drivers/md/multipath.c

   1 /*
   2  * multipath.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  * MULTIPATH management functions.
   9  *
  10  * derived from raid1.c.
  11  *
  12  * This program is free software; you can redistribute it and/or modify
  13  * it under the terms of the GNU General Public License as published by
  14  * the Free Software Foundation; either version 2, or (at your option)
  15  * any later version.
  16  *
  17  * You should have received a copy of the GNU General Public License
  18  * (for example /usr/src/linux/COPYING); if not, write to the Free
  19  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #include <linux/module.h>
  23 #include <linux/slab.h>
  24 #include <linux/spinlock.h>
  25 #include <linux/raid/multipath.h>
  26 #include <linux/buffer_head.h>
  27 #include <asm/atomic.h>
  28
  29 #define MAJOR_NR MD_MAJOR
  30 #define MD_DRIVER
  31 #define MD_PERSONALITY
  32
  33 #define MAX_WORK_PER_DISK 128
  34
  35 #define NR_RESERVED_BUFS        32
  36
  37
  38 static mdk_personality_t multipath_personality;
  39
  40
  41 static void *mp_pool_alloc(gfp_t gfp_flags, void *data)
  42 {
  43         struct multipath_bh *mpb;
  44         mpb = kmalloc(sizeof(*mpb), gfp_flags);
  45         if (mpb)
  46                 memset(mpb, 0, sizeof(*mpb));
  47         return mpb;
  48 }
  49
  50 static void mp_pool_free(void *mpb, void *data)
  51 {
  52         kfree(mpb);
  53 }
  54
  55 static int multipath_map (multipath_conf_t *conf)
  56 {
  57         int i, disks = conf->raid_disks;
  58
  59         /*
  60          * Later we do read balancing on the read side
  61          * now we use the first available disk.
  62          */
  63
  64         rcu_read_lock();
  65         for (i = 0; i < disks; i++) {
  66                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
  67                 if (rdev && test_bit(In_sync, &rdev->flags)) {
  68                         atomic_inc(&rdev->nr_pending);
  69                         rcu_read_unlock();
  70                         return i;
  71                 }
  72         }
  73         rcu_read_unlock();
  74
  75         printk(KERN_ERR "multipath_map(): no more operational IO paths?\n");
  76         return (-1);
  77 }
  78
  79 static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
  80 {
  81         unsigned long flags;
  82         mddev_t *mddev = mp_bh->mddev;
  83         multipath_conf_t *conf = mddev_to_conf(mddev);
  84
  85         spin_lock_irqsave(&conf->device_lock, flags);
  86         list_add(&mp_bh->retry_list, &conf->retry_list);
  87         spin_unlock_irqrestore(&conf->device_lock, flags);
  88         md_wakeup_thread(mddev->thread);
  89 }
  90
  91
  92 /*
  93  * multipath_end_bh_io() is called when we have finished servicing a multipathed
  94  * operation and are ready to return a success/failure code to the buffer
  95  * cache layer.
  96  */
  97 static void multipath_end_bh_io (struct multipath_bh *mp_bh, int err)
  98 {
  99         struct bio *bio = mp_bh->master_bio;
 100         multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
 101
 102         bio_endio(bio, bio->bi_size, err);
 103         mempool_free(mp_bh, conf->pool);
 104 }
 105
 106 static int multipath_end_request(struct bio *bio, unsigned int bytes_done,
 107                                  int error)
 108 {
 109         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 110         struct multipath_bh * mp_bh = (struct multipath_bh *)(bio->bi_private);
 111         multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
 112         mdk_rdev_t *rdev = conf->multipaths[mp_bh->path].rdev;
 113
 114         if (bio->bi_size)
 115                 return 1;
 116
 117         if (uptodate)
 118                 multipath_end_bh_io(mp_bh, 0);
 119         else if (!bio_rw_ahead(bio)) {
 120                 /*
 121                  * oops, IO error:
 122                  */
 123                 char b[BDEVNAME_SIZE];
 124                 md_error (mp_bh->mddev, rdev);
 125                 printk(KERN_ERR "multipath: %s: rescheduling sector %llu\n",
 126                        bdevname(rdev->bdev,b),
 127                        (unsigned long long)bio->bi_sector);
 128                 multipath_reschedule_retry(mp_bh);
 129         } else
 130                 multipath_end_bh_io(mp_bh, error);
 131         rdev_dec_pending(rdev, conf->mddev);
 132         return 0;
 133 }
 134
 135 static void unplug_slaves(mddev_t *mddev)
 136 {
 137         multipath_conf_t *conf = mddev_to_conf(mddev);
 138         int i;
 139
 140         rcu_read_lock();
 141         for (i=0; i<mddev->raid_disks; i++) {
 142                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
 143                 if (rdev && !test_bit(Faulty, &rdev->flags)
 144                     && atomic_read(&rdev->nr_pending)) {
 145                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
 146
 147                         atomic_inc(&rdev->nr_pending);
 148                         rcu_read_unlock();
 149
 150                         if (r_queue->unplug_fn)
 151                                 r_queue->unplug_fn(r_queue);
 152
 153                         rdev_dec_pending(rdev, mddev);
 154                         rcu_read_lock();
 155                 }
 156         }
 157         rcu_read_unlock();
 158 }
 159
 160 static void multipath_unplug(request_queue_t *q)
 161 {
 162         unplug_slaves(q->queuedata);
 163 }
 164
 165
 166 static int multipath_make_request (request_queue_t *q, struct bio * bio)
 167 {
 168         mddev_t *mddev = q->queuedata;
 169         multipath_conf_t *conf = mddev_to_conf(mddev);
 170         struct multipath_bh * mp_bh;
 171         struct multipath_info *multipath;
 172         const int rw = bio_data_dir(bio);
 173
 174         if (unlikely(bio_barrier(bio))) {
 175                 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
 176                 return 0;
 177         }
 178
 179         mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
 180
 181         mp_bh->master_bio = bio;
 182         mp_bh->mddev = mddev;
 183
 184         disk_stat_inc(mddev->gendisk, ios[rw]);
 185         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
 186
 187         mp_bh->path = multipath_map(conf);
 188         if (mp_bh->path < 0) {
 189                 bio_endio(bio, bio->bi_size, -EIO);
 190                 mempool_free(mp_bh, conf->pool);
 191                 return 0;
 192         }
 193         multipath = conf->multipaths + mp_bh->path;
 194
 195         mp_bh->bio = *bio;
 196         mp_bh->bio.bi_sector += multipath->rdev->data_offset;
 197         mp_bh->bio.bi_bdev = multipath->rdev->bdev;
 198         mp_bh->bio.bi_rw |= (1 << BIO_RW_FAILFAST);
 199         mp_bh->bio.bi_end_io = multipath_end_request;
 200         mp_bh->bio.bi_private = mp_bh;
 201         generic_make_request(&mp_bh->bio);
 202         return 0;
 203 }
 204
 205 static void multipath_status (struct seq_file *seq, mddev_t *mddev)
 206 {
 207         multipath_conf_t *conf = mddev_to_conf(mddev);
 208         int i;
 209
 210         seq_printf (seq, " [%d/%d] [", conf->raid_disks,
 211                                                  conf->working_disks);
 212         for (i = 0; i < conf->raid_disks; i++)
 213                 seq_printf (seq, "%s",
 214                                conf->multipaths[i].rdev &&
 215                                test_bit(In_sync, &conf->multipaths[i].rdev->flags) ? "U" : "_");
 216         seq_printf (seq, "]");
 217 }
 218
 219 static int multipath_issue_flush(request_queue_t *q, struct gendisk *disk,
 220                                  sector_t *error_sector)
 221 {
 222         mddev_t *mddev = q->queuedata;
 223         multipath_conf_t *conf = mddev_to_conf(mddev);
 224         int i, ret = 0;
 225
 226         rcu_read_lock();
 227         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
 228                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
 229                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 230                         struct block_device *bdev = rdev->bdev;
 231                         request_queue_t *r_queue = bdev_get_queue(bdev);
 232
 233                         if (!r_queue->issue_flush_fn)
 234                                 ret = -EOPNOTSUPP;
 235                         else {
 236                                 atomic_inc(&rdev->nr_pending);
 237                                 rcu_read_unlock();
 238                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
 239                                                               error_sector);
 240                                 rdev_dec_pending(rdev, mddev);
 241                                 rcu_read_lock();
 242                         }
 243                 }
 244         }
 245         rcu_read_unlock();
 246         return ret;
 247 }
 248
 249 /*
 250  * Careful, this can execute in IRQ contexts as well!
 251  */
 252 static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
 253 {
 254         multipath_conf_t *conf = mddev_to_conf(mddev);
 255
 256         if (conf->working_disks <= 1) {
 257                 /*
 258                  * Uh oh, we can do nothing if this is our last path, but
 259                  * first check if this is a queued request for a device
 260                  * which has just failed.
 261                  */
 262                 printk(KERN_ALERT
 263                         "multipath: only one IO path left and IO error.\n");
 264                 /* leave it active... it's all we have */
 265         } else {
 266                 /*
 267                  * Mark disk as unusable
 268                  */
 269                 if (!test_bit(Faulty, &rdev->flags)) {
 270                         char b[BDEVNAME_SIZE];
 271                         clear_bit(In_sync, &rdev->flags);
 272                         set_bit(Faulty, &rdev->flags);
 273                         mddev->sb_dirty = 1;
 274                         conf->working_disks--;
 275                         printk(KERN_ALERT "multipath: IO failure on %s,"
 276                                 " disabling IO path. \n Operation continuing"
 277                                 " on %d IO paths.\n",
 278                                 bdevname (rdev->bdev,b),
 279                                 conf->working_disks);
 280                 }
 281         }
 282 }
 283
 284 static void print_multipath_conf (multipath_conf_t *conf)
 285 {
 286         int i;
 287         struct multipath_info *tmp;
 288
 289         printk("MULTIPATH conf printout:\n");
 290         if (!conf) {
 291                 printk("(conf==NULL)\n");
 292                 return;
 293         }
 294         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 295                          conf->raid_disks);
 296
 297         for (i = 0; i < conf->raid_disks; i++) {
 298                 char b[BDEVNAME_SIZE];
 299                 tmp = conf->multipaths + i;
 300                 if (tmp->rdev)
 301                         printk(" disk%d, o:%d, dev:%s\n",
 302                                 i,!test_bit(Faulty, &tmp->rdev->flags),
 303                                bdevname(tmp->rdev->bdev,b));
 304         }
 305 }
 306
 307
 308 static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 309 {
 310         multipath_conf_t *conf = mddev->private;
 311         int found = 0;
 312         int path;
 313         struct multipath_info *p;
 314
 315         print_multipath_conf(conf);
 316
 317         for (path=0; path<mddev->raid_disks; path++)
 318                 if ((p=conf->multipaths+path)->rdev == NULL) {
 319                         blk_queue_stack_limits(mddev->queue,
 320                                                rdev->bdev->bd_disk->queue);
 321
 322                 /* as we don't honour merge_bvec_fn, we must never risk
 323                  * violating it, so limit ->max_sector to one PAGE, as
 324                  * a one page request is never in violation.
 325                  * (Note: it is very unlikely that a device with
 326                  * merge_bvec_fn will be involved in multipath.)
 327                  */
 328                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
 329                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
 330                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
 331
 332                         conf->working_disks++;
 333                         rdev->raid_disk = path;
 334                         set_bit(In_sync, &rdev->flags);
 335                         rcu_assign_pointer(p->rdev, rdev);
 336                         found = 1;
 337                 }
 338
 339         print_multipath_conf(conf);
 340         return found;
 341 }
 342
 343 static int multipath_remove_disk(mddev_t *mddev, int number)
 344 {
 345         multipath_conf_t *conf = mddev->private;
 346         int err = 0;
 347         mdk_rdev_t *rdev;
 348         struct multipath_info *p = conf->multipaths + number;
 349
 350         print_multipath_conf(conf);
 351
 352         rdev = p->rdev;
 353         if (rdev) {
 354                 if (test_bit(In_sync, &rdev->flags) ||
 355                     atomic_read(&rdev->nr_pending)) {
 356                         printk(KERN_ERR "hot-remove-disk, slot %d is identified"                                " but is still operational!\n", number);
 357                         err = -EBUSY;
 358                         goto abort;
 359                 }
 360                 p->rdev = NULL;
 361                 synchronize_rcu();
 362                 if (atomic_read(&rdev->nr_pending)) {
 363                         /* lost the race, try later */
 364                         err = -EBUSY;
 365                         p->rdev = rdev;
 366                 }
 367         }
 368 abort:
 369
 370         print_multipath_conf(conf);
 371         return err;
 372 }
 373
 374
 375
 376 /*
 377  * This is a kernel thread which:
 378  *
 379  *      1.      Retries failed read operations on working multipaths.
 380  *      2.      Updates the raid superblock when problems encounter.
 381  *      3.      Performs writes following reads for array syncronising.
 382  */
 383
 384 static void multipathd (mddev_t *mddev)
 385 {
 386         struct multipath_bh *mp_bh;
 387         struct bio *bio;
 388         unsigned long flags;
 389         multipath_conf_t *conf = mddev_to_conf(mddev);
 390         struct list_head *head = &conf->retry_list;
 391
 392         md_check_recovery(mddev);
 393         for (;;) {
 394                 char b[BDEVNAME_SIZE];
 395                 spin_lock_irqsave(&conf->device_lock, flags);
 396                 if (list_empty(head))
 397                         break;
 398                 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
 399                 list_del(head->prev);
 400                 spin_unlock_irqrestore(&conf->device_lock, flags);
 401
 402                 bio = &mp_bh->bio;
 403                 bio->bi_sector = mp_bh->master_bio->bi_sector;
 404
 405                 if ((mp_bh->path = multipath_map (conf))<0) {
 406                         printk(KERN_ALERT "multipath: %s: unrecoverable IO read"
 407                                 " error for block %llu\n",
 408                                 bdevname(bio->bi_bdev,b),
 409                                 (unsigned long long)bio->bi_sector);
 410                         multipath_end_bh_io(mp_bh, -EIO);
 411                 } else {
 412                         printk(KERN_ERR "multipath: %s: redirecting sector %llu"
 413                                 " to another IO path\n",
 414                                 bdevname(bio->bi_bdev,b),
 415                                 (unsigned long long)bio->bi_sector);
 416                         *bio = *(mp_bh->master_bio);
 417                         bio->bi_sector += conf->multipaths[mp_bh->path].rdev->data_offset;
 418                         bio->bi_bdev = conf->multipaths[mp_bh->path].rdev->bdev;
 419                         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
 420                         bio->bi_end_io = multipath_end_request;
 421                         bio->bi_private = mp_bh;
 422                         generic_make_request(bio);
 423                 }
 424         }
 425         spin_unlock_irqrestore(&conf->device_lock, flags);
 426 }
 427
 428 static int multipath_run (mddev_t *mddev)
 429 {
 430         multipath_conf_t *conf;
 431         int disk_idx;
 432         struct multipath_info *disk;
 433         mdk_rdev_t *rdev;
 434         struct list_head *tmp;
 435
 436         if (mddev->level != LEVEL_MULTIPATH) {
 437                 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
 438                        mdname(mddev), mddev->level);
 439                 goto out;
 440         }
 441         /*
 442          * copy the already verified devices into our private MULTIPATH
 443          * bookkeeping area. [whatever we allocate in multipath_run(),
 444          * should be freed in multipath_stop()]
 445          */
 446
 447         conf = kmalloc(sizeof(multipath_conf_t), GFP_KERNEL);
 448         mddev->private = conf;
 449         if (!conf) {
 450                 printk(KERN_ERR
 451                         "multipath: couldn't allocate memory for %s\n",
 452                         mdname(mddev));
 453                 goto out;
 454         }
 455         memset(conf, 0, sizeof(*conf));
 456
 457         conf->multipaths = kmalloc(sizeof(struct multipath_info)*mddev->raid_disks,
 458                                    GFP_KERNEL);
 459         if (!conf->multipaths) {
 460                 printk(KERN_ERR
 461                         "multipath: couldn't allocate memory for %s\n",
 462                         mdname(mddev));
 463                 goto out_free_conf;
 464         }
 465         memset(conf->multipaths, 0, sizeof(struct multipath_info)*mddev->raid_disks);
 466
 467         conf->working_disks = 0;
 468         ITERATE_RDEV(mddev,rdev,tmp) {
 469                 disk_idx = rdev->raid_disk;
 470                 if (disk_idx < 0 ||
 471                     disk_idx >= mddev->raid_disks)
 472                         continue;
 473
 474                 disk = conf->multipaths + disk_idx;
 475                 disk->rdev = rdev;
 476
 477                 blk_queue_stack_limits(mddev->queue,
 478                                        rdev->bdev->bd_disk->queue);
 479                 /* as we don't honour merge_bvec_fn, we must never risk
 480                  * violating it, not that we ever expect a device with
 481                  * a merge_bvec_fn to be involved in multipath */
 482                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
 483                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
 484                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
 485
 486                 if (!test_bit(Faulty, &rdev->flags))
 487                         conf->working_disks++;
 488         }
 489
 490         conf->raid_disks = mddev->raid_disks;
 491         mddev->sb_dirty = 1;
 492         conf->mddev = mddev;
 493         spin_lock_init(&conf->device_lock);
 494         INIT_LIST_HEAD(&conf->retry_list);
 495
 496         if (!conf->working_disks) {
 497                 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
 498                         mdname(mddev));
 499                 goto out_free_conf;
 500         }
 501         mddev->degraded = conf->raid_disks = conf->working_disks;
 502
 503         conf->pool = mempool_create(NR_RESERVED_BUFS,
 504                                     mp_pool_alloc, mp_pool_free,
 505                                     NULL);
 506         if (conf->pool == NULL) {
 507                 printk(KERN_ERR
 508                         "multipath: couldn't allocate memory for %s\n",
 509                         mdname(mddev));
 510                 goto out_free_conf;
 511         }
 512
 513         {
 514                 mddev->thread = md_register_thread(multipathd, mddev, "%s_multipath");
 515                 if (!mddev->thread) {
 516                         printk(KERN_ERR "multipath: couldn't allocate thread"
 517                                 " for %s\n", mdname(mddev));
 518                         goto out_free_conf;
 519                 }
 520         }
 521
 522         printk(KERN_INFO
 523                 "multipath: array %s active with %d out of %d IO paths\n",
 524                 mdname(mddev), conf->working_disks, mddev->raid_disks);
 525         /*
 526          * Ok, everything is just fine now
 527          */
 528         mddev->array_size = mddev->size;
 529
 530         mddev->queue->unplug_fn = multipath_unplug;
 531         mddev->queue->issue_flush_fn = multipath_issue_flush;
 532
 533         return 0;
 534
 535 out_free_conf:
 536         if (conf->pool)
 537                 mempool_destroy(conf->pool);
 538         kfree(conf->multipaths);
 539         kfree(conf);
 540         mddev->private = NULL;
 541 out:
 542         return -EIO;
 543 }
 544
 545
 546 static int multipath_stop (mddev_t *mddev)
 547 {
 548         multipath_conf_t *conf = mddev_to_conf(mddev);
 549
 550         md_unregister_thread(mddev->thread);
 551         mddev->thread = NULL;
 552         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
 553         mempool_destroy(conf->pool);
 554         kfree(conf->multipaths);
 555         kfree(conf);
 556         mddev->private = NULL;
 557         return 0;
 558 }
 559
 560 static mdk_personality_t multipath_personality=
 561 {
 562         .name           = "multipath",
 563         .owner          = THIS_MODULE,
 564         .make_request   = multipath_make_request,
 565         .run            = multipath_run,
 566         .stop           = multipath_stop,
 567         .status         = multipath_status,
 568         .error_handler  = multipath_error,
 569         .hot_add_disk   = multipath_add_disk,
 570         .hot_remove_disk= multipath_remove_disk,
 571 };
 572
 573 static int __init multipath_init (void)
 574 {
 575         return register_md_personality (MULTIPATH, &multipath_personality);
 576 }
 577
 578 static void __exit multipath_exit (void)
 579 {
 580         unregister_md_personality (MULTIPATH);
 581 }
 582
 583 module_init(multipath_init);
 584 module_exit(multipath_exit);
 585 MODULE_LICENSE("GPL");
 586 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */