blob: 97a56aaaef6d90c56e211c7b193530fb93c561a4 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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
38static mdk_personality_t multipath_personality;
39
40
Al Virodd0fc662005-10-07 07:46:04 +010041static void *mp_pool_alloc(gfp_t gfp_flags, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -070042{
43 struct multipath_bh *mpb;
NeilBrown9ffae0c2006-01-06 00:20:32 -080044 mpb = kzalloc(sizeof(*mpb), gfp_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -070045 return mpb;
46}
47
48static void mp_pool_free(void *mpb, void *data)
49{
50 kfree(mpb);
51}
52
53static int multipath_map (multipath_conf_t *conf)
54{
55 int i, disks = conf->raid_disks;
56
57 /*
58 * Later we do read balancing on the read side
59 * now we use the first available disk.
60 */
61
62 rcu_read_lock();
63 for (i = 0; i < disks; i++) {
Suzanne Woodd6065f72005-11-08 21:39:27 -080064 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
NeilBrownb2d444d2005-11-08 21:39:31 -080065 if (rdev && test_bit(In_sync, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -070066 atomic_inc(&rdev->nr_pending);
67 rcu_read_unlock();
68 return i;
69 }
70 }
71 rcu_read_unlock();
72
73 printk(KERN_ERR "multipath_map(): no more operational IO paths?\n");
74 return (-1);
75}
76
77static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
78{
79 unsigned long flags;
80 mddev_t *mddev = mp_bh->mddev;
81 multipath_conf_t *conf = mddev_to_conf(mddev);
82
83 spin_lock_irqsave(&conf->device_lock, flags);
84 list_add(&mp_bh->retry_list, &conf->retry_list);
85 spin_unlock_irqrestore(&conf->device_lock, flags);
86 md_wakeup_thread(mddev->thread);
87}
88
89
90/*
91 * multipath_end_bh_io() is called when we have finished servicing a multipathed
92 * operation and are ready to return a success/failure code to the buffer
93 * cache layer.
94 */
95static void multipath_end_bh_io (struct multipath_bh *mp_bh, int err)
96{
97 struct bio *bio = mp_bh->master_bio;
98 multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
99
100 bio_endio(bio, bio->bi_size, err);
101 mempool_free(mp_bh, conf->pool);
102}
103
Adrian Bunk75c96f82005-05-05 16:16:09 -0700104static int multipath_end_request(struct bio *bio, unsigned int bytes_done,
105 int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106{
107 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
108 struct multipath_bh * mp_bh = (struct multipath_bh *)(bio->bi_private);
109 multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
110 mdk_rdev_t *rdev = conf->multipaths[mp_bh->path].rdev;
111
112 if (bio->bi_size)
113 return 1;
114
115 if (uptodate)
116 multipath_end_bh_io(mp_bh, 0);
117 else if (!bio_rw_ahead(bio)) {
118 /*
119 * oops, IO error:
120 */
121 char b[BDEVNAME_SIZE];
122 md_error (mp_bh->mddev, rdev);
123 printk(KERN_ERR "multipath: %s: rescheduling sector %llu\n",
124 bdevname(rdev->bdev,b),
125 (unsigned long long)bio->bi_sector);
126 multipath_reschedule_retry(mp_bh);
127 } else
128 multipath_end_bh_io(mp_bh, error);
129 rdev_dec_pending(rdev, conf->mddev);
130 return 0;
131}
132
133static void unplug_slaves(mddev_t *mddev)
134{
135 multipath_conf_t *conf = mddev_to_conf(mddev);
136 int i;
137
138 rcu_read_lock();
139 for (i=0; i<mddev->raid_disks; i++) {
Suzanne Woodd6065f72005-11-08 21:39:27 -0800140 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
NeilBrownb2d444d2005-11-08 21:39:31 -0800141 if (rdev && !test_bit(Faulty, &rdev->flags)
142 && atomic_read(&rdev->nr_pending)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143 request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
144
145 atomic_inc(&rdev->nr_pending);
146 rcu_read_unlock();
147
148 if (r_queue->unplug_fn)
149 r_queue->unplug_fn(r_queue);
150
151 rdev_dec_pending(rdev, mddev);
152 rcu_read_lock();
153 }
154 }
155 rcu_read_unlock();
156}
157
158static void multipath_unplug(request_queue_t *q)
159{
160 unplug_slaves(q->queuedata);
161}
162
163
164static int multipath_make_request (request_queue_t *q, struct bio * bio)
165{
166 mddev_t *mddev = q->queuedata;
167 multipath_conf_t *conf = mddev_to_conf(mddev);
168 struct multipath_bh * mp_bh;
169 struct multipath_info *multipath;
Jens Axboea3623572005-11-01 09:26:16 +0100170 const int rw = bio_data_dir(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171
NeilBrowne5dcdd82005-09-09 16:23:41 -0700172 if (unlikely(bio_barrier(bio))) {
173 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
174 return 0;
175 }
176
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177 mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
178
179 mp_bh->master_bio = bio;
180 mp_bh->mddev = mddev;
181
Jens Axboea3623572005-11-01 09:26:16 +0100182 disk_stat_inc(mddev->gendisk, ios[rw]);
183 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184
185 mp_bh->path = multipath_map(conf);
186 if (mp_bh->path < 0) {
187 bio_endio(bio, bio->bi_size, -EIO);
188 mempool_free(mp_bh, conf->pool);
189 return 0;
190 }
191 multipath = conf->multipaths + mp_bh->path;
192
193 mp_bh->bio = *bio;
194 mp_bh->bio.bi_sector += multipath->rdev->data_offset;
195 mp_bh->bio.bi_bdev = multipath->rdev->bdev;
196 mp_bh->bio.bi_rw |= (1 << BIO_RW_FAILFAST);
197 mp_bh->bio.bi_end_io = multipath_end_request;
198 mp_bh->bio.bi_private = mp_bh;
199 generic_make_request(&mp_bh->bio);
200 return 0;
201}
202
203static void multipath_status (struct seq_file *seq, mddev_t *mddev)
204{
205 multipath_conf_t *conf = mddev_to_conf(mddev);
206 int i;
207
208 seq_printf (seq, " [%d/%d] [", conf->raid_disks,
209 conf->working_disks);
210 for (i = 0; i < conf->raid_disks; i++)
211 seq_printf (seq, "%s",
212 conf->multipaths[i].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -0800213 test_bit(In_sync, &conf->multipaths[i].rdev->flags) ? "U" : "_");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214 seq_printf (seq, "]");
215}
216
217static int multipath_issue_flush(request_queue_t *q, struct gendisk *disk,
218 sector_t *error_sector)
219{
220 mddev_t *mddev = q->queuedata;
221 multipath_conf_t *conf = mddev_to_conf(mddev);
222 int i, ret = 0;
223
224 rcu_read_lock();
225 for (i=0; i<mddev->raid_disks && ret == 0; i++) {
Suzanne Woodd6065f72005-11-08 21:39:27 -0800226 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
NeilBrownb2d444d2005-11-08 21:39:31 -0800227 if (rdev && !test_bit(Faulty, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228 struct block_device *bdev = rdev->bdev;
229 request_queue_t *r_queue = bdev_get_queue(bdev);
230
231 if (!r_queue->issue_flush_fn)
232 ret = -EOPNOTSUPP;
233 else {
234 atomic_inc(&rdev->nr_pending);
235 rcu_read_unlock();
236 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
237 error_sector);
238 rdev_dec_pending(rdev, mddev);
239 rcu_read_lock();
240 }
241 }
242 }
243 rcu_read_unlock();
244 return ret;
245}
246
247/*
248 * Careful, this can execute in IRQ contexts as well!
249 */
250static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
251{
252 multipath_conf_t *conf = mddev_to_conf(mddev);
253
254 if (conf->working_disks <= 1) {
255 /*
256 * Uh oh, we can do nothing if this is our last path, but
257 * first check if this is a queued request for a device
258 * which has just failed.
259 */
260 printk(KERN_ALERT
261 "multipath: only one IO path left and IO error.\n");
262 /* leave it active... it's all we have */
263 } else {
264 /*
265 * Mark disk as unusable
266 */
NeilBrownb2d444d2005-11-08 21:39:31 -0800267 if (!test_bit(Faulty, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268 char b[BDEVNAME_SIZE];
NeilBrownb2d444d2005-11-08 21:39:31 -0800269 clear_bit(In_sync, &rdev->flags);
270 set_bit(Faulty, &rdev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271 mddev->sb_dirty = 1;
272 conf->working_disks--;
273 printk(KERN_ALERT "multipath: IO failure on %s,"
274 " disabling IO path. \n Operation continuing"
275 " on %d IO paths.\n",
276 bdevname (rdev->bdev,b),
277 conf->working_disks);
278 }
279 }
280}
281
282static void print_multipath_conf (multipath_conf_t *conf)
283{
284 int i;
285 struct multipath_info *tmp;
286
287 printk("MULTIPATH conf printout:\n");
288 if (!conf) {
289 printk("(conf==NULL)\n");
290 return;
291 }
292 printk(" --- wd:%d rd:%d\n", conf->working_disks,
293 conf->raid_disks);
294
295 for (i = 0; i < conf->raid_disks; i++) {
296 char b[BDEVNAME_SIZE];
297 tmp = conf->multipaths + i;
298 if (tmp->rdev)
299 printk(" disk%d, o:%d, dev:%s\n",
NeilBrownb2d444d2005-11-08 21:39:31 -0800300 i,!test_bit(Faulty, &tmp->rdev->flags),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700301 bdevname(tmp->rdev->bdev,b));
302 }
303}
304
305
306static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
307{
308 multipath_conf_t *conf = mddev->private;
309 int found = 0;
310 int path;
311 struct multipath_info *p;
312
313 print_multipath_conf(conf);
314
315 for (path=0; path<mddev->raid_disks; path++)
316 if ((p=conf->multipaths+path)->rdev == NULL) {
317 blk_queue_stack_limits(mddev->queue,
318 rdev->bdev->bd_disk->queue);
319
320 /* as we don't honour merge_bvec_fn, we must never risk
321 * violating it, so limit ->max_sector to one PAGE, as
322 * a one page request is never in violation.
323 * (Note: it is very unlikely that a device with
324 * merge_bvec_fn will be involved in multipath.)
325 */
326 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
327 mddev->queue->max_sectors > (PAGE_SIZE>>9))
328 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
329
330 conf->working_disks++;
331 rdev->raid_disk = path;
NeilBrownb2d444d2005-11-08 21:39:31 -0800332 set_bit(In_sync, &rdev->flags);
Suzanne Woodd6065f72005-11-08 21:39:27 -0800333 rcu_assign_pointer(p->rdev, rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334 found = 1;
335 }
336
337 print_multipath_conf(conf);
338 return found;
339}
340
341static int multipath_remove_disk(mddev_t *mddev, int number)
342{
343 multipath_conf_t *conf = mddev->private;
344 int err = 0;
345 mdk_rdev_t *rdev;
346 struct multipath_info *p = conf->multipaths + number;
347
348 print_multipath_conf(conf);
349
350 rdev = p->rdev;
351 if (rdev) {
NeilBrownb2d444d2005-11-08 21:39:31 -0800352 if (test_bit(In_sync, &rdev->flags) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353 atomic_read(&rdev->nr_pending)) {
354 printk(KERN_ERR "hot-remove-disk, slot %d is identified" " but is still operational!\n", number);
355 err = -EBUSY;
356 goto abort;
357 }
358 p->rdev = NULL;
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -0700359 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360 if (atomic_read(&rdev->nr_pending)) {
361 /* lost the race, try later */
362 err = -EBUSY;
363 p->rdev = rdev;
364 }
365 }
366abort:
367
368 print_multipath_conf(conf);
369 return err;
370}
371
372
373
374/*
375 * This is a kernel thread which:
376 *
377 * 1. Retries failed read operations on working multipaths.
378 * 2. Updates the raid superblock when problems encounter.
379 * 3. Performs writes following reads for array syncronising.
380 */
381
382static void multipathd (mddev_t *mddev)
383{
384 struct multipath_bh *mp_bh;
385 struct bio *bio;
386 unsigned long flags;
387 multipath_conf_t *conf = mddev_to_conf(mddev);
388 struct list_head *head = &conf->retry_list;
389
390 md_check_recovery(mddev);
391 for (;;) {
392 char b[BDEVNAME_SIZE];
393 spin_lock_irqsave(&conf->device_lock, flags);
394 if (list_empty(head))
395 break;
396 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
397 list_del(head->prev);
398 spin_unlock_irqrestore(&conf->device_lock, flags);
399
400 bio = &mp_bh->bio;
401 bio->bi_sector = mp_bh->master_bio->bi_sector;
402
403 if ((mp_bh->path = multipath_map (conf))<0) {
404 printk(KERN_ALERT "multipath: %s: unrecoverable IO read"
405 " error for block %llu\n",
406 bdevname(bio->bi_bdev,b),
407 (unsigned long long)bio->bi_sector);
408 multipath_end_bh_io(mp_bh, -EIO);
409 } else {
410 printk(KERN_ERR "multipath: %s: redirecting sector %llu"
411 " to another IO path\n",
412 bdevname(bio->bi_bdev,b),
413 (unsigned long long)bio->bi_sector);
414 *bio = *(mp_bh->master_bio);
415 bio->bi_sector += conf->multipaths[mp_bh->path].rdev->data_offset;
416 bio->bi_bdev = conf->multipaths[mp_bh->path].rdev->bdev;
417 bio->bi_rw |= (1 << BIO_RW_FAILFAST);
418 bio->bi_end_io = multipath_end_request;
419 bio->bi_private = mp_bh;
420 generic_make_request(bio);
421 }
422 }
423 spin_unlock_irqrestore(&conf->device_lock, flags);
424}
425
426static 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 struct list_head *tmp;
433
434 if (mddev->level != LEVEL_MULTIPATH) {
435 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
436 mdname(mddev), mddev->level);
437 goto out;
438 }
439 /*
440 * copy the already verified devices into our private MULTIPATH
441 * bookkeeping area. [whatever we allocate in multipath_run(),
442 * should be freed in multipath_stop()]
443 */
444
NeilBrown9ffae0c2006-01-06 00:20:32 -0800445 conf = kzalloc(sizeof(multipath_conf_t), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 mddev->private = conf;
447 if (!conf) {
448 printk(KERN_ERR
449 "multipath: couldn't allocate memory for %s\n",
450 mdname(mddev));
451 goto out;
452 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453
NeilBrown9ffae0c2006-01-06 00:20:32 -0800454 conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455 GFP_KERNEL);
456 if (!conf->multipaths) {
457 printk(KERN_ERR
458 "multipath: couldn't allocate memory for %s\n",
459 mdname(mddev));
460 goto out_free_conf;
461 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463 conf->working_disks = 0;
464 ITERATE_RDEV(mddev,rdev,tmp) {
465 disk_idx = rdev->raid_disk;
466 if (disk_idx < 0 ||
467 disk_idx >= mddev->raid_disks)
468 continue;
469
470 disk = conf->multipaths + disk_idx;
471 disk->rdev = rdev;
472
473 blk_queue_stack_limits(mddev->queue,
474 rdev->bdev->bd_disk->queue);
475 /* as we don't honour merge_bvec_fn, we must never risk
476 * violating it, not that we ever expect a device with
477 * a merge_bvec_fn to be involved in multipath */
478 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
479 mddev->queue->max_sectors > (PAGE_SIZE>>9))
480 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
481
NeilBrownb2d444d2005-11-08 21:39:31 -0800482 if (!test_bit(Faulty, &rdev->flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483 conf->working_disks++;
484 }
485
486 conf->raid_disks = mddev->raid_disks;
487 mddev->sb_dirty = 1;
488 conf->mddev = mddev;
489 spin_lock_init(&conf->device_lock);
490 INIT_LIST_HEAD(&conf->retry_list);
491
492 if (!conf->working_disks) {
493 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
494 mdname(mddev));
495 goto out_free_conf;
496 }
497 mddev->degraded = conf->raid_disks = conf->working_disks;
498
499 conf->pool = mempool_create(NR_RESERVED_BUFS,
500 mp_pool_alloc, mp_pool_free,
501 NULL);
502 if (conf->pool == NULL) {
503 printk(KERN_ERR
504 "multipath: couldn't allocate memory for %s\n",
505 mdname(mddev));
506 goto out_free_conf;
507 }
508
509 {
510 mddev->thread = md_register_thread(multipathd, mddev, "%s_multipath");
511 if (!mddev->thread) {
512 printk(KERN_ERR "multipath: couldn't allocate thread"
513 " for %s\n", mdname(mddev));
514 goto out_free_conf;
515 }
516 }
517
518 printk(KERN_INFO
519 "multipath: array %s active with %d out of %d IO paths\n",
520 mdname(mddev), conf->working_disks, mddev->raid_disks);
521 /*
522 * Ok, everything is just fine now
523 */
524 mddev->array_size = mddev->size;
NeilBrown7a5febe2005-05-16 21:53:16 -0700525
526 mddev->queue->unplug_fn = multipath_unplug;
527 mddev->queue->issue_flush_fn = multipath_issue_flush;
528
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529 return 0;
530
531out_free_conf:
532 if (conf->pool)
533 mempool_destroy(conf->pool);
Jesper Juhl990a8ba2005-06-21 17:17:30 -0700534 kfree(conf->multipaths);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535 kfree(conf);
536 mddev->private = NULL;
537out:
538 return -EIO;
539}
540
541
542static int multipath_stop (mddev_t *mddev)
543{
544 multipath_conf_t *conf = mddev_to_conf(mddev);
545
546 md_unregister_thread(mddev->thread);
547 mddev->thread = NULL;
548 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
549 mempool_destroy(conf->pool);
550 kfree(conf->multipaths);
551 kfree(conf);
552 mddev->private = NULL;
553 return 0;
554}
555
556static mdk_personality_t multipath_personality=
557{
558 .name = "multipath",
559 .owner = THIS_MODULE,
560 .make_request = multipath_make_request,
561 .run = multipath_run,
562 .stop = multipath_stop,
563 .status = multipath_status,
564 .error_handler = multipath_error,
565 .hot_add_disk = multipath_add_disk,
566 .hot_remove_disk= multipath_remove_disk,
567};
568
569static int __init multipath_init (void)
570{
571 return register_md_personality (MULTIPATH, &multipath_personality);
572}
573
574static void __exit multipath_exit (void)
575{
576 unregister_md_personality (MULTIPATH);
577}
578
579module_init(multipath_init);
580module_exit(multipath_exit);
581MODULE_LICENSE("GPL");
582MODULE_ALIAS("md-personality-7"); /* MULTIPATH */