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