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 err)
 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, err);
 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, 0);
 123         else if (!bio_rw_ahead(bio)) {
 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         } else
 134                 multipath_end_bh_io(mp_bh, error);
 135         rdev_dec_pending(rdev, conf->mddev);
 136         return 0;
 137 }
 138
 139 static void unplug_slaves(mddev_t *mddev)
 140 {
 141         multipath_conf_t *conf = mddev_to_conf(mddev);
 142         int i;
 143         unsigned long flags;
 144
 145         spin_lock_irqsave(&conf->device_lock, flags);
 146         for (i=0; i<mddev->raid_disks; i++) {
 147                 mdk_rdev_t *rdev = conf->multipaths[i].rdev;
 148                 if (rdev && !rdev->faulty) {
 149                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
 150
 151                         atomic_inc(&rdev->nr_pending);
 152                         spin_unlock_irqrestore(&conf->device_lock, flags);
 153
 154                         if (r_queue->unplug_fn)
 155                                 r_queue->unplug_fn(r_queue);
 156
 157                         spin_lock_irqsave(&conf->device_lock, flags);
 158                         atomic_dec(&rdev->nr_pending);
 159                 }
 160         }
 161         spin_unlock_irqrestore(&conf->device_lock, flags);
 162 }
 163 static void multipath_unplug(request_queue_t *q)
 164 {
 165         unplug_slaves(q->queuedata);
 166 }
 167
 168
 169 static int multipath_make_request (request_queue_t *q, struct bio * bio)
 170 {
 171         mddev_t *mddev = q->queuedata;
 172         multipath_conf_t *conf = mddev_to_conf(mddev);
 173         struct multipath_bh * mp_bh;
 174         struct multipath_info *multipath;
 175
 176         mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
 177
 178         mp_bh->master_bio = bio;
 179         mp_bh->mddev = mddev;
 180
 181         if (bio_data_dir(bio)==WRITE) {
 182                 disk_stat_inc(mddev->gendisk, writes);
 183                 disk_stat_add(mddev->gendisk, write_sectors, bio_sectors(bio));
 184         } else {
 185                 disk_stat_inc(mddev->gendisk, reads);
 186                 disk_stat_add(mddev->gendisk, read_sectors, bio_sectors(bio));
 187         }
 188
 189         mp_bh->path = multipath_map(conf);
 190         if (mp_bh->path < 0) {
 191                 bio_endio(bio, bio->bi_size, -EIO);
 192                 mempool_free(mp_bh, conf->pool);
 193                 return 0;
 194         }
 195         multipath = conf->multipaths + mp_bh->path;
 196
 197         mp_bh->bio = *bio;
 198         mp_bh->bio.bi_bdev = multipath->rdev->bdev;
 199         mp_bh->bio.bi_rw |= (1 << BIO_RW_FAILFAST);
 200         mp_bh->bio.bi_end_io = multipath_end_request;
 201         mp_bh->bio.bi_private = mp_bh;
 202         generic_make_request(&mp_bh->bio);
 203         return 0;
 204 }
 205
 206 static void multipath_status (struct seq_file *seq, mddev_t *mddev)
 207 {
 208         multipath_conf_t *conf = mddev_to_conf(mddev);
 209         int i;
 210
 211         seq_printf (seq, " [%d/%d] [", conf->raid_disks,
 212                                                  conf->working_disks);
 213         for (i = 0; i < conf->raid_disks; i++)
 214                 seq_printf (seq, "%s",
 215                                conf->multipaths[i].rdev &&
 216                                conf->multipaths[i].rdev->in_sync ? "U" : "_");
 217         seq_printf (seq, "]");
 218 }
 219
 220 static int multipath_issue_flush(request_queue_t *q, struct gendisk *disk,
 221                                  sector_t *error_sector)
 222 {
 223         mddev_t *mddev = q->queuedata;
 224         multipath_conf_t *conf = mddev_to_conf(mddev);
 225         int i, ret = 0;
 226
 227         for (i=0; i<mddev->raid_disks; i++) {
 228                 mdk_rdev_t *rdev = conf->multipaths[i].rdev;
 229                 if (rdev && !rdev->faulty) {
 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                                 break;
 236                         }
 237
 238                         ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk, error_sector);
 239                         if (ret)
 240                                 break;
 241                 }
 242         }
 243         return ret;
 244 }
 245
 246 /*
 247  * Careful, this can execute in IRQ contexts as well!
 248  */
 249 static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
 250 {
 251         multipath_conf_t *conf = mddev_to_conf(mddev);
 252
 253         if (conf->working_disks <= 1) {
 254                 /*
 255                  * Uh oh, we can do nothing if this is our last path, but
 256                  * first check if this is a queued request for a device
 257                  * which has just failed.
 258                  */
 259                 printk(KERN_ALERT
 260                         "multipath: only one IO path left and IO error.\n");
 261                 /* leave it active... it's all we have */
 262         } else {
 263                 /*
 264                  * Mark disk as unusable
 265                  */
 266                 if (!rdev->faulty) {
 267                         char b[BDEVNAME_SIZE];
 268                         rdev->in_sync = 0;
 269                         rdev->faulty = 1;
 270                         mddev->sb_dirty = 1;
 271                         conf->working_disks--;
 272                         printk(KERN_ALERT "multipath: IO failure on %s,"
 273                                 " disabling IO path. \n Operation continuing"
 274                                 " on %d IO paths.\n",
 275                                 bdevname (rdev->bdev,b),
 276                                 conf->working_disks);
 277                 }
 278         }
 279 }
 280
 281 static void print_multipath_conf (multipath_conf_t *conf)
 282 {
 283         int i;
 284         struct multipath_info *tmp;
 285
 286         printk("MULTIPATH conf printout:\n");
 287         if (!conf) {
 288                 printk("(conf==NULL)\n");
 289                 return;
 290         }
 291         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 292                          conf->raid_disks);
 293
 294         for (i = 0; i < conf->raid_disks; i++) {
 295                 char b[BDEVNAME_SIZE];
 296                 tmp = conf->multipaths + i;
 297                 if (tmp->rdev)
 298                         printk(" disk%d, o:%d, dev:%s\n",
 299                                 i,!tmp->rdev->faulty,
 300                                bdevname(tmp->rdev->bdev,b));
 301         }
 302 }
 303
 304
 305 static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 306 {
 307         multipath_conf_t *conf = mddev->private;
 308         int found = 0;
 309         int path;
 310         struct multipath_info *p;
 311
 312         print_multipath_conf(conf);
 313         spin_lock_irq(&conf->device_lock);
 314         for (path=0; path<mddev->raid_disks; path++)
 315                 if ((p=conf->multipaths+path)->rdev == NULL) {
 316                         p->rdev = rdev;
 317                         blk_queue_stack_limits(mddev->queue,
 318                                                rdev->bdev->bd_disk->queue);
 319
 320                 /* as we don't honour merge_bvec_fn, we must never risk
 321                  * violating it, so limit ->max_sector to one PAGE, as
 322                  * a one page request is never in violation.
 323                  * (Note: it is very unlikely that a device with
 324                  * merge_bvec_fn will be involved in multipath.)
 325                  */
 326                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
 327                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
 328                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
 329
 330                         conf->working_disks++;
 331                         rdev->raid_disk = path;
 332                         rdev->in_sync = 1;
 333                         found = 1;
 334                 }
 335         spin_unlock_irq(&conf->device_lock);
 336
 337         print_multipath_conf(conf);
 338         return found;
 339 }
 340
 341 static int multipath_remove_disk(mddev_t *mddev, int number)
 342 {
 343         multipath_conf_t *conf = mddev->private;
 344         int err = 1;
 345         struct multipath_info *p = conf->multipaths + number;
 346
 347         print_multipath_conf(conf);
 348         spin_lock_irq(&conf->device_lock);
 349
 350         if (p->rdev) {
 351                 if (p->rdev->in_sync ||
 352                     atomic_read(&p->rdev->nr_pending)) {
 353                         printk(KERN_ERR "hot-remove-disk, slot %d is identified"                                " but is still operational!\n", number);
 354                         err = -EBUSY;
 355                         goto abort;
 356                 }
 357                 p->rdev = NULL;
 358                 err = 0;
 359         }
 360         if (err)
 361                 MD_BUG();
 362 abort:
 363         spin_unlock_irq(&conf->device_lock);
 364
 365         print_multipath_conf(conf);
 366         return err;
 367 }
 368
 369
 370
 371 /*
 372  * This is a kernel thread which:
 373  *
 374  *      1.      Retries failed read operations on working multipaths.
 375  *      2.      Updates the raid superblock when problems encounter.
 376  *      3.      Performs writes following reads for array syncronising.
 377  */
 378
 379 static void multipathd (mddev_t *mddev)
 380 {
 381         struct multipath_bh *mp_bh;
 382         struct bio *bio;
 383         unsigned long flags;
 384         multipath_conf_t *conf = mddev_to_conf(mddev);
 385
 386         md_check_recovery(mddev);
 387         for (;;) {
 388                 char b[BDEVNAME_SIZE];
 389                 spin_lock_irqsave(&retry_list_lock, flags);
 390                 mp_bh = multipath_retry_list;
 391                 if (!mp_bh)
 392                         break;
 393                 multipath_retry_list = mp_bh->next_mp;
 394                 spin_unlock_irqrestore(&retry_list_lock, flags);
 395
 396                 mddev = mp_bh->mddev;
 397                 bio = &mp_bh->bio;
 398                 bio->bi_sector = mp_bh->master_bio->bi_sector;
 399
 400                 if ((mp_bh->path = multipath_map (conf))<0) {
 401                         printk(KERN_ALERT "multipath: %s: unrecoverable IO read"
 402                                 " error for block %llu\n",
 403                                 bdevname(bio->bi_bdev,b),
 404                                 (unsigned long long)bio->bi_sector);
 405                         multipath_end_bh_io(mp_bh, -EIO);
 406                 } else {
 407                         printk(KERN_ERR "multipath: %s: redirecting sector %llu"
 408                                 " to another IO path\n",
 409                                 bdevname(bio->bi_bdev,b),
 410                                 (unsigned long long)bio->bi_sector);
 411                         *bio = *(mp_bh->master_bio);
 412                         bio->bi_bdev = conf->multipaths[mp_bh->path].rdev->bdev;
 413                         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
 414                         bio->bi_end_io = multipath_end_request;
 415                         bio->bi_private = mp_bh;
 416                         generic_make_request(bio);
 417                 }
 418         }
 419         spin_unlock_irqrestore(&retry_list_lock, flags);
 420 }
 421
 422 static int multipath_run (mddev_t *mddev)
 423 {
 424         multipath_conf_t *conf;
 425         int disk_idx;
 426         struct multipath_info *disk;
 427         mdk_rdev_t *rdev;
 428         struct list_head *tmp;
 429
 430         if (mddev->level != LEVEL_MULTIPATH) {
 431                 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
 432                        mdname(mddev), mddev->level);
 433                 goto out;
 434         }
 435         /*
 436          * copy the already verified devices into our private MULTIPATH
 437          * bookkeeping area. [whatever we allocate in multipath_run(),
 438          * should be freed in multipath_stop()]
 439          */
 440
 441         conf = kmalloc(sizeof(multipath_conf_t), GFP_KERNEL);
 442         mddev->private = conf;
 443         if (!conf) {
 444                 printk(KERN_ERR
 445                         "multipath: couldn't allocate memory for %s\n",
 446                         mdname(mddev));
 447                 goto out;
 448         }
 449         memset(conf, 0, sizeof(*conf));
 450
 451         conf->multipaths = kmalloc(sizeof(struct multipath_info)*mddev->raid_disks,
 452                                    GFP_KERNEL);
 453         if (!conf->multipaths) {
 454                 printk(KERN_ERR
 455                         "multipath: couldn't allocate memory for %s\n",
 456                         mdname(mddev));
 457                 goto out_free_conf;
 458         }
 459         memset(conf->multipaths, 0, sizeof(struct multipath_info)*mddev->raid_disks);
 460
 461         mddev->queue->unplug_fn = multipath_unplug;
 462
 463         mddev->queue->issue_flush_fn = multipath_issue_flush;
 464
 465         conf->working_disks = 0;
 466         ITERATE_RDEV(mddev,rdev,tmp) {
 467                 disk_idx = rdev->raid_disk;
 468                 if (disk_idx < 0 ||
 469                     disk_idx >= mddev->raid_disks)
 470                         continue;
 471
 472                 disk = conf->multipaths + disk_idx;
 473                 disk->rdev = rdev;
 474
 475                 blk_queue_stack_limits(mddev->queue,
 476                                        rdev->bdev->bd_disk->queue);
 477                 /* as we don't honour merge_bvec_fn, we must never risk
 478                  * violating it, not that we ever expect a device with
 479                  * a merge_bvec_fn to be involved in multipath */
 480                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
 481                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
 482                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
 483
 484                 if (!rdev->faulty)
 485                         conf->working_disks++;
 486         }
 487
 488         conf->raid_disks = mddev->raid_disks;
 489         mddev->sb_dirty = 1;
 490         conf->mddev = mddev;
 491         conf->device_lock = SPIN_LOCK_UNLOCKED;
 492
 493         if (!conf->working_disks) {
 494                 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
 495                         mdname(mddev));
 496                 goto out_free_conf;
 497         }
 498         mddev->degraded = conf->raid_disks = conf->working_disks;
 499
 500         conf->pool = mempool_create(NR_RESERVED_BUFS,
 501                                     mp_pool_alloc, mp_pool_free,
 502                                     NULL);
 503         if (conf->pool == NULL) {
 504                 printk(KERN_ERR
 505                         "multipath: couldn't allocate memory for %s\n",
 506                         mdname(mddev));
 507                 goto out_free_conf;
 508         }
 509
 510         {
 511                 mddev->thread = md_register_thread(multipathd, mddev, "%s_multipath");
 512                 if (!mddev->thread) {
 513                         printk(KERN_ERR "multipath: couldn't allocate thread"
 514                                 " for %s\n", mdname(mddev));
 515                         goto out_free_conf;
 516                 }
 517         }
 518
 519         printk(KERN_INFO
 520                 "multipath: array %s active with %d out of %d IO paths\n",
 521                 mdname(mddev), conf->working_disks, mddev->raid_disks);
 522         /*
 523          * Ok, everything is just fine now
 524          */
 525         mddev->array_size = mddev->size;
 526         return 0;
 527
 528 out_free_conf:
 529         if (conf->pool)
 530                 mempool_destroy(conf->pool);
 531         if (conf->multipaths)
 532                 kfree(conf->multipaths);
 533         kfree(conf);
 534         mddev->private = NULL;
 535 out:
 536         return -EIO;
 537 }
 538
 539
 540 static int multipath_stop (mddev_t *mddev)
 541 {
 542         multipath_conf_t *conf = mddev_to_conf(mddev);
 543
 544         md_unregister_thread(mddev->thread);
 545         mddev->thread = NULL;
 546         mempool_destroy(conf->pool);
 547         kfree(conf->multipaths);
 548         kfree(conf);
 549         mddev->private = NULL;
 550         return 0;
 551 }
 552
 553 static mdk_personality_t multipath_personality=
 554 {
 555         .name           = "multipath",
 556         .owner          = THIS_MODULE,
 557         .make_request   = multipath_make_request,
 558         .run            = multipath_run,
 559         .stop           = multipath_stop,
 560         .status         = multipath_status,
 561         .error_handler  = multipath_error,
 562         .hot_add_disk   = multipath_add_disk,
 563         .hot_remove_disk= multipath_remove_disk,
 564 };
 565
 566 static int __init multipath_init (void)
 567 {
 568         return register_md_personality (MULTIPATH, &multipath_personality);
 569 }
 570
 571 static void __exit multipath_exit (void)
 572 {
 573         unregister_md_personality (MULTIPATH);
 574 }
 575
 576 module_init(multipath_init);
 577 module_exit(multipath_exit);
 578 MODULE_LICENSE("GPL");
 579 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */