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/blkdev.h>
  23 #include <linux/raid/md_u.h>
  24 #include <linux/seq_file.h>
  25 #include "md.h"
  26 #include "multipath.h"
  27
  28 #define MAX_WORK_PER_DISK 128
  29
  30 #define NR_RESERVED_BUFS        32
  31
  32
  33 static int multipath_map (multipath_conf_t *conf)
  34 {
  35         int i, disks = conf->raid_disks;
  36
  37         /*
  38          * Later we do read balancing on the read side
  39          * now we use the first available disk.
  40          */
  41
  42         rcu_read_lock();
  43         for (i = 0; i < disks; i++) {
  44                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
  45                 if (rdev && test_bit(In_sync, &rdev->flags)) {
  46                         atomic_inc(&rdev->nr_pending);
  47                         rcu_read_unlock();
  48                         return i;
  49                 }
  50         }
  51         rcu_read_unlock();
  52
  53         printk(KERN_ERR "multipath_map(): no more operational IO paths?\n");
  54         return (-1);
  55 }
  56
  57 static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
  58 {
  59         unsigned long flags;
  60         mddev_t *mddev = mp_bh->mddev;
  61         multipath_conf_t *conf = mddev->private;
  62
  63         spin_lock_irqsave(&conf->device_lock, flags);
  64         list_add(&mp_bh->retry_list, &conf->retry_list);
  65         spin_unlock_irqrestore(&conf->device_lock, flags);
  66         md_wakeup_thread(mddev->thread);
  67 }
  68
  69
  70 /*
  71  * multipath_end_bh_io() is called when we have finished servicing a multipathed
  72  * operation and are ready to return a success/failure code to the buffer
  73  * cache layer.
  74  */
  75 static void multipath_end_bh_io (struct multipath_bh *mp_bh, int err)
  76 {
  77         struct bio *bio = mp_bh->master_bio;
  78         multipath_conf_t *conf = mp_bh->mddev->private;
  79
  80         bio_endio(bio, err);
  81         mempool_free(mp_bh, conf->pool);
  82 }
  83
  84 static void multipath_end_request(struct bio *bio, int error)
  85 {
  86         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
  87         struct multipath_bh * mp_bh = (struct multipath_bh *)(bio->bi_private);
  88         multipath_conf_t *conf = mp_bh->mddev->private;
  89         mdk_rdev_t *rdev = conf->multipaths[mp_bh->path].rdev;
  90
  91         if (uptodate)
  92                 multipath_end_bh_io(mp_bh, 0);
  93         else if (!bio_rw_flagged(bio, BIO_RW_AHEAD)) {
  94                 /*
  95                  * oops, IO error:
  96                  */
  97                 char b[BDEVNAME_SIZE];
  98                 md_error (mp_bh->mddev, rdev);
  99                 printk(KERN_ERR "multipath: %s: rescheduling sector %llu\n",
 100                        bdevname(rdev->bdev,b),
 101                        (unsigned long long)bio->bi_sector);
 102                 multipath_reschedule_retry(mp_bh);
 103         } else
 104                 multipath_end_bh_io(mp_bh, error);
 105         rdev_dec_pending(rdev, conf->mddev);
 106 }
 107
 108 static void unplug_slaves(mddev_t *mddev)
 109 {
 110         multipath_conf_t *conf = mddev->private;
 111         int i;
 112
 113         rcu_read_lock();
 114         for (i=0; i<mddev->raid_disks; i++) {
 115                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
 116                 if (rdev && !test_bit(Faulty, &rdev->flags)
 117                     && atomic_read(&rdev->nr_pending)) {
 118                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
 119
 120                         atomic_inc(&rdev->nr_pending);
 121                         rcu_read_unlock();
 122
 123                         blk_unplug(r_queue);
 124
 125                         rdev_dec_pending(rdev, mddev);
 126                         rcu_read_lock();
 127                 }
 128         }
 129         rcu_read_unlock();
 130 }
 131
 132 static void multipath_unplug(struct request_queue *q)
 133 {
 134         unplug_slaves(q->queuedata);
 135 }
 136
 137
 138 static int multipath_make_request (struct request_queue *q, struct bio * bio)
 139 {
 140         mddev_t *mddev = q->queuedata;
 141         multipath_conf_t *conf = mddev->private;
 142         struct multipath_bh * mp_bh;
 143         struct multipath_info *multipath;
 144         const int rw = bio_data_dir(bio);
 145         int cpu;
 146
 147         if (unlikely(bio_rw_flagged(bio, BIO_RW_BARRIER))) {
 148                 md_barrier_request(mddev, bio);
 149                 return 0;
 150         }
 151
 152         mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
 153
 154         mp_bh->master_bio = bio;
 155         mp_bh->mddev = mddev;
 156
 157         cpu = part_stat_lock();
 158         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
 159         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
 160                       bio_sectors(bio));
 161         part_stat_unlock();
 162
 163         mp_bh->path = multipath_map(conf);
 164         if (mp_bh->path < 0) {
 165                 bio_endio(bio, -EIO);
 166                 mempool_free(mp_bh, conf->pool);
 167                 return 0;
 168         }
 169         multipath = conf->multipaths + mp_bh->path;
 170
 171         mp_bh->bio = *bio;
 172         mp_bh->bio.bi_sector += multipath->rdev->data_offset;
 173         mp_bh->bio.bi_bdev = multipath->rdev->bdev;
 174         mp_bh->bio.bi_rw |= (1 << BIO_RW_FAILFAST_TRANSPORT);
 175         mp_bh->bio.bi_end_io = multipath_end_request;
 176         mp_bh->bio.bi_private = mp_bh;
 177         generic_make_request(&mp_bh->bio);
 178         return 0;
 179 }
 180
 181 static void multipath_status (struct seq_file *seq, mddev_t *mddev)
 182 {
 183         multipath_conf_t *conf = mddev->private;
 184         int i;
 185
 186         seq_printf (seq, " [%d/%d] [", conf->raid_disks,
 187                                                  conf->working_disks);
 188         for (i = 0; i < conf->raid_disks; i++)
 189                 seq_printf (seq, "%s",
 190                                conf->multipaths[i].rdev &&
 191                                test_bit(In_sync, &conf->multipaths[i].rdev->flags) ? "U" : "_");
 192         seq_printf (seq, "]");
 193 }
 194
 195 static int multipath_congested(void *data, int bits)
 196 {
 197         mddev_t *mddev = data;
 198         multipath_conf_t *conf = mddev->private;
 199         int i, ret = 0;
 200
 201         if (mddev_congested(mddev, bits))
 202                 return 1;
 203
 204         rcu_read_lock();
 205         for (i = 0; i < mddev->raid_disks ; i++) {
 206                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
 207                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 208                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 209
 210                         ret |= bdi_congested(&q->backing_dev_info, bits);
 211                         /* Just like multipath_map, we just check the
 212                          * first available device
 213                          */
 214                         break;
 215                 }
 216         }
 217         rcu_read_unlock();
 218         return ret;
 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->private;
 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 (!test_bit(Faulty, &rdev->flags)) {
 242                         char b[BDEVNAME_SIZE];
 243                         clear_bit(In_sync, &rdev->flags);
 244                         set_bit(Faulty, &rdev->flags);
 245                         set_bit(MD_CHANGE_DEVS, &mddev->flags);
 246                         conf->working_disks--;
 247                         mddev->degraded++;
 248                         printk(KERN_ALERT "multipath: IO failure on %s,"
 249                                 " disabling IO path.\n"
 250                                 "multipath: Operation continuing"
 251                                 " on %d IO paths.\n",
 252                                 bdevname (rdev->bdev,b),
 253                                 conf->working_disks);
 254                 }
 255         }
 256 }
 257
 258 static void print_multipath_conf (multipath_conf_t *conf)
 259 {
 260         int i;
 261         struct multipath_info *tmp;
 262
 263         printk("MULTIPATH conf printout:\n");
 264         if (!conf) {
 265                 printk("(conf==NULL)\n");
 266                 return;
 267         }
 268         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 269                          conf->raid_disks);
 270
 271         for (i = 0; i < conf->raid_disks; i++) {
 272                 char b[BDEVNAME_SIZE];
 273                 tmp = conf->multipaths + i;
 274                 if (tmp->rdev)
 275                         printk(" disk%d, o:%d, dev:%s\n",
 276                                 i,!test_bit(Faulty, &tmp->rdev->flags),
 277                                bdevname(tmp->rdev->bdev,b));
 278         }
 279 }
 280
 281
 282 static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 283 {
 284         multipath_conf_t *conf = mddev->private;
 285         struct request_queue *q;
 286         int err = -EEXIST;
 287         int path;
 288         struct multipath_info *p;
 289         int first = 0;
 290         int last = mddev->raid_disks - 1;
 291
 292         if (rdev->raid_disk >= 0)
 293                 first = last = rdev->raid_disk;
 294
 295         print_multipath_conf(conf);
 296
 297         for (path = first; path <= last; path++)
 298                 if ((p=conf->multipaths+path)->rdev == NULL) {
 299                         q = rdev->bdev->bd_disk->queue;
 300                         disk_stack_limits(mddev->gendisk, rdev->bdev,
 301                                           rdev->data_offset << 9);
 302
 303                 /* as we don't honour merge_bvec_fn, we must never risk
 304                  * violating it, so limit ->max_segments to one, lying
 305                  * within a single page.
 306                  * (Note: it is very unlikely that a device with
 307                  * merge_bvec_fn will be involved in multipath.)
 308                  */
 309                         if (q->merge_bvec_fn) {
 310                                 blk_queue_max_segments(mddev->queue, 1);
 311                                 blk_queue_segment_boundary(mddev->queue,
 312                                                            PAGE_CACHE_SIZE - 1);
 313                         }
 314
 315                         conf->working_disks++;
 316                         mddev->degraded--;
 317                         rdev->raid_disk = path;
 318                         set_bit(In_sync, &rdev->flags);
 319                         rcu_assign_pointer(p->rdev, rdev);
 320                         err = 0;
 321                         md_integrity_add_rdev(rdev, mddev);
 322                         break;
 323                 }
 324
 325         print_multipath_conf(conf);
 326
 327         return err;
 328 }
 329
 330 static int multipath_remove_disk(mddev_t *mddev, int number)
 331 {
 332         multipath_conf_t *conf = mddev->private;
 333         int err = 0;
 334         mdk_rdev_t *rdev;
 335         struct multipath_info *p = conf->multipaths + number;
 336
 337         print_multipath_conf(conf);
 338
 339         rdev = p->rdev;
 340         if (rdev) {
 341                 if (test_bit(In_sync, &rdev->flags) ||
 342                     atomic_read(&rdev->nr_pending)) {
 343                         printk(KERN_ERR "hot-remove-disk, slot %d is identified"
 344                                " but is still operational!\n", number);
 345                         err = -EBUSY;
 346                         goto abort;
 347                 }
 348                 p->rdev = NULL;
 349                 synchronize_rcu();
 350                 if (atomic_read(&rdev->nr_pending)) {
 351                         /* lost the race, try later */
 352                         err = -EBUSY;
 353                         p->rdev = rdev;
 354                         goto abort;
 355                 }
 356                 md_integrity_register(mddev);
 357         }
 358 abort:
 359
 360         print_multipath_conf(conf);
 361         return err;
 362 }
 363
 364
 365
 366 /*
 367  * This is a kernel thread which:
 368  *
 369  *      1.      Retries failed read operations on working multipaths.
 370  *      2.      Updates the raid superblock when problems encounter.
 371  *      3.      Performs writes following reads for array syncronising.
 372  */
 373
 374 static void multipathd (mddev_t *mddev)
 375 {
 376         struct multipath_bh *mp_bh;
 377         struct bio *bio;
 378         unsigned long flags;
 379         multipath_conf_t *conf = mddev->private;
 380         struct list_head *head = &conf->retry_list;
 381
 382         md_check_recovery(mddev);
 383         for (;;) {
 384                 char b[BDEVNAME_SIZE];
 385                 spin_lock_irqsave(&conf->device_lock, flags);
 386                 if (list_empty(head))
 387                         break;
 388                 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
 389                 list_del(head->prev);
 390                 spin_unlock_irqrestore(&conf->device_lock, flags);
 391
 392                 bio = &mp_bh->bio;
 393                 bio->bi_sector = mp_bh->master_bio->bi_sector;
 394
 395                 if ((mp_bh->path = multipath_map (conf))<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, -EIO);
 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 = *(mp_bh->master_bio);
 407                         bio->bi_sector += conf->multipaths[mp_bh->path].rdev->data_offset;
 408                         bio->bi_bdev = conf->multipaths[mp_bh->path].rdev->bdev;
 409                         bio->bi_rw |= (1 << BIO_RW_FAILFAST_TRANSPORT);
 410                         bio->bi_end_io = multipath_end_request;
 411                         bio->bi_private = mp_bh;
 412                         generic_make_request(bio);
 413                 }
 414         }
 415         spin_unlock_irqrestore(&conf->device_lock, flags);
 416 }
 417
 418 static sector_t multipath_size(mddev_t *mddev, sector_t sectors, int raid_disks)
 419 {
 420         WARN_ONCE(sectors || raid_disks,
 421                   "%s does not support generic reshape\n", __func__);
 422
 423         return mddev->dev_sectors;
 424 }
 425
 426 static int multipath_run (mddev_t *mddev)
 427 {
 428         multipath_conf_t *conf;
 429         int disk_idx;
 430         struct multipath_info *disk;
 431         mdk_rdev_t *rdev;
 432
 433         if (md_check_no_bitmap(mddev))
 434                 return -EINVAL;
 435
 436         if (mddev->level != LEVEL_MULTIPATH) {
 437                 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
 438                        mdname(mddev), mddev->level);
 439                 goto out;
 440         }
 441         /*
 442          * copy the already verified devices into our private MULTIPATH
 443          * bookkeeping area. [whatever we allocate in multipath_run(),
 444          * should be freed in multipath_stop()]
 445          */
 446         mddev->queue->queue_lock = &mddev->queue->__queue_lock;
 447
 448         conf = kzalloc(sizeof(multipath_conf_t), GFP_KERNEL);
 449         mddev->private = conf;
 450         if (!conf) {
 451                 printk(KERN_ERR
 452                         "multipath: couldn't allocate memory for %s\n",
 453                         mdname(mddev));
 454                 goto out;
 455         }
 456
 457         conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
 458                                    GFP_KERNEL);
 459         if (!conf->multipaths) {
 460                 printk(KERN_ERR
 461                         "multipath: couldn't allocate memory for %s\n",
 462                         mdname(mddev));
 463                 goto out_free_conf;
 464         }
 465
 466         conf->working_disks = 0;
 467         list_for_each_entry(rdev, &mddev->disks, same_set) {
 468                 disk_idx = rdev->raid_disk;
 469                 if (disk_idx < 0 ||
 470                     disk_idx >= mddev->raid_disks)
 471                         continue;
 472
 473                 disk = conf->multipaths + disk_idx;
 474                 disk->rdev = rdev;
 475                 disk_stack_limits(mddev->gendisk, rdev->bdev,
 476                                   rdev->data_offset << 9);
 477
 478                 /* as we don't honour merge_bvec_fn, we must never risk
 479                  * violating it, not that we ever expect a device with
 480                  * a merge_bvec_fn to be involved in multipath */
 481                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
 482                         blk_queue_max_segments(mddev->queue, 1);
 483                         blk_queue_segment_boundary(mddev->queue,
 484                                                    PAGE_CACHE_SIZE - 1);
 485                 }
 486
 487                 if (!test_bit(Faulty, &rdev->flags))
 488                         conf->working_disks++;
 489         }
 490
 491         conf->raid_disks = mddev->raid_disks;
 492         conf->mddev = mddev;
 493         spin_lock_init(&conf->device_lock);
 494         INIT_LIST_HEAD(&conf->retry_list);
 495
 496         if (!conf->working_disks) {
 497                 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
 498                         mdname(mddev));
 499                 goto out_free_conf;
 500         }
 501         mddev->degraded = conf->raid_disks - conf->working_disks;
 502
 503         conf->pool = mempool_create_kmalloc_pool(NR_RESERVED_BUFS,
 504                                                  sizeof(struct multipath_bh));
 505         if (conf->pool == NULL) {
 506                 printk(KERN_ERR
 507                         "multipath: couldn't allocate memory for %s\n",
 508                         mdname(mddev));
 509                 goto out_free_conf;
 510         }
 511
 512         {
 513                 mddev->thread = md_register_thread(multipathd, mddev, NULL);
 514                 if (!mddev->thread) {
 515                         printk(KERN_ERR "multipath: couldn't allocate thread"
 516                                 " for %s\n", mdname(mddev));
 517                         goto out_free_conf;
 518                 }
 519         }
 520
 521         printk(KERN_INFO
 522                 "multipath: array %s active with %d out of %d IO paths\n",
 523                 mdname(mddev), conf->working_disks, mddev->raid_disks);
 524         /*
 525          * Ok, everything is just fine now
 526          */
 527         md_set_array_sectors(mddev, multipath_size(mddev, 0, 0));
 528
 529         mddev->queue->unplug_fn = multipath_unplug;
 530         mddev->queue->backing_dev_info.congested_fn = multipath_congested;
 531         mddev->queue->backing_dev_info.congested_data = mddev;
 532         md_integrity_register(mddev);
 533         return 0;
 534
 535 out_free_conf:
 536         if (conf->pool)
 537                 mempool_destroy(conf->pool);
 538         kfree(conf->multipaths);
 539         kfree(conf);
 540         mddev->private = NULL;
 541 out:
 542         return -EIO;
 543 }
 544
 545
 546 static int multipath_stop (mddev_t *mddev)
 547 {
 548         multipath_conf_t *conf = mddev->private;
 549
 550         md_unregister_thread(mddev->thread);
 551         mddev->thread = NULL;
 552         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
 553         mempool_destroy(conf->pool);
 554         kfree(conf->multipaths);
 555         kfree(conf);
 556         mddev->private = NULL;
 557         return 0;
 558 }
 559
 560 static struct mdk_personality multipath_personality =
 561 {
 562         .name           = "multipath",
 563         .level          = LEVEL_MULTIPATH,
 564         .owner          = THIS_MODULE,
 565         .make_request   = multipath_make_request,
 566         .run            = multipath_run,
 567         .stop           = multipath_stop,
 568         .status         = multipath_status,
 569         .error_handler  = multipath_error,
 570         .hot_add_disk   = multipath_add_disk,
 571         .hot_remove_disk= multipath_remove_disk,
 572         .size           = multipath_size,
 573 };
 574
 575 static int __init multipath_init (void)
 576 {
 577         return register_md_personality (&multipath_personality);
 578 }
 579
 580 static void __exit multipath_exit (void)
 581 {
 582         unregister_md_personality (&multipath_personality);
 583 }
 584
 585 module_init(multipath_init);
 586 module_exit(multipath_exit);
 587 MODULE_LICENSE("GPL");
 588 MODULE_DESCRIPTION("simple multi-path personality for MD");
 589 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
 590 MODULE_ALIAS("md-multipath");
 591 MODULE_ALIAS("md-level--4");