Merge branch 'for-2.6.39/stack-plug' into for-2.6.39/core
[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->working_disks);
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
190         if (conf->working_disks <= 1) {
191                 /*
192                  * Uh oh, we can do nothing if this is our last path, but
193                  * first check if this is a queued request for a device
194                  * which has just failed.
195                  */
196                 printk(KERN_ALERT 
197                         "multipath: only one IO path left and IO error.\n");
198                 /* leave it active... it's all we have */
199         } else {
200                 /*
201                  * Mark disk as unusable
202                  */
203                 if (!test_bit(Faulty, &rdev->flags)) {
204                         char b[BDEVNAME_SIZE];
205                         clear_bit(In_sync, &rdev->flags);
206                         set_bit(Faulty, &rdev->flags);
207                         set_bit(MD_CHANGE_DEVS, &mddev->flags);
208                         conf->working_disks--;
209                         mddev->degraded++;
210                         printk(KERN_ALERT "multipath: IO failure on %s,"
211                                 " disabling IO path.\n"
212                                 "multipath: Operation continuing"
213                                 " on %d IO paths.\n",
214                                 bdevname (rdev->bdev,b),
215                                 conf->working_disks);
216                 }
217         }
218 }
219
220 static void print_multipath_conf (multipath_conf_t *conf)
221 {
222         int i;
223         struct multipath_info *tmp;
224
225         printk("MULTIPATH conf printout:\n");
226         if (!conf) {
227                 printk("(conf==NULL)\n");
228                 return;
229         }
230         printk(" --- wd:%d rd:%d\n", conf->working_disks,
231                          conf->raid_disks);
232
233         for (i = 0; i < conf->raid_disks; i++) {
234                 char b[BDEVNAME_SIZE];
235                 tmp = conf->multipaths + i;
236                 if (tmp->rdev)
237                         printk(" disk%d, o:%d, dev:%s\n",
238                                 i,!test_bit(Faulty, &tmp->rdev->flags),
239                                bdevname(tmp->rdev->bdev,b));
240         }
241 }
242
243
244 static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
245 {
246         multipath_conf_t *conf = mddev->private;
247         struct request_queue *q;
248         int err = -EEXIST;
249         int path;
250         struct multipath_info *p;
251         int first = 0;
252         int last = mddev->raid_disks - 1;
253
254         if (rdev->raid_disk >= 0)
255                 first = last = rdev->raid_disk;
256
257         print_multipath_conf(conf);
258
259         for (path = first; path <= last; path++)
260                 if ((p=conf->multipaths+path)->rdev == NULL) {
261                         q = rdev->bdev->bd_disk->queue;
262                         disk_stack_limits(mddev->gendisk, rdev->bdev,
263                                           rdev->data_offset << 9);
264
265                 /* as we don't honour merge_bvec_fn, we must never risk
266                  * violating it, so limit ->max_segments to one, lying
267                  * within a single page.
268                  * (Note: it is very unlikely that a device with
269                  * merge_bvec_fn will be involved in multipath.)
270                  */
271                         if (q->merge_bvec_fn) {
272                                 blk_queue_max_segments(mddev->queue, 1);
273                                 blk_queue_segment_boundary(mddev->queue,
274                                                            PAGE_CACHE_SIZE - 1);
275                         }
276
277                         conf->working_disks++;
278                         mddev->degraded--;
279                         rdev->raid_disk = path;
280                         set_bit(In_sync, &rdev->flags);
281                         rcu_assign_pointer(p->rdev, rdev);
282                         err = 0;
283                         md_integrity_add_rdev(rdev, mddev);
284                         break;
285                 }
286
287         print_multipath_conf(conf);
288
289         return err;
290 }
291
292 static int multipath_remove_disk(mddev_t *mddev, int number)
293 {
294         multipath_conf_t *conf = mddev->private;
295         int err = 0;
296         mdk_rdev_t *rdev;
297         struct multipath_info *p = conf->multipaths + number;
298
299         print_multipath_conf(conf);
300
301         rdev = p->rdev;
302         if (rdev) {
303                 if (test_bit(In_sync, &rdev->flags) ||
304                     atomic_read(&rdev->nr_pending)) {
305                         printk(KERN_ERR "hot-remove-disk, slot %d is identified"
306                                " but is still operational!\n", number);
307                         err = -EBUSY;
308                         goto abort;
309                 }
310                 p->rdev = NULL;
311                 synchronize_rcu();
312                 if (atomic_read(&rdev->nr_pending)) {
313                         /* lost the race, try later */
314                         err = -EBUSY;
315                         p->rdev = rdev;
316                         goto abort;
317                 }
318                 md_integrity_register(mddev);
319         }
320 abort:
321
322         print_multipath_conf(conf);
323         return err;
324 }
325
326
327
328 /*
329  * This is a kernel thread which:
330  *
331  *      1.      Retries failed read operations on working multipaths.
332  *      2.      Updates the raid superblock when problems encounter.
333  *      3.      Performs writes following reads for array syncronising.
334  */
335
336 static void multipathd (mddev_t *mddev)
337 {
338         struct multipath_bh *mp_bh;
339         struct bio *bio;
340         unsigned long flags;
341         multipath_conf_t *conf = mddev->private;
342         struct list_head *head = &conf->retry_list;
343
344         md_check_recovery(mddev);
345         for (;;) {
346                 char b[BDEVNAME_SIZE];
347                 spin_lock_irqsave(&conf->device_lock, flags);
348                 if (list_empty(head))
349                         break;
350                 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
351                 list_del(head->prev);
352                 spin_unlock_irqrestore(&conf->device_lock, flags);
353
354                 bio = &mp_bh->bio;
355                 bio->bi_sector = mp_bh->master_bio->bi_sector;
356                 
357                 if ((mp_bh->path = multipath_map (conf))<0) {
358                         printk(KERN_ALERT "multipath: %s: unrecoverable IO read"
359                                 " error for block %llu\n",
360                                 bdevname(bio->bi_bdev,b),
361                                 (unsigned long long)bio->bi_sector);
362                         multipath_end_bh_io(mp_bh, -EIO);
363                 } else {
364                         printk(KERN_ERR "multipath: %s: redirecting sector %llu"
365                                 " to another IO path\n",
366                                 bdevname(bio->bi_bdev,b),
367                                 (unsigned long long)bio->bi_sector);
368                         *bio = *(mp_bh->master_bio);
369                         bio->bi_sector += conf->multipaths[mp_bh->path].rdev->data_offset;
370                         bio->bi_bdev = conf->multipaths[mp_bh->path].rdev->bdev;
371                         bio->bi_rw |= REQ_FAILFAST_TRANSPORT;
372                         bio->bi_end_io = multipath_end_request;
373                         bio->bi_private = mp_bh;
374                         generic_make_request(bio);
375                 }
376         }
377         spin_unlock_irqrestore(&conf->device_lock, flags);
378 }
379
380 static sector_t multipath_size(mddev_t *mddev, sector_t sectors, int raid_disks)
381 {
382         WARN_ONCE(sectors || raid_disks,
383                   "%s does not support generic reshape\n", __func__);
384
385         return mddev->dev_sectors;
386 }
387
388 static int multipath_run (mddev_t *mddev)
389 {
390         multipath_conf_t *conf;
391         int disk_idx;
392         struct multipath_info *disk;
393         mdk_rdev_t *rdev;
394
395         if (md_check_no_bitmap(mddev))
396                 return -EINVAL;
397
398         if (mddev->level != LEVEL_MULTIPATH) {
399                 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
400                        mdname(mddev), mddev->level);
401                 goto out;
402         }
403         /*
404          * copy the already verified devices into our private MULTIPATH
405          * bookkeeping area. [whatever we allocate in multipath_run(),
406          * should be freed in multipath_stop()]
407          */
408
409         conf = kzalloc(sizeof(multipath_conf_t), GFP_KERNEL);
410         mddev->private = conf;
411         if (!conf) {
412                 printk(KERN_ERR 
413                         "multipath: couldn't allocate memory for %s\n",
414                         mdname(mddev));
415                 goto out;
416         }
417
418         conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
419                                    GFP_KERNEL);
420         if (!conf->multipaths) {
421                 printk(KERN_ERR 
422                         "multipath: couldn't allocate memory for %s\n",
423                         mdname(mddev));
424                 goto out_free_conf;
425         }
426
427         conf->working_disks = 0;
428         list_for_each_entry(rdev, &mddev->disks, same_set) {
429                 disk_idx = rdev->raid_disk;
430                 if (disk_idx < 0 ||
431                     disk_idx >= mddev->raid_disks)
432                         continue;
433
434                 disk = conf->multipaths + disk_idx;
435                 disk->rdev = rdev;
436                 disk_stack_limits(mddev->gendisk, rdev->bdev,
437                                   rdev->data_offset << 9);
438
439                 /* as we don't honour merge_bvec_fn, we must never risk
440                  * violating it, not that we ever expect a device with
441                  * a merge_bvec_fn to be involved in multipath */
442                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
443                         blk_queue_max_segments(mddev->queue, 1);
444                         blk_queue_segment_boundary(mddev->queue,
445                                                    PAGE_CACHE_SIZE - 1);
446                 }
447
448                 if (!test_bit(Faulty, &rdev->flags))
449                         conf->working_disks++;
450         }
451
452         conf->raid_disks = mddev->raid_disks;
453         conf->mddev = mddev;
454         spin_lock_init(&conf->device_lock);
455         INIT_LIST_HEAD(&conf->retry_list);
456
457         if (!conf->working_disks) {
458                 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
459                         mdname(mddev));
460                 goto out_free_conf;
461         }
462         mddev->degraded = conf->raid_disks - conf->working_disks;
463
464         conf->pool = mempool_create_kmalloc_pool(NR_RESERVED_BUFS,
465                                                  sizeof(struct multipath_bh));
466         if (conf->pool == NULL) {
467                 printk(KERN_ERR 
468                         "multipath: couldn't allocate memory for %s\n",
469                         mdname(mddev));
470                 goto out_free_conf;
471         }
472
473         {
474                 mddev->thread = md_register_thread(multipathd, mddev, NULL);
475                 if (!mddev->thread) {
476                         printk(KERN_ERR "multipath: couldn't allocate thread"
477                                 " for %s\n", mdname(mddev));
478                         goto out_free_conf;
479                 }
480         }
481
482         printk(KERN_INFO 
483                 "multipath: array %s active with %d out of %d IO paths\n",
484                 mdname(mddev), conf->working_disks, mddev->raid_disks);
485         /*
486          * Ok, everything is just fine now
487          */
488         md_set_array_sectors(mddev, multipath_size(mddev, 0, 0));
489
490         mddev->queue->backing_dev_info.congested_fn = multipath_congested;
491         mddev->queue->backing_dev_info.congested_data = mddev;
492         md_integrity_register(mddev);
493         return 0;
494
495 out_free_conf:
496         if (conf->pool)
497                 mempool_destroy(conf->pool);
498         kfree(conf->multipaths);
499         kfree(conf);
500         mddev->private = NULL;
501 out:
502         return -EIO;
503 }
504
505
506 static int multipath_stop (mddev_t *mddev)
507 {
508         multipath_conf_t *conf = mddev->private;
509
510         md_unregister_thread(mddev->thread);
511         mddev->thread = NULL;
512         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
513         mempool_destroy(conf->pool);
514         kfree(conf->multipaths);
515         kfree(conf);
516         mddev->private = NULL;
517         return 0;
518 }
519
520 static struct mdk_personality multipath_personality =
521 {
522         .name           = "multipath",
523         .level          = LEVEL_MULTIPATH,
524         .owner          = THIS_MODULE,
525         .make_request   = multipath_make_request,
526         .run            = multipath_run,
527         .stop           = multipath_stop,
528         .status         = multipath_status,
529         .error_handler  = multipath_error,
530         .hot_add_disk   = multipath_add_disk,
531         .hot_remove_disk= multipath_remove_disk,
532         .size           = multipath_size,
533 };
534
535 static int __init multipath_init (void)
536 {
537         return register_md_personality (&multipath_personality);
538 }
539
540 static void __exit multipath_exit (void)
541 {
542         unregister_md_personality (&multipath_personality);
543 }
544
545 module_init(multipath_init);
546 module_exit(multipath_exit);
547 MODULE_LICENSE("GPL");
548 MODULE_DESCRIPTION("simple multi-path personality for MD");
549 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
550 MODULE_ALIAS("md-multipath");
551 MODULE_ALIAS("md-level--4");