drivers/md/md-multipath.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * multipath.c : Multiple Devices driver for Linux
   4  *
   5  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   6  *
   7  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   8  *
   9  * MULTIPATH management functions.
  10  *
  11  * derived from raid1.c.
  12  */
  13
  14 #include <linux/blkdev.h>
  15 #include <linux/module.h>
  16 #include <linux/raid/md_u.h>
  17 #include <linux/seq_file.h>
  18 #include <linux/slab.h>
  19 #include "md.h"
  20 #include "md-multipath.h"
  21
  22 #define MAX_WORK_PER_DISK 128
  23
  24 #define NR_RESERVED_BUFS        32
  25
  26 static int multipath_map (struct mpconf *conf)
  27 {
  28         int i, disks = conf->raid_disks;
  29
  30         /*
  31          * Later we do read balancing on the read side
  32          * now we use the first available disk.
  33          */
  34
  35         rcu_read_lock();
  36         for (i = 0; i < disks; i++) {
  37                 struct md_rdev *rdev = rcu_dereference(conf->multipaths[i].rdev);
  38                 if (rdev && test_bit(In_sync, &rdev->flags) &&
  39                     !test_bit(Faulty, &rdev->flags)) {
  40                         atomic_inc(&rdev->nr_pending);
  41                         rcu_read_unlock();
  42                         return i;
  43                 }
  44         }
  45         rcu_read_unlock();
  46
  47         pr_crit_ratelimited("multipath_map(): no more operational IO paths?\n");
  48         return (-1);
  49 }
  50
  51 static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
  52 {
  53         unsigned long flags;
  54         struct mddev *mddev = mp_bh->mddev;
  55         struct mpconf *conf = mddev->private;
  56
  57         spin_lock_irqsave(&conf->device_lock, flags);
  58         list_add(&mp_bh->retry_list, &conf->retry_list);
  59         spin_unlock_irqrestore(&conf->device_lock, flags);
  60         md_wakeup_thread(mddev->thread);
  61 }
  62
  63 /*
  64  * multipath_end_bh_io() is called when we have finished servicing a multipathed
  65  * operation and are ready to return a success/failure code to the buffer
  66  * cache layer.
  67  */
  68 static void multipath_end_bh_io(struct multipath_bh *mp_bh, blk_status_t status)
  69 {
  70         struct bio *bio = mp_bh->master_bio;
  71         struct mpconf *conf = mp_bh->mddev->private;
  72
  73         bio->bi_status = status;
  74         bio_endio(bio);
  75         mempool_free(mp_bh, &conf->pool);
  76 }
  77
  78 static void multipath_end_request(struct bio *bio)
  79 {
  80         struct multipath_bh *mp_bh = bio->bi_private;
  81         struct mpconf *conf = mp_bh->mddev->private;
  82         struct md_rdev *rdev = conf->multipaths[mp_bh->path].rdev;
  83
  84         if (!bio->bi_status)
  85                 multipath_end_bh_io(mp_bh, 0);
  86         else if (!(bio->bi_opf & REQ_RAHEAD)) {
  87                 /*
  88                  * oops, IO error:
  89                  */
  90                 char b[BDEVNAME_SIZE];
  91                 md_error (mp_bh->mddev, rdev);
  92                 pr_info("multipath: %s: rescheduling sector %llu\n",
  93                         bdevname(rdev->bdev,b),
  94                         (unsigned long long)bio->bi_iter.bi_sector);
  95                 multipath_reschedule_retry(mp_bh);
  96         } else
  97                 multipath_end_bh_io(mp_bh, bio->bi_status);
  98         rdev_dec_pending(rdev, conf->mddev);
  99 }
 100
 101 static bool multipath_make_request(struct mddev *mddev, struct bio * bio)
 102 {
 103         struct mpconf *conf = mddev->private;
 104         struct multipath_bh * mp_bh;
 105         struct multipath_info *multipath;
 106
 107         if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
 108                 md_flush_request(mddev, bio);
 109                 return true;
 110         }
 111
 112         mp_bh = mempool_alloc(&conf->pool, GFP_NOIO);
 113
 114         mp_bh->master_bio = bio;
 115         mp_bh->mddev = mddev;
 116
 117         mp_bh->path = multipath_map(conf);
 118         if (mp_bh->path < 0) {
 119                 bio_io_error(bio);
 120                 mempool_free(mp_bh, &conf->pool);
 121                 return true;
 122         }
 123         multipath = conf->multipaths + mp_bh->path;
 124
 125         bio_init(&mp_bh->bio, NULL, 0);
 126         __bio_clone_fast(&mp_bh->bio, bio);
 127
 128         mp_bh->bio.bi_iter.bi_sector += multipath->rdev->data_offset;
 129         bio_set_dev(&mp_bh->bio, multipath->rdev->bdev);
 130         mp_bh->bio.bi_opf |= REQ_FAILFAST_TRANSPORT;
 131         mp_bh->bio.bi_end_io = multipath_end_request;
 132         mp_bh->bio.bi_private = mp_bh;
 133         mddev_check_writesame(mddev, &mp_bh->bio);
 134         mddev_check_write_zeroes(mddev, &mp_bh->bio);
 135         generic_make_request(&mp_bh->bio);
 136         return true;
 137 }
 138
 139 static void multipath_status(struct seq_file *seq, struct mddev *mddev)
 140 {
 141         struct mpconf *conf = mddev->private;
 142         int i;
 143
 144         seq_printf (seq, " [%d/%d] [", conf->raid_disks,
 145                     conf->raid_disks - mddev->degraded);
 146         rcu_read_lock();
 147         for (i = 0; i < conf->raid_disks; i++) {
 148                 struct md_rdev *rdev = rcu_dereference(conf->multipaths[i].rdev);
 149                 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
 150         }
 151         rcu_read_unlock();
 152         seq_putc(seq, ']');
 153 }
 154
 155 static int multipath_congested(struct mddev *mddev, int bits)
 156 {
 157         struct mpconf *conf = mddev->private;
 158         int i, ret = 0;
 159
 160         rcu_read_lock();
 161         for (i = 0; i < mddev->raid_disks ; i++) {
 162                 struct md_rdev *rdev = rcu_dereference(conf->multipaths[i].rdev);
 163                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 164                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 165
 166                         ret |= bdi_congested(q->backing_dev_info, bits);
 167                         /* Just like multipath_map, we just check the
 168                          * first available device
 169                          */
 170                         break;
 171                 }
 172         }
 173         rcu_read_unlock();
 174         return ret;
 175 }
 176
 177 /*
 178  * Careful, this can execute in IRQ contexts as well!
 179  */
 180 static void multipath_error (struct mddev *mddev, struct md_rdev *rdev)
 181 {
 182         struct mpconf *conf = mddev->private;
 183         char b[BDEVNAME_SIZE];
 184
 185         if (conf->raid_disks - mddev->degraded <= 1) {
 186                 /*
 187                  * Uh oh, we can do nothing if this is our last path, but
 188                  * first check if this is a queued request for a device
 189                  * which has just failed.
 190                  */
 191                 pr_warn("multipath: only one IO path left and IO error.\n");
 192                 /* leave it active... it's all we have */
 193                 return;
 194         }
 195         /*
 196          * Mark disk as unusable
 197          */
 198         if (test_and_clear_bit(In_sync, &rdev->flags)) {
 199                 unsigned long flags;
 200                 spin_lock_irqsave(&conf->device_lock, flags);
 201                 mddev->degraded++;
 202                 spin_unlock_irqrestore(&conf->device_lock, flags);
 203         }
 204         set_bit(Faulty, &rdev->flags);
 205         set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
 206         pr_err("multipath: IO failure on %s, disabling IO path.\n"
 207                "multipath: Operation continuing on %d IO paths.\n",
 208                bdevname(rdev->bdev, b),
 209                conf->raid_disks - mddev->degraded);
 210 }
 211
 212 static void print_multipath_conf (struct mpconf *conf)
 213 {
 214         int i;
 215         struct multipath_info *tmp;
 216
 217         pr_debug("MULTIPATH conf printout:\n");
 218         if (!conf) {
 219                 pr_debug("(conf==NULL)\n");
 220                 return;
 221         }
 222         pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
 223                  conf->raid_disks);
 224
 225         for (i = 0; i < conf->raid_disks; i++) {
 226                 char b[BDEVNAME_SIZE];
 227                 tmp = conf->multipaths + i;
 228                 if (tmp->rdev)
 229                         pr_debug(" disk%d, o:%d, dev:%s\n",
 230                                  i,!test_bit(Faulty, &tmp->rdev->flags),
 231                                  bdevname(tmp->rdev->bdev,b));
 232         }
 233 }
 234
 235 static int multipath_add_disk(struct mddev *mddev, struct md_rdev *rdev)
 236 {
 237         struct mpconf *conf = mddev->private;
 238         int err = -EEXIST;
 239         int path;
 240         struct multipath_info *p;
 241         int first = 0;
 242         int last = mddev->raid_disks - 1;
 243
 244         if (rdev->raid_disk >= 0)
 245                 first = last = rdev->raid_disk;
 246
 247         print_multipath_conf(conf);
 248
 249         for (path = first; path <= last; path++)
 250                 if ((p=conf->multipaths+path)->rdev == NULL) {
 251                         disk_stack_limits(mddev->gendisk, rdev->bdev,
 252                                           rdev->data_offset << 9);
 253
 254                         err = md_integrity_add_rdev(rdev, mddev);
 255                         if (err)
 256                                 break;
 257                         spin_lock_irq(&conf->device_lock);
 258                         mddev->degraded--;
 259                         rdev->raid_disk = path;
 260                         set_bit(In_sync, &rdev->flags);
 261                         spin_unlock_irq(&conf->device_lock);
 262                         rcu_assign_pointer(p->rdev, rdev);
 263                         err = 0;
 264                         break;
 265                 }
 266
 267         print_multipath_conf(conf);
 268
 269         return err;
 270 }
 271
 272 static int multipath_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
 273 {
 274         struct mpconf *conf = mddev->private;
 275         int err = 0;
 276         int number = rdev->raid_disk;
 277         struct multipath_info *p = conf->multipaths + number;
 278
 279         print_multipath_conf(conf);
 280
 281         if (rdev == p->rdev) {
 282                 if (test_bit(In_sync, &rdev->flags) ||
 283                     atomic_read(&rdev->nr_pending)) {
 284                         pr_warn("hot-remove-disk, slot %d is identified but is still operational!\n", number);
 285                         err = -EBUSY;
 286                         goto abort;
 287                 }
 288                 p->rdev = NULL;
 289                 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
 290                         synchronize_rcu();
 291                         if (atomic_read(&rdev->nr_pending)) {
 292                                 /* lost the race, try later */
 293                                 err = -EBUSY;
 294                                 p->rdev = rdev;
 295                                 goto abort;
 296                         }
 297                 }
 298                 err = md_integrity_register(mddev);
 299         }
 300 abort:
 301
 302         print_multipath_conf(conf);
 303         return err;
 304 }
 305
 306 /*
 307  * This is a kernel thread which:
 308  *
 309  *      1.      Retries failed read operations on working multipaths.
 310  *      2.      Updates the raid superblock when problems encounter.
 311  *      3.      Performs writes following reads for array syncronising.
 312  */
 313
 314 static void multipathd(struct md_thread *thread)
 315 {
 316         struct mddev *mddev = thread->mddev;
 317         struct multipath_bh *mp_bh;
 318         struct bio *bio;
 319         unsigned long flags;
 320         struct mpconf *conf = mddev->private;
 321         struct list_head *head = &conf->retry_list;
 322
 323         md_check_recovery(mddev);
 324         for (;;) {
 325                 char b[BDEVNAME_SIZE];
 326                 spin_lock_irqsave(&conf->device_lock, flags);
 327                 if (list_empty(head))
 328                         break;
 329                 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
 330                 list_del(head->prev);
 331                 spin_unlock_irqrestore(&conf->device_lock, flags);
 332
 333                 bio = &mp_bh->bio;
 334                 bio->bi_iter.bi_sector = mp_bh->master_bio->bi_iter.bi_sector;
 335
 336                 if ((mp_bh->path = multipath_map (conf))<0) {
 337                         pr_err("multipath: %s: unrecoverable IO read error for block %llu\n",
 338                                bio_devname(bio, b),
 339                                (unsigned long long)bio->bi_iter.bi_sector);
 340                         multipath_end_bh_io(mp_bh, BLK_STS_IOERR);
 341                 } else {
 342                         pr_err("multipath: %s: redirecting sector %llu to another IO path\n",
 343                                bio_devname(bio, b),
 344                                (unsigned long long)bio->bi_iter.bi_sector);
 345                         *bio = *(mp_bh->master_bio);
 346                         bio->bi_iter.bi_sector +=
 347                                 conf->multipaths[mp_bh->path].rdev->data_offset;
 348                         bio_set_dev(bio, conf->multipaths[mp_bh->path].rdev->bdev);
 349                         bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
 350                         bio->bi_end_io = multipath_end_request;
 351                         bio->bi_private = mp_bh;
 352                         generic_make_request(bio);
 353                 }
 354         }
 355         spin_unlock_irqrestore(&conf->device_lock, flags);
 356 }
 357
 358 static sector_t multipath_size(struct mddev *mddev, sector_t sectors, int raid_disks)
 359 {
 360         WARN_ONCE(sectors || raid_disks,
 361                   "%s does not support generic reshape\n", __func__);
 362
 363         return mddev->dev_sectors;
 364 }
 365
 366 static int multipath_run (struct mddev *mddev)
 367 {
 368         struct mpconf *conf;
 369         int disk_idx;
 370         struct multipath_info *disk;
 371         struct md_rdev *rdev;
 372         int working_disks;
 373         int ret;
 374
 375         if (md_check_no_bitmap(mddev))
 376                 return -EINVAL;
 377
 378         if (mddev->level != LEVEL_MULTIPATH) {
 379                 pr_warn("multipath: %s: raid level not set to multipath IO (%d)\n",
 380                         mdname(mddev), mddev->level);
 381                 goto out;
 382         }
 383         /*
 384          * copy the already verified devices into our private MULTIPATH
 385          * bookkeeping area. [whatever we allocate in multipath_run(),
 386          * should be freed in multipath_free()]
 387          */
 388
 389         conf = kzalloc(sizeof(struct mpconf), GFP_KERNEL);
 390         mddev->private = conf;
 391         if (!conf)
 392                 goto out;
 393
 394         conf->multipaths = kcalloc(mddev->raid_disks,
 395                                    sizeof(struct multipath_info),
 396                                    GFP_KERNEL);
 397         if (!conf->multipaths)
 398                 goto out_free_conf;
 399
 400         working_disks = 0;
 401         rdev_for_each(rdev, mddev) {
 402                 disk_idx = rdev->raid_disk;
 403                 if (disk_idx < 0 ||
 404                     disk_idx >= mddev->raid_disks)
 405                         continue;
 406
 407                 disk = conf->multipaths + disk_idx;
 408                 disk->rdev = rdev;
 409                 disk_stack_limits(mddev->gendisk, rdev->bdev,
 410                                   rdev->data_offset << 9);
 411
 412                 if (!test_bit(Faulty, &rdev->flags))
 413                         working_disks++;
 414         }
 415
 416         conf->raid_disks = mddev->raid_disks;
 417         conf->mddev = mddev;
 418         spin_lock_init(&conf->device_lock);
 419         INIT_LIST_HEAD(&conf->retry_list);
 420
 421         if (!working_disks) {
 422                 pr_warn("multipath: no operational IO paths for %s\n",
 423                         mdname(mddev));
 424                 goto out_free_conf;
 425         }
 426         mddev->degraded = conf->raid_disks - working_disks;
 427
 428         ret = mempool_init_kmalloc_pool(&conf->pool, NR_RESERVED_BUFS,
 429                                         sizeof(struct multipath_bh));
 430         if (ret)
 431                 goto out_free_conf;
 432
 433         mddev->thread = md_register_thread(multipathd, mddev,
 434                                            "multipath");
 435         if (!mddev->thread)
 436                 goto out_free_conf;
 437
 438         pr_info("multipath: array %s active with %d out of %d IO paths\n",
 439                 mdname(mddev), conf->raid_disks - mddev->degraded,
 440                 mddev->raid_disks);
 441         /*
 442          * Ok, everything is just fine now
 443          */
 444         md_set_array_sectors(mddev, multipath_size(mddev, 0, 0));
 445
 446         if (md_integrity_register(mddev))
 447                 goto out_free_conf;
 448
 449         return 0;
 450
 451 out_free_conf:
 452         mempool_exit(&conf->pool);
 453         kfree(conf->multipaths);
 454         kfree(conf);
 455         mddev->private = NULL;
 456 out:
 457         return -EIO;
 458 }
 459
 460 static void multipath_free(struct mddev *mddev, void *priv)
 461 {
 462         struct mpconf *conf = priv;
 463
 464         mempool_exit(&conf->pool);
 465         kfree(conf->multipaths);
 466         kfree(conf);
 467 }
 468
 469 static struct md_personality multipath_personality =
 470 {
 471         .name           = "multipath",
 472         .level          = LEVEL_MULTIPATH,
 473         .owner          = THIS_MODULE,
 474         .make_request   = multipath_make_request,
 475         .run            = multipath_run,
 476         .free           = multipath_free,
 477         .status         = multipath_status,
 478         .error_handler  = multipath_error,
 479         .hot_add_disk   = multipath_add_disk,
 480         .hot_remove_disk= multipath_remove_disk,
 481         .size           = multipath_size,
 482         .congested      = multipath_congested,
 483 };
 484
 485 static int __init multipath_init (void)
 486 {
 487         return register_md_personality (&multipath_personality);
 488 }
 489
 490 static void __exit multipath_exit (void)
 491 {
 492         unregister_md_personality (&multipath_personality);
 493 }
 494
 495 module_init(multipath_init);
 496 module_exit(multipath_exit);
 497 MODULE_LICENSE("GPL");
 498 MODULE_DESCRIPTION("simple multi-path personality for MD");
 499 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
 500 MODULE_ALIAS("md-multipath");
 501 MODULE_ALIAS("md-level--4");