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/module.h>
  24 #include <linux/raid/md_u.h>
  25 #include <linux/seq_file.h>
  26 #include <linux/slab.h>
  27 #include "md.h"
  28 #include "multipath.h"
  29
  30 #define MAX_WORK_PER_DISK 128
  31
  32 #define NR_RESERVED_BUFS        32
  33
  34 static int multipath_map (struct mpconf *conf)
  35 {
  36         int i, disks = conf->raid_disks;
  37
  38         /*
  39          * Later we do read balancing on the read side
  40          * now we use the first available disk.
  41          */
  42
  43         rcu_read_lock();
  44         for (i = 0; i < disks; i++) {
  45                 struct md_rdev *rdev = rcu_dereference(conf->multipaths[i].rdev);
  46                 if (rdev && test_bit(In_sync, &rdev->flags)) {
  47                         atomic_inc(&rdev->nr_pending);
  48                         rcu_read_unlock();
  49                         return i;
  50                 }
  51         }
  52         rcu_read_unlock();
  53
  54         printk(KERN_ERR "multipath_map(): no more operational IO paths?\n");
  55         return (-1);
  56 }
  57
  58 static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
  59 {
  60         unsigned long flags;
  61         struct mddev *mddev = mp_bh->mddev;
  62         struct mpconf *conf = mddev->private;
  63
  64         spin_lock_irqsave(&conf->device_lock, flags);
  65         list_add(&mp_bh->retry_list, &conf->retry_list);
  66         spin_unlock_irqrestore(&conf->device_lock, flags);
  67         md_wakeup_thread(mddev->thread);
  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         struct mpconf *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 = bio->bi_private;
  88         struct mpconf *conf = mp_bh->mddev->private;
  89         struct md_rdev *rdev = conf->multipaths[mp_bh->path].rdev;
  90
  91         if (uptodate)
  92                 multipath_end_bh_io(mp_bh, 0);
  93         else if (!(bio->bi_rw & REQ_RAHEAD)) {
  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_iter.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 multipath_make_request(struct mddev *mddev, struct bio * bio)
 109 {
 110         struct mpconf *conf = mddev->private;
 111         struct multipath_bh * mp_bh;
 112         struct multipath_info *multipath;
 113
 114         if (unlikely(bio->bi_rw & REQ_FLUSH)) {
 115                 md_flush_request(mddev, bio);
 116                 return;
 117         }
 118
 119         mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
 120
 121         mp_bh->master_bio = bio;
 122         mp_bh->mddev = mddev;
 123
 124         mp_bh->path = multipath_map(conf);
 125         if (mp_bh->path < 0) {
 126                 bio_endio(bio, -EIO);
 127                 mempool_free(mp_bh, conf->pool);
 128                 return;
 129         }
 130         multipath = conf->multipaths + mp_bh->path;
 131
 132         mp_bh->bio = *bio;
 133         mp_bh->bio.bi_iter.bi_sector += multipath->rdev->data_offset;
 134         mp_bh->bio.bi_bdev = multipath->rdev->bdev;
 135         mp_bh->bio.bi_rw |= REQ_FAILFAST_TRANSPORT;
 136         mp_bh->bio.bi_end_io = multipath_end_request;
 137         mp_bh->bio.bi_private = mp_bh;
 138         generic_make_request(&mp_bh->bio);
 139         return;
 140 }
 141
 142 static void multipath_status (struct seq_file *seq, struct mddev *mddev)
 143 {
 144         struct mpconf *conf = mddev->private;
 145         int i;
 146
 147         seq_printf (seq, " [%d/%d] [", conf->raid_disks,
 148                     conf->raid_disks - mddev->degraded);
 149         for (i = 0; i < conf->raid_disks; i++)
 150                 seq_printf (seq, "%s",
 151                                conf->multipaths[i].rdev &&
 152                                test_bit(In_sync, &conf->multipaths[i].rdev->flags) ? "U" : "_");
 153         seq_printf (seq, "]");
 154 }
 155
 156 static int multipath_congested(struct mddev *mddev, int bits)
 157 {
 158         struct mpconf *conf = mddev->private;
 159         int i, ret = 0;
 160
 161         rcu_read_lock();
 162         for (i = 0; i < mddev->raid_disks ; i++) {
 163                 struct md_rdev *rdev = rcu_dereference(conf->multipaths[i].rdev);
 164                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 165                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 166
 167                         ret |= bdi_congested(&q->backing_dev_info, bits);
 168                         /* Just like multipath_map, we just check the
 169                          * first available device
 170                          */
 171                         break;
 172                 }
 173         }
 174         rcu_read_unlock();
 175         return ret;
 176 }
 177
 178 /*
 179  * Careful, this can execute in IRQ contexts as well!
 180  */
 181 static void multipath_error (struct mddev *mddev, struct md_rdev *rdev)
 182 {
 183         struct mpconf *conf = mddev->private;
 184         char b[BDEVNAME_SIZE];
 185
 186         if (conf->raid_disks - mddev->degraded <= 1) {
 187                 /*
 188                  * Uh oh, we can do nothing if this is our last path, but
 189                  * first check if this is a queued request for a device
 190                  * which has just failed.
 191                  */
 192                 printk(KERN_ALERT
 193                        "multipath: only one IO path left and IO error.\n");
 194                 /* leave it active... it's all we have */
 195                 return;
 196         }
 197         /*
 198          * Mark disk as unusable
 199          */
 200         if (test_and_clear_bit(In_sync, &rdev->flags)) {
 201                 unsigned long flags;
 202                 spin_lock_irqsave(&conf->device_lock, flags);
 203                 mddev->degraded++;
 204                 spin_unlock_irqrestore(&conf->device_lock, flags);
 205         }
 206         set_bit(Faulty, &rdev->flags);
 207         set_bit(MD_CHANGE_DEVS, &mddev->flags);
 208         printk(KERN_ALERT "multipath: IO failure on %s,"
 209                " disabling IO path.\n"
 210                "multipath: Operation continuing"
 211                " on %d IO paths.\n",
 212                bdevname(rdev->bdev, b),
 213                conf->raid_disks - mddev->degraded);
 214 }
 215
 216 static void print_multipath_conf (struct mpconf *conf)
 217 {
 218         int i;
 219         struct multipath_info *tmp;
 220
 221         printk("MULTIPATH conf printout:\n");
 222         if (!conf) {
 223                 printk("(conf==NULL)\n");
 224                 return;
 225         }
 226         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
 227                          conf->raid_disks);
 228
 229         for (i = 0; i < conf->raid_disks; i++) {
 230                 char b[BDEVNAME_SIZE];
 231                 tmp = conf->multipaths + i;
 232                 if (tmp->rdev)
 233                         printk(" disk%d, o:%d, dev:%s\n",
 234                                 i,!test_bit(Faulty, &tmp->rdev->flags),
 235                                bdevname(tmp->rdev->bdev,b));
 236         }
 237 }
 238
 239 static int multipath_add_disk(struct mddev *mddev, struct md_rdev *rdev)
 240 {
 241         struct mpconf *conf = mddev->private;
 242         struct request_queue *q;
 243         int err = -EEXIST;
 244         int path;
 245         struct multipath_info *p;
 246         int first = 0;
 247         int last = mddev->raid_disks - 1;
 248
 249         if (rdev->raid_disk >= 0)
 250                 first = last = rdev->raid_disk;
 251
 252         print_multipath_conf(conf);
 253
 254         for (path = first; path <= last; path++)
 255                 if ((p=conf->multipaths+path)->rdev == NULL) {
 256                         q = rdev->bdev->bd_disk->queue;
 257                         disk_stack_limits(mddev->gendisk, rdev->bdev,
 258                                           rdev->data_offset << 9);
 259
 260                 /* as we don't honour merge_bvec_fn, we must never risk
 261                  * violating it, so limit ->max_segments to one, lying
 262                  * within a single page.
 263                  * (Note: it is very unlikely that a device with
 264                  * merge_bvec_fn will be involved in multipath.)
 265                  */
 266                         if (q->merge_bvec_fn) {
 267                                 blk_queue_max_segments(mddev->queue, 1);
 268                                 blk_queue_segment_boundary(mddev->queue,
 269                                                            PAGE_CACHE_SIZE - 1);
 270                         }
 271
 272                         spin_lock_irq(&conf->device_lock);
 273                         mddev->degraded--;
 274                         rdev->raid_disk = path;
 275                         set_bit(In_sync, &rdev->flags);
 276                         spin_unlock_irq(&conf->device_lock);
 277                         rcu_assign_pointer(p->rdev, rdev);
 278                         err = 0;
 279                         md_integrity_add_rdev(rdev, mddev);
 280                         break;
 281                 }
 282
 283         print_multipath_conf(conf);
 284
 285         return err;
 286 }
 287
 288 static int multipath_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
 289 {
 290         struct mpconf *conf = mddev->private;
 291         int err = 0;
 292         int number = rdev->raid_disk;
 293         struct multipath_info *p = conf->multipaths + number;
 294
 295         print_multipath_conf(conf);
 296
 297         if (rdev == p->rdev) {
 298                 if (test_bit(In_sync, &rdev->flags) ||
 299                     atomic_read(&rdev->nr_pending)) {
 300                         printk(KERN_ERR "hot-remove-disk, slot %d is identified"
 301                                " but is still operational!\n", number);
 302                         err = -EBUSY;
 303                         goto abort;
 304                 }
 305                 p->rdev = NULL;
 306                 synchronize_rcu();
 307                 if (atomic_read(&rdev->nr_pending)) {
 308                         /* lost the race, try later */
 309                         err = -EBUSY;
 310                         p->rdev = rdev;
 311                         goto abort;
 312                 }
 313                 err = md_integrity_register(mddev);
 314         }
 315 abort:
 316
 317         print_multipath_conf(conf);
 318         return err;
 319 }
 320
 321 /*
 322  * This is a kernel thread which:
 323  *
 324  *      1.      Retries failed read operations on working multipaths.
 325  *      2.      Updates the raid superblock when problems encounter.
 326  *      3.      Performs writes following reads for array syncronising.
 327  */
 328
 329 static void multipathd(struct md_thread *thread)
 330 {
 331         struct mddev *mddev = thread->mddev;
 332         struct multipath_bh *mp_bh;
 333         struct bio *bio;
 334         unsigned long flags;
 335         struct mpconf *conf = mddev->private;
 336         struct list_head *head = &conf->retry_list;
 337
 338         md_check_recovery(mddev);
 339         for (;;) {
 340                 char b[BDEVNAME_SIZE];
 341                 spin_lock_irqsave(&conf->device_lock, flags);
 342                 if (list_empty(head))
 343                         break;
 344                 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
 345                 list_del(head->prev);
 346                 spin_unlock_irqrestore(&conf->device_lock, flags);
 347
 348                 bio = &mp_bh->bio;
 349                 bio->bi_iter.bi_sector = mp_bh->master_bio->bi_iter.bi_sector;
 350
 351                 if ((mp_bh->path = multipath_map (conf))<0) {
 352                         printk(KERN_ALERT "multipath: %s: unrecoverable IO read"
 353                                 " error for block %llu\n",
 354                                 bdevname(bio->bi_bdev,b),
 355                                 (unsigned long long)bio->bi_iter.bi_sector);
 356                         multipath_end_bh_io(mp_bh, -EIO);
 357                 } else {
 358                         printk(KERN_ERR "multipath: %s: redirecting sector %llu"
 359                                 " to another IO path\n",
 360                                 bdevname(bio->bi_bdev,b),
 361                                 (unsigned long long)bio->bi_iter.bi_sector);
 362                         *bio = *(mp_bh->master_bio);
 363                         bio->bi_iter.bi_sector +=
 364                                 conf->multipaths[mp_bh->path].rdev->data_offset;
 365                         bio->bi_bdev = conf->multipaths[mp_bh->path].rdev->bdev;
 366                         bio->bi_rw |= REQ_FAILFAST_TRANSPORT;
 367                         bio->bi_end_io = multipath_end_request;
 368                         bio->bi_private = mp_bh;
 369                         generic_make_request(bio);
 370                 }
 371         }
 372         spin_unlock_irqrestore(&conf->device_lock, flags);
 373 }
 374
 375 static sector_t multipath_size(struct mddev *mddev, sector_t sectors, int raid_disks)
 376 {
 377         WARN_ONCE(sectors || raid_disks,
 378                   "%s does not support generic reshape\n", __func__);
 379
 380         return mddev->dev_sectors;
 381 }
 382
 383 static int multipath_run (struct mddev *mddev)
 384 {
 385         struct mpconf *conf;
 386         int disk_idx;
 387         struct multipath_info *disk;
 388         struct md_rdev *rdev;
 389         int working_disks;
 390
 391         if (md_check_no_bitmap(mddev))
 392                 return -EINVAL;
 393
 394         if (mddev->level != LEVEL_MULTIPATH) {
 395                 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
 396                        mdname(mddev), mddev->level);
 397                 goto out;
 398         }
 399         /*
 400          * copy the already verified devices into our private MULTIPATH
 401          * bookkeeping area. [whatever we allocate in multipath_run(),
 402          * should be freed in multipath_free()]
 403          */
 404
 405         conf = kzalloc(sizeof(struct mpconf), GFP_KERNEL);
 406         mddev->private = conf;
 407         if (!conf) {
 408                 printk(KERN_ERR
 409                         "multipath: couldn't allocate memory for %s\n",
 410                         mdname(mddev));
 411                 goto out;
 412         }
 413
 414         conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
 415                                    GFP_KERNEL);
 416         if (!conf->multipaths) {
 417                 printk(KERN_ERR
 418                         "multipath: couldn't allocate memory for %s\n",
 419                         mdname(mddev));
 420                 goto out_free_conf;
 421         }
 422
 423         working_disks = 0;
 424         rdev_for_each(rdev, mddev) {
 425                 disk_idx = rdev->raid_disk;
 426                 if (disk_idx < 0 ||
 427                     disk_idx >= mddev->raid_disks)
 428                         continue;
 429
 430                 disk = conf->multipaths + disk_idx;
 431                 disk->rdev = rdev;
 432                 disk_stack_limits(mddev->gendisk, rdev->bdev,
 433                                   rdev->data_offset << 9);
 434
 435                 /* as we don't honour merge_bvec_fn, we must never risk
 436                  * violating it, not that we ever expect a device with
 437                  * a merge_bvec_fn to be involved in multipath */
 438                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
 439                         blk_queue_max_segments(mddev->queue, 1);
 440                         blk_queue_segment_boundary(mddev->queue,
 441                                                    PAGE_CACHE_SIZE - 1);
 442                 }
 443
 444                 if (!test_bit(Faulty, &rdev->flags))
 445                         working_disks++;
 446         }
 447
 448         conf->raid_disks = mddev->raid_disks;
 449         conf->mddev = mddev;
 450         spin_lock_init(&conf->device_lock);
 451         INIT_LIST_HEAD(&conf->retry_list);
 452
 453         if (!working_disks) {
 454                 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
 455                         mdname(mddev));
 456                 goto out_free_conf;
 457         }
 458         mddev->degraded = conf->raid_disks - working_disks;
 459
 460         conf->pool = mempool_create_kmalloc_pool(NR_RESERVED_BUFS,
 461                                                  sizeof(struct multipath_bh));
 462         if (conf->pool == NULL) {
 463                 printk(KERN_ERR
 464                         "multipath: couldn't allocate memory for %s\n",
 465                         mdname(mddev));
 466                 goto out_free_conf;
 467         }
 468
 469         {
 470                 mddev->thread = md_register_thread(multipathd, mddev,
 471                                                    "multipath");
 472                 if (!mddev->thread) {
 473                         printk(KERN_ERR "multipath: couldn't allocate thread"
 474                                 " for %s\n", mdname(mddev));
 475                         goto out_free_conf;
 476                 }
 477         }
 478
 479         printk(KERN_INFO
 480                 "multipath: array %s active with %d out of %d IO paths\n",
 481                 mdname(mddev), conf->raid_disks - mddev->degraded,
 482                mddev->raid_disks);
 483         /*
 484          * Ok, everything is just fine now
 485          */
 486         md_set_array_sectors(mddev, multipath_size(mddev, 0, 0));
 487
 488         if (md_integrity_register(mddev))
 489                 goto out_free_conf;
 490
 491         return 0;
 492
 493 out_free_conf:
 494         if (conf->pool)
 495                 mempool_destroy(conf->pool);
 496         kfree(conf->multipaths);
 497         kfree(conf);
 498         mddev->private = NULL;
 499 out:
 500         return -EIO;
 501 }
 502
 503 static void multipath_free(struct mddev *mddev, void *priv)
 504 {
 505         struct mpconf *conf = priv;
 506
 507         mempool_destroy(conf->pool);
 508         kfree(conf->multipaths);
 509         kfree(conf);
 510 }
 511
 512 static struct md_personality multipath_personality =
 513 {
 514         .name           = "multipath",
 515         .level          = LEVEL_MULTIPATH,
 516         .owner          = THIS_MODULE,
 517         .make_request   = multipath_make_request,
 518         .run            = multipath_run,
 519         .free           = multipath_free,
 520         .status         = multipath_status,
 521         .error_handler  = multipath_error,
 522         .hot_add_disk   = multipath_add_disk,
 523         .hot_remove_disk= multipath_remove_disk,
 524         .size           = multipath_size,
 525         .congested      = multipath_congested,
 526 };
 527
 528 static int __init multipath_init (void)
 529 {
 530         return register_md_personality (&multipath_personality);
 531 }
 532
 533 static void __exit multipath_exit (void)
 534 {
 535         unregister_md_personality (&multipath_personality);
 536 }
 537
 538 module_init(multipath_init);
 539 module_exit(multipath_exit);
 540 MODULE_LICENSE("GPL");
 541 MODULE_DESCRIPTION("simple multi-path personality for MD");
 542 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
 543 MODULE_ALIAS("md-multipath");
 544 MODULE_ALIAS("md-level--4");