blob: 9339e67fcc79a9f961d016d3f80291ce012323bb [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>
Paul Gortmaker056075c2011-07-03 13:58:33 -040023#include <linux/module.h>
NeilBrownbff61972009-03-31 14:33:13 +110024#include <linux/raid/md_u.h>
NeilBrownbff61972009-03-31 14:33:13 +110025#include <linux/seq_file.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090026#include <linux/slab.h>
NeilBrown43b2e5d2009-03-31 14:33:13 +110027#include "md.h"
Christoph Hellwigef740c32009-03-31 14:27:03 +110028#include "multipath.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070029
30#define MAX_WORK_PER_DISK 128
31
32#define NR_RESERVED_BUFS 32
33
34
NeilBrown69724e22011-10-11 16:48:57 +110035static int multipath_map (struct mpconf *conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -070036{
37 int i, disks = conf->raid_disks;
38
39 /*
40 * Later we do read balancing on the read side
41 * now we use the first available disk.
42 */
43
44 rcu_read_lock();
45 for (i = 0; i < disks; i++) {
NeilBrown3cb03002011-10-11 16:45:26 +110046 struct md_rdev *rdev = rcu_dereference(conf->multipaths[i].rdev);
NeilBrownb2d444d2005-11-08 21:39:31 -080047 if (rdev && test_bit(In_sync, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -070048 atomic_inc(&rdev->nr_pending);
49 rcu_read_unlock();
50 return i;
51 }
52 }
53 rcu_read_unlock();
54
55 printk(KERN_ERR "multipath_map(): no more operational IO paths?\n");
56 return (-1);
57}
58
59static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
60{
61 unsigned long flags;
NeilBrownfd01b882011-10-11 16:47:53 +110062 struct mddev *mddev = mp_bh->mddev;
NeilBrown69724e22011-10-11 16:48:57 +110063 struct mpconf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -070064
65 spin_lock_irqsave(&conf->device_lock, flags);
66 list_add(&mp_bh->retry_list, &conf->retry_list);
67 spin_unlock_irqrestore(&conf->device_lock, flags);
68 md_wakeup_thread(mddev->thread);
69}
70
71
72/*
73 * multipath_end_bh_io() is called when we have finished servicing a multipathed
74 * operation and are ready to return a success/failure code to the buffer
75 * cache layer.
76 */
77static void multipath_end_bh_io (struct multipath_bh *mp_bh, int err)
78{
79 struct bio *bio = mp_bh->master_bio;
NeilBrown69724e22011-10-11 16:48:57 +110080 struct mpconf *conf = mp_bh->mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -070081
NeilBrown6712ecf2007-09-27 12:47:43 +020082 bio_endio(bio, err);
Linus Torvalds1da177e2005-04-16 15:20:36 -070083 mempool_free(mp_bh, conf->pool);
84}
85
NeilBrown6712ecf2007-09-27 12:47:43 +020086static void multipath_end_request(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -070087{
88 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
H Hartley Sweeten7b928132010-03-08 16:02:40 +110089 struct multipath_bh *mp_bh = bio->bi_private;
NeilBrown69724e22011-10-11 16:48:57 +110090 struct mpconf *conf = mp_bh->mddev->private;
NeilBrown3cb03002011-10-11 16:45:26 +110091 struct md_rdev *rdev = conf->multipaths[mp_bh->path].rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -070092
Linus Torvalds1da177e2005-04-16 15:20:36 -070093 if (uptodate)
94 multipath_end_bh_io(mp_bh, 0);
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +020095 else if (!(bio->bi_rw & REQ_RAHEAD)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -070096 /*
97 * oops, IO error:
98 */
99 char b[BDEVNAME_SIZE];
100 md_error (mp_bh->mddev, rdev);
101 printk(KERN_ERR "multipath: %s: rescheduling sector %llu\n",
102 bdevname(rdev->bdev,b),
103 (unsigned long long)bio->bi_sector);
104 multipath_reschedule_retry(mp_bh);
105 } else
106 multipath_end_bh_io(mp_bh, error);
107 rdev_dec_pending(rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108}
109
Linus Torvaldsb4fdcb02011-11-04 17:06:58 -0700110static void multipath_make_request(struct mddev *mddev, struct bio * bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111{
NeilBrown69724e22011-10-11 16:48:57 +1100112 struct mpconf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113 struct multipath_bh * mp_bh;
114 struct multipath_info *multipath;
115
Tejun Heoe9c74692010-09-03 11:56:18 +0200116 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
117 md_flush_request(mddev, bio);
Christoph Hellwig5a7bbad2011-09-12 12:12:01 +0200118 return;
NeilBrowne5dcdd82005-09-09 16:23:41 -0700119 }
120
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121 mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
122
123 mp_bh->master_bio = bio;
124 mp_bh->mddev = mddev;
125
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126 mp_bh->path = multipath_map(conf);
127 if (mp_bh->path < 0) {
NeilBrown6712ecf2007-09-27 12:47:43 +0200128 bio_endio(bio, -EIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129 mempool_free(mp_bh, conf->pool);
Christoph Hellwig5a7bbad2011-09-12 12:12:01 +0200130 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131 }
132 multipath = conf->multipaths + mp_bh->path;
133
134 mp_bh->bio = *bio;
135 mp_bh->bio.bi_sector += multipath->rdev->data_offset;
136 mp_bh->bio.bi_bdev = multipath->rdev->bdev;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200137 mp_bh->bio.bi_rw |= REQ_FAILFAST_TRANSPORT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138 mp_bh->bio.bi_end_io = multipath_end_request;
139 mp_bh->bio.bi_private = mp_bh;
140 generic_make_request(&mp_bh->bio);
Christoph Hellwig5a7bbad2011-09-12 12:12:01 +0200141 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142}
143
NeilBrownfd01b882011-10-11 16:47:53 +1100144static void multipath_status (struct seq_file *seq, struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145{
NeilBrown69724e22011-10-11 16:48:57 +1100146 struct mpconf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147 int i;
148
149 seq_printf (seq, " [%d/%d] [", conf->raid_disks,
NeilBrown92f861a2011-05-11 14:38:02 +1000150 conf->raid_disks - mddev->degraded);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151 for (i = 0; i < conf->raid_disks; i++)
152 seq_printf (seq, "%s",
153 conf->multipaths[i].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -0800154 test_bit(In_sync, &conf->multipaths[i].rdev->flags) ? "U" : "_");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155 seq_printf (seq, "]");
156}
157
NeilBrown0d129222006-10-03 01:15:54 -0700158static int multipath_congested(void *data, int bits)
159{
NeilBrownfd01b882011-10-11 16:47:53 +1100160 struct mddev *mddev = data;
NeilBrown69724e22011-10-11 16:48:57 +1100161 struct mpconf *conf = mddev->private;
NeilBrown0d129222006-10-03 01:15:54 -0700162 int i, ret = 0;
163
NeilBrown3fa841d2009-09-23 18:10:29 +1000164 if (mddev_congested(mddev, bits))
165 return 1;
166
NeilBrown0d129222006-10-03 01:15:54 -0700167 rcu_read_lock();
168 for (i = 0; i < mddev->raid_disks ; i++) {
NeilBrown3cb03002011-10-11 16:45:26 +1100169 struct md_rdev *rdev = rcu_dereference(conf->multipaths[i].rdev);
NeilBrown0d129222006-10-03 01:15:54 -0700170 if (rdev && !test_bit(Faulty, &rdev->flags)) {
Jens Axboe165125e2007-07-24 09:28:11 +0200171 struct request_queue *q = bdev_get_queue(rdev->bdev);
NeilBrown0d129222006-10-03 01:15:54 -0700172
173 ret |= bdi_congested(&q->backing_dev_info, bits);
174 /* Just like multipath_map, we just check the
175 * first available device
176 */
177 break;
178 }
179 }
180 rcu_read_unlock();
181 return ret;
182}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183
184/*
185 * Careful, this can execute in IRQ contexts as well!
186 */
NeilBrownfd01b882011-10-11 16:47:53 +1100187static void multipath_error (struct mddev *mddev, struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188{
NeilBrown69724e22011-10-11 16:48:57 +1100189 struct mpconf *conf = mddev->private;
NeilBrown6f8d0c72011-05-11 14:38:44 +1000190 char b[BDEVNAME_SIZE];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191
NeilBrown92f861a2011-05-11 14:38:02 +1000192 if (conf->raid_disks - mddev->degraded <= 1) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193 /*
194 * Uh oh, we can do nothing if this is our last path, but
195 * first check if this is a queued request for a device
196 * which has just failed.
197 */
198 printk(KERN_ALERT
NeilBrown6f8d0c72011-05-11 14:38:44 +1000199 "multipath: only one IO path left and IO error.\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200 /* leave it active... it's all we have */
NeilBrown6f8d0c72011-05-11 14:38:44 +1000201 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202 }
NeilBrown6f8d0c72011-05-11 14:38:44 +1000203 /*
204 * Mark disk as unusable
205 */
206 if (test_and_clear_bit(In_sync, &rdev->flags)) {
207 unsigned long flags;
208 spin_lock_irqsave(&conf->device_lock, flags);
209 mddev->degraded++;
210 spin_unlock_irqrestore(&conf->device_lock, flags);
211 }
212 set_bit(Faulty, &rdev->flags);
213 set_bit(MD_CHANGE_DEVS, &mddev->flags);
214 printk(KERN_ALERT "multipath: IO failure on %s,"
215 " disabling IO path.\n"
216 "multipath: Operation continuing"
217 " on %d IO paths.\n",
218 bdevname(rdev->bdev, b),
219 conf->raid_disks - mddev->degraded);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220}
221
NeilBrown69724e22011-10-11 16:48:57 +1100222static void print_multipath_conf (struct mpconf *conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223{
224 int i;
225 struct multipath_info *tmp;
226
227 printk("MULTIPATH conf printout:\n");
228 if (!conf) {
229 printk("(conf==NULL)\n");
230 return;
231 }
NeilBrown92f861a2011-05-11 14:38:02 +1000232 printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700233 conf->raid_disks);
234
235 for (i = 0; i < conf->raid_disks; i++) {
236 char b[BDEVNAME_SIZE];
237 tmp = conf->multipaths + i;
238 if (tmp->rdev)
239 printk(" disk%d, o:%d, dev:%s\n",
NeilBrownb2d444d2005-11-08 21:39:31 -0800240 i,!test_bit(Faulty, &tmp->rdev->flags),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700241 bdevname(tmp->rdev->bdev,b));
242 }
243}
244
245
NeilBrownfd01b882011-10-11 16:47:53 +1100246static int multipath_add_disk(struct mddev *mddev, struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247{
NeilBrown69724e22011-10-11 16:48:57 +1100248 struct mpconf *conf = mddev->private;
Jesper Juhl352d7682006-01-09 20:54:44 -0800249 struct request_queue *q;
Neil Brown199050e2008-06-28 08:31:33 +1000250 int err = -EEXIST;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251 int path;
252 struct multipath_info *p;
Neil Brown6c2fce22008-06-28 08:31:31 +1000253 int first = 0;
254 int last = mddev->raid_disks - 1;
255
256 if (rdev->raid_disk >= 0)
257 first = last = rdev->raid_disk;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258
259 print_multipath_conf(conf);
260
Neil Brown6c2fce22008-06-28 08:31:31 +1000261 for (path = first; path <= last; path++)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262 if ((p=conf->multipaths+path)->rdev == NULL) {
Jesper Juhl352d7682006-01-09 20:54:44 -0800263 q = rdev->bdev->bd_disk->queue;
Martin K. Petersen8f6c2e42009-07-01 11:13:45 +1000264 disk_stack_limits(mddev->gendisk, rdev->bdev,
265 rdev->data_offset << 9);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266
267 /* as we don't honour merge_bvec_fn, we must never risk
NeilBrown627a2d32010-03-08 16:44:38 +1100268 * violating it, so limit ->max_segments to one, lying
269 * within a single page.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270 * (Note: it is very unlikely that a device with
271 * merge_bvec_fn will be involved in multipath.)
272 */
NeilBrown627a2d32010-03-08 16:44:38 +1100273 if (q->merge_bvec_fn) {
274 blk_queue_max_segments(mddev->queue, 1);
275 blk_queue_segment_boundary(mddev->queue,
276 PAGE_CACHE_SIZE - 1);
277 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278
NeilBrown6f8d0c72011-05-11 14:38:44 +1000279 spin_lock_irq(&conf->device_lock);
NeilBrown750a8f32006-10-28 10:38:31 -0700280 mddev->degraded--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281 rdev->raid_disk = path;
NeilBrownb2d444d2005-11-08 21:39:31 -0800282 set_bit(In_sync, &rdev->flags);
NeilBrown6f8d0c72011-05-11 14:38:44 +1000283 spin_unlock_irq(&conf->device_lock);
Suzanne Woodd6065f72005-11-08 21:39:27 -0800284 rcu_assign_pointer(p->rdev, rdev);
Neil Brown199050e2008-06-28 08:31:33 +1000285 err = 0;
Andre Nollac5e7112009-08-03 10:59:47 +1000286 md_integrity_add_rdev(rdev, mddev);
Neil Brown199050e2008-06-28 08:31:33 +1000287 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288 }
289
290 print_multipath_conf(conf);
Neil Brown199050e2008-06-28 08:31:33 +1000291
292 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293}
294
NeilBrownb8321b62011-12-23 10:17:51 +1100295static int multipath_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296{
NeilBrown69724e22011-10-11 16:48:57 +1100297 struct mpconf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298 int err = 0;
NeilBrownb8321b62011-12-23 10:17:51 +1100299 int number = rdev->raid_disk;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300 struct multipath_info *p = conf->multipaths + number;
301
302 print_multipath_conf(conf);
303
NeilBrownb8321b62011-12-23 10:17:51 +1100304 if (rdev == p->rdev) {
NeilBrownb2d444d2005-11-08 21:39:31 -0800305 if (test_bit(In_sync, &rdev->flags) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306 atomic_read(&rdev->nr_pending)) {
NeilBrowndfc70642008-05-23 13:04:39 -0700307 printk(KERN_ERR "hot-remove-disk, slot %d is identified"
308 " but is still operational!\n", number);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309 err = -EBUSY;
310 goto abort;
311 }
312 p->rdev = NULL;
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -0700313 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314 if (atomic_read(&rdev->nr_pending)) {
315 /* lost the race, try later */
316 err = -EBUSY;
317 p->rdev = rdev;
Andre Nollac5e7112009-08-03 10:59:47 +1000318 goto abort;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700319 }
Martin K. Petersena91a2782011-03-17 11:11:05 +0100320 err = md_integrity_register(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700321 }
322abort:
323
324 print_multipath_conf(conf);
325 return err;
326}
327
328
329
330/*
331 * This is a kernel thread which:
332 *
333 * 1. Retries failed read operations on working multipaths.
334 * 2. Updates the raid superblock when problems encounter.
335 * 3. Performs writes following reads for array syncronising.
336 */
337
NeilBrownfd01b882011-10-11 16:47:53 +1100338static void multipathd (struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339{
340 struct multipath_bh *mp_bh;
341 struct bio *bio;
342 unsigned long flags;
NeilBrown69724e22011-10-11 16:48:57 +1100343 struct mpconf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344 struct list_head *head = &conf->retry_list;
345
346 md_check_recovery(mddev);
347 for (;;) {
348 char b[BDEVNAME_SIZE];
349 spin_lock_irqsave(&conf->device_lock, flags);
350 if (list_empty(head))
351 break;
352 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
353 list_del(head->prev);
354 spin_unlock_irqrestore(&conf->device_lock, flags);
355
356 bio = &mp_bh->bio;
357 bio->bi_sector = mp_bh->master_bio->bi_sector;
358
359 if ((mp_bh->path = multipath_map (conf))<0) {
360 printk(KERN_ALERT "multipath: %s: unrecoverable IO read"
361 " error for block %llu\n",
362 bdevname(bio->bi_bdev,b),
363 (unsigned long long)bio->bi_sector);
364 multipath_end_bh_io(mp_bh, -EIO);
365 } else {
366 printk(KERN_ERR "multipath: %s: redirecting sector %llu"
367 " to another IO path\n",
368 bdevname(bio->bi_bdev,b),
369 (unsigned long long)bio->bi_sector);
370 *bio = *(mp_bh->master_bio);
371 bio->bi_sector += conf->multipaths[mp_bh->path].rdev->data_offset;
372 bio->bi_bdev = conf->multipaths[mp_bh->path].rdev->bdev;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200373 bio->bi_rw |= REQ_FAILFAST_TRANSPORT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374 bio->bi_end_io = multipath_end_request;
375 bio->bi_private = mp_bh;
376 generic_make_request(bio);
377 }
378 }
379 spin_unlock_irqrestore(&conf->device_lock, flags);
380}
381
NeilBrownfd01b882011-10-11 16:47:53 +1100382static sector_t multipath_size(struct mddev *mddev, sector_t sectors, int raid_disks)
Dan Williams80c3a6c2009-03-17 18:10:40 -0700383{
384 WARN_ONCE(sectors || raid_disks,
385 "%s does not support generic reshape\n", __func__);
386
387 return mddev->dev_sectors;
388}
389
NeilBrownfd01b882011-10-11 16:47:53 +1100390static int multipath_run (struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391{
NeilBrown69724e22011-10-11 16:48:57 +1100392 struct mpconf *conf;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393 int disk_idx;
394 struct multipath_info *disk;
NeilBrown3cb03002011-10-11 16:45:26 +1100395 struct md_rdev *rdev;
NeilBrown92f861a2011-05-11 14:38:02 +1000396 int working_disks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397
Andre Noll0894cc32009-06-18 08:49:23 +1000398 if (md_check_no_bitmap(mddev))
399 return -EINVAL;
400
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401 if (mddev->level != LEVEL_MULTIPATH) {
402 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
403 mdname(mddev), mddev->level);
404 goto out;
405 }
406 /*
407 * copy the already verified devices into our private MULTIPATH
408 * bookkeeping area. [whatever we allocate in multipath_run(),
409 * should be freed in multipath_stop()]
410 */
411
NeilBrown69724e22011-10-11 16:48:57 +1100412 conf = kzalloc(sizeof(struct mpconf), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413 mddev->private = conf;
414 if (!conf) {
415 printk(KERN_ERR
416 "multipath: couldn't allocate memory for %s\n",
417 mdname(mddev));
418 goto out;
419 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420
NeilBrown9ffae0c2006-01-06 00:20:32 -0800421 conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422 GFP_KERNEL);
423 if (!conf->multipaths) {
424 printk(KERN_ERR
425 "multipath: couldn't allocate memory for %s\n",
426 mdname(mddev));
427 goto out_free_conf;
428 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429
NeilBrown92f861a2011-05-11 14:38:02 +1000430 working_disks = 0;
NeilBrowndafb20f2012-03-19 12:46:39 +1100431 rdev_for_each(rdev, mddev) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432 disk_idx = rdev->raid_disk;
433 if (disk_idx < 0 ||
434 disk_idx >= mddev->raid_disks)
435 continue;
436
437 disk = conf->multipaths + disk_idx;
438 disk->rdev = rdev;
Martin K. Petersen8f6c2e42009-07-01 11:13:45 +1000439 disk_stack_limits(mddev->gendisk, rdev->bdev,
440 rdev->data_offset << 9);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442 /* as we don't honour merge_bvec_fn, we must never risk
443 * violating it, not that we ever expect a device with
444 * a merge_bvec_fn to be involved in multipath */
NeilBrown627a2d32010-03-08 16:44:38 +1100445 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
446 blk_queue_max_segments(mddev->queue, 1);
447 blk_queue_segment_boundary(mddev->queue,
448 PAGE_CACHE_SIZE - 1);
449 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450
NeilBrownb2d444d2005-11-08 21:39:31 -0800451 if (!test_bit(Faulty, &rdev->flags))
NeilBrown92f861a2011-05-11 14:38:02 +1000452 working_disks++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453 }
454
455 conf->raid_disks = mddev->raid_disks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 conf->mddev = mddev;
457 spin_lock_init(&conf->device_lock);
458 INIT_LIST_HEAD(&conf->retry_list);
459
NeilBrown92f861a2011-05-11 14:38:02 +1000460 if (!working_disks) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
462 mdname(mddev));
463 goto out_free_conf;
464 }
NeilBrown92f861a2011-05-11 14:38:02 +1000465 mddev->degraded = conf->raid_disks - working_disks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466
Sage Weilbbba8092009-09-21 17:02:55 -0700467 conf->pool = mempool_create_kmalloc_pool(NR_RESERVED_BUFS,
Matthew Dobson26b6e052006-03-26 01:37:48 -0800468 sizeof(struct multipath_bh));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469 if (conf->pool == NULL) {
470 printk(KERN_ERR
471 "multipath: couldn't allocate memory for %s\n",
472 mdname(mddev));
473 goto out_free_conf;
474 }
475
476 {
NeilBrown0da3c612009-09-23 18:09:45 +1000477 mddev->thread = md_register_thread(multipathd, mddev, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478 if (!mddev->thread) {
479 printk(KERN_ERR "multipath: couldn't allocate thread"
480 " for %s\n", mdname(mddev));
481 goto out_free_conf;
482 }
483 }
484
485 printk(KERN_INFO
486 "multipath: array %s active with %d out of %d IO paths\n",
NeilBrown92f861a2011-05-11 14:38:02 +1000487 mdname(mddev), conf->raid_disks - mddev->degraded,
488 mddev->raid_disks);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489 /*
490 * Ok, everything is just fine now
491 */
Dan Williams1f403622009-03-31 14:59:03 +1100492 md_set_array_sectors(mddev, multipath_size(mddev, 0, 0));
NeilBrown7a5febe2005-05-16 21:53:16 -0700493
NeilBrown0d129222006-10-03 01:15:54 -0700494 mddev->queue->backing_dev_info.congested_fn = multipath_congested;
495 mddev->queue->backing_dev_info.congested_data = mddev;
Martin K. Petersena91a2782011-03-17 11:11:05 +0100496
497 if (md_integrity_register(mddev))
498 goto out_free_conf;
499
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500 return 0;
501
502out_free_conf:
503 if (conf->pool)
504 mempool_destroy(conf->pool);
Jesper Juhl990a8ba2005-06-21 17:17:30 -0700505 kfree(conf->multipaths);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506 kfree(conf);
507 mddev->private = NULL;
508out:
509 return -EIO;
510}
511
512
NeilBrownfd01b882011-10-11 16:47:53 +1100513static int multipath_stop (struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514{
NeilBrown69724e22011-10-11 16:48:57 +1100515 struct mpconf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516
NeilBrown01f96c02011-09-21 15:30:20 +1000517 md_unregister_thread(&mddev->thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
519 mempool_destroy(conf->pool);
520 kfree(conf->multipaths);
521 kfree(conf);
522 mddev->private = NULL;
523 return 0;
524}
525
NeilBrown84fc4b52011-10-11 16:49:58 +1100526static struct md_personality multipath_personality =
Linus Torvalds1da177e2005-04-16 15:20:36 -0700527{
528 .name = "multipath",
NeilBrown2604b702006-01-06 00:20:36 -0800529 .level = LEVEL_MULTIPATH,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530 .owner = THIS_MODULE,
531 .make_request = multipath_make_request,
532 .run = multipath_run,
533 .stop = multipath_stop,
534 .status = multipath_status,
535 .error_handler = multipath_error,
536 .hot_add_disk = multipath_add_disk,
537 .hot_remove_disk= multipath_remove_disk,
Dan Williams80c3a6c2009-03-17 18:10:40 -0700538 .size = multipath_size,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539};
540
541static int __init multipath_init (void)
542{
NeilBrown2604b702006-01-06 00:20:36 -0800543 return register_md_personality (&multipath_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544}
545
546static void __exit multipath_exit (void)
547{
NeilBrown2604b702006-01-06 00:20:36 -0800548 unregister_md_personality (&multipath_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549}
550
551module_init(multipath_init);
552module_exit(multipath_exit);
553MODULE_LICENSE("GPL");
NeilBrown0efb9e62009-12-14 12:49:58 +1100554MODULE_DESCRIPTION("simple multi-path personality for MD");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
NeilBrownd9d166c2006-01-06 00:20:51 -0800556MODULE_ALIAS("md-multipath");
NeilBrown2604b702006-01-06 00:20:36 -0800557MODULE_ALIAS("md-level--4");