block: convert blkdev_issue_flush() to use empty barriers
[linux-2.6.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/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
38 static int multipath_map (multipath_conf_t *conf)
39 {
40         int i, disks = conf->raid_disks;
41
42         /*
43          * Later we do read balancing on the read side 
44          * now we use the first available disk.
45          */
46
47         rcu_read_lock();
48         for (i = 0; i < disks; i++) {
49                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
50                 if (rdev && test_bit(In_sync, &rdev->flags)) {
51                         atomic_inc(&rdev->nr_pending);
52                         rcu_read_unlock();
53                         return i;
54                 }
55         }
56         rcu_read_unlock();
57
58         printk(KERN_ERR "multipath_map(): no more operational IO paths?\n");
59         return (-1);
60 }
61
62 static void multipath_reschedule_retry (struct multipath_bh *mp_bh)
63 {
64         unsigned long flags;
65         mddev_t *mddev = mp_bh->mddev;
66         multipath_conf_t *conf = mddev_to_conf(mddev);
67
68         spin_lock_irqsave(&conf->device_lock, flags);
69         list_add(&mp_bh->retry_list, &conf->retry_list);
70         spin_unlock_irqrestore(&conf->device_lock, flags);
71         md_wakeup_thread(mddev->thread);
72 }
73
74
75 /*
76  * multipath_end_bh_io() is called when we have finished servicing a multipathed
77  * operation and are ready to return a success/failure code to the buffer
78  * cache layer.
79  */
80 static void multipath_end_bh_io (struct multipath_bh *mp_bh, int err)
81 {
82         struct bio *bio = mp_bh->master_bio;
83         multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
84
85         bio_endio(bio, err);
86         mempool_free(mp_bh, conf->pool);
87 }
88
89 static void multipath_end_request(struct bio *bio, int error)
90 {
91         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
92         struct multipath_bh * mp_bh = (struct multipath_bh *)(bio->bi_private);
93         multipath_conf_t *conf = mddev_to_conf(mp_bh->mddev);
94         mdk_rdev_t *rdev = conf->multipaths[mp_bh->path].rdev;
95
96         if (uptodate)
97                 multipath_end_bh_io(mp_bh, 0);
98         else if (!bio_rw_ahead(bio)) {
99                 /*
100                  * oops, IO error:
101                  */
102                 char b[BDEVNAME_SIZE];
103                 md_error (mp_bh->mddev, rdev);
104                 printk(KERN_ERR "multipath: %s: rescheduling sector %llu\n", 
105                        bdevname(rdev->bdev,b), 
106                        (unsigned long long)bio->bi_sector);
107                 multipath_reschedule_retry(mp_bh);
108         } else
109                 multipath_end_bh_io(mp_bh, error);
110         rdev_dec_pending(rdev, conf->mddev);
111 }
112
113 static void unplug_slaves(mddev_t *mddev)
114 {
115         multipath_conf_t *conf = mddev_to_conf(mddev);
116         int i;
117
118         rcu_read_lock();
119         for (i=0; i<mddev->raid_disks; i++) {
120                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
121                 if (rdev && !test_bit(Faulty, &rdev->flags)
122                     && atomic_read(&rdev->nr_pending)) {
123                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
124
125                         atomic_inc(&rdev->nr_pending);
126                         rcu_read_unlock();
127
128                         if (r_queue->unplug_fn)
129                                 r_queue->unplug_fn(r_queue);
130
131                         rdev_dec_pending(rdev, mddev);
132                         rcu_read_lock();
133                 }
134         }
135         rcu_read_unlock();
136 }
137
138 static void multipath_unplug(struct request_queue *q)
139 {
140         unplug_slaves(q->queuedata);
141 }
142
143
144 static int multipath_make_request (struct request_queue *q, struct bio * bio)
145 {
146         mddev_t *mddev = q->queuedata;
147         multipath_conf_t *conf = mddev_to_conf(mddev);
148         struct multipath_bh * mp_bh;
149         struct multipath_info *multipath;
150         const int rw = bio_data_dir(bio);
151
152         if (unlikely(bio_barrier(bio))) {
153                 bio_endio(bio, -EOPNOTSUPP);
154                 return 0;
155         }
156
157         mp_bh = mempool_alloc(conf->pool, GFP_NOIO);
158
159         mp_bh->master_bio = bio;
160         mp_bh->mddev = mddev;
161
162         disk_stat_inc(mddev->gendisk, ios[rw]);
163         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
164
165         mp_bh->path = multipath_map(conf);
166         if (mp_bh->path < 0) {
167                 bio_endio(bio, -EIO);
168                 mempool_free(mp_bh, conf->pool);
169                 return 0;
170         }
171         multipath = conf->multipaths + mp_bh->path;
172
173         mp_bh->bio = *bio;
174         mp_bh->bio.bi_sector += multipath->rdev->data_offset;
175         mp_bh->bio.bi_bdev = multipath->rdev->bdev;
176         mp_bh->bio.bi_rw |= (1 << BIO_RW_FAILFAST);
177         mp_bh->bio.bi_end_io = multipath_end_request;
178         mp_bh->bio.bi_private = mp_bh;
179         generic_make_request(&mp_bh->bio);
180         return 0;
181 }
182
183 static void multipath_status (struct seq_file *seq, mddev_t *mddev)
184 {
185         multipath_conf_t *conf = mddev_to_conf(mddev);
186         int i;
187         
188         seq_printf (seq, " [%d/%d] [", conf->raid_disks,
189                                                  conf->working_disks);
190         for (i = 0; i < conf->raid_disks; i++)
191                 seq_printf (seq, "%s",
192                                conf->multipaths[i].rdev && 
193                                test_bit(In_sync, &conf->multipaths[i].rdev->flags) ? "U" : "_");
194         seq_printf (seq, "]");
195 }
196
197 static int multipath_congested(void *data, int bits)
198 {
199         mddev_t *mddev = data;
200         multipath_conf_t *conf = mddev_to_conf(mddev);
201         int i, ret = 0;
202
203         rcu_read_lock();
204         for (i = 0; i < mddev->raid_disks ; i++) {
205                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
206                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
207                         struct request_queue *q = bdev_get_queue(rdev->bdev);
208
209                         ret |= bdi_congested(&q->backing_dev_info, bits);
210                         /* Just like multipath_map, we just check the
211                          * first available device
212                          */
213                         break;
214                 }
215         }
216         rcu_read_unlock();
217         return ret;
218 }
219
220 /*
221  * Careful, this can execute in IRQ contexts as well!
222  */
223 static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
224 {
225         multipath_conf_t *conf = mddev_to_conf(mddev);
226
227         if (conf->working_disks <= 1) {
228                 /*
229                  * Uh oh, we can do nothing if this is our last path, but
230                  * first check if this is a queued request for a device
231                  * which has just failed.
232                  */
233                 printk(KERN_ALERT 
234                         "multipath: only one IO path left and IO error.\n");
235                 /* leave it active... it's all we have */
236         } else {
237                 /*
238                  * Mark disk as unusable
239                  */
240                 if (!test_bit(Faulty, &rdev->flags)) {
241                         char b[BDEVNAME_SIZE];
242                         clear_bit(In_sync, &rdev->flags);
243                         set_bit(Faulty, &rdev->flags);
244                         set_bit(MD_CHANGE_DEVS, &mddev->flags);
245                         conf->working_disks--;
246                         mddev->degraded++;
247                         printk(KERN_ALERT "multipath: IO failure on %s,"
248                                 " disabling IO path. \n Operation continuing"
249                                 " on %d IO paths.\n",
250                                 bdevname (rdev->bdev,b),
251                                 conf->working_disks);
252                 }
253         }
254 }
255
256 static void print_multipath_conf (multipath_conf_t *conf)
257 {
258         int i;
259         struct multipath_info *tmp;
260
261         printk("MULTIPATH conf printout:\n");
262         if (!conf) {
263                 printk("(conf==NULL)\n");
264                 return;
265         }
266         printk(" --- wd:%d rd:%d\n", conf->working_disks,
267                          conf->raid_disks);
268
269         for (i = 0; i < conf->raid_disks; i++) {
270                 char b[BDEVNAME_SIZE];
271                 tmp = conf->multipaths + i;
272                 if (tmp->rdev)
273                         printk(" disk%d, o:%d, dev:%s\n",
274                                 i,!test_bit(Faulty, &tmp->rdev->flags),
275                                bdevname(tmp->rdev->bdev,b));
276         }
277 }
278
279
280 static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
281 {
282         multipath_conf_t *conf = mddev->private;
283         struct request_queue *q;
284         int found = 0;
285         int path;
286         struct multipath_info *p;
287
288         print_multipath_conf(conf);
289
290         for (path=0; path<mddev->raid_disks; path++) 
291                 if ((p=conf->multipaths+path)->rdev == NULL) {
292                         q = rdev->bdev->bd_disk->queue;
293                         blk_queue_stack_limits(mddev->queue, q);
294
295                 /* as we don't honour merge_bvec_fn, we must never risk
296                  * violating it, so limit ->max_sector to one PAGE, as
297                  * a one page request is never in violation.
298                  * (Note: it is very unlikely that a device with
299                  * merge_bvec_fn will be involved in multipath.)
300                  */
301                         if (q->merge_bvec_fn &&
302                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
303                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
304
305                         conf->working_disks++;
306                         mddev->degraded--;
307                         rdev->raid_disk = path;
308                         set_bit(In_sync, &rdev->flags);
309                         rcu_assign_pointer(p->rdev, rdev);
310                         found = 1;
311                 }
312
313         print_multipath_conf(conf);
314         return found;
315 }
316
317 static int multipath_remove_disk(mddev_t *mddev, int number)
318 {
319         multipath_conf_t *conf = mddev->private;
320         int err = 0;
321         mdk_rdev_t *rdev;
322         struct multipath_info *p = conf->multipaths + number;
323
324         print_multipath_conf(conf);
325
326         rdev = p->rdev;
327         if (rdev) {
328                 if (test_bit(In_sync, &rdev->flags) ||
329                     atomic_read(&rdev->nr_pending)) {
330                         printk(KERN_ERR "hot-remove-disk, slot %d is identified"                                " but is still operational!\n", number);
331                         err = -EBUSY;
332                         goto abort;
333                 }
334                 p->rdev = NULL;
335                 synchronize_rcu();
336                 if (atomic_read(&rdev->nr_pending)) {
337                         /* lost the race, try later */
338                         err = -EBUSY;
339                         p->rdev = rdev;
340                 }
341         }
342 abort:
343
344         print_multipath_conf(conf);
345         return err;
346 }
347
348
349
350 /*
351  * This is a kernel thread which:
352  *
353  *      1.      Retries failed read operations on working multipaths.
354  *      2.      Updates the raid superblock when problems encounter.
355  *      3.      Performs writes following reads for array syncronising.
356  */
357
358 static void multipathd (mddev_t *mddev)
359 {
360         struct multipath_bh *mp_bh;
361         struct bio *bio;
362         unsigned long flags;
363         multipath_conf_t *conf = mddev_to_conf(mddev);
364         struct list_head *head = &conf->retry_list;
365
366         md_check_recovery(mddev);
367         for (;;) {
368                 char b[BDEVNAME_SIZE];
369                 spin_lock_irqsave(&conf->device_lock, flags);
370                 if (list_empty(head))
371                         break;
372                 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
373                 list_del(head->prev);
374                 spin_unlock_irqrestore(&conf->device_lock, flags);
375
376                 bio = &mp_bh->bio;
377                 bio->bi_sector = mp_bh->master_bio->bi_sector;
378                 
379                 if ((mp_bh->path = multipath_map (conf))<0) {
380                         printk(KERN_ALERT "multipath: %s: unrecoverable IO read"
381                                 " error for block %llu\n",
382                                 bdevname(bio->bi_bdev,b),
383                                 (unsigned long long)bio->bi_sector);
384                         multipath_end_bh_io(mp_bh, -EIO);
385                 } else {
386                         printk(KERN_ERR "multipath: %s: redirecting sector %llu"
387                                 " to another IO path\n",
388                                 bdevname(bio->bi_bdev,b),
389                                 (unsigned long long)bio->bi_sector);
390                         *bio = *(mp_bh->master_bio);
391                         bio->bi_sector += conf->multipaths[mp_bh->path].rdev->data_offset;
392                         bio->bi_bdev = conf->multipaths[mp_bh->path].rdev->bdev;
393                         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
394                         bio->bi_end_io = multipath_end_request;
395                         bio->bi_private = mp_bh;
396                         generic_make_request(bio);
397                 }
398         }
399         spin_unlock_irqrestore(&conf->device_lock, flags);
400 }
401
402 static int multipath_run (mddev_t *mddev)
403 {
404         multipath_conf_t *conf;
405         int disk_idx;
406         struct multipath_info *disk;
407         mdk_rdev_t *rdev;
408         struct list_head *tmp;
409
410         if (mddev->level != LEVEL_MULTIPATH) {
411                 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
412                        mdname(mddev), mddev->level);
413                 goto out;
414         }
415         /*
416          * copy the already verified devices into our private MULTIPATH
417          * bookkeeping area. [whatever we allocate in multipath_run(),
418          * should be freed in multipath_stop()]
419          */
420
421         conf = kzalloc(sizeof(multipath_conf_t), GFP_KERNEL);
422         mddev->private = conf;
423         if (!conf) {
424                 printk(KERN_ERR 
425                         "multipath: couldn't allocate memory for %s\n",
426                         mdname(mddev));
427                 goto out;
428         }
429
430         conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
431                                    GFP_KERNEL);
432         if (!conf->multipaths) {
433                 printk(KERN_ERR 
434                         "multipath: couldn't allocate memory for %s\n",
435                         mdname(mddev));
436                 goto out_free_conf;
437         }
438
439         conf->working_disks = 0;
440         ITERATE_RDEV(mddev,rdev,tmp) {
441                 disk_idx = rdev->raid_disk;
442                 if (disk_idx < 0 ||
443                     disk_idx >= mddev->raid_disks)
444                         continue;
445
446                 disk = conf->multipaths + disk_idx;
447                 disk->rdev = rdev;
448
449                 blk_queue_stack_limits(mddev->queue,
450                                        rdev->bdev->bd_disk->queue);
451                 /* as we don't honour merge_bvec_fn, we must never risk
452                  * violating it, not that we ever expect a device with
453                  * a merge_bvec_fn to be involved in multipath */
454                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
455                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
456                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
457
458                 if (!test_bit(Faulty, &rdev->flags))
459                         conf->working_disks++;
460         }
461
462         conf->raid_disks = mddev->raid_disks;
463         conf->mddev = mddev;
464         spin_lock_init(&conf->device_lock);
465         INIT_LIST_HEAD(&conf->retry_list);
466
467         if (!conf->working_disks) {
468                 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
469                         mdname(mddev));
470                 goto out_free_conf;
471         }
472         mddev->degraded = conf->raid_disks - conf->working_disks;
473
474         conf->pool = mempool_create_kzalloc_pool(NR_RESERVED_BUFS,
475                                                  sizeof(struct multipath_bh));
476         if (conf->pool == NULL) {
477                 printk(KERN_ERR 
478                         "multipath: couldn't allocate memory for %s\n",
479                         mdname(mddev));
480                 goto out_free_conf;
481         }
482
483         {
484                 mddev->thread = md_register_thread(multipathd, mddev, "%s_multipath");
485                 if (!mddev->thread) {
486                         printk(KERN_ERR "multipath: couldn't allocate thread"
487                                 " for %s\n", mdname(mddev));
488                         goto out_free_conf;
489                 }
490         }
491
492         printk(KERN_INFO 
493                 "multipath: array %s active with %d out of %d IO paths\n",
494                 mdname(mddev), conf->working_disks, mddev->raid_disks);
495         /*
496          * Ok, everything is just fine now
497          */
498         mddev->array_size = mddev->size;
499
500         mddev->queue->unplug_fn = multipath_unplug;
501         mddev->queue->backing_dev_info.congested_fn = multipath_congested;
502         mddev->queue->backing_dev_info.congested_data = mddev;
503
504         return 0;
505
506 out_free_conf:
507         if (conf->pool)
508                 mempool_destroy(conf->pool);
509         kfree(conf->multipaths);
510         kfree(conf);
511         mddev->private = NULL;
512 out:
513         return -EIO;
514 }
515
516
517 static int multipath_stop (mddev_t *mddev)
518 {
519         multipath_conf_t *conf = mddev_to_conf(mddev);
520
521         md_unregister_thread(mddev->thread);
522         mddev->thread = NULL;
523         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
524         mempool_destroy(conf->pool);
525         kfree(conf->multipaths);
526         kfree(conf);
527         mddev->private = NULL;
528         return 0;
529 }
530
531 static struct mdk_personality multipath_personality =
532 {
533         .name           = "multipath",
534         .level          = LEVEL_MULTIPATH,
535         .owner          = THIS_MODULE,
536         .make_request   = multipath_make_request,
537         .run            = multipath_run,
538         .stop           = multipath_stop,
539         .status         = multipath_status,
540         .error_handler  = multipath_error,
541         .hot_add_disk   = multipath_add_disk,
542         .hot_remove_disk= multipath_remove_disk,
543 };
544
545 static int __init multipath_init (void)
546 {
547         return register_md_personality (&multipath_personality);
548 }
549
550 static void __exit multipath_exit (void)
551 {
552         unregister_md_personality (&multipath_personality);
553 }
554
555 module_init(multipath_init);
556 module_exit(multipath_exit);
557 MODULE_LICENSE("GPL");
558 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
559 MODULE_ALIAS("md-multipath");
560 MODULE_ALIAS("md-level--4");