blob: 410fb60699ac1d2b2ca4af62bca7ed89b2af60d7 [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
NeilBrownbff61972009-03-31 14:33:13 +110022#include <linux/blkdev.h>
23#include <linux/raid/md_u.h>
NeilBrownbff61972009-03-31 14:33:13 +110024#include <linux/seq_file.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090025#include <linux/slab.h>
NeilBrown43b2e5d2009-03-31 14:33:13 +110026#include "md.h"
Christoph Hellwigef740c32009-03-31 14:27:03 +110027#include "multipath.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070028
29#define MAX_WORK_PER_DISK 128
30
31#define NR_RESERVED_BUFS 32
32
33
Linus Torvalds1da177e2005-04-16 15:20:36 -070034static int multipath_map (multipath_conf_t *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++) {
Suzanne Woodd6065f72005-11-08 21:39:27 -080045 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
NeilBrownb2d444d2005-11-08 21:39:31 -080046 if (rdev && test_bit(In_sync, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -070047 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
58static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
59{
60 unsigned long flags;
61 mddev_t *mddev = mp_bh->mddev;
NeilBrown070ec552009-06-16 16:54:21 +100062 multipath_conf_t *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -070063
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/*
72 * multipath_end_bh_io() is called when we have finished servicing a multipathed
73 * operation and are ready to return a success/failure code to the buffer
74 * cache layer.
75 */
76static void multipath_end_bh_io (struct multipath_bh *mp_bh, int err)
77{
78 struct bio *bio = mp_bh->master_bio;
NeilBrown070ec552009-06-16 16:54:21 +100079 multipath_conf_t *conf = mp_bh->mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -070080
NeilBrown6712ecf2007-09-27 12:47:43 +020081 bio_endio(bio, err);
Linus Torvalds1da177e2005-04-16 15:20:36 -070082 mempool_free(mp_bh, conf->pool);
83}
84
NeilBrown6712ecf2007-09-27 12:47:43 +020085static void multipath_end_request(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -070086{
87 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
H Hartley Sweeten7b928132010-03-08 16:02:40 +110088 struct multipath_bh *mp_bh = bio->bi_private;
NeilBrown070ec552009-06-16 16:54:21 +100089 multipath_conf_t *conf = mp_bh->mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -070090 mdk_rdev_t *rdev = conf->multipaths[mp_bh->path].rdev;
91
Linus Torvalds1da177e2005-04-16 15:20:36 -070092 if (uptodate)
93 multipath_end_bh_io(mp_bh, 0);
Jens Axboe1f98a132009-09-11 14:32:04 +020094 else if (!bio_rw_flagged(bio, BIO_RW_AHEAD)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -070095 /*
96 * oops, IO error:
97 */
98 char b[BDEVNAME_SIZE];
99 md_error (mp_bh->mddev, rdev);
100 printk(KERN_ERR "multipath: %s: rescheduling sector %llu\n",
101 bdevname(rdev->bdev,b),
102 (unsigned long long)bio->bi_sector);
103 multipath_reschedule_retry(mp_bh);
104 } else
105 multipath_end_bh_io(mp_bh, error);
106 rdev_dec_pending(rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107}
108
109static void unplug_slaves(mddev_t *mddev)
110{
NeilBrown070ec552009-06-16 16:54:21 +1000111 multipath_conf_t *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700112 int i;
113
114 rcu_read_lock();
115 for (i=0; i<mddev->raid_disks; i++) {
Suzanne Woodd6065f72005-11-08 21:39:27 -0800116 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
NeilBrownb2d444d2005-11-08 21:39:31 -0800117 if (rdev && !test_bit(Faulty, &rdev->flags)
118 && atomic_read(&rdev->nr_pending)) {
Jens Axboe165125e2007-07-24 09:28:11 +0200119 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120
121 atomic_inc(&rdev->nr_pending);
122 rcu_read_unlock();
123
Alan D. Brunelle2ad8b1e2007-11-07 14:26:56 -0500124 blk_unplug(r_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125
126 rdev_dec_pending(rdev, mddev);
127 rcu_read_lock();
128 }
129 }
130 rcu_read_unlock();
131}
132
Jens Axboe165125e2007-07-24 09:28:11 +0200133static void multipath_unplug(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134{
135 unplug_slaves(q->queuedata);
136}
137
138
NeilBrown21a52c62010-04-01 15:02:13 +1100139static int multipath_make_request(mddev_t *mddev, struct bio * bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700140{
NeilBrown070ec552009-06-16 16:54:21 +1000141 multipath_conf_t *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142 struct multipath_bh * mp_bh;
143 struct multipath_info *multipath;
144
Jens Axboe1f98a132009-09-11 14:32:04 +0200145 if (unlikely(bio_rw_flagged(bio, BIO_RW_BARRIER))) {
NeilBrowna2826aa2009-12-14 12:49:49 +1100146 md_barrier_request(mddev, bio);
NeilBrowne5dcdd82005-09-09 16:23:41 -0700147 return 0;
148 }
149
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150 mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
151
152 mp_bh->master_bio = bio;
153 mp_bh->mddev = mddev;
154
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155 mp_bh->path = multipath_map(conf);
156 if (mp_bh->path < 0) {
NeilBrown6712ecf2007-09-27 12:47:43 +0200157 bio_endio(bio, -EIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158 mempool_free(mp_bh, conf->pool);
159 return 0;
160 }
161 multipath = conf->multipaths + mp_bh->path;
162
163 mp_bh->bio = *bio;
164 mp_bh->bio.bi_sector += multipath->rdev->data_offset;
165 mp_bh->bio.bi_bdev = multipath->rdev->bdev;
Mike Christie6000a362008-08-19 18:45:30 -0500166 mp_bh->bio.bi_rw |= (1 << BIO_RW_FAILFAST_TRANSPORT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167 mp_bh->bio.bi_end_io = multipath_end_request;
168 mp_bh->bio.bi_private = mp_bh;
169 generic_make_request(&mp_bh->bio);
170 return 0;
171}
172
173static void multipath_status (struct seq_file *seq, mddev_t *mddev)
174{
NeilBrown070ec552009-06-16 16:54:21 +1000175 multipath_conf_t *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176 int i;
177
178 seq_printf (seq, " [%d/%d] [", conf->raid_disks,
179 conf->working_disks);
180 for (i = 0; i < conf->raid_disks; i++)
181 seq_printf (seq, "%s",
182 conf->multipaths[i].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -0800183 test_bit(In_sync, &conf->multipaths[i].rdev->flags) ? "U" : "_");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184 seq_printf (seq, "]");
185}
186
NeilBrown0d129222006-10-03 01:15:54 -0700187static int multipath_congested(void *data, int bits)
188{
189 mddev_t *mddev = data;
NeilBrown070ec552009-06-16 16:54:21 +1000190 multipath_conf_t *conf = mddev->private;
NeilBrown0d129222006-10-03 01:15:54 -0700191 int i, ret = 0;
192
NeilBrown3fa841d2009-09-23 18:10:29 +1000193 if (mddev_congested(mddev, bits))
194 return 1;
195
NeilBrown0d129222006-10-03 01:15:54 -0700196 rcu_read_lock();
197 for (i = 0; i < mddev->raid_disks ; i++) {
198 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
199 if (rdev && !test_bit(Faulty, &rdev->flags)) {
Jens Axboe165125e2007-07-24 09:28:11 +0200200 struct request_queue *q = bdev_get_queue(rdev->bdev);
NeilBrown0d129222006-10-03 01:15:54 -0700201
202 ret |= bdi_congested(&q->backing_dev_info, bits);
203 /* Just like multipath_map, we just check the
204 * first available device
205 */
206 break;
207 }
208 }
209 rcu_read_unlock();
210 return ret;
211}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212
213/*
214 * Careful, this can execute in IRQ contexts as well!
215 */
216static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
217{
NeilBrown070ec552009-06-16 16:54:21 +1000218 multipath_conf_t *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219
220 if (conf->working_disks <= 1) {
221 /*
222 * Uh oh, we can do nothing if this is our last path, but
223 * first check if this is a queued request for a device
224 * which has just failed.
225 */
226 printk(KERN_ALERT
227 "multipath: only one IO path left and IO error.\n");
228 /* leave it active... it's all we have */
229 } else {
230 /*
231 * Mark disk as unusable
232 */
NeilBrownb2d444d2005-11-08 21:39:31 -0800233 if (!test_bit(Faulty, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234 char b[BDEVNAME_SIZE];
NeilBrownb2d444d2005-11-08 21:39:31 -0800235 clear_bit(In_sync, &rdev->flags);
236 set_bit(Faulty, &rdev->flags);
NeilBrown850b2b42006-10-03 01:15:46 -0700237 set_bit(MD_CHANGE_DEVS, &mddev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700238 conf->working_disks--;
NeilBrown750a8f32006-10-28 10:38:31 -0700239 mddev->degraded++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700240 printk(KERN_ALERT "multipath: IO failure on %s,"
Nick Andrewd7a420c2008-04-28 02:15:55 -0700241 " disabling IO path.\n"
242 "multipath: Operation continuing"
Linus Torvalds1da177e2005-04-16 15:20:36 -0700243 " on %d IO paths.\n",
244 bdevname (rdev->bdev,b),
245 conf->working_disks);
246 }
247 }
248}
249
250static void print_multipath_conf (multipath_conf_t *conf)
251{
252 int i;
253 struct multipath_info *tmp;
254
255 printk("MULTIPATH conf printout:\n");
256 if (!conf) {
257 printk("(conf==NULL)\n");
258 return;
259 }
260 printk(" --- wd:%d rd:%d\n", conf->working_disks,
261 conf->raid_disks);
262
263 for (i = 0; i < conf->raid_disks; i++) {
264 char b[BDEVNAME_SIZE];
265 tmp = conf->multipaths + i;
266 if (tmp->rdev)
267 printk(" disk%d, o:%d, dev:%s\n",
NeilBrownb2d444d2005-11-08 21:39:31 -0800268 i,!test_bit(Faulty, &tmp->rdev->flags),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269 bdevname(tmp->rdev->bdev,b));
270 }
271}
272
273
274static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
275{
276 multipath_conf_t *conf = mddev->private;
Jesper Juhl352d7682006-01-09 20:54:44 -0800277 struct request_queue *q;
Neil Brown199050e2008-06-28 08:31:33 +1000278 int err = -EEXIST;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279 int path;
280 struct multipath_info *p;
Neil Brown6c2fce22008-06-28 08:31:31 +1000281 int first = 0;
282 int last = mddev->raid_disks - 1;
283
284 if (rdev->raid_disk >= 0)
285 first = last = rdev->raid_disk;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286
287 print_multipath_conf(conf);
288
Neil Brown6c2fce22008-06-28 08:31:31 +1000289 for (path = first; path <= last; path++)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290 if ((p=conf->multipaths+path)->rdev == NULL) {
Jesper Juhl352d7682006-01-09 20:54:44 -0800291 q = rdev->bdev->bd_disk->queue;
Martin K. Petersen8f6c2e42009-07-01 11:13:45 +1000292 disk_stack_limits(mddev->gendisk, rdev->bdev,
293 rdev->data_offset << 9);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294
295 /* as we don't honour merge_bvec_fn, we must never risk
NeilBrown627a2d32010-03-08 16:44:38 +1100296 * violating it, so limit ->max_segments to one, lying
297 * within a single page.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298 * (Note: it is very unlikely that a device with
299 * merge_bvec_fn will be involved in multipath.)
300 */
NeilBrown627a2d32010-03-08 16:44:38 +1100301 if (q->merge_bvec_fn) {
302 blk_queue_max_segments(mddev->queue, 1);
303 blk_queue_segment_boundary(mddev->queue,
304 PAGE_CACHE_SIZE - 1);
305 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306
307 conf->working_disks++;
NeilBrown750a8f32006-10-28 10:38:31 -0700308 mddev->degraded--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309 rdev->raid_disk = path;
NeilBrownb2d444d2005-11-08 21:39:31 -0800310 set_bit(In_sync, &rdev->flags);
Suzanne Woodd6065f72005-11-08 21:39:27 -0800311 rcu_assign_pointer(p->rdev, rdev);
Neil Brown199050e2008-06-28 08:31:33 +1000312 err = 0;
Andre Nollac5e7112009-08-03 10:59:47 +1000313 md_integrity_add_rdev(rdev, mddev);
Neil Brown199050e2008-06-28 08:31:33 +1000314 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315 }
316
317 print_multipath_conf(conf);
Neil Brown199050e2008-06-28 08:31:33 +1000318
319 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320}
321
322static int multipath_remove_disk(mddev_t *mddev, int number)
323{
324 multipath_conf_t *conf = mddev->private;
325 int err = 0;
326 mdk_rdev_t *rdev;
327 struct multipath_info *p = conf->multipaths + number;
328
329 print_multipath_conf(conf);
330
331 rdev = p->rdev;
332 if (rdev) {
NeilBrownb2d444d2005-11-08 21:39:31 -0800333 if (test_bit(In_sync, &rdev->flags) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334 atomic_read(&rdev->nr_pending)) {
NeilBrowndfc70642008-05-23 13:04:39 -0700335 printk(KERN_ERR "hot-remove-disk, slot %d is identified"
336 " but is still operational!\n", number);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337 err = -EBUSY;
338 goto abort;
339 }
340 p->rdev = NULL;
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -0700341 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342 if (atomic_read(&rdev->nr_pending)) {
343 /* lost the race, try later */
344 err = -EBUSY;
345 p->rdev = rdev;
Andre Nollac5e7112009-08-03 10:59:47 +1000346 goto abort;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347 }
Andre Nollac5e7112009-08-03 10:59:47 +1000348 md_integrity_register(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700349 }
350abort:
351
352 print_multipath_conf(conf);
353 return err;
354}
355
356
357
358/*
359 * This is a kernel thread which:
360 *
361 * 1. Retries failed read operations on working multipaths.
362 * 2. Updates the raid superblock when problems encounter.
363 * 3. Performs writes following reads for array syncronising.
364 */
365
366static void multipathd (mddev_t *mddev)
367{
368 struct multipath_bh *mp_bh;
369 struct bio *bio;
370 unsigned long flags;
NeilBrown070ec552009-06-16 16:54:21 +1000371 multipath_conf_t *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372 struct list_head *head = &conf->retry_list;
373
374 md_check_recovery(mddev);
375 for (;;) {
376 char b[BDEVNAME_SIZE];
377 spin_lock_irqsave(&conf->device_lock, flags);
378 if (list_empty(head))
379 break;
380 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
381 list_del(head->prev);
382 spin_unlock_irqrestore(&conf->device_lock, flags);
383
384 bio = &mp_bh->bio;
385 bio->bi_sector = mp_bh->master_bio->bi_sector;
386
387 if ((mp_bh->path = multipath_map (conf))<0) {
388 printk(KERN_ALERT "multipath: %s: unrecoverable IO read"
389 " error for block %llu\n",
390 bdevname(bio->bi_bdev,b),
391 (unsigned long long)bio->bi_sector);
392 multipath_end_bh_io(mp_bh, -EIO);
393 } else {
394 printk(KERN_ERR "multipath: %s: redirecting sector %llu"
395 " to another IO path\n",
396 bdevname(bio->bi_bdev,b),
397 (unsigned long long)bio->bi_sector);
398 *bio = *(mp_bh->master_bio);
399 bio->bi_sector += conf->multipaths[mp_bh->path].rdev->data_offset;
400 bio->bi_bdev = conf->multipaths[mp_bh->path].rdev->bdev;
Mike Christie6000a362008-08-19 18:45:30 -0500401 bio->bi_rw |= (1 << BIO_RW_FAILFAST_TRANSPORT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402 bio->bi_end_io = multipath_end_request;
403 bio->bi_private = mp_bh;
404 generic_make_request(bio);
405 }
406 }
407 spin_unlock_irqrestore(&conf->device_lock, flags);
408}
409
Dan Williams80c3a6c2009-03-17 18:10:40 -0700410static sector_t multipath_size(mddev_t *mddev, sector_t sectors, int raid_disks)
411{
412 WARN_ONCE(sectors || raid_disks,
413 "%s does not support generic reshape\n", __func__);
414
415 return mddev->dev_sectors;
416}
417
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418static int multipath_run (mddev_t *mddev)
419{
420 multipath_conf_t *conf;
421 int disk_idx;
422 struct multipath_info *disk;
423 mdk_rdev_t *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424
Andre Noll0894cc32009-06-18 08:49:23 +1000425 if (md_check_no_bitmap(mddev))
426 return -EINVAL;
427
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428 if (mddev->level != LEVEL_MULTIPATH) {
429 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
430 mdname(mddev), mddev->level);
431 goto out;
432 }
433 /*
434 * copy the already verified devices into our private MULTIPATH
435 * bookkeeping area. [whatever we allocate in multipath_run(),
436 * should be freed in multipath_stop()]
437 */
Neil Browne7e72bf2008-05-14 16:05:54 -0700438 mddev->queue->queue_lock = &mddev->queue->__queue_lock;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439
NeilBrown9ffae0c2006-01-06 00:20:32 -0800440 conf = kzalloc(sizeof(multipath_conf_t), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441 mddev->private = conf;
442 if (!conf) {
443 printk(KERN_ERR
444 "multipath: couldn't allocate memory for %s\n",
445 mdname(mddev));
446 goto out;
447 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448
NeilBrown9ffae0c2006-01-06 00:20:32 -0800449 conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450 GFP_KERNEL);
451 if (!conf->multipaths) {
452 printk(KERN_ERR
453 "multipath: couldn't allocate memory for %s\n",
454 mdname(mddev));
455 goto out_free_conf;
456 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458 conf->working_disks = 0;
Cheng Renquan159ec1f2009-01-09 08:31:08 +1100459 list_for_each_entry(rdev, &mddev->disks, same_set) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460 disk_idx = rdev->raid_disk;
461 if (disk_idx < 0 ||
462 disk_idx >= mddev->raid_disks)
463 continue;
464
465 disk = conf->multipaths + disk_idx;
466 disk->rdev = rdev;
Martin K. Petersen8f6c2e42009-07-01 11:13:45 +1000467 disk_stack_limits(mddev->gendisk, rdev->bdev,
468 rdev->data_offset << 9);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470 /* as we don't honour merge_bvec_fn, we must never risk
471 * violating it, not that we ever expect a device with
472 * a merge_bvec_fn to be involved in multipath */
NeilBrown627a2d32010-03-08 16:44:38 +1100473 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
474 blk_queue_max_segments(mddev->queue, 1);
475 blk_queue_segment_boundary(mddev->queue,
476 PAGE_CACHE_SIZE - 1);
477 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478
NeilBrownb2d444d2005-11-08 21:39:31 -0800479 if (!test_bit(Faulty, &rdev->flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480 conf->working_disks++;
481 }
482
483 conf->raid_disks = mddev->raid_disks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484 conf->mddev = mddev;
485 spin_lock_init(&conf->device_lock);
486 INIT_LIST_HEAD(&conf->retry_list);
487
488 if (!conf->working_disks) {
489 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
490 mdname(mddev));
491 goto out_free_conf;
492 }
NeilBrown2e333e82006-10-21 10:24:07 -0700493 mddev->degraded = conf->raid_disks - conf->working_disks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494
Sage Weilbbba8092009-09-21 17:02:55 -0700495 conf->pool = mempool_create_kmalloc_pool(NR_RESERVED_BUFS,
Matthew Dobson26b6e052006-03-26 01:37:48 -0800496 sizeof(struct multipath_bh));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497 if (conf->pool == NULL) {
498 printk(KERN_ERR
499 "multipath: couldn't allocate memory for %s\n",
500 mdname(mddev));
501 goto out_free_conf;
502 }
503
504 {
NeilBrown0da3c612009-09-23 18:09:45 +1000505 mddev->thread = md_register_thread(multipathd, mddev, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506 if (!mddev->thread) {
507 printk(KERN_ERR "multipath: couldn't allocate thread"
508 " for %s\n", mdname(mddev));
509 goto out_free_conf;
510 }
511 }
512
513 printk(KERN_INFO
514 "multipath: array %s active with %d out of %d IO paths\n",
515 mdname(mddev), conf->working_disks, mddev->raid_disks);
516 /*
517 * Ok, everything is just fine now
518 */
Dan Williams1f403622009-03-31 14:59:03 +1100519 md_set_array_sectors(mddev, multipath_size(mddev, 0, 0));
NeilBrown7a5febe2005-05-16 21:53:16 -0700520
521 mddev->queue->unplug_fn = multipath_unplug;
NeilBrown0d129222006-10-03 01:15:54 -0700522 mddev->queue->backing_dev_info.congested_fn = multipath_congested;
523 mddev->queue->backing_dev_info.congested_data = mddev;
Andre Nollac5e7112009-08-03 10:59:47 +1000524 md_integrity_register(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525 return 0;
526
527out_free_conf:
528 if (conf->pool)
529 mempool_destroy(conf->pool);
Jesper Juhl990a8ba2005-06-21 17:17:30 -0700530 kfree(conf->multipaths);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531 kfree(conf);
532 mddev->private = NULL;
533out:
534 return -EIO;
535}
536
537
538static int multipath_stop (mddev_t *mddev)
539{
NeilBrown070ec552009-06-16 16:54:21 +1000540 multipath_conf_t *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541
542 md_unregister_thread(mddev->thread);
543 mddev->thread = NULL;
544 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
545 mempool_destroy(conf->pool);
546 kfree(conf->multipaths);
547 kfree(conf);
548 mddev->private = NULL;
549 return 0;
550}
551
NeilBrown2604b702006-01-06 00:20:36 -0800552static struct mdk_personality multipath_personality =
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553{
554 .name = "multipath",
NeilBrown2604b702006-01-06 00:20:36 -0800555 .level = LEVEL_MULTIPATH,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556 .owner = THIS_MODULE,
557 .make_request = multipath_make_request,
558 .run = multipath_run,
559 .stop = multipath_stop,
560 .status = multipath_status,
561 .error_handler = multipath_error,
562 .hot_add_disk = multipath_add_disk,
563 .hot_remove_disk= multipath_remove_disk,
Dan Williams80c3a6c2009-03-17 18:10:40 -0700564 .size = multipath_size,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565};
566
567static int __init multipath_init (void)
568{
NeilBrown2604b702006-01-06 00:20:36 -0800569 return register_md_personality (&multipath_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570}
571
572static void __exit multipath_exit (void)
573{
NeilBrown2604b702006-01-06 00:20:36 -0800574 unregister_md_personality (&multipath_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575}
576
577module_init(multipath_init);
578module_exit(multipath_exit);
579MODULE_LICENSE("GPL");
NeilBrown0efb9e62009-12-14 12:49:58 +1100580MODULE_DESCRIPTION("simple multi-path personality for MD");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
NeilBrownd9d166c2006-01-06 00:20:51 -0800582MODULE_ALIAS("md-multipath");
NeilBrown2604b702006-01-06 00:20:36 -0800583MODULE_ALIAS("md-level--4");