block: move stats from disk to part0
[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                         blk_unplug(r_queue);
129
130                         rdev_dec_pending(rdev, mddev);
131                         rcu_read_lock();
132                 }
133         }
134         rcu_read_unlock();
135 }
136
137 static void multipath_unplug(struct request_queue *q)
138 {
139         unplug_slaves(q->queuedata);
140 }
141
142
143 static int multipath_make_request (struct request_queue *q, struct bio * bio)
144 {
145         mddev_t *mddev = q->queuedata;
146         multipath_conf_t *conf = mddev_to_conf(mddev);
147         struct multipath_bh * mp_bh;
148         struct multipath_info *multipath;
149         const int rw = bio_data_dir(bio);
150         int cpu;
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         cpu = part_stat_lock();
163         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
164         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
165                       bio_sectors(bio));
166         part_stat_unlock();
167
168         mp_bh->path = multipath_map(conf);
169         if (mp_bh->path < 0) {
170                 bio_endio(bio, -EIO);
171                 mempool_free(mp_bh, conf->pool);
172                 return 0;
173         }
174         multipath = conf->multipaths + mp_bh->path;
175
176         mp_bh->bio = *bio;
177         mp_bh->bio.bi_sector += multipath->rdev->data_offset;
178         mp_bh->bio.bi_bdev = multipath->rdev->bdev;
179         mp_bh->bio.bi_rw |= (1 << BIO_RW_FAILFAST);
180         mp_bh->bio.bi_end_io = multipath_end_request;
181         mp_bh->bio.bi_private = mp_bh;
182         generic_make_request(&mp_bh->bio);
183         return 0;
184 }
185
186 static void multipath_status (struct seq_file *seq, mddev_t *mddev)
187 {
188         multipath_conf_t *conf = mddev_to_conf(mddev);
189         int i;
190         
191         seq_printf (seq, " [%d/%d] [", conf->raid_disks,
192                                                  conf->working_disks);
193         for (i = 0; i < conf->raid_disks; i++)
194                 seq_printf (seq, "%s",
195                                conf->multipaths[i].rdev && 
196                                test_bit(In_sync, &conf->multipaths[i].rdev->flags) ? "U" : "_");
197         seq_printf (seq, "]");
198 }
199
200 static int multipath_congested(void *data, int bits)
201 {
202         mddev_t *mddev = data;
203         multipath_conf_t *conf = mddev_to_conf(mddev);
204         int i, ret = 0;
205
206         rcu_read_lock();
207         for (i = 0; i < mddev->raid_disks ; i++) {
208                 mdk_rdev_t *rdev = rcu_dereference(conf->multipaths[i].rdev);
209                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
210                         struct request_queue *q = bdev_get_queue(rdev->bdev);
211
212                         ret |= bdi_congested(&q->backing_dev_info, bits);
213                         /* Just like multipath_map, we just check the
214                          * first available device
215                          */
216                         break;
217                 }
218         }
219         rcu_read_unlock();
220         return ret;
221 }
222
223 /*
224  * Careful, this can execute in IRQ contexts as well!
225  */
226 static void multipath_error (mddev_t *mddev, mdk_rdev_t *rdev)
227 {
228         multipath_conf_t *conf = mddev_to_conf(mddev);
229
230         if (conf->working_disks <= 1) {
231                 /*
232                  * Uh oh, we can do nothing if this is our last path, but
233                  * first check if this is a queued request for a device
234                  * which has just failed.
235                  */
236                 printk(KERN_ALERT 
237                         "multipath: only one IO path left and IO error.\n");
238                 /* leave it active... it's all we have */
239         } else {
240                 /*
241                  * Mark disk as unusable
242                  */
243                 if (!test_bit(Faulty, &rdev->flags)) {
244                         char b[BDEVNAME_SIZE];
245                         clear_bit(In_sync, &rdev->flags);
246                         set_bit(Faulty, &rdev->flags);
247                         set_bit(MD_CHANGE_DEVS, &mddev->flags);
248                         conf->working_disks--;
249                         mddev->degraded++;
250                         printk(KERN_ALERT "multipath: IO failure on %s,"
251                                 " disabling IO path.\n"
252                                 "multipath: Operation continuing"
253                                 " on %d IO paths.\n",
254                                 bdevname (rdev->bdev,b),
255                                 conf->working_disks);
256                 }
257         }
258 }
259
260 static void print_multipath_conf (multipath_conf_t *conf)
261 {
262         int i;
263         struct multipath_info *tmp;
264
265         printk("MULTIPATH conf printout:\n");
266         if (!conf) {
267                 printk("(conf==NULL)\n");
268                 return;
269         }
270         printk(" --- wd:%d rd:%d\n", conf->working_disks,
271                          conf->raid_disks);
272
273         for (i = 0; i < conf->raid_disks; i++) {
274                 char b[BDEVNAME_SIZE];
275                 tmp = conf->multipaths + i;
276                 if (tmp->rdev)
277                         printk(" disk%d, o:%d, dev:%s\n",
278                                 i,!test_bit(Faulty, &tmp->rdev->flags),
279                                bdevname(tmp->rdev->bdev,b));
280         }
281 }
282
283
284 static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
285 {
286         multipath_conf_t *conf = mddev->private;
287         struct request_queue *q;
288         int err = -EEXIST;
289         int path;
290         struct multipath_info *p;
291         int first = 0;
292         int last = mddev->raid_disks - 1;
293
294         if (rdev->raid_disk >= 0)
295                 first = last = rdev->raid_disk;
296
297         print_multipath_conf(conf);
298
299         for (path = first; path <= last; path++)
300                 if ((p=conf->multipaths+path)->rdev == NULL) {
301                         q = rdev->bdev->bd_disk->queue;
302                         blk_queue_stack_limits(mddev->queue, q);
303
304                 /* as we don't honour merge_bvec_fn, we must never risk
305                  * violating it, so limit ->max_sector to one PAGE, as
306                  * a one page request is never in violation.
307                  * (Note: it is very unlikely that a device with
308                  * merge_bvec_fn will be involved in multipath.)
309                  */
310                         if (q->merge_bvec_fn &&
311                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
312                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
313
314                         conf->working_disks++;
315                         mddev->degraded--;
316                         rdev->raid_disk = path;
317                         set_bit(In_sync, &rdev->flags);
318                         rcu_assign_pointer(p->rdev, rdev);
319                         err = 0;
320                         break;
321                 }
322
323         print_multipath_conf(conf);
324
325         return err;
326 }
327
328 static int multipath_remove_disk(mddev_t *mddev, int number)
329 {
330         multipath_conf_t *conf = mddev->private;
331         int err = 0;
332         mdk_rdev_t *rdev;
333         struct multipath_info *p = conf->multipaths + number;
334
335         print_multipath_conf(conf);
336
337         rdev = p->rdev;
338         if (rdev) {
339                 if (test_bit(In_sync, &rdev->flags) ||
340                     atomic_read(&rdev->nr_pending)) {
341                         printk(KERN_ERR "hot-remove-disk, slot %d is identified"
342                                " but is still operational!\n", number);
343                         err = -EBUSY;
344                         goto abort;
345                 }
346                 p->rdev = NULL;
347                 synchronize_rcu();
348                 if (atomic_read(&rdev->nr_pending)) {
349                         /* lost the race, try later */
350                         err = -EBUSY;
351                         p->rdev = rdev;
352                 }
353         }
354 abort:
355
356         print_multipath_conf(conf);
357         return err;
358 }
359
360
361
362 /*
363  * This is a kernel thread which:
364  *
365  *      1.      Retries failed read operations on working multipaths.
366  *      2.      Updates the raid superblock when problems encounter.
367  *      3.      Performs writes following reads for array syncronising.
368  */
369
370 static void multipathd (mddev_t *mddev)
371 {
372         struct multipath_bh *mp_bh;
373         struct bio *bio;
374         unsigned long flags;
375         multipath_conf_t *conf = mddev_to_conf(mddev);
376         struct list_head *head = &conf->retry_list;
377
378         md_check_recovery(mddev);
379         for (;;) {
380                 char b[BDEVNAME_SIZE];
381                 spin_lock_irqsave(&conf->device_lock, flags);
382                 if (list_empty(head))
383                         break;
384                 mp_bh = list_entry(head->prev, struct multipath_bh, retry_list);
385                 list_del(head->prev);
386                 spin_unlock_irqrestore(&conf->device_lock, flags);
387
388                 bio = &mp_bh->bio;
389                 bio->bi_sector = mp_bh->master_bio->bi_sector;
390                 
391                 if ((mp_bh->path = multipath_map (conf))<0) {
392                         printk(KERN_ALERT "multipath: %s: unrecoverable IO read"
393                                 " error for block %llu\n",
394                                 bdevname(bio->bi_bdev,b),
395                                 (unsigned long long)bio->bi_sector);
396                         multipath_end_bh_io(mp_bh, -EIO);
397                 } else {
398                         printk(KERN_ERR "multipath: %s: redirecting sector %llu"
399                                 " to another IO path\n",
400                                 bdevname(bio->bi_bdev,b),
401                                 (unsigned long long)bio->bi_sector);
402                         *bio = *(mp_bh->master_bio);
403                         bio->bi_sector += conf->multipaths[mp_bh->path].rdev->data_offset;
404                         bio->bi_bdev = conf->multipaths[mp_bh->path].rdev->bdev;
405                         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
406                         bio->bi_end_io = multipath_end_request;
407                         bio->bi_private = mp_bh;
408                         generic_make_request(bio);
409                 }
410         }
411         spin_unlock_irqrestore(&conf->device_lock, flags);
412 }
413
414 static int multipath_run (mddev_t *mddev)
415 {
416         multipath_conf_t *conf;
417         int disk_idx;
418         struct multipath_info *disk;
419         mdk_rdev_t *rdev;
420         struct list_head *tmp;
421
422         if (mddev->level != LEVEL_MULTIPATH) {
423                 printk("multipath: %s: raid level not set to multipath IO (%d)\n",
424                        mdname(mddev), mddev->level);
425                 goto out;
426         }
427         /*
428          * copy the already verified devices into our private MULTIPATH
429          * bookkeeping area. [whatever we allocate in multipath_run(),
430          * should be freed in multipath_stop()]
431          */
432         mddev->queue->queue_lock = &mddev->queue->__queue_lock;
433
434         conf = kzalloc(sizeof(multipath_conf_t), GFP_KERNEL);
435         mddev->private = conf;
436         if (!conf) {
437                 printk(KERN_ERR 
438                         "multipath: couldn't allocate memory for %s\n",
439                         mdname(mddev));
440                 goto out;
441         }
442
443         conf->multipaths = kzalloc(sizeof(struct multipath_info)*mddev->raid_disks,
444                                    GFP_KERNEL);
445         if (!conf->multipaths) {
446                 printk(KERN_ERR 
447                         "multipath: couldn't allocate memory for %s\n",
448                         mdname(mddev));
449                 goto out_free_conf;
450         }
451
452         conf->working_disks = 0;
453         rdev_for_each(rdev, tmp, mddev) {
454                 disk_idx = rdev->raid_disk;
455                 if (disk_idx < 0 ||
456                     disk_idx >= mddev->raid_disks)
457                         continue;
458
459                 disk = conf->multipaths + disk_idx;
460                 disk->rdev = rdev;
461
462                 blk_queue_stack_limits(mddev->queue,
463                                        rdev->bdev->bd_disk->queue);
464                 /* as we don't honour merge_bvec_fn, we must never risk
465                  * violating it, not that we ever expect a device with
466                  * a merge_bvec_fn to be involved in multipath */
467                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
468                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
469                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
470
471                 if (!test_bit(Faulty, &rdev->flags))
472                         conf->working_disks++;
473         }
474
475         conf->raid_disks = mddev->raid_disks;
476         conf->mddev = mddev;
477         spin_lock_init(&conf->device_lock);
478         INIT_LIST_HEAD(&conf->retry_list);
479
480         if (!conf->working_disks) {
481                 printk(KERN_ERR "multipath: no operational IO paths for %s\n",
482                         mdname(mddev));
483                 goto out_free_conf;
484         }
485         mddev->degraded = conf->raid_disks - conf->working_disks;
486
487         conf->pool = mempool_create_kzalloc_pool(NR_RESERVED_BUFS,
488                                                  sizeof(struct multipath_bh));
489         if (conf->pool == NULL) {
490                 printk(KERN_ERR 
491                         "multipath: couldn't allocate memory for %s\n",
492                         mdname(mddev));
493                 goto out_free_conf;
494         }
495
496         {
497                 mddev->thread = md_register_thread(multipathd, mddev, "%s_multipath");
498                 if (!mddev->thread) {
499                         printk(KERN_ERR "multipath: couldn't allocate thread"
500                                 " for %s\n", mdname(mddev));
501                         goto out_free_conf;
502                 }
503         }
504
505         printk(KERN_INFO 
506                 "multipath: array %s active with %d out of %d IO paths\n",
507                 mdname(mddev), conf->working_disks, mddev->raid_disks);
508         /*
509          * Ok, everything is just fine now
510          */
511         mddev->array_sectors = mddev->size * 2;
512
513         mddev->queue->unplug_fn = multipath_unplug;
514         mddev->queue->backing_dev_info.congested_fn = multipath_congested;
515         mddev->queue->backing_dev_info.congested_data = mddev;
516
517         return 0;
518
519 out_free_conf:
520         if (conf->pool)
521                 mempool_destroy(conf->pool);
522         kfree(conf->multipaths);
523         kfree(conf);
524         mddev->private = NULL;
525 out:
526         return -EIO;
527 }
528
529
530 static int multipath_stop (mddev_t *mddev)
531 {
532         multipath_conf_t *conf = mddev_to_conf(mddev);
533
534         md_unregister_thread(mddev->thread);
535         mddev->thread = NULL;
536         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
537         mempool_destroy(conf->pool);
538         kfree(conf->multipaths);
539         kfree(conf);
540         mddev->private = NULL;
541         return 0;
542 }
543
544 static struct mdk_personality multipath_personality =
545 {
546         .name           = "multipath",
547         .level          = LEVEL_MULTIPATH,
548         .owner          = THIS_MODULE,
549         .make_request   = multipath_make_request,
550         .run            = multipath_run,
551         .stop           = multipath_stop,
552         .status         = multipath_status,
553         .error_handler  = multipath_error,
554         .hot_add_disk   = multipath_add_disk,
555         .hot_remove_disk= multipath_remove_disk,
556 };
557
558 static int __init multipath_init (void)
559 {
560         return register_md_personality (&multipath_personality);
561 }
562
563 static void __exit multipath_exit (void)
564 {
565         unregister_md_personality (&multipath_personality);
566 }
567
568 module_init(multipath_init);
569 module_exit(multipath_exit);
570 MODULE_LICENSE("GPL");
571 MODULE_ALIAS("md-personality-7"); /* MULTIPATH */
572 MODULE_ALIAS("md-multipath");
573 MODULE_ALIAS("md-level--4");