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