3535c23af288e8eaac45042eb2e1909af8f0759a
[linux-3.10.git] / drivers / md / multipath.c
1 /*
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/blkdev.h>
23 #include <linux/raid/md_u.h>
24 #include <linux/seq_file.h>
25 #include <linux/slab.h>
26 #include "md.h"
27 #include "multipath.h"
28
29 #define MAX_WORK_PER_DISK 128
30
31 #define NR_RESERVED_BUFS        32
32
33
34 static 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++) {
45                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
46                 if (rdev && test_bit(In_sync, &rdev->flags)) {
47                         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
58 static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
59 {
60         unsigned long flags;
61         mddev_t *mddev = mp_bh->mddev;
62         multipath_conf_t *conf = mddev->private;
63
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  */
76 static void multipath_end_bh_io (struct multipath_bh *mp_bh, int err)
77 {
78         struct bio *bio = mp_bh->master_bio;
79         multipath_conf_t *conf = mp_bh->mddev->private;
80
81         bio_endio(bio, err);
82         mempool_free(mp_bh, conf->pool);
83 }
84
85 static void multipath_end_request(struct bio *bio, int error)
86 {
87         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
88         struct multipath_bh *mp_bh = bio->bi_private;
89         multipath_conf_t *conf = mp_bh->mddev->private;
90         mdk_rdev_t *rdev = conf->multipaths[mp_bh->path].rdev;
91
92         if (uptodate)
93                 multipath_end_bh_io(mp_bh, 0);
94         else if (!(bio->bi_rw & REQ_RAHEAD)) {
95                 /*
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);
107 }
108
109 static int multipath_make_request(mddev_t *mddev, struct bio * bio)
110 {
111         multipath_conf_t *conf = mddev->private;
112         struct multipath_bh * mp_bh;
113         struct multipath_info *multipath;
114
115         if (unlikely(bio->bi_rw & REQ_FLUSH)) {
116                 md_flush_request(mddev, bio);
117                 return 0;
118         }
119
120         mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
121
122         mp_bh->master_bio = bio;
123         mp_bh->mddev = mddev;
124
125         mp_bh->path = multipath_map(conf);
126         if (mp_bh->path < 0) {
127                 bio_endio(bio, -EIO);
128                 mempool_free(mp_bh, conf->pool);
129                 return 0;
130         }
131         multipath = conf->multipaths + mp_bh->path;
132
133         mp_bh->bio = *bio;
134         mp_bh->bio.bi_sector += multipath->rdev->data_offset;
135         mp_bh->bio.bi_bdev = multipath->rdev->bdev;
136         mp_bh->bio.bi_rw |= REQ_FAILFAST_TRANSPORT;
137         mp_bh->bio.bi_end_io = multipath_end_request;
138         mp_bh->bio.bi_private = mp_bh;
139         generic_make_request(&mp_bh->bio);
140         return 0;
141 }
142
143 static void multipath_status (struct seq_file *seq, mddev_t *mddev)
144 {
145         multipath_conf_t *conf = mddev->private;
146         int i;
147         
148         seq_printf (seq, " [%d/%d] [", conf->raid_disks,
149                     conf->raid_disks - mddev->degraded);
150         for (i = 0; i < conf->raid_disks; i++)
151                 seq_printf (seq, "%s",
152                                conf->multipaths[i].rdev && 
153                                test_bit(In_sync, &conf->multipaths[i].rdev->flags) ? "U" : "_");
154         seq_printf (seq, "]");
155 }
156
157 static int multipath_congested(void *data, int bits)
158 {
159         mddev_t *mddev = data;
160         multipath_conf_t *conf = mddev->private;
161         int i, ret = 0;
162
163         if (mddev_congested(mddev, bits))
164                 return 1;
165
166         rcu_read_lock();
167         for (i = 0; i < mddev->raid_disks ; i++) {
168                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
169                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
170                         struct request_queue *q = bdev_get_queue(rdev->bdev);
171
172                         ret |= bdi_congested(&q->backing_dev_info, bits);
173                         /* Just like multipath_map, we just check the
174                          * first available device
175                          */
176                         break;
177                 }
178         }
179         rcu_read_unlock();
180         return ret;
181 }
182
183 /*
184  * Careful, this can execute in IRQ contexts as well!
185  */
186 static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
187 {
188         multipath_conf_t *conf = mddev->private;
189         char b[BDEVNAME_SIZE];
190
191         if (conf->raid_disks - mddev->degraded <= 1) {
192                 /*
193                  * Uh oh, we can do nothing if this is our last path, but
194                  * first check if this is a queued request for a device
195                  * which has just failed.
196                  */
197                 printk(KERN_ALERT 
198                        "multipath: only one IO path left and IO error.\n");
199                 /* leave it active... it's all we have */
200                 return;
201         }
202         /*
203          * Mark disk as unusable
204          */
205         if (test_and_clear_bit(In_sync, &rdev->flags)) {
206                 unsigned long flags;
207                 spin_lock_irqsave(&conf->device_lock, flags);
208                 mddev->degraded++;
209                 spin_unlock_irqrestore(&conf->device_lock, flags);
210         }
211         set_bit(Faulty, &rdev->flags);
212         set_bit(MD_CHANGE_DEVS, &mddev->flags);
213         printk(KERN_ALERT "multipath: IO failure on %s,"
214                " disabling IO path.\n"
215                "multipath: Operation continuing"
216                " on %d IO paths.\n",
217                bdevname(rdev->bdev, b),
218                conf->raid_disks - mddev->degraded);
219 }
220
221 static void print_multipath_conf (multipath_conf_t *conf)
222 {
223         int i;
224         struct multipath_info *tmp;
225
226         printk("MULTIPATH conf printout:\n");
227         if (!conf) {
228                 printk("(conf==NULL)\n");
229                 return;
230         }
231         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
232                          conf->raid_disks);
233
234         for (i = 0; i < conf->raid_disks; i++) {
235                 char b[BDEVNAME_SIZE];
236                 tmp = conf->multipaths + i;
237                 if (tmp->rdev)
238                         printk(" disk%d, o:%d, dev:%s\n",
239                                 i,!test_bit(Faulty, &tmp->rdev->flags),
240                                bdevname(tmp->rdev->bdev,b));
241         }
242 }
243
244
245 static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
246 {
247         multipath_conf_t *conf = mddev->private;
248         struct request_queue *q;
249         int err = -EEXIST;
250         int path;
251         struct multipath_info *p;
252         int first = 0;
253         int last = mddev->raid_disks - 1;
254
255         if (rdev->raid_disk >= 0)
256                 first = last = rdev->raid_disk;
257
258         print_multipath_conf(conf);
259
260         for (path = first; path <= last; path++)
261                 if ((p=conf->multipaths+path)->rdev == NULL) {
262                         q = rdev->bdev->bd_disk->queue;
263                         disk_stack_limits(mddev->gendisk, rdev->bdev,
264                                           rdev->data_offset << 9);
265
266                 /* as we don't honour merge_bvec_fn, we must never risk
267                  * violating it, so limit ->max_segments to one, lying
268                  * within a single page.
269                  * (Note: it is very unlikely that a device with
270                  * merge_bvec_fn will be involved in multipath.)
271                  */
272                         if (q->merge_bvec_fn) {
273                                 blk_queue_max_segments(mddev->queue, 1);
274                                 blk_queue_segment_boundary(mddev->queue,
275                                                            PAGE_CACHE_SIZE - 1);
276                         }
277
278                         spin_lock_irq(&conf->device_lock);
279                         mddev->degraded--;
280                         rdev->raid_disk = path;
281                         set_bit(In_sync, &rdev->flags);
282                         spin_unlock_irq(&conf->device_lock);
283                         rcu_assign_pointer(p->rdev, rdev);
284                         err = 0;
285                         md_integrity_add_rdev(rdev, mddev);
286                         break;
287                 }
288
289         print_multipath_conf(conf);
290
291         return err;
292 }
293
294 static int multipath_remove_disk(mddev_t *mddev, int number)
295 {
296         multipath_conf_t *conf = mddev->private;
297         int err = 0;
298         mdk_rdev_t *rdev;
299         struct multipath_info *p = conf->multipaths + number;
300
301         print_multipath_conf(conf);
302
303         rdev = p->rdev;
304         if (rdev) {
305                 if (test_bit(In_sync, &rdev->flags) ||
306                     atomic_read(&rdev->nr_pending)) {
307                         printk(KERN_ERR "hot-remove-disk, slot %d is identified"
308                                " but is still operational!\n", number);
309                         err = -EBUSY;
310                         goto abort;
311                 }
312                 p->rdev = NULL;
313                 synchronize_rcu();
314                 if (atomic_read(&rdev->nr_pending)) {
315                         /* lost the race, try later */
316                         err = -EBUSY;
317                         p->rdev = rdev;
318                         goto abort;
319                 }
320                 err = md_integrity_register(mddev);
321         }
322 abort:
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
338 static void multipathd (mddev_t *mddev)
339 {
340         struct multipath_bh *mp_bh;
341         struct bio *bio;
342         unsigned long flags;
343         multipath_conf_t *conf = mddev->private;
344         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;
373                         bio->bi_rw |= REQ_FAILFAST_TRANSPORT;
374                         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
382 static sector_t multipath_size(mddev_t *mddev, sector_t sectors, int raid_disks)
383 {
384         WARN_ONCE(sectors || raid_disks,
385                   "%s does not support generic reshape\n", __func__);
386
387         return mddev->dev_sectors;
388 }
389
390 static int multipath_run (mddev_t *mddev)
391 {
392         multipath_conf_t *conf;
393         int disk_idx;
394         struct multipath_info *disk;
395         mdk_rdev_t *rdev;
396         int working_disks;
397
398         if (md_check_no_bitmap(mddev))
399                 return -EINVAL;
400
401         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
412         conf = kzalloc(sizeof(multipath_conf_t), GFP_KERNEL);
413         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         }
420
421         conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
422                                    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         }
429
430         working_disks = 0;
431         list_for_each_entry(rdev, &mddev->disks, same_set) {
432                 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;
439                 disk_stack_limits(mddev->gendisk, rdev->bdev,
440                                   rdev->data_offset << 9);
441
442                 /* 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 */
445                 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                 }
450
451                 if (!test_bit(Faulty, &rdev->flags))
452                         working_disks++;
453         }
454
455         conf->raid_disks = mddev->raid_disks;
456         conf->mddev = mddev;
457         spin_lock_init(&conf->device_lock);
458         INIT_LIST_HEAD(&conf->retry_list);
459
460         if (!working_disks) {
461                 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
462                         mdname(mddev));
463                 goto out_free_conf;
464         }
465         mddev->degraded = conf->raid_disks - working_disks;
466
467         conf->pool = mempool_create_kmalloc_pool(NR_RESERVED_BUFS,
468                                                  sizeof(struct multipath_bh));
469         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         {
477                 mddev->thread = md_register_thread(multipathd, mddev, NULL);
478                 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",
487                 mdname(mddev), conf->raid_disks - mddev->degraded,
488                mddev->raid_disks);
489         /*
490          * Ok, everything is just fine now
491          */
492         md_set_array_sectors(mddev, multipath_size(mddev, 0, 0));
493
494         mddev->queue->backing_dev_info.congested_fn = multipath_congested;
495         mddev->queue->backing_dev_info.congested_data = mddev;
496
497         if (md_integrity_register(mddev))
498                 goto out_free_conf;
499
500         return 0;
501
502 out_free_conf:
503         if (conf->pool)
504                 mempool_destroy(conf->pool);
505         kfree(conf->multipaths);
506         kfree(conf);
507         mddev->private = NULL;
508 out:
509         return -EIO;
510 }
511
512
513 static int multipath_stop (mddev_t *mddev)
514 {
515         multipath_conf_t *conf = mddev->private;
516
517         md_unregister_thread(mddev->thread);
518         mddev->thread = NULL;
519         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
520         mempool_destroy(conf->pool);
521         kfree(conf->multipaths);
522         kfree(conf);
523         mddev->private = NULL;
524         return 0;
525 }
526
527 static struct mdk_personality multipath_personality =
528 {
529         .name           = "multipath",
530         .level          = LEVEL_MULTIPATH,
531         .owner          = THIS_MODULE,
532         .make_request   = multipath_make_request,
533         .run            = multipath_run,
534         .stop           = multipath_stop,
535         .status         = multipath_status,
536         .error_handler  = multipath_error,
537         .hot_add_disk   = multipath_add_disk,
538         .hot_remove_disk= multipath_remove_disk,
539         .size           = multipath_size,
540 };
541
542 static int __init multipath_init (void)
543 {
544         return register_md_personality (&multipath_personality);
545 }
546
547 static void __exit multipath_exit (void)
548 {
549         unregister_md_personality (&multipath_personality);
550 }
551
552 module_init(multipath_init);
553 module_exit(multipath_exit);
554 MODULE_LICENSE("GPL");
555 MODULE_DESCRIPTION("simple multi-path personality for MD");
556 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
557 MODULE_ALIAS("md-multipath");
558 MODULE_ALIAS("md-level--4");