2 * multipath.c : Multiple Devices driver for Linux
4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
8 * MULTIPATH management functions.
10 * derived from raid1.c.
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)
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.
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>
29 #define MAJOR_NR MD_MAJOR
31 #define MD_PERSONALITY
33 #define MAX_WORK_PER_DISK 128
35 #define NR_RESERVED_BUFS 32
38 static int multipath_map (multipath_conf_t *conf)
40 int i, disks = conf->raid_disks;
43 * Later we do read balancing on the read side
44 * now we use the first available disk.
48 for (i = 0; i < disks; i++) {
49 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
50 if (rdev && test_bit(In_sync, &rdev->flags)) {
51 atomic_inc(&rdev->nr_pending);
58 printk(KERN_ERR "multipath_map(): no more operational IO paths?\n");
62 static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
65 mddev_t *mddev = mp_bh->mddev;
66 multipath_conf_t *conf = mddev_to_conf(mddev);
68 spin_lock_irqsave(&conf->device_lock, flags);
69 list_add(&mp_bh->retry_list, &conf->retry_list);
70 spin_unlock_irqrestore(&conf->device_lock, flags);
71 md_wakeup_thread(mddev->thread);
76 * multipath_end_bh_io() is called when we have finished servicing a multipathed
77 * operation and are ready to return a success/failure code to the buffer
80 static void multipath_end_bh_io (struct multipath_bh *mp_bh, int err)
82 struct bio *bio = mp_bh->master_bio;
83 multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
86 mempool_free(mp_bh, conf->pool);
89 static void multipath_end_request(struct bio *bio, int error)
91 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
92 struct multipath_bh * mp_bh = (struct multipath_bh *)(bio->bi_private);
93 multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
94 mdk_rdev_t *rdev = conf->multipaths[mp_bh->path].rdev;
97 multipath_end_bh_io(mp_bh, 0);
98 else if (!bio_rw_ahead(bio)) {
102 char b[BDEVNAME_SIZE];
103 md_error (mp_bh->mddev, rdev);
104 printk(KERN_ERR "multipath: %s: rescheduling sector %llu\n",
105 bdevname(rdev->bdev,b),
106 (unsigned long long)bio->bi_sector);
107 multipath_reschedule_retry(mp_bh);
109 multipath_end_bh_io(mp_bh, error);
110 rdev_dec_pending(rdev, conf->mddev);
113 static void unplug_slaves(mddev_t *mddev)
115 multipath_conf_t *conf = mddev_to_conf(mddev);
119 for (i=0; i<mddev->raid_disks; i++) {
120 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
121 if (rdev && !test_bit(Faulty, &rdev->flags)
122 && atomic_read(&rdev->nr_pending)) {
123 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
125 atomic_inc(&rdev->nr_pending);
128 if (r_queue->unplug_fn)
129 r_queue->unplug_fn(r_queue);
131 rdev_dec_pending(rdev, mddev);
138 static void multipath_unplug(struct request_queue *q)
140 unplug_slaves(q->queuedata);
144 static int multipath_make_request (struct request_queue *q, struct bio * bio)
146 mddev_t *mddev = q->queuedata;
147 multipath_conf_t *conf = mddev_to_conf(mddev);
148 struct multipath_bh * mp_bh;
149 struct multipath_info *multipath;
150 const int rw = bio_data_dir(bio);
152 if (unlikely(bio_barrier(bio))) {
153 bio_endio(bio, -EOPNOTSUPP);
157 mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
159 mp_bh->master_bio = bio;
160 mp_bh->mddev = mddev;
162 disk_stat_inc(mddev->gendisk, ios[rw]);
163 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
165 mp_bh->path = multipath_map(conf);
166 if (mp_bh->path < 0) {
167 bio_endio(bio, -EIO);
168 mempool_free(mp_bh, conf->pool);
171 multipath = conf->multipaths + mp_bh->path;
174 mp_bh->bio.bi_sector += multipath->rdev->data_offset;
175 mp_bh->bio.bi_bdev = multipath->rdev->bdev;
176 mp_bh->bio.bi_rw |= (1 << BIO_RW_FAILFAST);
177 mp_bh->bio.bi_end_io = multipath_end_request;
178 mp_bh->bio.bi_private = mp_bh;
179 generic_make_request(&mp_bh->bio);
183 static void multipath_status (struct seq_file *seq, mddev_t *mddev)
185 multipath_conf_t *conf = mddev_to_conf(mddev);
188 seq_printf (seq, " [%d/%d] [", conf->raid_disks,
189 conf->working_disks);
190 for (i = 0; i < conf->raid_disks; i++)
191 seq_printf (seq, "%s",
192 conf->multipaths[i].rdev &&
193 test_bit(In_sync, &conf->multipaths[i].rdev->flags) ? "U" : "_");
194 seq_printf (seq, "]");
197 static int multipath_issue_flush(struct request_queue *q, struct gendisk *disk,
198 sector_t *error_sector)
200 mddev_t *mddev = q->queuedata;
201 multipath_conf_t *conf = mddev_to_conf(mddev);
205 for (i=0; i<mddev->raid_disks && ret == 0; i++) {
206 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
207 if (rdev && !test_bit(Faulty, &rdev->flags)) {
208 struct block_device *bdev = rdev->bdev;
209 struct request_queue *r_queue = bdev_get_queue(bdev);
211 if (!r_queue->issue_flush_fn)
214 atomic_inc(&rdev->nr_pending);
216 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
218 rdev_dec_pending(rdev, mddev);
226 static int multipath_congested(void *data, int bits)
228 mddev_t *mddev = data;
229 multipath_conf_t *conf = mddev_to_conf(mddev);
233 for (i = 0; i < mddev->raid_disks ; i++) {
234 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
235 if (rdev && !test_bit(Faulty, &rdev->flags)) {
236 struct request_queue *q = bdev_get_queue(rdev->bdev);
238 ret |= bdi_congested(&q->backing_dev_info, bits);
239 /* Just like multipath_map, we just check the
240 * first available device
250 * Careful, this can execute in IRQ contexts as well!
252 static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
254 multipath_conf_t *conf = mddev_to_conf(mddev);
256 if (conf->working_disks <= 1) {
258 * Uh oh, we can do nothing if this is our last path, but
259 * first check if this is a queued request for a device
260 * which has just failed.
263 "multipath: only one IO path left and IO error.\n");
264 /* leave it active... it's all we have */
267 * Mark disk as unusable
269 if (!test_bit(Faulty, &rdev->flags)) {
270 char b[BDEVNAME_SIZE];
271 clear_bit(In_sync, &rdev->flags);
272 set_bit(Faulty, &rdev->flags);
273 set_bit(MD_CHANGE_DEVS, &mddev->flags);
274 conf->working_disks--;
276 printk(KERN_ALERT "multipath: IO failure on %s,"
277 " disabling IO path. \n Operation continuing"
278 " on %d IO paths.\n",
279 bdevname (rdev->bdev,b),
280 conf->working_disks);
285 static void print_multipath_conf (multipath_conf_t *conf)
288 struct multipath_info *tmp;
290 printk("MULTIPATH conf printout:\n");
292 printk("(conf==NULL)\n");
295 printk(" --- wd:%d rd:%d\n", conf->working_disks,
298 for (i = 0; i < conf->raid_disks; i++) {
299 char b[BDEVNAME_SIZE];
300 tmp = conf->multipaths + i;
302 printk(" disk%d, o:%d, dev:%s\n",
303 i,!test_bit(Faulty, &tmp->rdev->flags),
304 bdevname(tmp->rdev->bdev,b));
309 static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
311 multipath_conf_t *conf = mddev->private;
312 struct request_queue *q;
315 struct multipath_info *p;
317 print_multipath_conf(conf);
319 for (path=0; path<mddev->raid_disks; path++)
320 if ((p=conf->multipaths+path)->rdev == NULL) {
321 q = rdev->bdev->bd_disk->queue;
322 blk_queue_stack_limits(mddev->queue, q);
324 /* as we don't honour merge_bvec_fn, we must never risk
325 * violating it, so limit ->max_sector to one PAGE, as
326 * a one page request is never in violation.
327 * (Note: it is very unlikely that a device with
328 * merge_bvec_fn will be involved in multipath.)
330 if (q->merge_bvec_fn &&
331 mddev->queue->max_sectors > (PAGE_SIZE>>9))
332 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
334 conf->working_disks++;
336 rdev->raid_disk = path;
337 set_bit(In_sync, &rdev->flags);
338 rcu_assign_pointer(p->rdev, rdev);
342 print_multipath_conf(conf);
346 static int multipath_remove_disk(mddev_t *mddev, int number)
348 multipath_conf_t *conf = mddev->private;
351 struct multipath_info *p = conf->multipaths + number;
353 print_multipath_conf(conf);
357 if (test_bit(In_sync, &rdev->flags) ||
358 atomic_read(&rdev->nr_pending)) {
359 printk(KERN_ERR "hot-remove-disk, slot %d is identified" " but is still operational!\n", number);
365 if (atomic_read(&rdev->nr_pending)) {
366 /* lost the race, try later */
373 print_multipath_conf(conf);
380 * This is a kernel thread which:
382 * 1. Retries failed read operations on working multipaths.
383 * 2. Updates the raid superblock when problems encounter.
384 * 3. Performs writes following reads for array syncronising.
387 static void multipathd (mddev_t *mddev)
389 struct multipath_bh *mp_bh;
392 multipath_conf_t *conf = mddev_to_conf(mddev);
393 struct list_head *head = &conf->retry_list;
395 md_check_recovery(mddev);
397 char b[BDEVNAME_SIZE];
398 spin_lock_irqsave(&conf->device_lock, flags);
399 if (list_empty(head))
401 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
402 list_del(head->prev);
403 spin_unlock_irqrestore(&conf->device_lock, flags);
406 bio->bi_sector = mp_bh->master_bio->bi_sector;
408 if ((mp_bh->path = multipath_map (conf))<0) {
409 printk(KERN_ALERT "multipath: %s: unrecoverable IO read"
410 " error for block %llu\n",
411 bdevname(bio->bi_bdev,b),
412 (unsigned long long)bio->bi_sector);
413 multipath_end_bh_io(mp_bh, -EIO);
415 printk(KERN_ERR "multipath: %s: redirecting sector %llu"
416 " to another IO path\n",
417 bdevname(bio->bi_bdev,b),
418 (unsigned long long)bio->bi_sector);
419 *bio = *(mp_bh->master_bio);
420 bio->bi_sector += conf->multipaths[mp_bh->path].rdev->data_offset;
421 bio->bi_bdev = conf->multipaths[mp_bh->path].rdev->bdev;
422 bio->bi_rw |= (1 << BIO_RW_FAILFAST);
423 bio->bi_end_io = multipath_end_request;
424 bio->bi_private = mp_bh;
425 generic_make_request(bio);
428 spin_unlock_irqrestore(&conf->device_lock, flags);
431 static int multipath_run (mddev_t *mddev)
433 multipath_conf_t *conf;
435 struct multipath_info *disk;
437 struct list_head *tmp;
439 if (mddev->level != LEVEL_MULTIPATH) {
440 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
441 mdname(mddev), mddev->level);
445 * copy the already verified devices into our private MULTIPATH
446 * bookkeeping area. [whatever we allocate in multipath_run(),
447 * should be freed in multipath_stop()]
450 conf = kzalloc(sizeof(multipath_conf_t), GFP_KERNEL);
451 mddev->private = conf;
454 "multipath: couldn't allocate memory for %s\n",
459 conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
461 if (!conf->multipaths) {
463 "multipath: couldn't allocate memory for %s\n",
468 conf->working_disks = 0;
469 ITERATE_RDEV(mddev,rdev,tmp) {
470 disk_idx = rdev->raid_disk;
472 disk_idx >= mddev->raid_disks)
475 disk = conf->multipaths + disk_idx;
478 blk_queue_stack_limits(mddev->queue,
479 rdev->bdev->bd_disk->queue);
480 /* as we don't honour merge_bvec_fn, we must never risk
481 * violating it, not that we ever expect a device with
482 * a merge_bvec_fn to be involved in multipath */
483 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
484 mddev->queue->max_sectors > (PAGE_SIZE>>9))
485 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
487 if (!test_bit(Faulty, &rdev->flags))
488 conf->working_disks++;
491 conf->raid_disks = mddev->raid_disks;
493 spin_lock_init(&conf->device_lock);
494 INIT_LIST_HEAD(&conf->retry_list);
496 if (!conf->working_disks) {
497 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
501 mddev->degraded = conf->raid_disks - conf->working_disks;
503 conf->pool = mempool_create_kzalloc_pool(NR_RESERVED_BUFS,
504 sizeof(struct multipath_bh));
505 if (conf->pool == NULL) {
507 "multipath: couldn't allocate memory for %s\n",
513 mddev->thread = md_register_thread(multipathd, mddev, "%s_multipath");
514 if (!mddev->thread) {
515 printk(KERN_ERR "multipath: couldn't allocate thread"
516 " for %s\n", mdname(mddev));
522 "multipath: array %s active with %d out of %d IO paths\n",
523 mdname(mddev), conf->working_disks, mddev->raid_disks);
525 * Ok, everything is just fine now
527 mddev->array_size = mddev->size;
529 mddev->queue->unplug_fn = multipath_unplug;
530 mddev->queue->issue_flush_fn = multipath_issue_flush;
531 mddev->queue->backing_dev_info.congested_fn = multipath_congested;
532 mddev->queue->backing_dev_info.congested_data = mddev;
538 mempool_destroy(conf->pool);
539 kfree(conf->multipaths);
541 mddev->private = NULL;
547 static int multipath_stop (mddev_t *mddev)
549 multipath_conf_t *conf = mddev_to_conf(mddev);
551 md_unregister_thread(mddev->thread);
552 mddev->thread = NULL;
553 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
554 mempool_destroy(conf->pool);
555 kfree(conf->multipaths);
557 mddev->private = NULL;
561 static struct mdk_personality multipath_personality =
564 .level = LEVEL_MULTIPATH,
565 .owner = THIS_MODULE,
566 .make_request = multipath_make_request,
567 .run = multipath_run,
568 .stop = multipath_stop,
569 .status = multipath_status,
570 .error_handler = multipath_error,
571 .hot_add_disk = multipath_add_disk,
572 .hot_remove_disk= multipath_remove_disk,
575 static int __init multipath_init (void)
577 return register_md_personality (&multipath_personality);
580 static void __exit multipath_exit (void)
582 unregister_md_personality (&multipath_personality);
585 module_init(multipath_init);
586 module_exit(multipath_exit);
587 MODULE_LICENSE("GPL");
588 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
589 MODULE_ALIAS("md-multipath");
590 MODULE_ALIAS("md-level--4");