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