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 unplug_slaves(mddev_t *mddev);
  41
  42
  43 static void * r1bio_pool_alloc(int gfp_flags, void *data)
  44 {
  45         mddev_t *mddev = data;
  46         r1bio_t *r1_bio;
  47
  48         /* allocate a r1bio with room for raid_disks entries in the bios array */
  49         r1_bio = kmalloc(sizeof(r1bio_t) + sizeof(struct bio*)*mddev->raid_disks,
  50                          gfp_flags);
  51         if (r1_bio)
  52                 memset(r1_bio, 0, sizeof(*r1_bio) + sizeof(struct bio*)*mddev->raid_disks);
  53         else
  54                 unplug_slaves(mddev);
  55
  56         return r1_bio;
  57 }
  58
  59 static void r1bio_pool_free(void *r1_bio, void *data)
  60 {
  61         kfree(r1_bio);
  62 }
  63
  64 #define RESYNC_BLOCK_SIZE (64*1024)
  65 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  66 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  67 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  68 #define RESYNC_WINDOW (2048*1024)
  69
  70 static void * r1buf_pool_alloc(int gfp_flags, void *data)
  71 {
  72         conf_t *conf = data;
  73         struct page *page;
  74         r1bio_t *r1_bio;
  75         struct bio *bio;
  76         int i, j;
  77
  78         r1_bio = r1bio_pool_alloc(gfp_flags, conf->mddev);
  79         if (!r1_bio) {
  80                 unplug_slaves(conf->mddev);
  81                 return NULL;
  82         }
  83
  84         /*
  85          * Allocate bios : 1 for reading, n-1 for writing
  86          */
  87         for (j = conf->raid_disks ; j-- ; ) {
  88                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
  89                 if (!bio)
  90                         goto out_free_bio;
  91                 r1_bio->bios[j] = bio;
  92         }
  93         /*
  94          * Allocate RESYNC_PAGES data pages and attach them to
  95          * the first bio;
  96          */
  97         bio = r1_bio->bios[0];
  98         for (i = 0; i < RESYNC_PAGES; i++) {
  99                 page = alloc_page(gfp_flags);
 100                 if (unlikely(!page))
 101                         goto out_free_pages;
 102
 103                 bio->bi_io_vec[i].bv_page = page;
 104         }
 105
 106         r1_bio->master_bio = bio;
 107
 108         return r1_bio;
 109
 110 out_free_pages:
 111         for ( ; i > 0 ; i--)
 112                 __free_page(bio->bi_io_vec[i-1].bv_page);
 113 out_free_bio:
 114         while ( ++j < conf->raid_disks )
 115                 bio_put(r1_bio->bios[j]);
 116         r1bio_pool_free(r1_bio, conf->mddev);
 117         return NULL;
 118 }
 119
 120 static void r1buf_pool_free(void *__r1_bio, void *data)
 121 {
 122         int i;
 123         conf_t *conf = data;
 124         r1bio_t *r1bio = __r1_bio;
 125         struct bio *bio = r1bio->bios[0];
 126
 127         for (i = 0; i < RESYNC_PAGES; i++) {
 128                 __free_page(bio->bi_io_vec[i].bv_page);
 129                 bio->bi_io_vec[i].bv_page = NULL;
 130         }
 131         for (i=0 ; i < conf->raid_disks; i++)
 132                 bio_put(r1bio->bios[i]);
 133
 134         r1bio_pool_free(r1bio, conf->mddev);
 135 }
 136
 137 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
 138 {
 139         int i;
 140
 141         for (i = 0; i < conf->raid_disks; i++) {
 142                 struct bio **bio = r1_bio->bios + i;
 143                 if (*bio)
 144                         bio_put(*bio);
 145                 *bio = NULL;
 146         }
 147 }
 148
 149 static inline void free_r1bio(r1bio_t *r1_bio)
 150 {
 151         unsigned long flags;
 152
 153         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 154
 155         /*
 156          * Wake up any possible resync thread that waits for the device
 157          * to go idle.
 158          */
 159         spin_lock_irqsave(&conf->resync_lock, flags);
 160         if (!--conf->nr_pending) {
 161                 wake_up(&conf->wait_idle);
 162                 wake_up(&conf->wait_resume);
 163         }
 164         spin_unlock_irqrestore(&conf->resync_lock, flags);
 165
 166         put_all_bios(conf, r1_bio);
 167         mempool_free(r1_bio, conf->r1bio_pool);
 168 }
 169
 170 static inline void put_buf(r1bio_t *r1_bio)
 171 {
 172         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 173         unsigned long flags;
 174
 175         mempool_free(r1_bio, conf->r1buf_pool);
 176
 177         spin_lock_irqsave(&conf->resync_lock, flags);
 178         if (!conf->barrier)
 179                 BUG();
 180         --conf->barrier;
 181         wake_up(&conf->wait_resume);
 182         wake_up(&conf->wait_idle);
 183
 184         if (!--conf->nr_pending) {
 185                 wake_up(&conf->wait_idle);
 186                 wake_up(&conf->wait_resume);
 187         }
 188         spin_unlock_irqrestore(&conf->resync_lock, flags);
 189 }
 190
 191 static int map(mddev_t *mddev, mdk_rdev_t **rdevp)
 192 {
 193         conf_t *conf = mddev_to_conf(mddev);
 194         int i, disks = conf->raid_disks;
 195
 196         /*
 197          * Later we do read balancing on the read side
 198          * now we use the first available disk.
 199          */
 200
 201         spin_lock_irq(&conf->device_lock);
 202         for (i = 0; i < disks; i++) {
 203                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
 204                 if (rdev && rdev->in_sync) {
 205                         *rdevp = rdev;
 206                         atomic_inc(&rdev->nr_pending);
 207                         spin_unlock_irq(&conf->device_lock);
 208                         return 0;
 209                 }
 210         }
 211         spin_unlock_irq(&conf->device_lock);
 212
 213         printk(KERN_ERR "raid1_map(): huh, no more operational devices?\n");
 214         return -1;
 215 }
 216
 217 static void reschedule_retry(r1bio_t *r1_bio)
 218 {
 219         unsigned long flags;
 220         mddev_t *mddev = r1_bio->mddev;
 221
 222         spin_lock_irqsave(&retry_list_lock, flags);
 223         list_add(&r1_bio->retry_list, &retry_list_head);
 224         spin_unlock_irqrestore(&retry_list_lock, flags);
 225
 226         md_wakeup_thread(mddev->thread);
 227 }
 228
 229 /*
 230  * raid_end_bio_io() is called when we have finished servicing a mirrored
 231  * operation and are ready to return a success/failure code to the buffer
 232  * cache layer.
 233  */
 234 static void raid_end_bio_io(r1bio_t *r1_bio)
 235 {
 236         struct bio *bio = r1_bio->master_bio;
 237
 238         bio_endio(bio, bio->bi_size,
 239                 test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
 240         free_r1bio(r1_bio);
 241 }
 242
 243 /*
 244  * Update disk head position estimator based on IRQ completion info.
 245  */
 246 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
 247 {
 248         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 249
 250         conf->mirrors[disk].head_position =
 251                 r1_bio->sector + (r1_bio->sectors);
 252 }
 253
 254 static int raid1_end_read_request(struct bio *bio, unsigned int bytes_done, int error)
 255 {
 256         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 257         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 258         int mirror;
 259         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 260
 261         if (bio->bi_size)
 262                 return 1;
 263
 264         mirror = r1_bio->read_disk;
 265         /*
 266          * this branch is our 'one mirror IO has finished' event handler:
 267          */
 268         if (!uptodate)
 269                 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 270         else
 271                 /*
 272                  * Set R1BIO_Uptodate in our master bio, so that
 273                  * we will return a good error code for to the higher
 274                  * levels even if IO on some other mirrored buffer fails.
 275                  *
 276                  * The 'master' represents the composite IO operation to
 277                  * user-side. So if something waits for IO, then it will
 278                  * wait for the 'master' bio.
 279                  */
 280                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 281
 282         update_head_pos(mirror, r1_bio);
 283
 284         /*
 285          * we have only one bio on the read side
 286          */
 287         if (uptodate)
 288                 raid_end_bio_io(r1_bio);
 289         else {
 290                 /*
 291                  * oops, read error:
 292                  */
 293                 char b[BDEVNAME_SIZE];
 294                 printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
 295                        bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
 296                 reschedule_retry(r1_bio);
 297         }
 298
 299         atomic_dec(&conf->mirrors[mirror].rdev->nr_pending);
 300         return 0;
 301 }
 302
 303 static int raid1_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
 304 {
 305         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 306         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 307         int mirror;
 308         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 309
 310         if (bio->bi_size)
 311                 return 1;
 312
 313         for (mirror = 0; mirror < conf->raid_disks; mirror++)
 314                 if (r1_bio->bios[mirror] == bio)
 315                         break;
 316
 317         /*
 318          * this branch is our 'one mirror IO has finished' event handler:
 319          */
 320         if (!uptodate)
 321                 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 322         else
 323                 /*
 324                  * Set R1BIO_Uptodate in our master bio, so that
 325                  * we will return a good error code for to the higher
 326                  * levels even if IO on some other mirrored buffer fails.
 327                  *
 328                  * The 'master' represents the composite IO operation to
 329                  * user-side. So if something waits for IO, then it will
 330                  * wait for the 'master' bio.
 331                  */
 332                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 333
 334         update_head_pos(mirror, r1_bio);
 335
 336         /*
 337          *
 338          * Let's see if all mirrored write operations have finished
 339          * already.
 340          */
 341         if (atomic_dec_and_test(&r1_bio->remaining)) {
 342                 md_write_end(r1_bio->mddev);
 343                 raid_end_bio_io(r1_bio);
 344         }
 345
 346         atomic_dec(&conf->mirrors[mirror].rdev->nr_pending);
 347         return 0;
 348 }
 349
 350
 351 /*
 352  * This routine returns the disk from which the requested read should
 353  * be done. There is a per-array 'next expected sequential IO' sector
 354  * number - if this matches on the next IO then we use the last disk.
 355  * There is also a per-disk 'last know head position' sector that is
 356  * maintained from IRQ contexts, both the normal and the resync IO
 357  * completion handlers update this position correctly. If there is no
 358  * perfect sequential match then we pick the disk whose head is closest.
 359  *
 360  * If there are 2 mirrors in the same 2 devices, performance degrades
 361  * because position is mirror, not device based.
 362  *
 363  * The rdev for the device selected will have nr_pending incremented.
 364  */
 365 static int read_balance(conf_t *conf, struct bio *bio, r1bio_t *r1_bio)
 366 {
 367         const unsigned long this_sector = r1_bio->sector;
 368         int new_disk = conf->last_used, disk = new_disk;
 369         const int sectors = bio->bi_size >> 9;
 370         sector_t new_distance, current_distance;
 371
 372         spin_lock_irq(&conf->device_lock);
 373         /*
 374          * Check if it if we can balance. We can balance on the whole
 375          * device if no resync is going on, or below the resync window.
 376          * We take the first readable disk when above the resync window.
 377          */
 378         if (!conf->mddev->in_sync && (this_sector + sectors >= conf->next_resync)) {
 379                 /* make sure that disk is operational */
 380                 new_disk = 0;
 381
 382                 while (!conf->mirrors[new_disk].rdev ||
 383                        !conf->mirrors[new_disk].rdev->in_sync) {
 384                         new_disk++;
 385                         if (new_disk == conf->raid_disks) {
 386                                 new_disk = 0;
 387                                 break;
 388                         }
 389                 }
 390                 goto rb_out;
 391         }
 392
 393
 394         /* make sure the disk is operational */
 395         while (!conf->mirrors[new_disk].rdev ||
 396                !conf->mirrors[new_disk].rdev->in_sync) {
 397                 if (new_disk <= 0)
 398                         new_disk = conf->raid_disks;
 399                 new_disk--;
 400                 if (new_disk == disk) {
 401                         new_disk = conf->last_used;
 402                         goto rb_out;
 403                 }
 404         }
 405         disk = new_disk;
 406         /* now disk == new_disk == starting point for search */
 407
 408         /*
 409          * Don't change to another disk for sequential reads:
 410          */
 411         if (conf->next_seq_sect == this_sector)
 412                 goto rb_out;
 413         if (this_sector == conf->mirrors[new_disk].head_position)
 414                 goto rb_out;
 415
 416         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
 417
 418         /* Find the disk whose head is closest */
 419
 420         do {
 421                 if (disk <= 0)
 422                         disk = conf->raid_disks;
 423                 disk--;
 424
 425                 if (!conf->mirrors[disk].rdev ||
 426                     !conf->mirrors[disk].rdev->in_sync)
 427                         continue;
 428
 429                 if (!atomic_read(&conf->mirrors[disk].rdev->nr_pending)) {
 430                         new_disk = disk;
 431                         break;
 432                 }
 433                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
 434                 if (new_distance < current_distance) {
 435                         current_distance = new_distance;
 436                         new_disk = disk;
 437                 }
 438         } while (disk != conf->last_used);
 439
 440 rb_out:
 441         r1_bio->read_disk = new_disk;
 442         conf->next_seq_sect = this_sector + sectors;
 443
 444         conf->last_used = new_disk;
 445
 446         if (conf->mirrors[new_disk].rdev)
 447                 atomic_inc(&conf->mirrors[new_disk].rdev->nr_pending);
 448         spin_unlock_irq(&conf->device_lock);
 449
 450         return new_disk;
 451 }
 452
 453 static void unplug_slaves(mddev_t *mddev)
 454 {
 455         conf_t *conf = mddev_to_conf(mddev);
 456         int i;
 457         unsigned long flags;
 458
 459         spin_lock_irqsave(&conf->device_lock, flags);
 460         for (i=0; i<mddev->raid_disks; i++) {
 461                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
 462                 if (rdev && atomic_read(&rdev->nr_pending)) {
 463                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
 464
 465                         atomic_inc(&rdev->nr_pending);
 466                         spin_unlock_irqrestore(&conf->device_lock, flags);
 467
 468                         if (r_queue->unplug_fn)
 469                                 r_queue->unplug_fn(r_queue);
 470
 471                         spin_lock_irqsave(&conf->device_lock, flags);
 472                         atomic_dec(&rdev->nr_pending);
 473                 }
 474         }
 475         spin_unlock_irqrestore(&conf->device_lock, flags);
 476 }
 477 static void raid1_unplug(request_queue_t *q)
 478 {
 479         unplug_slaves(q->queuedata);
 480 }
 481
 482 /*
 483  * Throttle resync depth, so that we can both get proper overlapping of
 484  * requests, but are still able to handle normal requests quickly.
 485  */
 486 #define RESYNC_DEPTH 32
 487
 488 static void device_barrier(conf_t *conf, sector_t sect)
 489 {
 490         spin_lock_irq(&conf->resync_lock);
 491         wait_event_lock_irq(conf->wait_idle, !waitqueue_active(&conf->wait_resume),
 492                             conf->resync_lock, unplug_slaves(conf->mddev));
 493
 494         if (!conf->barrier++) {
 495                 wait_event_lock_irq(conf->wait_idle, !conf->nr_pending,
 496                                     conf->resync_lock, unplug_slaves(conf->mddev));
 497                 if (conf->nr_pending)
 498                         BUG();
 499         }
 500         wait_event_lock_irq(conf->wait_resume, conf->barrier < RESYNC_DEPTH,
 501                             conf->resync_lock, unplug_slaves(conf->mddev));
 502         conf->next_resync = sect;
 503         spin_unlock_irq(&conf->resync_lock);
 504 }
 505
 506 static int make_request(request_queue_t *q, struct bio * bio)
 507 {
 508         mddev_t *mddev = q->queuedata;
 509         conf_t *conf = mddev_to_conf(mddev);
 510         mirror_info_t *mirror;
 511         r1bio_t *r1_bio;
 512         struct bio *read_bio;
 513         int i, disks = conf->raid_disks;
 514
 515         /*
 516          * Register the new request and wait if the reconstruction
 517          * thread has put up a bar for new requests.
 518          * Continue immediately if no resync is active currently.
 519          */
 520         spin_lock_irq(&conf->resync_lock);
 521         wait_event_lock_irq(conf->wait_resume, !conf->barrier, conf->resync_lock, );
 522         conf->nr_pending++;
 523         spin_unlock_irq(&conf->resync_lock);
 524
 525         if (bio_data_dir(bio)==WRITE) {
 526                 disk_stat_inc(mddev->gendisk, writes);
 527                 disk_stat_add(mddev->gendisk, write_sectors, bio_sectors(bio));
 528         } else {
 529                 disk_stat_inc(mddev->gendisk, reads);
 530                 disk_stat_add(mddev->gendisk, read_sectors, bio_sectors(bio));
 531         }
 532
 533         /*
 534          * make_request() can abort the operation when READA is being
 535          * used and no empty request is available.
 536          *
 537          */
 538         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 539
 540         r1_bio->master_bio = bio;
 541         r1_bio->sectors = bio->bi_size >> 9;
 542
 543         r1_bio->mddev = mddev;
 544         r1_bio->sector = bio->bi_sector;
 545
 546         if (bio_data_dir(bio) == READ) {
 547                 /*
 548                  * read balancing logic:
 549                  */
 550                 mirror = conf->mirrors + read_balance(conf, bio, r1_bio);
 551
 552                 read_bio = bio_clone(bio, GFP_NOIO);
 553
 554                 r1_bio->bios[r1_bio->read_disk] = read_bio;
 555
 556                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 557                 read_bio->bi_bdev = mirror->rdev->bdev;
 558                 read_bio->bi_end_io = raid1_end_read_request;
 559                 read_bio->bi_rw = READ;
 560                 read_bio->bi_private = r1_bio;
 561
 562                 generic_make_request(read_bio);
 563                 return 0;
 564         }
 565
 566         /*
 567          * WRITE:
 568          */
 569         /* first select target devices under spinlock and
 570          * inc refcount on their rdev.  Record them by setting
 571          * bios[x] to bio
 572          */
 573         spin_lock_irq(&conf->device_lock);
 574         for (i = 0;  i < disks; i++) {
 575                 if (conf->mirrors[i].rdev &&
 576                     !conf->mirrors[i].rdev->faulty) {
 577                         atomic_inc(&conf->mirrors[i].rdev->nr_pending);
 578                         r1_bio->bios[i] = bio;
 579                 } else
 580                         r1_bio->bios[i] = NULL;
 581         }
 582         spin_unlock_irq(&conf->device_lock);
 583
 584         atomic_set(&r1_bio->remaining, 1);
 585         md_write_start(mddev);
 586         for (i = 0; i < disks; i++) {
 587                 struct bio *mbio;
 588                 if (!r1_bio->bios[i])
 589                         continue;
 590
 591                 mbio = bio_clone(bio, GFP_NOIO);
 592                 r1_bio->bios[i] = mbio;
 593
 594                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 595                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 596                 mbio->bi_end_io = raid1_end_write_request;
 597                 mbio->bi_rw = WRITE;
 598                 mbio->bi_private = r1_bio;
 599
 600                 atomic_inc(&r1_bio->remaining);
 601                 generic_make_request(mbio);
 602         }
 603
 604         if (atomic_dec_and_test(&r1_bio->remaining)) {
 605                 md_write_end(mddev);
 606                 raid_end_bio_io(r1_bio);
 607         }
 608
 609         return 0;
 610 }
 611
 612 static void status(struct seq_file *seq, mddev_t *mddev)
 613 {
 614         conf_t *conf = mddev_to_conf(mddev);
 615         int i;
 616
 617         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 618                                                 conf->working_disks);
 619         for (i = 0; i < conf->raid_disks; i++)
 620                 seq_printf(seq, "%s",
 621                               conf->mirrors[i].rdev &&
 622                               conf->mirrors[i].rdev->in_sync ? "U" : "_");
 623         seq_printf(seq, "]");
 624 }
 625
 626
 627 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 628 {
 629         char b[BDEVNAME_SIZE];
 630         conf_t *conf = mddev_to_conf(mddev);
 631
 632         /*
 633          * If it is not operational, then we have already marked it as dead
 634          * else if it is the last working disks, ignore the error, let the
 635          * next level up know.
 636          * else mark the drive as failed
 637          */
 638         if (rdev->in_sync
 639             && conf->working_disks == 1)
 640                 /*
 641                  * Don't fail the drive, act as though we were just a
 642                  * normal single drive
 643                  */
 644                 return;
 645         if (rdev->in_sync) {
 646                 mddev->degraded++;
 647                 conf->working_disks--;
 648                 /*
 649                  * if recovery is running, make sure it aborts.
 650                  */
 651                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 652         }
 653         rdev->in_sync = 0;
 654         rdev->faulty = 1;
 655         mddev->sb_dirty = 1;
 656         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device. \n"
 657                 "       Operation continuing on %d devices\n",
 658                 bdevname(rdev->bdev,b), conf->working_disks);
 659 }
 660
 661 static void print_conf(conf_t *conf)
 662 {
 663         int i;
 664         mirror_info_t *tmp;
 665
 666         printk("RAID1 conf printout:\n");
 667         if (!conf) {
 668                 printk("(!conf)\n");
 669                 return;
 670         }
 671         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 672                 conf->raid_disks);
 673
 674         for (i = 0; i < conf->raid_disks; i++) {
 675                 char b[BDEVNAME_SIZE];
 676                 tmp = conf->mirrors + i;
 677                 if (tmp->rdev)
 678                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
 679                                 i, !tmp->rdev->in_sync, !tmp->rdev->faulty,
 680                                 bdevname(tmp->rdev->bdev,b));
 681         }
 682 }
 683
 684 static void close_sync(conf_t *conf)
 685 {
 686         spin_lock_irq(&conf->resync_lock);
 687         wait_event_lock_irq(conf->wait_resume, !conf->barrier,
 688                             conf->resync_lock,  unplug_slaves(conf->mddev));
 689         spin_unlock_irq(&conf->resync_lock);
 690
 691         if (conf->barrier) BUG();
 692         if (waitqueue_active(&conf->wait_idle)) BUG();
 693
 694         mempool_destroy(conf->r1buf_pool);
 695         conf->r1buf_pool = NULL;
 696 }
 697
 698 static int raid1_spare_active(mddev_t *mddev)
 699 {
 700         int i;
 701         conf_t *conf = mddev->private;
 702         mirror_info_t *tmp;
 703
 704         spin_lock_irq(&conf->device_lock);
 705         /*
 706          * Find all failed disks within the RAID1 configuration
 707          * and mark them readable
 708          */
 709         for (i = 0; i < conf->raid_disks; i++) {
 710                 tmp = conf->mirrors + i;
 711                 if (tmp->rdev
 712                     && !tmp->rdev->faulty
 713                     && !tmp->rdev->in_sync) {
 714                         conf->working_disks++;
 715                         mddev->degraded--;
 716                         tmp->rdev->in_sync = 1;
 717                 }
 718         }
 719         spin_unlock_irq(&conf->device_lock);
 720
 721         print_conf(conf);
 722         return 0;
 723 }
 724
 725
 726 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 727 {
 728         conf_t *conf = mddev->private;
 729         int found = 0;
 730         int mirror;
 731         mirror_info_t *p;
 732
 733         spin_lock_irq(&conf->device_lock);
 734         for (mirror=0; mirror < mddev->raid_disks; mirror++)
 735                 if ( !(p=conf->mirrors+mirror)->rdev) {
 736                         p->rdev = rdev;
 737
 738                         blk_queue_stack_limits(mddev->queue,
 739                                                rdev->bdev->bd_disk->queue);
 740                         /* as we don't honour merge_bvec_fn, we must never risk
 741                          * violating it, so limit ->max_sector to one PAGE, as
 742                          * a one page request is never in violation.
 743                          */
 744                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
 745                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
 746                                 mddev->queue->max_sectors = (PAGE_SIZE>>9);
 747
 748                         p->head_position = 0;
 749                         rdev->raid_disk = mirror;
 750                         found = 1;
 751                         break;
 752                 }
 753         spin_unlock_irq(&conf->device_lock);
 754
 755         print_conf(conf);
 756         return found;
 757 }
 758
 759 static int raid1_remove_disk(mddev_t *mddev, int number)
 760 {
 761         conf_t *conf = mddev->private;
 762         int err = 1;
 763         mirror_info_t *p = conf->mirrors+ number;
 764
 765         print_conf(conf);
 766         spin_lock_irq(&conf->device_lock);
 767         if (p->rdev) {
 768                 if (p->rdev->in_sync ||
 769                     atomic_read(&p->rdev->nr_pending)) {
 770                         err = -EBUSY;
 771                         goto abort;
 772                 }
 773                 p->rdev = NULL;
 774                 err = 0;
 775         }
 776         if (err)
 777                 MD_BUG();
 778 abort:
 779         spin_unlock_irq(&conf->device_lock);
 780
 781         print_conf(conf);
 782         return err;
 783 }
 784
 785
 786 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
 787 {
 788         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 789         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 790         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 791
 792         if (bio->bi_size)
 793                 return 1;
 794
 795         if (r1_bio->bios[r1_bio->read_disk] != bio)
 796                 BUG();
 797         update_head_pos(r1_bio->read_disk, r1_bio);
 798         /*
 799          * we have read a block, now it needs to be re-written,
 800          * or re-read if the read failed.
 801          * We don't do much here, just schedule handling by raid1d
 802          */
 803         if (!uptodate)
 804                 md_error(r1_bio->mddev,
 805                          conf->mirrors[r1_bio->read_disk].rdev);
 806         else
 807                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 808         atomic_dec(&conf->mirrors[r1_bio->read_disk].rdev->nr_pending);
 809         reschedule_retry(r1_bio);
 810         return 0;
 811 }
 812
 813 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
 814 {
 815         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 816         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 817         mddev_t *mddev = r1_bio->mddev;
 818         conf_t *conf = mddev_to_conf(mddev);
 819         int i;
 820         int mirror=0;
 821
 822         if (bio->bi_size)
 823                 return 1;
 824
 825         for (i = 0; i < conf->raid_disks; i++)
 826                 if (r1_bio->bios[i] == bio) {
 827                         mirror = i;
 828                         break;
 829                 }
 830         if (!uptodate)
 831                 md_error(mddev, conf->mirrors[mirror].rdev);
 832         update_head_pos(mirror, r1_bio);
 833
 834         if (atomic_dec_and_test(&r1_bio->remaining)) {
 835                 md_done_sync(mddev, r1_bio->sectors, uptodate);
 836                 put_buf(r1_bio);
 837         }
 838         atomic_dec(&conf->mirrors[mirror].rdev->nr_pending);
 839         return 0;
 840 }
 841
 842 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
 843 {
 844         conf_t *conf = mddev_to_conf(mddev);
 845         int i;
 846         int disks = conf->raid_disks;
 847         struct bio *bio, *wbio;
 848
 849         bio = r1_bio->bios[r1_bio->read_disk];
 850
 851         /*
 852          * schedule writes
 853          */
 854         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 855                 /*
 856                  * There is no point trying a read-for-reconstruct as
 857                  * reconstruct is about to be aborted
 858                  */
 859                 char b[BDEVNAME_SIZE];
 860                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
 861                         " for block %llu\n",
 862                         bdevname(bio->bi_bdev,b),
 863                         (unsigned long long)r1_bio->sector);
 864                 md_done_sync(mddev, r1_bio->sectors, 0);
 865                 put_buf(r1_bio);
 866                 return;
 867         }
 868
 869         atomic_set(&r1_bio->remaining, 1);
 870         for (i = 0; i < disks ; i++) {
 871                 wbio = r1_bio->bios[i];
 872                 if (wbio->bi_end_io != end_sync_write)
 873                         continue;
 874
 875                 atomic_inc(&conf->mirrors[i].rdev->nr_pending);
 876                 atomic_inc(&r1_bio->remaining);
 877                 md_sync_acct(conf->mirrors[i].rdev, wbio->bi_size >> 9);
 878                 generic_make_request(wbio);
 879         }
 880
 881         if (atomic_dec_and_test(&r1_bio->remaining)) {
 882                 md_done_sync(mddev, r1_bio->sectors, 1);
 883                 put_buf(r1_bio);
 884         }
 885 }
 886
 887 /*
 888  * This is a kernel thread which:
 889  *
 890  *      1.      Retries failed read operations on working mirrors.
 891  *      2.      Updates the raid superblock when problems encounter.
 892  *      3.      Performs writes following reads for array syncronising.
 893  */
 894
 895 static void raid1d(mddev_t *mddev)
 896 {
 897         struct list_head *head = &retry_list_head;
 898         r1bio_t *r1_bio;
 899         struct bio *bio;
 900         unsigned long flags;
 901         conf_t *conf = mddev_to_conf(mddev);
 902         int unplug=0;
 903         mdk_rdev_t *rdev;
 904
 905         md_check_recovery(mddev);
 906         md_handle_safemode(mddev);
 907
 908         for (;;) {
 909                 char b[BDEVNAME_SIZE];
 910                 spin_lock_irqsave(&retry_list_lock, flags);
 911                 if (list_empty(head))
 912                         break;
 913                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
 914                 list_del(head->prev);
 915                 spin_unlock_irqrestore(&retry_list_lock, flags);
 916
 917                 mddev = r1_bio->mddev;
 918                 conf = mddev_to_conf(mddev);
 919                 bio = r1_bio->master_bio;
 920                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
 921                         sync_request_write(mddev, r1_bio);
 922                         unplug = 1;
 923                 } else {
 924                         if (map(mddev, &rdev) == -1) {
 925                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
 926                                        " read error for block %llu\n",
 927                                        bdevname(bio->bi_bdev,b),
 928                                        (unsigned long long)r1_bio->sector);
 929                                 raid_end_bio_io(r1_bio);
 930                         } else {
 931                                 printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
 932                                        " another mirror\n",
 933                                        bdevname(rdev->bdev,b),
 934                                        (unsigned long long)r1_bio->sector);
 935                                 bio->bi_bdev = rdev->bdev;
 936                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
 937                                 bio->bi_rw = READ;
 938                                 unplug = 1;
 939                                 generic_make_request(bio);
 940                         }
 941                 }
 942         }
 943         spin_unlock_irqrestore(&retry_list_lock, flags);
 944         if (unplug)
 945                 unplug_slaves(mddev);
 946 }
 947
 948
 949 static int init_resync(conf_t *conf)
 950 {
 951         int buffs;
 952
 953         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
 954         if (conf->r1buf_pool)
 955                 BUG();
 956         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free, conf);
 957         if (!conf->r1buf_pool)
 958                 return -ENOMEM;
 959         conf->next_resync = 0;
 960         return 0;
 961 }
 962
 963 /*
 964  * perform a "sync" on one "block"
 965  *
 966  * We need to make sure that no normal I/O request - particularly write
 967  * requests - conflict with active sync requests.
 968  *
 969  * This is achieved by tracking pending requests and a 'barrier' concept
 970  * that can be installed to exclude normal IO requests.
 971  */
 972
 973 static int sync_request(mddev_t *mddev, sector_t sector_nr, int go_faster)
 974 {
 975         conf_t *conf = mddev_to_conf(mddev);
 976         mirror_info_t *mirror;
 977         r1bio_t *r1_bio;
 978         struct bio *bio;
 979         sector_t max_sector, nr_sectors;
 980         int disk;
 981         int i;
 982
 983         if (!conf->r1buf_pool)
 984                 if (init_resync(conf))
 985                         return -ENOMEM;
 986
 987         max_sector = mddev->size << 1;
 988         if (sector_nr >= max_sector) {
 989                 close_sync(conf);
 990                 return 0;
 991         }
 992
 993         /*
 994          * If there is non-resync activity waiting for us then
 995          * put in a delay to throttle resync.
 996          */
 997         if (!go_faster && waitqueue_active(&conf->wait_resume))
 998                 schedule_timeout(HZ);
 999         device_barrier(conf, sector_nr + RESYNC_SECTORS);
1000
1001         /*
1002          * If reconstructing, and >1 working disc,
1003          * could dedicate one to rebuild and others to
1004          * service read requests ..
1005          */
1006         disk = conf->last_used;
1007         /* make sure disk is operational */
1008         spin_lock_irq(&conf->device_lock);
1009         while (conf->mirrors[disk].rdev == NULL ||
1010                !conf->mirrors[disk].rdev->in_sync) {
1011                 if (disk <= 0)
1012                         disk = conf->raid_disks;
1013                 disk--;
1014                 if (disk == conf->last_used)
1015                         break;
1016         }
1017         conf->last_used = disk;
1018         atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
1019         spin_unlock_irq(&conf->device_lock);
1020
1021         mirror = conf->mirrors + disk;
1022
1023         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1024
1025         spin_lock_irq(&conf->resync_lock);
1026         conf->nr_pending++;
1027         spin_unlock_irq(&conf->resync_lock);
1028
1029         r1_bio->mddev = mddev;
1030         r1_bio->sector = sector_nr;
1031         set_bit(R1BIO_IsSync, &r1_bio->state);
1032         r1_bio->read_disk = disk;
1033
1034         for (i=0; i < conf->raid_disks; i++) {
1035                 bio = r1_bio->bios[i];
1036
1037                 /* take from bio_init */
1038                 bio->bi_next = NULL;
1039                 bio->bi_flags |= 1 << BIO_UPTODATE;
1040                 bio->bi_rw = 0;
1041                 bio->bi_vcnt = 0;
1042                 bio->bi_idx = 0;
1043                 bio->bi_phys_segments = 0;
1044                 bio->bi_hw_segments = 0;
1045                 bio->bi_size = 0;
1046                 bio->bi_end_io = NULL;
1047                 bio->bi_private = NULL;
1048
1049                 if (i == disk) {
1050                         bio->bi_rw = READ;
1051                         bio->bi_end_io = end_sync_read;
1052                 } else if (conf->mirrors[i].rdev &&
1053                            !conf->mirrors[i].rdev->faulty &&
1054                            (!conf->mirrors[i].rdev->in_sync ||
1055                             sector_nr + RESYNC_SECTORS > mddev->recovery_cp)) {
1056                         bio->bi_rw = WRITE;
1057                         bio->bi_end_io = end_sync_write;
1058                 } else
1059                         continue;
1060                 bio->bi_sector = sector_nr + conf->mirrors[i].rdev->data_offset;
1061                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1062                 bio->bi_private = r1_bio;
1063         }
1064         nr_sectors = 0;
1065         do {
1066                 struct page *page;
1067                 int len = PAGE_SIZE;
1068                 if (sector_nr + (len>>9) > max_sector)
1069                         len = (max_sector - sector_nr) << 9;
1070                 if (len == 0)
1071                         break;
1072                 for (i=0 ; i < conf->raid_disks; i++) {
1073                         bio = r1_bio->bios[i];
1074                         if (bio->bi_end_io) {
1075                                 page = r1_bio->bios[0]->bi_io_vec[bio->bi_vcnt].bv_page;
1076                                 if (bio_add_page(bio, page, len, 0) == 0) {
1077                                         /* stop here */
1078                                         r1_bio->bios[0]->bi_io_vec[bio->bi_vcnt].bv_page = page;
1079                                         while (i > 0) {
1080                                                 i--;
1081                                                 bio = r1_bio->bios[i];
1082                                                 if (bio->bi_end_io==NULL) continue;
1083                                                 /* remove last page from this bio */
1084                                                 bio->bi_vcnt--;
1085                                                 bio->bi_size -= len;
1086                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1087                                         }
1088                                         goto bio_full;
1089                                 }
1090                         }
1091                 }
1092                 nr_sectors += len>>9;
1093                 sector_nr += len>>9;
1094         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1095  bio_full:
1096         bio = r1_bio->bios[disk];
1097         r1_bio->sectors = nr_sectors;
1098
1099         md_sync_acct(mirror->rdev, nr_sectors);
1100
1101         generic_make_request(bio);
1102
1103         return nr_sectors;
1104 }
1105
1106 static int run(mddev_t *mddev)
1107 {
1108         conf_t *conf;
1109         int i, j, disk_idx;
1110         mirror_info_t *disk;
1111         mdk_rdev_t *rdev;
1112         struct list_head *tmp;
1113
1114         if (mddev->level != 1) {
1115                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1116                        mdname(mddev), mddev->level);
1117                 goto out;
1118         }
1119         /*
1120          * copy the already verified devices into our private RAID1
1121          * bookkeeping area. [whatever we allocate in run(),
1122          * should be freed in stop()]
1123          */
1124         conf = kmalloc(sizeof(conf_t), GFP_KERNEL);
1125         mddev->private = conf;
1126         if (!conf) {
1127                 printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1128                         mdname(mddev));
1129                 goto out;
1130         }
1131         memset(conf, 0, sizeof(*conf));
1132         conf->mirrors = kmalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1133                                  GFP_KERNEL);
1134         if (!conf->mirrors) {
1135                 printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1136                        mdname(mddev));
1137                 goto out_free_conf;
1138         }
1139         memset(conf->mirrors, 0, sizeof(struct mirror_info)*mddev->raid_disks);
1140
1141         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1142                                                 r1bio_pool_free, mddev);
1143         if (!conf->r1bio_pool) {
1144                 printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1145                         mdname(mddev));
1146                 goto out_free_conf;
1147         }
1148         mddev->queue->unplug_fn = raid1_unplug;
1149
1150
1151         ITERATE_RDEV(mddev, rdev, tmp) {
1152                 disk_idx = rdev->raid_disk;
1153                 if (disk_idx >= mddev->raid_disks
1154                     || disk_idx < 0)
1155                         continue;
1156                 disk = conf->mirrors + disk_idx;
1157
1158                 disk->rdev = rdev;
1159
1160                 blk_queue_stack_limits(mddev->queue,
1161                                        rdev->bdev->bd_disk->queue);
1162                 /* as we don't honour merge_bvec_fn, we must never risk
1163                  * violating it, so limit ->max_sector to one PAGE, as
1164                  * a one page request is never in violation.
1165                  */
1166                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1167                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1168                         mddev->queue->max_sectors = (PAGE_SIZE>>9);
1169
1170                 disk->head_position = 0;
1171                 if (!rdev->faulty && rdev->in_sync)
1172                         conf->working_disks++;
1173         }
1174         conf->raid_disks = mddev->raid_disks;
1175         conf->mddev = mddev;
1176         conf->device_lock = SPIN_LOCK_UNLOCKED;
1177         if (conf->working_disks == 1)
1178                 mddev->recovery_cp = MaxSector;
1179
1180         conf->resync_lock = SPIN_LOCK_UNLOCKED;
1181         init_waitqueue_head(&conf->wait_idle);
1182         init_waitqueue_head(&conf->wait_resume);
1183
1184         if (!conf->working_disks) {
1185                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
1186                         mdname(mddev));
1187                 goto out_free_conf;
1188         }
1189
1190         mddev->degraded = 0;
1191         for (i = 0; i < conf->raid_disks; i++) {
1192
1193                 disk = conf->mirrors + i;
1194
1195                 if (!disk->rdev) {
1196                         disk->head_position = 0;
1197                         mddev->degraded++;
1198                 }
1199         }
1200
1201         /*
1202          * find the first working one and use it as a starting point
1203          * to read balancing.
1204          */
1205         for (j = 0; j < conf->raid_disks &&
1206                      (!conf->mirrors[j].rdev ||
1207                       !conf->mirrors[j].rdev->in_sync) ; j++)
1208                 /* nothing */;
1209         conf->last_used = j;
1210
1211
1212
1213         {
1214                 mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
1215                 if (!mddev->thread) {
1216                         printk(KERN_ERR
1217                                 "raid1: couldn't allocate thread for %s\n",
1218                                 mdname(mddev));
1219                         goto out_free_conf;
1220                 }
1221         }
1222         printk(KERN_INFO
1223                 "raid1: raid set %s active with %d out of %d mirrors\n",
1224                 mdname(mddev), mddev->raid_disks - mddev->degraded,
1225                 mddev->raid_disks);
1226         /*
1227          * Ok, everything is just fine now
1228          */
1229         mddev->array_size = mddev->size;
1230
1231         return 0;
1232
1233 out_free_conf:
1234         if (conf->r1bio_pool)
1235                 mempool_destroy(conf->r1bio_pool);
1236         if (conf->mirrors)
1237                 kfree(conf->mirrors);
1238         kfree(conf);
1239         mddev->private = NULL;
1240 out:
1241         return -EIO;
1242 }
1243
1244 static int stop(mddev_t *mddev)
1245 {
1246         conf_t *conf = mddev_to_conf(mddev);
1247
1248         md_unregister_thread(mddev->thread);
1249         mddev->thread = NULL;
1250         if (conf->r1bio_pool)
1251                 mempool_destroy(conf->r1bio_pool);
1252         if (conf->mirrors)
1253                 kfree(conf->mirrors);
1254         kfree(conf);
1255         mddev->private = NULL;
1256         return 0;
1257 }
1258
1259 static mdk_personality_t raid1_personality =
1260 {
1261         .name           = "raid1",
1262         .owner          = THIS_MODULE,
1263         .make_request   = make_request,
1264         .run            = run,
1265         .stop           = stop,
1266         .status         = status,
1267         .error_handler  = error,
1268         .hot_add_disk   = raid1_add_disk,
1269         .hot_remove_disk= raid1_remove_disk,
1270         .spare_active   = raid1_spare_active,
1271         .sync_request   = sync_request,
1272 };
1273
1274 static int __init raid_init(void)
1275 {
1276         return register_md_personality(RAID1, &raid1_personality);
1277 }
1278
1279 static void raid_exit(void)
1280 {
1281         unregister_md_personality(RAID1);
1282 }
1283
1284 module_init(raid_init);
1285 module_exit(raid_exit);
1286 MODULE_LICENSE("GPL");
1287 MODULE_ALIAS("md-personality-3"); /* RAID1 */