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