)
 623 {
 624         int i;
 625         mirror_info_t *tmp;
 626
 627         printk("RAID1 conf printout:\n");
 628         if (!conf) {
 629                 printk("(!conf)\n");
 630                 return;
 631         }
 632         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 633                 conf->raid_disks);
 634
 635         for (i = 0; i < conf->raid_disks; i++) {
 636                 char b[BDEVNAME_SIZE];
 637                 tmp = conf->mirrors + i;
 638                 if (tmp->rdev)
 639                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
 640                                 i, !tmp->rdev->in_sync, !tmp->rdev->faulty,
 641                                 bdevname(tmp->rdev->bdev,b));
 642         }
 643 }
 644
 645 static void close_sync(conf_t *conf)
 646 {
 647         spin_lock_irq(&conf->resync_lock);
 648         wait_event_lock_irq(conf->wait_resume, !conf->barrier, conf->resync_lock);
 649         spin_unlock_irq(&conf->resync_lock);
 650
 651         if (conf->barrier) BUG();
 652         if (waitqueue_active(&conf->wait_idle)) BUG();
 653
 654         mempool_destroy(conf->r1buf_pool);
 655         conf->r1buf_pool = NULL;
 656 }
 657
 658 static int raid1_spare_active(mddev_t *mddev)
 659 {
 660         int i;
 661         conf_t *conf = mddev->private;
 662         mirror_info_t *tmp;
 663
 664         spin_lock_irq(&conf->device_lock);
 665         /*
 666          * Find all failed disks within the RAID1 configuration
 667          * and mark them readable
 668          */
 669         for (i = 0; i < conf->raid_disks; i++) {
 670                 tmp = conf->mirrors + i;
 671                 if (tmp->rdev
 672                     && !tmp->rdev->faulty
 673                     && !tmp->rdev->in_sync) {
 674                         conf->working_disks++;
 675                         mddev->degraded--;
 676                         tmp->rdev->in_sync = 1;
 677                 }
 678         }
 679         spin_unlock_irq(&conf->device_lock);
 680
 681         print_conf(conf);
 682         return 0;
 683 }
 684
 685
 686 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 687 {
 688         conf_t *conf = mddev->private;
 689         int found = 0;
 690         int mirror;
 691         mirror_info_t *p;
 692
 693         spin_lock_irq(&conf->device_lock);
 694         for (mirror=0; mirror < mddev->raid_disks; mirror++)
 695                 if ( !(p=conf->mirrors+mirror)->rdev) {
 696                         p->rdev = rdev;
 697
 698                         blk_queue_stack_limits(mddev->queue,
 699                                                rdev->bdev->bd_disk->queue);
 700                         /* as we don't honour merge_bvec_fn, we must never risk
 701                          * violating it, so limit ->max_sector to one PAGE, as
 702                          * a one page request is never in violation.
 703                          */
 704                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
 705                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
 706                                 mddev->queue->max_sectors = (PAGE_SIZE>>9);
 707
 708                         p->head_position = 0;
 709                         rdev->raid_disk = mirror;
 710                         found = 1;
 711                         break;
 712                 }
 713         spin_unlock_irq(&conf->device_lock);
 714
 715         print_conf(conf);
 716         return found;
 717 }
 718
 719 static int raid1_remove_disk(mddev_t *mddev, int number)
 720 {
 721         conf_t *conf = mddev->private;
 722         int err = 1;
 723         mirror_info_t *p = conf->mirrors+ number;
 724
 725         print_conf(conf);
 726         spin_lock_irq(&conf->device_lock);
 727         if (p->rdev) {
 728                 if (p->rdev->in_sync ||
 729                     atomic_read(&p->rdev->nr_pending)) {
 730                         err = -EBUSY;
 731                         goto abort;
 732                 }
 733                 p->rdev = NULL;
 734                 err = 0;
 735         }
 736         if (err)
 737                 MD_BUG();
 738 abort:
 739         spin_unlock_irq(&conf->device_lock);
 740
 741         print_conf(conf);
 742         return err;
 743 }
 744
 745
 746 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
 747 {
 748         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 749         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 750         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 751
 752         if (bio->bi_size)
 753                 return 1;
 754
 755         if (r1_bio->bios[r1_bio->read_disk] != bio)
 756                 BUG();
 757         update_head_pos(r1_bio->read_disk, r1_bio);
 758         /*
 759          * we have read a block, now it needs to be re-written,
 760          * or re-read if the read failed.
 761          * We don't do much here, just schedule handling by raid1d
 762          */
 763         if (!uptodate)
 764                 md_error(r1_bio->mddev,
 765                          conf->mirrors[r1_bio->read_disk].rdev);
 766         else
 767                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 768         atomic_dec(&conf->mirrors[r1_bio->read_disk].rdev->nr_pending);
 769         reschedule_retry(r1_bio);
 770         return 0;
 771 }
 772
 773 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
 774 {
 775         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 776         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 777         mddev_t *mddev = r1_bio->mddev;
 778         conf_t *conf = mddev_to_conf(mddev);
 779         int i;
 780         int mirror=0;
 781
 782         if (bio->bi_size)
 783                 return 1;
 784
 785         for (i = 0; i < conf->raid_disks; i++)
 786                 if (r1_bio->bios[i] == bio) {
 787                         mirror = i;
 788                         break;
 789                 }
 790         if (!uptodate)
 791                 md_error(mddev, conf->mirrors[mirror].rdev);
 792         update_head_pos(mirror, r1_bio);
 793
 794         if (atomic_dec_and_test(&r1_bio->remaining)) {
 795                 md_done_sync(mddev, r1_bio->sectors, uptodate);
 796                 put_buf(r1_bio);
 797         }
 798         atomic_dec(&conf->mirrors[mirror].rdev->nr_pending);
 799         return 0;
 800 }
 801
 802 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
 803 {
 804         conf_t *conf = mddev_to_conf(mddev);
 805         int i;
 806         int disks = conf->raid_disks;
 807         struct bio *bio, *wbio;
 808
 809         bio = r1_bio->bios[r1_bio->read_disk];
 810
 811         /*
 812          * schedule writes
 813          */
 814         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 815                 /*
 816                  * There is no point trying a read-for-reconstruct as
 817                  * reconstruct is about to be aborted
 818                  */
 819                 char b[BDEVNAME_SIZE];
 820                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
 821                         " for block %llu\n",
 822                         bdevname(bio->bi_bdev,b),
 823                         (unsigned long long)r1_bio->sector);
 824                 md_done_sync(mddev, r1_bio->sectors, 0);
 825                 put_buf(r1_bio);
 826                 return;
 827         }
 828
 829         atomic_set(&r1_bio->remaining, 1);
 830         for (i = 0; i < disks ; i++) {
 831                 wbio = r1_bio->bios[i];
 832                 if (wbio->bi_end_io != end_sync_write)
 833                         continue;
 834
 835                 atomic_inc(&conf->mirrors[i].rdev->nr_pending);
 836                 atomic_inc(&r1_bio->remaining);
 837                 md_sync_acct(conf->mirrors[i].rdev, wbio->bi_size >> 9);
 838                 generic_make_request(wbio);
 839         }
 840
 841         if (atomic_dec_and_test(&r1_bio->remaining)) {
 842                 md_done_sync(mddev, r1_bio->sectors, 1);
 843                 put_buf(r1_bio);
 844         }
 845 }
 846
 847 /*
 848  * This is a kernel thread which:
 849  *
 850  *      1.      Retries failed read operations on working mirrors.
 851  *      2.      Updates the raid superblock when problems encounter.
 852  *      3.      Performs writes following reads for array syncronising.
 853  */
 854
 855 static void raid1d(mddev_t *mddev)
 856 {
 857         struct list_head *head = &retry_list_head;
 858         r1bio_t *r1_bio;
 859         struct bio *bio;
 860         unsigned long flags;
 861         conf_t *conf = mddev_to_conf(mddev);
 862         mdk_rdev_t *rdev;
 863
 864         md_check_recovery(mddev);
 865         md_handle_safemode(mddev);
 866
 867         for (;;) {
 868                 char b[BDEVNAME_SIZE];
 869                 spin_lock_irqsave(&retry_list_lock, flags);
 870                 if (list_empty(head))
 871                         break;
 872                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
 873                 list_del(head->prev);
 874                 spin_unlock_irqrestore(&retry_list_lock, flags);
 875
 876                 mddev = r1_bio->mddev;
 877                 conf = mddev_to_conf(mddev);
 878                 bio = r1_bio->master_bio;
 879                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
 880                         sync_request_write(mddev, r1_bio);
 881                 } else {
 882                         if (map(mddev, &rdev) == -1) {
 883                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
 884                                        " read error for block %llu\n",
 885                                        bdevname(bio->bi_bdev,b),
 886                                        (unsigned long long)r1_bio->sector);
 887                                 raid_end_bio_io(r1_bio);
 888                         } else {
 889                                 printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
 890                                        " another mirror\n",
 891                                        bdevname(rdev->bdev,b),
 892                                        (unsigned long long)r1_bio->sector);
 893                                 bio->bi_bdev = rdev->bdev;
 894                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
 895                                 bio->bi_rw = READ;
 896
 897                                 generic_make_request(bio);
 898                         }
 899                 }
 900         }
 901         spin_unlock_irqrestore(&retry_list_lock, flags);
 902 }
 903
 904
 905 static int init_resync(conf_t *conf)
 906 {
 907         int buffs;
 908
 909         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
 910         if (conf->r1buf_pool)
 911                 BUG();
 912         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free, conf);
 913         if (!conf->r1buf_pool)
 914                 return -ENOMEM;
 915         conf->next_resync = 0;
 916         return 0;
 917 }
 918
 919 /*
 920  * perform a "sync" on one "block"
 921  *
 922  * We need to make sure that no normal I/O request - particularly write
 923  * requests - conflict with active sync requests.
 924  *
 925  * This is achieved by tracking pending requests and a 'barrier' concept
 926  * that can be installed to exclude normal IO requests.
 927  */
 928
 929 static int sync_request(mddev_t *mddev, sector_t sector_nr, int go_faster)
 930 {
 931         conf_t *conf = mddev_to_conf(mddev);
 932         mirror_info_t *mirror;
 933         r1bio_t *r1_bio;
 934         struct bio *bio;
 935         sector_t max_sector, nr_sectors;
 936         int disk;
 937         int i;
 938
 939         if (!conf->r1buf_pool)
 940                 if (init_resync(conf))
 941                         return -ENOMEM;
 942
 943         max_sector = mddev->size << 1;
 944         if (sector_nr >= max_sector) {
 945                 close_sync(conf);
 946                 return 0;
 947         }
 948
 949         /*
 950          * If there is non-resync activity waiting for us then
 951          * put in a delay to throttle resync.
 952          */
 953         if (!go_faster && waitqueue_active(&conf->wait_resume))
 954                 schedule_timeout(HZ);
 955         device_barrier(conf, sector_nr + RESYNC_SECTORS);
 956
 957         /*
 958          * If reconstructing, and >1 working disc,
 959          * could dedicate one to rebuild and others to
 960          * service read requests ..
 961          */
 962         disk = conf->last_used;
 963         /* make sure disk is operational */
 964         spin_lock_irq(&conf->device_lock);
 965         while (conf->mirrors[disk].rdev == NULL ||
 966                !conf->mirrors[disk].rdev->in_sync) {
 967                 if (disk <= 0)
 968                         disk = conf->raid_disks;
 969                 disk--;
 970                 if (disk == conf->last_used)
 971                         break;
 972         }
 973         conf->last_used = disk;
 974         atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
 975         spin_unlock_irq(&conf->device_lock);
 976
 977         mirror = conf->mirrors + disk;
 978
 979         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
 980
 981         spin_lock_irq(&conf->resync_lock);
 982         conf->nr_pending++;
 983         spin_unlock_irq(&conf->resync_lock);
 984
 985         r1_bio->mddev = mddev;
 986         r1_bio->sector = sector_nr;
 987         set_bit(R1BIO_IsSync, &r1_bio->state);
 988         r1_bio->read_disk = disk;
 989
 990         for (i=0; i < conf->raid_disks; i++) {
 991                 bio = r1_bio->bios[i];
 992
 993                 /* take from bio_init */
 994                 bio->bi_next = NULL;
 995                 bio->bi_flags |= 1 << BIO_UPTODATE;
 996                 bio->bi_rw = 0;
 997                 bio->bi_vcnt = 0;
 998                 bio->bi_idx = 0;
 999                 bio->bi_phys_segments = 0;
1000                 bio->bi_hw_segments = 0;
1001                 bio->bi_size = 0;
1002                 bio->bi_end_io = NULL;
1003                 bio->bi_private = NULL;
1004
1005                 if (i == disk) {
1006                         bio->bi_rw = READ;
1007                         bio->bi_end_io = end_sync_read;
1008                 } else if (conf->mirrors[i].rdev &&
1009                            !conf->mirrors[i].rdev->faulty &&
1010                            (!conf->mirrors[i].rdev->in_sync ||
1011                             sector_nr + RESYNC_SECTORS > mddev->recovery_cp)) {
1012                         bio->bi_rw = WRITE;
1013                         bio->bi_end_io = end_sync_write;
1014                 } else
1015                         continue;
1016                 bio->bi_sector = sector_nr + conf->mirrors[i].rdev->data_offset;
1017                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1018                 bio->bi_private = r1_bio;
1019         }
1020         nr_sectors = 0;
1021         do {
1022                 struct page *page;
1023                 int len = PAGE_SIZE;
1024                 if (sector_nr + (len>>9) > max_sector)
1025                         len = (max_sector - sector_nr) << 9;
1026                 if (len == 0)
1027                         break;
1028                 for (i=0 ; i < conf->raid_disks; i++) {
1029                         bio = r1_bio->bios[i];
1030                         if (bio->bi_end_io) {
1031                                 page = r1_bio->bios[0]->bi_io_vec[bio->bi_vcnt].bv_page;
1032                                 if (bio_add_page(bio, page, len, 0) == 0) {
1033                                         /* stop here */
1034                                         r1_bio->bios[0]->bi_io_vec[bio->bi_vcnt].bv_page = page;
1035                                         while (i > 0) {
1036                                                 i--;
1037                                                 bio = r1_bio->bios[i];
1038                                                 if (bio->bi_end_io==NULL) continue;
1039                                                 /* remove last page from this bio */
1040                                                 bio->bi_vcnt--;
1041                                                 bio->bi_size -= len;
1042                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1043                                         }
1044                                         goto bio_full;
1045                                 }
1046                         }
1047                 }
1048                 nr_sectors += len>>9;
1049                 sector_nr += len>>9;
1050         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1051  bio_full:
1052         bio = r1_bio->bios[disk];
1053         r1_bio->sectors = nr_sectors;
1054
1055         md_sync_acct(mirror->rdev, nr_sectors);
1056
1057         generic_make_request(bio);
1058
1059         return nr_sectors;
1060 }
1061
1062 static int run(mddev_t *mddev)
1063 {
1064         conf_t *conf;
1065         int i, j, disk_idx;
1066         mirror_info_t *disk;
1067         mdk_rdev_t *rdev;
1068         struct list_head *tmp;
1069
1070         if (mddev->level != 1) {
1071                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1072                        mdname(mddev), mddev->level);
1073                 goto out;
1074         }
1075         /*
1076          * copy the already verified devices into our private RAID1
1077          * bookkeeping area. [whatever we allocate in run(),
1078          * should be freed in stop()]
1079          */
1080         conf = kmalloc(sizeof(conf_t), GFP_KERNEL);
1081         mddev->private = conf;
1082         if (!conf) {
1083                 printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1084                         mdname(mddev));
1085                 goto out;
1086         }
1087         memset(conf, 0, sizeof(*conf));
1088         conf->mirrors = kmalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1089                                  GFP_KERNEL);
1090         if (!conf->mirrors) {
1091                 printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1092                        mdname(mddev));
1093                 goto out_free_conf;
1094         }
1095         memset(conf->mirrors, 0, sizeof(struct mirror_info)*mddev->raid_disks);
1096
1097         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1098                                                 r1bio_pool_free, mddev);
1099         if (!conf->r1bio_pool) {
1100                 printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1101                         mdname(mddev));
1102                 goto out_free_conf;
1103         }
1104
1105
1106         ITERATE_RDEV(mddev, rdev, tmp) {
1107                 disk_idx = rdev->raid_disk;
1108                 if (disk_idx >= mddev->raid_disks
1109                     || disk_idx < 0)
1110                         continue;
1111                 disk = conf->mirrors + disk_idx;
1112
1113                 disk->rdev = rdev;
1114
1115                 blk_queue_stack_limits(mddev->queue,
1116                                        rdev->bdev->bd_disk->queue);
1117                 /* as we don't honour merge_bvec_fn, we must never risk
1118                  * violating it, so limit ->max_sector to one PAGE, as
1119                  * a one page request is never in violation.
1120                  */
1121                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1122                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1123                         mddev->queue->max_sectors = (PAGE_SIZE>>9);
1124
1125                 disk->head_position = 0;
1126                 if (!rdev->faulty && rdev->in_sync)
1127                         conf->working_disks++;
1128         }
1129         conf->raid_disks = mddev->raid_disks;
1130         conf->mddev = mddev;
1131         conf->device_lock = SPIN_LOCK_UNLOCKED;
1132         if (conf->working_disks == 1)
1133                 mddev->recovery_cp = MaxSector;
1134
1135         conf->resync_lock = SPIN_LOCK_UNLOCKED;
1136         init_waitqueue_head(&conf->wait_idle);
1137         init_waitqueue_head(&conf->wait_resume);
1138
1139         if (!conf->working_disks) {
1140                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
1141                         mdname(mddev));
1142                 goto out_free_conf;
1143         }
1144
1145         mddev->degraded = 0;
1146         for (i = 0; i < conf->raid_disks; i++) {
1147
1148                 disk = conf->mirrors + i;
1149
1150                 if (!disk->rdev) {
1151                         disk->head_position = 0;
1152                         mddev->degraded++;
1153                 }
1154         }
1155
1156         /*
1157          * find the first working one and use it as a starting point
1158          * to read balancing.
1159          */
1160         for (j = 0; j < conf->raid_disks &&
1161                      (!conf->mirrors[j].rdev ||
1162                       !conf->mirrors[j].rdev->in_sync) ; j++)
1163                 /* nothing */;
1164         conf->last_used = j;
1165
1166
1167
1168         {
1169                 mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
1170                 if (!mddev->thread) {
1171                         printk(KERN_ERR
1172                                 "raid1: couldn't allocate thread for %s\n",
1173                                 mdname(mddev));
1174                         goto out_free_conf;
1175                 }
1176         }
1177         printk(KERN_INFO
1178                 "raid1: raid set %s active with %d out of %d mirrors\n",
1179                 mdname(mddev), mddev->raid_disks - mddev->degraded,
1180                 mddev->raid_disks);
1181         /*
1182          * Ok, everything is just fine now
1183          */
1184         mddev->array_size = mddev->size;
1185
1186         return 0;
1187
1188 out_free_conf:
1189         if (conf->r1bio_pool)
1190                 mempool_destroy(conf->r1bio_pool);
1191         if (conf->mirrors)
1192                 kfree(conf->mirrors);
1193         kfree(conf);
1194         mddev->private = NULL;
1195 out:
1196         return -EIO;
1197 }
1198
1199 static int stop(mddev_t *mddev)
1200 {
1201         conf_t *conf = mddev_to_conf(mddev);
1202
1203         md_unregister_thread(mddev->thread);
1204         mddev->thread = NULL;
1205         if (conf->r1bio_pool)
1206                 mempool_destroy(conf->r1bio_pool);
1207         if (conf->mirrors)
1208                 kfree(conf->mirrors);
1209         kfree(conf);
1210         mddev->private = NULL;
1211         return 0;
1212 }
1213
1214 static mdk_personality_t raid1_personality =
1215 {
1216         .name           = "raid1",
1217         .owner          = THIS_MODULE,
1218         .make_request   = make_request,
1219         .run            = run,
1220         .stop           = stop,
1221         .status         = status,
1222         .error_handler  = error,
1223         .hot_add_disk   = raid1_add_disk,
1224         .hot_remove_disk= raid1_remove_disk,
1225         .spare_active   = raid1_spare_active,
1226         .sync_request   = sync_request,
1227 };
1228
1229 static int __init raid_init(void)
1230 {
1231         return register_md_personality(RAID1, &raid1_personality);
1232 }
1233
1234 static void raid_exit(void)
1235 {
1236         unregister_md_personality(RAID1);
1237 }
1238
1239 module_init(raid_init);
1240 module_exit(raid_exit);
1241 MODULE_LICENSE("GPL");
1242 MODULE_ALIAS("md-personality-3"); /* RAID1 */