1/*
2 * Copyright (C) 2003 Sistina Software Limited.
3 * Copyright (C) 2005-2008 Red Hat, Inc. All rights reserved.
4 *
5 * This file is released under the GPL.
6 */
7
8#include "dm-bio-record.h"
9
10#include <linux/init.h>
11#include <linux/mempool.h>
12#include <linux/module.h>
13#include <linux/pagemap.h>
14#include <linux/slab.h>
15#include <linux/workqueue.h>
16#include <linux/device-mapper.h>
17#include <linux/dm-io.h>
18#include <linux/dm-dirty-log.h>
19#include <linux/dm-kcopyd.h>
20#include <linux/dm-region-hash.h>
21
22#define DM_MSG_PREFIX "raid1"
23
24#define MAX_RECOVERY 1 /* Maximum number of regions recovered in parallel. */
25
26#define MAX_NR_MIRRORS (DM_KCOPYD_MAX_REGIONS + 1)
27
28#define DM_RAID1_HANDLE_ERRORS 0x01
29#define DM_RAID1_KEEP_LOG 0x02
30#define errors_handled(p) ((p)->features & DM_RAID1_HANDLE_ERRORS)
31#define keep_log(p) ((p)->features & DM_RAID1_KEEP_LOG)
32
33static DECLARE_WAIT_QUEUE_HEAD(_kmirrord_recovery_stopped);
34
35/*-----------------------------------------------------------------
36 * Mirror set structures.
37 *---------------------------------------------------------------*/
38enum dm_raid1_error {
39 DM_RAID1_WRITE_ERROR,
40 DM_RAID1_FLUSH_ERROR,
41 DM_RAID1_SYNC_ERROR,
42 DM_RAID1_READ_ERROR
43};
44
45struct mirror {
46 struct mirror_set *ms;
47 atomic_t error_count;
48 unsigned long error_type;
49 struct dm_dev *dev;
50 sector_t offset;
51};
52
53struct mirror_set {
54 struct dm_target *ti;
55 struct list_head list;
56
57 uint64_t features;
58
59 spinlock_t lock; /* protects the lists */
60 struct bio_list reads;
61 struct bio_list writes;
62 struct bio_list failures;
63 struct bio_list holds; /* bios are waiting until suspend */
64
65 struct dm_region_hash *rh;
66 struct dm_kcopyd_client *kcopyd_client;
67 struct dm_io_client *io_client;
68
69 /* recovery */
70 region_t nr_regions;
71 int in_sync;
72 int log_failure;
73 int leg_failure;
74 atomic_t suspend;
75
76 atomic_t default_mirror; /* Default mirror */
77
78 struct workqueue_struct *kmirrord_wq;
79 struct work_struct kmirrord_work;
80 struct timer_list timer;
81 unsigned long timer_pending;
82
83 struct work_struct trigger_event;
84
85 unsigned nr_mirrors;
86 struct mirror mirror[0];
87};
88
89DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(raid1_resync_throttle,
90 "A percentage of time allocated for raid resynchronization");
91
92static void wakeup_mirrord(void *context)
93{
94 struct mirror_set *ms = context;
95
96 queue_work(ms->kmirrord_wq, &ms->kmirrord_work);
97}
98
99static void delayed_wake_fn(struct timer_list *t)
100{
101 struct mirror_set *ms = from_timer(ms, t, timer);
102
103 clear_bit(0, &ms->timer_pending);
104 wakeup_mirrord(ms);
105}
106
107static void delayed_wake(struct mirror_set *ms)
108{
109 if (test_and_set_bit(0, &ms->timer_pending))
110 return;
111
112 ms->timer.expires = jiffies + HZ / 5;
113 add_timer(&ms->timer);
114}
115
116static void wakeup_all_recovery_waiters(void *context)
117{
118 wake_up_all(&_kmirrord_recovery_stopped);
119}
120
121static void queue_bio(struct mirror_set *ms, struct bio *bio, int rw)
122{
123 unsigned long flags;
124 int should_wake = 0;
125 struct bio_list *bl;
126
127 bl = (rw == WRITE) ? &ms->writes : &ms->reads;
128 spin_lock_irqsave(&ms->lock, flags);
129 should_wake = !(bl->head);
130 bio_list_add(bl, bio);
131 spin_unlock_irqrestore(&ms->lock, flags);
132
133 if (should_wake)
134 wakeup_mirrord(ms);
135}
136
137static void dispatch_bios(void *context, struct bio_list *bio_list)
138{
139 struct mirror_set *ms = context;
140 struct bio *bio;
141
142 while ((bio = bio_list_pop(bio_list)))
143 queue_bio(ms, bio, WRITE);
144}
145
146struct dm_raid1_bio_record {
147 struct mirror *m;
148 /* if details->bi_disk == NULL, details were not saved */
149 struct dm_bio_details details;
150 region_t write_region;
151};
152
153/*
154 * Every mirror should look like this one.
155 */
156#define DEFAULT_MIRROR 0
157
158/*
159 * This is yucky. We squirrel the mirror struct away inside
160 * bi_next for read/write buffers. This is safe since the bh
161 * doesn't get submitted to the lower levels of block layer.
162 */
163static struct mirror *bio_get_m(struct bio *bio)
164{
165 return (struct mirror *) bio->bi_next;
166}
167
168static void bio_set_m(struct bio *bio, struct mirror *m)
169{
170 bio->bi_next = (struct bio *) m;
171}
172
173static struct mirror *get_default_mirror(struct mirror_set *ms)
174{
175 return &ms->mirror[atomic_read(&ms->default_mirror)];
176}
177
178static void set_default_mirror(struct mirror *m)
179{
180 struct mirror_set *ms = m->ms;
181 struct mirror *m0 = &(ms->mirror[0]);
182
183 atomic_set(&ms->default_mirror, m - m0);
184}
185
186static struct mirror *get_valid_mirror(struct mirror_set *ms)
187{
188 struct mirror *m;
189
190 for (m = ms->mirror; m < ms->mirror + ms->nr_mirrors; m++)
191 if (!atomic_read(&m->error_count))
192 return m;
193
194 return NULL;
195}
196
197/* fail_mirror
198 * @m: mirror device to fail
199 * @error_type: one of the enum's, DM_RAID1_*_ERROR
200 *
201 * If errors are being handled, record the type of
202 * error encountered for this device. If this type
203 * of error has already been recorded, we can return;
204 * otherwise, we must signal userspace by triggering
205 * an event. Additionally, if the device is the
206 * primary device, we must choose a new primary, but
207 * only if the mirror is in-sync.
208 *
209 * This function must not block.
210 */
211static void fail_mirror(struct mirror *m, enum dm_raid1_error error_type)
212{
213 struct mirror_set *ms = m->ms;
214 struct mirror *new;
215
216 ms->leg_failure = 1;
217
218 /*
219 * error_count is used for nothing more than a
220 * simple way to tell if a device has encountered
221 * errors.
222 */
223 atomic_inc(&m->error_count);
224
225 if (test_and_set_bit(error_type, &m->error_type))
226 return;
227
228 if (!errors_handled(ms))
229 return;
230
231 if (m != get_default_mirror(ms))
232 goto out;
233
234 if (!ms->in_sync && !keep_log(ms)) {
235 /*
236 * Better to issue requests to same failing device
237 * than to risk returning corrupt data.
238 */
239 DMERR("Primary mirror (%s) failed while out-of-sync: "
240 "Reads may fail.", m->dev->name);
241 goto out;
242 }
243
244 new = get_valid_mirror(ms);
245 if (new)
246 set_default_mirror(new);
247 else
248 DMWARN("All sides of mirror have failed.");
249
250out:
251 schedule_work(&ms->trigger_event);
252}
253
254static int mirror_flush(struct dm_target *ti)
255{
256 struct mirror_set *ms = ti->private;
257 unsigned long error_bits;
258
259 unsigned int i;
260 struct dm_io_region io[MAX_NR_MIRRORS];
261 struct mirror *m;
262 struct dm_io_request io_req = {
263 .bi_op = REQ_OP_WRITE,
264 .bi_op_flags = REQ_PREFLUSH | REQ_SYNC,
265 .mem.type = DM_IO_KMEM,
266 .mem.ptr.addr = NULL,
267 .client = ms->io_client,
268 };
269
270 for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++) {
271 io[i].bdev = m->dev->bdev;
272 io[i].sector = 0;
273 io[i].count = 0;
274 }
275
276 error_bits = -1;
277 dm_io(&io_req, ms->nr_mirrors, io, &error_bits);
278 if (unlikely(error_bits != 0)) {
279 for (i = 0; i < ms->nr_mirrors; i++)
280 if (test_bit(i, &error_bits))
281 fail_mirror(ms->mirror + i,
282 DM_RAID1_FLUSH_ERROR);
283 return -EIO;
284 }
285
286 return 0;
287}
288
289/*-----------------------------------------------------------------
290 * Recovery.
291 *
292 * When a mirror is first activated we may find that some regions
293 * are in the no-sync state. We have to recover these by
294 * recopying from the default mirror to all the others.
295 *---------------------------------------------------------------*/
296static void recovery_complete(int read_err, unsigned long write_err,
297 void *context)
298{
299 struct dm_region *reg = context;
300 struct mirror_set *ms = dm_rh_region_context(reg);
301 int m, bit = 0;
302
303 if (read_err) {
304 /* Read error means the failure of default mirror. */
305 DMERR_LIMIT("Unable to read primary mirror during recovery");
306 fail_mirror(get_default_mirror(ms), DM_RAID1_SYNC_ERROR);
307 }
308
309 if (write_err) {
310 DMERR_LIMIT("Write error during recovery (error = 0x%lx)",
311 write_err);
312 /*
313 * Bits correspond to devices (excluding default mirror).
314 * The default mirror cannot change during recovery.
315 */
316 for (m = 0; m < ms->nr_mirrors; m++) {
317 if (&ms->mirror[m] == get_default_mirror(ms))
318 continue;
319 if (test_bit(bit, &write_err))
320 fail_mirror(ms->mirror + m,
321 DM_RAID1_SYNC_ERROR);
322 bit++;
323 }
324 }
325
326 dm_rh_recovery_end(reg, !(read_err || write_err));
327}
328
329static void recover(struct mirror_set *ms, struct dm_region *reg)
330{
331 unsigned i;
332 struct dm_io_region from, to[DM_KCOPYD_MAX_REGIONS], *dest;
333 struct mirror *m;
334 unsigned long flags = 0;
335 region_t key = dm_rh_get_region_key(reg);
336 sector_t region_size = dm_rh_get_region_size(ms->rh);
337
338 /* fill in the source */
339 m = get_default_mirror(ms);
340 from.bdev = m->dev->bdev;
341 from.sector = m->offset + dm_rh_region_to_sector(ms->rh, key);
342 if (key == (ms->nr_regions - 1)) {
343 /*
344 * The final region may be smaller than
345 * region_size.
346 */
347 from.count = ms->ti->len & (region_size - 1);
348 if (!from.count)
349 from.count = region_size;
350 } else
351 from.count = region_size;
352
353 /* fill in the destinations */
354 for (i = 0, dest = to; i < ms->nr_mirrors; i++) {
355 if (&ms->mirror[i] == get_default_mirror(ms))
356 continue;
357
358 m = ms->mirror + i;
359 dest->bdev = m->dev->bdev;
360 dest->sector = m->offset + dm_rh_region_to_sector(ms->rh, key);
361 dest->count = from.count;
362 dest++;
363 }
364
365 /* hand to kcopyd */
366 if (!errors_handled(ms))
367 set_bit(DM_KCOPYD_IGNORE_ERROR, &flags);
368
369 dm_kcopyd_copy(ms->kcopyd_client, &from, ms->nr_mirrors - 1, to,
370 flags, recovery_complete, reg);
371}
372
373static void reset_ms_flags(struct mirror_set *ms)
374{
375 unsigned int m;
376
377 ms->leg_failure = 0;
378 for (m = 0; m < ms->nr_mirrors; m++) {
379 atomic_set(&(ms->mirror[m].error_count), 0);
380 ms->mirror[m].error_type = 0;
381 }
382}
383
384static void do_recovery(struct mirror_set *ms)
385{
386 struct dm_region *reg;
387 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
388
389 /*
390 * Start quiescing some regions.
391 */
392 dm_rh_recovery_prepare(ms->rh);
393
394 /*
395 * Copy any already quiesced regions.
396 */
397 while ((reg = dm_rh_recovery_start(ms->rh)))
398 recover(ms, reg);
399
400 /*
401 * Update the in sync flag.
402 */
403 if (!ms->in_sync &&
404 (log->type->get_sync_count(log) == ms->nr_regions)) {
405 /* the sync is complete */
406 dm_table_event(ms->ti->table);
407 ms->in_sync = 1;
408 reset_ms_flags(ms);
409 }
410}
411
412/*-----------------------------------------------------------------
413 * Reads
414 *---------------------------------------------------------------*/
415static struct mirror *choose_mirror(struct mirror_set *ms, sector_t sector)
416{
417 struct mirror *m = get_default_mirror(ms);
418
419 do {
420 if (likely(!atomic_read(&m->error_count)))
421 return m;
422
423 if (m-- == ms->mirror)
424 m += ms->nr_mirrors;
425 } while (m != get_default_mirror(ms));
426
427 return NULL;
428}
429
430static int default_ok(struct mirror *m)
431{
432 struct mirror *default_mirror = get_default_mirror(m->ms);
433
434 return !atomic_read(&default_mirror->error_count);
435}
436
437static int mirror_available(struct mirror_set *ms, struct bio *bio)
438{
439 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
440 region_t region = dm_rh_bio_to_region(ms->rh, bio);
441
442 if (log->type->in_sync(log, region, 0))
443 return choose_mirror(ms, bio->bi_iter.bi_sector) ? 1 : 0;
444
445 return 0;
446}
447
448/*
449 * remap a buffer to a particular mirror.
450 */
451static sector_t map_sector(struct mirror *m, struct bio *bio)
452{
453 if (unlikely(!bio->bi_iter.bi_size))
454 return 0;
455 return m->offset + dm_target_offset(m->ms->ti, bio->bi_iter.bi_sector);
456}
457
458static void map_bio(struct mirror *m, struct bio *bio)
459{
460 bio_set_dev(bio, m->dev->bdev);
461 bio->bi_iter.bi_sector = map_sector(m, bio);
462}
463
464static void map_region(struct dm_io_region *io, struct mirror *m,
465 struct bio *bio)
466{
467 io->bdev = m->dev->bdev;
468 io->sector = map_sector(m, bio);
469 io->count = bio_sectors(bio);
470}
471
472static void hold_bio(struct mirror_set *ms, struct bio *bio)
473{
474 /*
475 * Lock is required to avoid race condition during suspend
476 * process.
477 */
478 spin_lock_irq(&ms->lock);
479
480 if (atomic_read(&ms->suspend)) {
481 spin_unlock_irq(&ms->lock);
482
483 /*
484 * If device is suspended, complete the bio.
485 */
486 if (dm_noflush_suspending(ms->ti))
487 bio->bi_status = BLK_STS_DM_REQUEUE;
488 else
489 bio->bi_status = BLK_STS_IOERR;
490
491 bio_endio(bio);
492 return;
493 }
494
495 /*
496 * Hold bio until the suspend is complete.
497 */
498 bio_list_add(&ms->holds, bio);
499 spin_unlock_irq(&ms->lock);
500}
501
502/*-----------------------------------------------------------------
503 * Reads
504 *---------------------------------------------------------------*/
505static void read_callback(unsigned long error, void *context)
506{
507 struct bio *bio = context;
508 struct mirror *m;
509
510 m = bio_get_m(bio);
511 bio_set_m(bio, NULL);
512
513 if (likely(!error)) {
514 bio_endio(bio);
515 return;
516 }
517
518 fail_mirror(m, DM_RAID1_READ_ERROR);
519
520 if (likely(default_ok(m)) || mirror_available(m->ms, bio)) {
521 DMWARN_LIMIT("Read failure on mirror device %s. "
522 "Trying alternative device.",
523 m->dev->name);
524 queue_bio(m->ms, bio, bio_data_dir(bio));
525 return;
526 }
527
528 DMERR_LIMIT("Read failure on mirror device %s. Failing I/O.",
529 m->dev->name);
530 bio_io_error(bio);
531}
532
533/* Asynchronous read. */
534static void read_async_bio(struct mirror *m, struct bio *bio)
535{
536 struct dm_io_region io;
537 struct dm_io_request io_req = {
538 .bi_op = REQ_OP_READ,
539 .bi_op_flags = 0,
540 .mem.type = DM_IO_BIO,
541 .mem.ptr.bio = bio,
542 .notify.fn = read_callback,
543 .notify.context = bio,
544 .client = m->ms->io_client,
545 };
546
547 map_region(&io, m, bio);
548 bio_set_m(bio, m);
549 BUG_ON(dm_io(&io_req, 1, &io, NULL));
550}
551
552static inline int region_in_sync(struct mirror_set *ms, region_t region,
553 int may_block)
554{
555 int state = dm_rh_get_state(ms->rh, region, may_block);
556 return state == DM_RH_CLEAN || state == DM_RH_DIRTY;
557}
558
559static void do_reads(struct mirror_set *ms, struct bio_list *reads)
560{
561 region_t region;
562 struct bio *bio;
563 struct mirror *m;
564
565 while ((bio = bio_list_pop(reads))) {
566 region = dm_rh_bio_to_region(ms->rh, bio);
567 m = get_default_mirror(ms);
568
569 /*
570 * We can only read balance if the region is in sync.
571 */
572 if (likely(region_in_sync(ms, region, 1)))
573 m = choose_mirror(ms, bio->bi_iter.bi_sector);
574 else if (m && atomic_read(&m->error_count))
575 m = NULL;
576
577 if (likely(m))
578 read_async_bio(m, bio);
579 else
580 bio_io_error(bio);
581 }
582}
583
584/*-----------------------------------------------------------------
585 * Writes.
586 *
587 * We do different things with the write io depending on the
588 * state of the region that it's in:
589 *
590 * SYNC: increment pending, use kcopyd to write to *all* mirrors
591 * RECOVERING: delay the io until recovery completes
592 * NOSYNC: increment pending, just write to the default mirror
593 *---------------------------------------------------------------*/
594
595
596static void write_callback(unsigned long error, void *context)
597{
598 unsigned i;
599 struct bio *bio = (struct bio *) context;
600 struct mirror_set *ms;
601 int should_wake = 0;
602 unsigned long flags;
603
604 ms = bio_get_m(bio)->ms;
605 bio_set_m(bio, NULL);
606
607 /*
608 * NOTE: We don't decrement the pending count here,
609 * instead it is done by the targets endio function.
610 * This way we handle both writes to SYNC and NOSYNC
611 * regions with the same code.
612 */
613 if (likely(!error)) {
614 bio_endio(bio);
615 return;
616 }
617
618 /*
619 * If the bio is discard, return an error, but do not
620 * degrade the array.
621 */
622 if (bio_op(bio) == REQ_OP_DISCARD) {
623 bio->bi_status = BLK_STS_NOTSUPP;
624 bio_endio(bio);
625 return;
626 }
627
628 for (i = 0; i < ms->nr_mirrors; i++)
629 if (test_bit(i, &error))
630 fail_mirror(ms->mirror + i, DM_RAID1_WRITE_ERROR);
631
632 /*
633 * Need to raise event. Since raising
634 * events can block, we need to do it in
635 * the main thread.
636 */
637 spin_lock_irqsave(&ms->lock, flags);
638 if (!ms->failures.head)
639 should_wake = 1;
640 bio_list_add(&ms->failures, bio);
641 spin_unlock_irqrestore(&ms->lock, flags);
642 if (should_wake)
643 wakeup_mirrord(ms);
644}
645
646static void do_write(struct mirror_set *ms, struct bio *bio)
647{
648 unsigned int i;
649 struct dm_io_region io[MAX_NR_MIRRORS], *dest = io;
650 struct mirror *m;
651 struct dm_io_request io_req = {
652 .bi_op = REQ_OP_WRITE,
653 .bi_op_flags = bio->bi_opf & (REQ_FUA | REQ_PREFLUSH),
654 .mem.type = DM_IO_BIO,
655 .mem.ptr.bio = bio,
656 .notify.fn = write_callback,
657 .notify.context = bio,
658 .client = ms->io_client,
659 };
660
661 if (bio_op(bio) == REQ_OP_DISCARD) {
662 io_req.bi_op = REQ_OP_DISCARD;
663 io_req.mem.type = DM_IO_KMEM;
664 io_req.mem.ptr.addr = NULL;
665 }
666
667 for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++)
668 map_region(dest++, m, bio);
669
670 /*
671 * Use default mirror because we only need it to retrieve the reference
672 * to the mirror set in write_callback().
673 */
674 bio_set_m(bio, get_default_mirror(ms));
675
676 BUG_ON(dm_io(&io_req, ms->nr_mirrors, io, NULL));
677}
678
679static void do_writes(struct mirror_set *ms, struct bio_list *writes)
680{
681 int state;
682 struct bio *bio;
683 struct bio_list sync, nosync, recover, *this_list = NULL;
684 struct bio_list requeue;
685 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
686 region_t region;
687
688 if (!writes->head)
689 return;
690
691 /*
692 * Classify each write.
693 */
694 bio_list_init(&sync);
695 bio_list_init(&nosync);
696 bio_list_init(&recover);
697 bio_list_init(&requeue);
698
699 while ((bio = bio_list_pop(writes))) {
700 if ((bio->bi_opf & REQ_PREFLUSH) ||
701 (bio_op(bio) == REQ_OP_DISCARD)) {
702 bio_list_add(&sync, bio);
703 continue;
704 }
705
706 region = dm_rh_bio_to_region(ms->rh, bio);
707
708 if (log->type->is_remote_recovering &&
709 log->type->is_remote_recovering(log, region)) {
710 bio_list_add(&requeue, bio);
711 continue;
712 }
713
714 state = dm_rh_get_state(ms->rh, region, 1);
715 switch (state) {
716 case DM_RH_CLEAN:
717 case DM_RH_DIRTY:
718 this_list = &sync;
719 break;
720
721 case DM_RH_NOSYNC:
722 this_list = &nosync;
723 break;
724
725 case DM_RH_RECOVERING:
726 this_list = &recover;
727 break;
728 }
729
730 bio_list_add(this_list, bio);
731 }
732
733 /*
734 * Add bios that are delayed due to remote recovery
735 * back on to the write queue
736 */
737 if (unlikely(requeue.head)) {
738 spin_lock_irq(&ms->lock);
739 bio_list_merge(&ms->writes, &requeue);
740 spin_unlock_irq(&ms->lock);
741 delayed_wake(ms);
742 }
743
744 /*
745 * Increment the pending counts for any regions that will
746 * be written to (writes to recover regions are going to
747 * be delayed).
748 */
749 dm_rh_inc_pending(ms->rh, &sync);
750 dm_rh_inc_pending(ms->rh, &nosync);
751
752 /*
753 * If the flush fails on a previous call and succeeds here,
754 * we must not reset the log_failure variable. We need
755 * userspace interaction to do that.
756 */
757 ms->log_failure = dm_rh_flush(ms->rh) ? 1 : ms->log_failure;
758
759 /*
760 * Dispatch io.
761 */
762 if (unlikely(ms->log_failure) && errors_handled(ms)) {
763 spin_lock_irq(&ms->lock);
764 bio_list_merge(&ms->failures, &sync);
765 spin_unlock_irq(&ms->lock);
766 wakeup_mirrord(ms);
767 } else
768 while ((bio = bio_list_pop(&sync)))
769 do_write(ms, bio);
770
771 while ((bio = bio_list_pop(&recover)))
772 dm_rh_delay(ms->rh, bio);
773
774 while ((bio = bio_list_pop(&nosync))) {
775 if (unlikely(ms->leg_failure) && errors_handled(ms) && !keep_log(ms)) {
776 spin_lock_irq(&ms->lock);
777 bio_list_add(&ms->failures, bio);
778 spin_unlock_irq(&ms->lock);
779 wakeup_mirrord(ms);
780 } else {
781 map_bio(get_default_mirror(ms), bio);
782 generic_make_request(bio);
783 }
784 }
785}
786
787static void do_failures(struct mirror_set *ms, struct bio_list *failures)
788{
789 struct bio *bio;
790
791 if (likely(!failures->head))
792 return;
793
794 /*
795 * If the log has failed, unattempted writes are being
796 * put on the holds list. We can't issue those writes
797 * until a log has been marked, so we must store them.
798 *
799 * If a 'noflush' suspend is in progress, we can requeue
800 * the I/O's to the core. This give userspace a chance
801 * to reconfigure the mirror, at which point the core
802 * will reissue the writes. If the 'noflush' flag is
803 * not set, we have no choice but to return errors.
804 *
805 * Some writes on the failures list may have been
806 * submitted before the log failure and represent a
807 * failure to write to one of the devices. It is ok
808 * for us to treat them the same and requeue them
809 * as well.
810 */
811 while ((bio = bio_list_pop(failures))) {
812 if (!ms->log_failure) {
813 ms->in_sync = 0;
814 dm_rh_mark_nosync(ms->rh, bio);
815 }
816
817 /*
818 * If all the legs are dead, fail the I/O.
819 * If the device has failed and keep_log is enabled,
820 * fail the I/O.
821 *
822 * If we have been told to handle errors, and keep_log
823 * isn't enabled, hold the bio and wait for userspace to
824 * deal with the problem.
825 *
826 * Otherwise pretend that the I/O succeeded. (This would
827 * be wrong if the failed leg returned after reboot and
828 * got replicated back to the good legs.)
829 */
830 if (unlikely(!get_valid_mirror(ms) || (keep_log(ms) && ms->log_failure)))
831 bio_io_error(bio);
832 else if (errors_handled(ms) && !keep_log(ms))
833 hold_bio(ms, bio);
834 else
835 bio_endio(bio);
836 }
837}
838
839static void trigger_event(struct work_struct *work)
840{
841 struct mirror_set *ms =
842 container_of(work, struct mirror_set, trigger_event);
843
844 dm_table_event(ms->ti->table);
845}
846
847/*-----------------------------------------------------------------
848 * kmirrord
849 *---------------------------------------------------------------*/
850static void do_mirror(struct work_struct *work)
851{
852 struct mirror_set *ms = container_of(work, struct mirror_set,
853 kmirrord_work);
854 struct bio_list reads, writes, failures;
855 unsigned long flags;
856
857 spin_lock_irqsave(&ms->lock, flags);
858 reads = ms->reads;
859 writes = ms->writes;
860 failures = ms->failures;
861 bio_list_init(&ms->reads);
862 bio_list_init(&ms->writes);
863 bio_list_init(&ms->failures);
864 spin_unlock_irqrestore(&ms->lock, flags);
865
866 dm_rh_update_states(ms->rh, errors_handled(ms));
867 do_recovery(ms);
868 do_reads(ms, &reads);
869 do_writes(ms, &writes);
870 do_failures(ms, &failures);
871}
872
873/*-----------------------------------------------------------------
874 * Target functions
875 *---------------------------------------------------------------*/
876static struct mirror_set *alloc_context(unsigned int nr_mirrors,
877 uint32_t region_size,
878 struct dm_target *ti,
879 struct dm_dirty_log *dl)
880{
881 size_t len;
882 struct mirror_set *ms = NULL;
883
884 len = sizeof(*ms) + (sizeof(ms->mirror[0]) * nr_mirrors);
885
886 ms = kzalloc(len, GFP_KERNEL);
887 if (!ms) {
888 ti->error = "Cannot allocate mirror context";
889 return NULL;
890 }
891
892 spin_lock_init(&ms->lock);
893 bio_list_init(&ms->reads);
894 bio_list_init(&ms->writes);
895 bio_list_init(&ms->failures);
896 bio_list_init(&ms->holds);
897
898 ms->ti = ti;
899 ms->nr_mirrors = nr_mirrors;
900 ms->nr_regions = dm_sector_div_up(ti->len, region_size);
901 ms->in_sync = 0;
902 ms->log_failure = 0;
903 ms->leg_failure = 0;
904 atomic_set(&ms->suspend, 0);
905 atomic_set(&ms->default_mirror, DEFAULT_MIRROR);
906
907 ms->io_client = dm_io_client_create();
908 if (IS_ERR(ms->io_client)) {
909 ti->error = "Error creating dm_io client";
910 kfree(ms);
911 return NULL;
912 }
913
914 ms->rh = dm_region_hash_create(ms, dispatch_bios, wakeup_mirrord,
915 wakeup_all_recovery_waiters,
916 ms->ti->begin, MAX_RECOVERY,
917 dl, region_size, ms->nr_regions);
918 if (IS_ERR(ms->rh)) {
919 ti->error = "Error creating dirty region hash";
920 dm_io_client_destroy(ms->io_client);
921 kfree(ms);
922 return NULL;
923 }
924
925 return ms;
926}
927
928static void free_context(struct mirror_set *ms, struct dm_target *ti,
929 unsigned int m)
930{
931 while (m--)
932 dm_put_device(ti, ms->mirror[m].dev);
933
934 dm_io_client_destroy(ms->io_client);
935 dm_region_hash_destroy(ms->rh);
936 kfree(ms);
937}
938
939static int get_mirror(struct mirror_set *ms, struct dm_target *ti,
940 unsigned int mirror, char **argv)
941{
942 unsigned long long offset;
943 char dummy;
944 int ret;
945
946 if (sscanf(argv[1], "%llu%c", &offset, &dummy) != 1 ||
947 offset != (sector_t)offset) {
948 ti->error = "Invalid offset";
949 return -EINVAL;
950 }
951
952 ret = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table),
953 &ms->mirror[mirror].dev);
954 if (ret) {
955 ti->error = "Device lookup failure";
956 return ret;
957 }
958
959 ms->mirror[mirror].ms = ms;
960 atomic_set(&(ms->mirror[mirror].error_count), 0);
961 ms->mirror[mirror].error_type = 0;
962 ms->mirror[mirror].offset = offset;
963
964 return 0;
965}
966
967/*
968 * Create dirty log: log_type #log_params <log_params>
969 */
970static struct dm_dirty_log *create_dirty_log(struct dm_target *ti,
971 unsigned argc, char **argv,
972 unsigned *args_used)
973{
974 unsigned param_count;
975 struct dm_dirty_log *dl;
976 char dummy;
977
978 if (argc < 2) {
979 ti->error = "Insufficient mirror log arguments";
980 return NULL;
981 }
982
983 if (sscanf(argv[1], "%u%c", &param_count, &dummy) != 1) {
984 ti->error = "Invalid mirror log argument count";
985 return NULL;
986 }
987
988 *args_used = 2 + param_count;
989
990 if (argc < *args_used) {
991 ti->error = "Insufficient mirror log arguments";
992 return NULL;
993 }
994
995 dl = dm_dirty_log_create(argv[0], ti, mirror_flush, param_count,
996 argv + 2);
997 if (!dl) {
998 ti->error = "Error creating mirror dirty log";
999 return NULL;
1000 }
1001
1002 return dl;
1003}
1004
1005static int parse_features(struct mirror_set *ms, unsigned argc, char **argv,
1006 unsigned *args_used)
1007{
1008 unsigned num_features;
1009 struct dm_target *ti = ms->ti;
1010 char dummy;
1011 int i;
1012
1013 *args_used = 0;
1014
1015 if (!argc)
1016 return 0;
1017
1018 if (sscanf(argv[0], "%u%c", &num_features, &dummy) != 1) {
1019 ti->error = "Invalid number of features";
1020 return -EINVAL;
1021 }
1022
1023 argc--;
1024 argv++;
1025 (*args_used)++;
1026
1027 if (num_features > argc) {
1028 ti->error = "Not enough arguments to support feature count";
1029 return -EINVAL;
1030 }
1031
1032 for (i = 0; i < num_features; i++) {
1033 if (!strcmp("handle_errors", argv[0]))
1034 ms->features |= DM_RAID1_HANDLE_ERRORS;
1035 else if (!strcmp("keep_log", argv[0]))
1036 ms->features |= DM_RAID1_KEEP_LOG;
1037 else {
1038 ti->error = "Unrecognised feature requested";
1039 return -EINVAL;
1040 }
1041
1042 argc--;
1043 argv++;
1044 (*args_used)++;
1045 }
1046 if (!errors_handled(ms) && keep_log(ms)) {
1047 ti->error = "keep_log feature requires the handle_errors feature";
1048 return -EINVAL;
1049 }
1050
1051 return 0;
1052}
1053
1054/*
1055 * Construct a mirror mapping:
1056 *
1057 * log_type #log_params <log_params>
1058 * #mirrors [mirror_path offset]{2,}
1059 * [#features <features>]
1060 *
1061 * log_type is "core" or "disk"
1062 * #log_params is between 1 and 3
1063 *
1064 * If present, supported features are "handle_errors" and "keep_log".
1065 */
1066static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1067{
1068 int r;
1069 unsigned int nr_mirrors, m, args_used;
1070 struct mirror_set *ms;
1071 struct dm_dirty_log *dl;
1072 char dummy;
1073
1074 dl = create_dirty_log(ti, argc, argv, &args_used);
1075 if (!dl)
1076 return -EINVAL;
1077
1078 argv += args_used;
1079 argc -= args_used;
1080
1081 if (!argc || sscanf(argv[0], "%u%c", &nr_mirrors, &dummy) != 1 ||
1082 nr_mirrors < 2 || nr_mirrors > MAX_NR_MIRRORS) {
1083 ti->error = "Invalid number of mirrors";
1084 dm_dirty_log_destroy(dl);
1085 return -EINVAL;
1086 }
1087
1088 argv++, argc--;
1089
1090 if (argc < nr_mirrors * 2) {
1091 ti->error = "Too few mirror arguments";
1092 dm_dirty_log_destroy(dl);
1093 return -EINVAL;
1094 }
1095
1096 ms = alloc_context(nr_mirrors, dl->type->get_region_size(dl), ti, dl);
1097 if (!ms) {
1098 dm_dirty_log_destroy(dl);
1099 return -ENOMEM;
1100 }
1101
1102 /* Get the mirror parameter sets */
1103 for (m = 0; m < nr_mirrors; m++) {
1104 r = get_mirror(ms, ti, m, argv);
1105 if (r) {
1106 free_context(ms, ti, m);
1107 return r;
1108 }
1109 argv += 2;
1110 argc -= 2;
1111 }
1112
1113 ti->private = ms;
1114
1115 r = dm_set_target_max_io_len(ti, dm_rh_get_region_size(ms->rh));
1116 if (r)
1117 goto err_free_context;
1118
1119 ti->num_flush_bios = 1;
1120 ti->num_discard_bios = 1;
1121 ti->per_io_data_size = sizeof(struct dm_raid1_bio_record);
1122
1123 ms->kmirrord_wq = alloc_workqueue("kmirrord", WQ_MEM_RECLAIM, 0);
1124 if (!ms->kmirrord_wq) {
1125 DMERR("couldn't start kmirrord");
1126 r = -ENOMEM;
1127 goto err_free_context;
1128 }
1129 INIT_WORK(&ms->kmirrord_work, do_mirror);
1130 timer_setup(&ms->timer, delayed_wake_fn, 0);
1131 ms->timer_pending = 0;
1132 INIT_WORK(&ms->trigger_event, trigger_event);
1133
1134 r = parse_features(ms, argc, argv, &args_used);
1135 if (r)
1136 goto err_destroy_wq;
1137
1138 argv += args_used;
1139 argc -= args_used;
1140
1141 /*
1142 * Any read-balancing addition depends on the
1143 * DM_RAID1_HANDLE_ERRORS flag being present.
1144 * This is because the decision to balance depends
1145 * on the sync state of a region. If the above
1146 * flag is not present, we ignore errors; and
1147 * the sync state may be inaccurate.
1148 */
1149
1150 if (argc) {
1151 ti->error = "Too many mirror arguments";
1152 r = -EINVAL;
1153 goto err_destroy_wq;
1154 }
1155
1156 ms->kcopyd_client = dm_kcopyd_client_create(&dm_kcopyd_throttle);
1157 if (IS_ERR(ms->kcopyd_client)) {
1158 r = PTR_ERR(ms->kcopyd_client);
1159 goto err_destroy_wq;
1160 }
1161
1162 wakeup_mirrord(ms);
1163 return 0;
1164
1165err_destroy_wq:
1166 destroy_workqueue(ms->kmirrord_wq);
1167err_free_context:
1168 free_context(ms, ti, ms->nr_mirrors);
1169 return r;
1170}
1171
1172static void mirror_dtr(struct dm_target *ti)
1173{
1174 struct mirror_set *ms = (struct mirror_set *) ti->private;
1175
1176 del_timer_sync(&ms->timer);
1177 flush_workqueue(ms->kmirrord_wq);
1178 flush_work(&ms->trigger_event);
1179 dm_kcopyd_client_destroy(ms->kcopyd_client);
1180 destroy_workqueue(ms->kmirrord_wq);
1181 free_context(ms, ti, ms->nr_mirrors);
1182}
1183
1184/*
1185 * Mirror mapping function
1186 */
1187static int mirror_map(struct dm_target *ti, struct bio *bio)
1188{
1189 int r, rw = bio_data_dir(bio);
1190 struct mirror *m;
1191 struct mirror_set *ms = ti->private;
1192 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1193 struct dm_raid1_bio_record *bio_record =
1194 dm_per_bio_data(bio, sizeof(struct dm_raid1_bio_record));
1195
1196 bio_record->details.bi_disk = NULL;
1197
1198 if (rw == WRITE) {
1199 /* Save region for mirror_end_io() handler */
1200 bio_record->write_region = dm_rh_bio_to_region(ms->rh, bio);
1201 queue_bio(ms, bio, rw);
1202 return DM_MAPIO_SUBMITTED;
1203 }
1204
1205 r = log->type->in_sync(log, dm_rh_bio_to_region(ms->rh, bio), 0);
1206 if (r < 0 && r != -EWOULDBLOCK)
1207 return DM_MAPIO_KILL;
1208
1209 /*
1210 * If region is not in-sync queue the bio.
1211 */
1212 if (!r || (r == -EWOULDBLOCK)) {
1213 if (bio->bi_opf & REQ_RAHEAD)
1214 return DM_MAPIO_KILL;
1215
1216 queue_bio(ms, bio, rw);
1217 return DM_MAPIO_SUBMITTED;
1218 }
1219
1220 /*
1221 * The region is in-sync and we can perform reads directly.
1222 * Store enough information so we can retry if it fails.
1223 */
1224 m = choose_mirror(ms, bio->bi_iter.bi_sector);
1225 if (unlikely(!m))
1226 return DM_MAPIO_KILL;
1227
1228 dm_bio_record(&bio_record->details, bio);
1229 bio_record->m = m;
1230
1231 map_bio(m, bio);
1232
1233 return DM_MAPIO_REMAPPED;
1234}
1235
1236static int mirror_end_io(struct dm_target *ti, struct bio *bio,
1237 blk_status_t *error)
1238{
1239 int rw = bio_data_dir(bio);
1240 struct mirror_set *ms = (struct mirror_set *) ti->private;
1241 struct mirror *m = NULL;
1242 struct dm_bio_details *bd = NULL;
1243 struct dm_raid1_bio_record *bio_record =
1244 dm_per_bio_data(bio, sizeof(struct dm_raid1_bio_record));
1245
1246 /*
1247 * We need to dec pending if this was a write.
1248 */
1249 if (rw == WRITE) {
1250 if (!(bio->bi_opf & REQ_PREFLUSH) &&
1251 bio_op(bio) != REQ_OP_DISCARD)
1252 dm_rh_dec(ms->rh, bio_record->write_region);
1253 return DM_ENDIO_DONE;
1254 }
1255
1256 if (*error == BLK_STS_NOTSUPP)
1257 goto out;
1258
1259 if (bio->bi_opf & REQ_RAHEAD)
1260 goto out;
1261
1262 if (unlikely(*error)) {
1263 if (!bio_record->details.bi_disk) {
1264 /*
1265 * There wasn't enough memory to record necessary
1266 * information for a retry or there was no other
1267 * mirror in-sync.
1268 */
1269 DMERR_LIMIT("Mirror read failed.");
1270 return DM_ENDIO_DONE;
1271 }
1272
1273 m = bio_record->m;
1274
1275 DMERR("Mirror read failed from %s. Trying alternative device.",
1276 m->dev->name);
1277
1278 fail_mirror(m, DM_RAID1_READ_ERROR);
1279
1280 /*
1281 * A failed read is requeued for another attempt using an intact
1282 * mirror.
1283 */
1284 if (default_ok(m) || mirror_available(ms, bio)) {
1285 bd = &bio_record->details;
1286
1287 dm_bio_restore(bd, bio);
1288 bio_record->details.bi_disk = NULL;
1289 bio->bi_status = 0;
1290
1291 queue_bio(ms, bio, rw);
1292 return DM_ENDIO_INCOMPLETE;
1293 }
1294 DMERR("All replicated volumes dead, failing I/O");
1295 }
1296
1297out:
1298 bio_record->details.bi_disk = NULL;
1299
1300 return DM_ENDIO_DONE;
1301}
1302
1303static void mirror_presuspend(struct dm_target *ti)
1304{
1305 struct mirror_set *ms = (struct mirror_set *) ti->private;
1306 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1307
1308 struct bio_list holds;
1309 struct bio *bio;
1310
1311 atomic_set(&ms->suspend, 1);
1312
1313 /*
1314 * Process bios in the hold list to start recovery waiting
1315 * for bios in the hold list. After the process, no bio has
1316 * a chance to be added in the hold list because ms->suspend
1317 * is set.
1318 */
1319 spin_lock_irq(&ms->lock);
1320 holds = ms->holds;
1321 bio_list_init(&ms->holds);
1322 spin_unlock_irq(&ms->lock);
1323
1324 while ((bio = bio_list_pop(&holds)))
1325 hold_bio(ms, bio);
1326
1327 /*
1328 * We must finish up all the work that we've
1329 * generated (i.e. recovery work).
1330 */
1331 dm_rh_stop_recovery(ms->rh);
1332
1333 wait_event(_kmirrord_recovery_stopped,
1334 !dm_rh_recovery_in_flight(ms->rh));
1335
1336 if (log->type->presuspend && log->type->presuspend(log))
1337 /* FIXME: need better error handling */
1338 DMWARN("log presuspend failed");
1339
1340 /*
1341 * Now that recovery is complete/stopped and the
1342 * delayed bios are queued, we need to wait for
1343 * the worker thread to complete. This way,
1344 * we know that all of our I/O has been pushed.
1345 */
1346 flush_workqueue(ms->kmirrord_wq);
1347}
1348
1349static void mirror_postsuspend(struct dm_target *ti)
1350{
1351 struct mirror_set *ms = ti->private;
1352 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1353
1354 if (log->type->postsuspend && log->type->postsuspend(log))
1355 /* FIXME: need better error handling */
1356 DMWARN("log postsuspend failed");
1357}
1358
1359static void mirror_resume(struct dm_target *ti)
1360{
1361 struct mirror_set *ms = ti->private;
1362 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1363
1364 atomic_set(&ms->suspend, 0);
1365 if (log->type->resume && log->type->resume(log))
1366 /* FIXME: need better error handling */
1367 DMWARN("log resume failed");
1368 dm_rh_start_recovery(ms->rh);
1369}
1370
1371/*
1372 * device_status_char
1373 * @m: mirror device/leg we want the status of
1374 *
1375 * We return one character representing the most severe error
1376 * we have encountered.
1377 * A => Alive - No failures
1378 * D => Dead - A write failure occurred leaving mirror out-of-sync
1379 * S => Sync - A sychronization failure occurred, mirror out-of-sync
1380 * R => Read - A read failure occurred, mirror data unaffected
1381 *
1382 * Returns: <char>
1383 */
1384static char device_status_char(struct mirror *m)
1385{
1386 if (!atomic_read(&(m->error_count)))
1387 return 'A';
1388
1389 return (test_bit(DM_RAID1_FLUSH_ERROR, &(m->error_type))) ? 'F' :
1390 (test_bit(DM_RAID1_WRITE_ERROR, &(m->error_type))) ? 'D' :
1391 (test_bit(DM_RAID1_SYNC_ERROR, &(m->error_type))) ? 'S' :
1392 (test_bit(DM_RAID1_READ_ERROR, &(m->error_type))) ? 'R' : 'U';
1393}
1394
1395
1396static void mirror_status(struct dm_target *ti, status_type_t type,
1397 unsigned status_flags, char *result, unsigned maxlen)
1398{
1399 unsigned int m, sz = 0;
1400 int num_feature_args = 0;
1401 struct mirror_set *ms = (struct mirror_set *) ti->private;
1402 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1403 char buffer[MAX_NR_MIRRORS + 1];
1404
1405 switch (type) {
1406 case STATUSTYPE_INFO:
1407 DMEMIT("%d ", ms->nr_mirrors);
1408 for (m = 0; m < ms->nr_mirrors; m++) {
1409 DMEMIT("%s ", ms->mirror[m].dev->name);
1410 buffer[m] = device_status_char(&(ms->mirror[m]));
1411 }
1412 buffer[m] = '\0';
1413
1414 DMEMIT("%llu/%llu 1 %s ",
1415 (unsigned long long)log->type->get_sync_count(log),
1416 (unsigned long long)ms->nr_regions, buffer);
1417
1418 sz += log->type->status(log, type, result+sz, maxlen-sz);
1419
1420 break;
1421
1422 case STATUSTYPE_TABLE:
1423 sz = log->type->status(log, type, result, maxlen);
1424
1425 DMEMIT("%d", ms->nr_mirrors);
1426 for (m = 0; m < ms->nr_mirrors; m++)
1427 DMEMIT(" %s %llu", ms->mirror[m].dev->name,
1428 (unsigned long long)ms->mirror[m].offset);
1429
1430 num_feature_args += !!errors_handled(ms);
1431 num_feature_args += !!keep_log(ms);
1432 if (num_feature_args) {
1433 DMEMIT(" %d", num_feature_args);
1434 if (errors_handled(ms))
1435 DMEMIT(" handle_errors");
1436 if (keep_log(ms))
1437 DMEMIT(" keep_log");
1438 }
1439
1440 break;
1441 }
1442}
1443
1444static int mirror_iterate_devices(struct dm_target *ti,
1445 iterate_devices_callout_fn fn, void *data)
1446{
1447 struct mirror_set *ms = ti->private;
1448 int ret = 0;
1449 unsigned i;
1450
1451 for (i = 0; !ret && i < ms->nr_mirrors; i++)
1452 ret = fn(ti, ms->mirror[i].dev,
1453 ms->mirror[i].offset, ti->len, data);
1454
1455 return ret;
1456}
1457
1458static struct target_type mirror_target = {
1459 .name = "mirror",
1460 .version = {1, 14, 0},
1461 .module = THIS_MODULE,
1462 .ctr = mirror_ctr,
1463 .dtr = mirror_dtr,
1464 .map = mirror_map,
1465 .end_io = mirror_end_io,
1466 .presuspend = mirror_presuspend,
1467 .postsuspend = mirror_postsuspend,
1468 .resume = mirror_resume,
1469 .status = mirror_status,
1470 .iterate_devices = mirror_iterate_devices,
1471};
1472
1473static int __init dm_mirror_init(void)
1474{
1475 int r;
1476
1477 r = dm_register_target(&mirror_target);
1478 if (r < 0) {
1479 DMERR("Failed to register mirror target");
1480 goto bad_target;
1481 }
1482
1483 return 0;
1484
1485bad_target:
1486 return r;
1487}
1488
1489static void __exit dm_mirror_exit(void)
1490{
1491 dm_unregister_target(&mirror_target);
1492}
1493
1494/* Module hooks */
1495module_init(dm_mirror_init);
1496module_exit(dm_mirror_exit);
1497
1498MODULE_DESCRIPTION(DM_NAME " mirror target");
1499MODULE_AUTHOR("Joe Thornber");
1500MODULE_LICENSE("GPL");
1501