xref: /openbmc/linux/drivers/md/dm-raid1.c (revision b04b4f78)
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 #define DM_IO_PAGES 64
26 #define DM_KCOPYD_PAGES 64
27 
28 #define DM_RAID1_HANDLE_ERRORS 0x01
29 #define errors_handled(p)	((p)->features & DM_RAID1_HANDLE_ERRORS)
30 
31 static DECLARE_WAIT_QUEUE_HEAD(_kmirrord_recovery_stopped);
32 
33 /*-----------------------------------------------------------------
34  * Mirror set structures.
35  *---------------------------------------------------------------*/
36 enum dm_raid1_error {
37 	DM_RAID1_WRITE_ERROR,
38 	DM_RAID1_SYNC_ERROR,
39 	DM_RAID1_READ_ERROR
40 };
41 
42 struct mirror {
43 	struct mirror_set *ms;
44 	atomic_t error_count;
45 	unsigned long error_type;
46 	struct dm_dev *dev;
47 	sector_t offset;
48 };
49 
50 struct mirror_set {
51 	struct dm_target *ti;
52 	struct list_head list;
53 
54 	uint64_t features;
55 
56 	spinlock_t lock;	/* protects the lists */
57 	struct bio_list reads;
58 	struct bio_list writes;
59 	struct bio_list failures;
60 
61 	struct dm_region_hash *rh;
62 	struct dm_kcopyd_client *kcopyd_client;
63 	struct dm_io_client *io_client;
64 	mempool_t *read_record_pool;
65 
66 	/* recovery */
67 	region_t nr_regions;
68 	int in_sync;
69 	int log_failure;
70 	atomic_t suspend;
71 
72 	atomic_t default_mirror;	/* Default mirror */
73 
74 	struct workqueue_struct *kmirrord_wq;
75 	struct work_struct kmirrord_work;
76 	struct timer_list timer;
77 	unsigned long timer_pending;
78 
79 	struct work_struct trigger_event;
80 
81 	unsigned nr_mirrors;
82 	struct mirror mirror[0];
83 };
84 
85 static void wakeup_mirrord(void *context)
86 {
87 	struct mirror_set *ms = context;
88 
89 	queue_work(ms->kmirrord_wq, &ms->kmirrord_work);
90 }
91 
92 static void delayed_wake_fn(unsigned long data)
93 {
94 	struct mirror_set *ms = (struct mirror_set *) data;
95 
96 	clear_bit(0, &ms->timer_pending);
97 	wakeup_mirrord(ms);
98 }
99 
100 static void delayed_wake(struct mirror_set *ms)
101 {
102 	if (test_and_set_bit(0, &ms->timer_pending))
103 		return;
104 
105 	ms->timer.expires = jiffies + HZ / 5;
106 	ms->timer.data = (unsigned long) ms;
107 	ms->timer.function = delayed_wake_fn;
108 	add_timer(&ms->timer);
109 }
110 
111 static void wakeup_all_recovery_waiters(void *context)
112 {
113 	wake_up_all(&_kmirrord_recovery_stopped);
114 }
115 
116 static void queue_bio(struct mirror_set *ms, struct bio *bio, int rw)
117 {
118 	unsigned long flags;
119 	int should_wake = 0;
120 	struct bio_list *bl;
121 
122 	bl = (rw == WRITE) ? &ms->writes : &ms->reads;
123 	spin_lock_irqsave(&ms->lock, flags);
124 	should_wake = !(bl->head);
125 	bio_list_add(bl, bio);
126 	spin_unlock_irqrestore(&ms->lock, flags);
127 
128 	if (should_wake)
129 		wakeup_mirrord(ms);
130 }
131 
132 static void dispatch_bios(void *context, struct bio_list *bio_list)
133 {
134 	struct mirror_set *ms = context;
135 	struct bio *bio;
136 
137 	while ((bio = bio_list_pop(bio_list)))
138 		queue_bio(ms, bio, WRITE);
139 }
140 
141 #define MIN_READ_RECORDS 20
142 struct dm_raid1_read_record {
143 	struct mirror *m;
144 	struct dm_bio_details details;
145 };
146 
147 static struct kmem_cache *_dm_raid1_read_record_cache;
148 
149 /*
150  * Every mirror should look like this one.
151  */
152 #define DEFAULT_MIRROR 0
153 
154 /*
155  * This is yucky.  We squirrel the mirror struct away inside
156  * bi_next for read/write buffers.  This is safe since the bh
157  * doesn't get submitted to the lower levels of block layer.
158  */
159 static struct mirror *bio_get_m(struct bio *bio)
160 {
161 	return (struct mirror *) bio->bi_next;
162 }
163 
164 static void bio_set_m(struct bio *bio, struct mirror *m)
165 {
166 	bio->bi_next = (struct bio *) m;
167 }
168 
169 static struct mirror *get_default_mirror(struct mirror_set *ms)
170 {
171 	return &ms->mirror[atomic_read(&ms->default_mirror)];
172 }
173 
174 static void set_default_mirror(struct mirror *m)
175 {
176 	struct mirror_set *ms = m->ms;
177 	struct mirror *m0 = &(ms->mirror[0]);
178 
179 	atomic_set(&ms->default_mirror, m - m0);
180 }
181 
182 /* fail_mirror
183  * @m: mirror device to fail
184  * @error_type: one of the enum's, DM_RAID1_*_ERROR
185  *
186  * If errors are being handled, record the type of
187  * error encountered for this device.  If this type
188  * of error has already been recorded, we can return;
189  * otherwise, we must signal userspace by triggering
190  * an event.  Additionally, if the device is the
191  * primary device, we must choose a new primary, but
192  * only if the mirror is in-sync.
193  *
194  * This function must not block.
195  */
196 static void fail_mirror(struct mirror *m, enum dm_raid1_error error_type)
197 {
198 	struct mirror_set *ms = m->ms;
199 	struct mirror *new;
200 
201 	/*
202 	 * error_count is used for nothing more than a
203 	 * simple way to tell if a device has encountered
204 	 * errors.
205 	 */
206 	atomic_inc(&m->error_count);
207 
208 	if (test_and_set_bit(error_type, &m->error_type))
209 		return;
210 
211 	if (!errors_handled(ms))
212 		return;
213 
214 	if (m != get_default_mirror(ms))
215 		goto out;
216 
217 	if (!ms->in_sync) {
218 		/*
219 		 * Better to issue requests to same failing device
220 		 * than to risk returning corrupt data.
221 		 */
222 		DMERR("Primary mirror (%s) failed while out-of-sync: "
223 		      "Reads may fail.", m->dev->name);
224 		goto out;
225 	}
226 
227 	for (new = ms->mirror; new < ms->mirror + ms->nr_mirrors; new++)
228 		if (!atomic_read(&new->error_count)) {
229 			set_default_mirror(new);
230 			break;
231 		}
232 
233 	if (unlikely(new == ms->mirror + ms->nr_mirrors))
234 		DMWARN("All sides of mirror have failed.");
235 
236 out:
237 	schedule_work(&ms->trigger_event);
238 }
239 
240 /*-----------------------------------------------------------------
241  * Recovery.
242  *
243  * When a mirror is first activated we may find that some regions
244  * are in the no-sync state.  We have to recover these by
245  * recopying from the default mirror to all the others.
246  *---------------------------------------------------------------*/
247 static void recovery_complete(int read_err, unsigned long write_err,
248 			      void *context)
249 {
250 	struct dm_region *reg = context;
251 	struct mirror_set *ms = dm_rh_region_context(reg);
252 	int m, bit = 0;
253 
254 	if (read_err) {
255 		/* Read error means the failure of default mirror. */
256 		DMERR_LIMIT("Unable to read primary mirror during recovery");
257 		fail_mirror(get_default_mirror(ms), DM_RAID1_SYNC_ERROR);
258 	}
259 
260 	if (write_err) {
261 		DMERR_LIMIT("Write error during recovery (error = 0x%lx)",
262 			    write_err);
263 		/*
264 		 * Bits correspond to devices (excluding default mirror).
265 		 * The default mirror cannot change during recovery.
266 		 */
267 		for (m = 0; m < ms->nr_mirrors; m++) {
268 			if (&ms->mirror[m] == get_default_mirror(ms))
269 				continue;
270 			if (test_bit(bit, &write_err))
271 				fail_mirror(ms->mirror + m,
272 					    DM_RAID1_SYNC_ERROR);
273 			bit++;
274 		}
275 	}
276 
277 	dm_rh_recovery_end(reg, !(read_err || write_err));
278 }
279 
280 static int recover(struct mirror_set *ms, struct dm_region *reg)
281 {
282 	int r;
283 	unsigned i;
284 	struct dm_io_region from, to[DM_KCOPYD_MAX_REGIONS], *dest;
285 	struct mirror *m;
286 	unsigned long flags = 0;
287 	region_t key = dm_rh_get_region_key(reg);
288 	sector_t region_size = dm_rh_get_region_size(ms->rh);
289 
290 	/* fill in the source */
291 	m = get_default_mirror(ms);
292 	from.bdev = m->dev->bdev;
293 	from.sector = m->offset + dm_rh_region_to_sector(ms->rh, key);
294 	if (key == (ms->nr_regions - 1)) {
295 		/*
296 		 * The final region may be smaller than
297 		 * region_size.
298 		 */
299 		from.count = ms->ti->len & (region_size - 1);
300 		if (!from.count)
301 			from.count = region_size;
302 	} else
303 		from.count = region_size;
304 
305 	/* fill in the destinations */
306 	for (i = 0, dest = to; i < ms->nr_mirrors; i++) {
307 		if (&ms->mirror[i] == get_default_mirror(ms))
308 			continue;
309 
310 		m = ms->mirror + i;
311 		dest->bdev = m->dev->bdev;
312 		dest->sector = m->offset + dm_rh_region_to_sector(ms->rh, key);
313 		dest->count = from.count;
314 		dest++;
315 	}
316 
317 	/* hand to kcopyd */
318 	if (!errors_handled(ms))
319 		set_bit(DM_KCOPYD_IGNORE_ERROR, &flags);
320 
321 	r = dm_kcopyd_copy(ms->kcopyd_client, &from, ms->nr_mirrors - 1, to,
322 			   flags, recovery_complete, reg);
323 
324 	return r;
325 }
326 
327 static void do_recovery(struct mirror_set *ms)
328 {
329 	struct dm_region *reg;
330 	struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
331 	int r;
332 
333 	/*
334 	 * Start quiescing some regions.
335 	 */
336 	dm_rh_recovery_prepare(ms->rh);
337 
338 	/*
339 	 * Copy any already quiesced regions.
340 	 */
341 	while ((reg = dm_rh_recovery_start(ms->rh))) {
342 		r = recover(ms, reg);
343 		if (r)
344 			dm_rh_recovery_end(reg, 0);
345 	}
346 
347 	/*
348 	 * Update the in sync flag.
349 	 */
350 	if (!ms->in_sync &&
351 	    (log->type->get_sync_count(log) == ms->nr_regions)) {
352 		/* the sync is complete */
353 		dm_table_event(ms->ti->table);
354 		ms->in_sync = 1;
355 	}
356 }
357 
358 /*-----------------------------------------------------------------
359  * Reads
360  *---------------------------------------------------------------*/
361 static struct mirror *choose_mirror(struct mirror_set *ms, sector_t sector)
362 {
363 	struct mirror *m = get_default_mirror(ms);
364 
365 	do {
366 		if (likely(!atomic_read(&m->error_count)))
367 			return m;
368 
369 		if (m-- == ms->mirror)
370 			m += ms->nr_mirrors;
371 	} while (m != get_default_mirror(ms));
372 
373 	return NULL;
374 }
375 
376 static int default_ok(struct mirror *m)
377 {
378 	struct mirror *default_mirror = get_default_mirror(m->ms);
379 
380 	return !atomic_read(&default_mirror->error_count);
381 }
382 
383 static int mirror_available(struct mirror_set *ms, struct bio *bio)
384 {
385 	struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
386 	region_t region = dm_rh_bio_to_region(ms->rh, bio);
387 
388 	if (log->type->in_sync(log, region, 0))
389 		return choose_mirror(ms,  bio->bi_sector) ? 1 : 0;
390 
391 	return 0;
392 }
393 
394 /*
395  * remap a buffer to a particular mirror.
396  */
397 static sector_t map_sector(struct mirror *m, struct bio *bio)
398 {
399 	return m->offset + (bio->bi_sector - m->ms->ti->begin);
400 }
401 
402 static void map_bio(struct mirror *m, struct bio *bio)
403 {
404 	bio->bi_bdev = m->dev->bdev;
405 	bio->bi_sector = map_sector(m, bio);
406 }
407 
408 static void map_region(struct dm_io_region *io, struct mirror *m,
409 		       struct bio *bio)
410 {
411 	io->bdev = m->dev->bdev;
412 	io->sector = map_sector(m, bio);
413 	io->count = bio->bi_size >> 9;
414 }
415 
416 /*-----------------------------------------------------------------
417  * Reads
418  *---------------------------------------------------------------*/
419 static void read_callback(unsigned long error, void *context)
420 {
421 	struct bio *bio = context;
422 	struct mirror *m;
423 
424 	m = bio_get_m(bio);
425 	bio_set_m(bio, NULL);
426 
427 	if (likely(!error)) {
428 		bio_endio(bio, 0);
429 		return;
430 	}
431 
432 	fail_mirror(m, DM_RAID1_READ_ERROR);
433 
434 	if (likely(default_ok(m)) || mirror_available(m->ms, bio)) {
435 		DMWARN_LIMIT("Read failure on mirror device %s.  "
436 			     "Trying alternative device.",
437 			     m->dev->name);
438 		queue_bio(m->ms, bio, bio_rw(bio));
439 		return;
440 	}
441 
442 	DMERR_LIMIT("Read failure on mirror device %s.  Failing I/O.",
443 		    m->dev->name);
444 	bio_endio(bio, -EIO);
445 }
446 
447 /* Asynchronous read. */
448 static void read_async_bio(struct mirror *m, struct bio *bio)
449 {
450 	struct dm_io_region io;
451 	struct dm_io_request io_req = {
452 		.bi_rw = READ,
453 		.mem.type = DM_IO_BVEC,
454 		.mem.ptr.bvec = bio->bi_io_vec + bio->bi_idx,
455 		.notify.fn = read_callback,
456 		.notify.context = bio,
457 		.client = m->ms->io_client,
458 	};
459 
460 	map_region(&io, m, bio);
461 	bio_set_m(bio, m);
462 	BUG_ON(dm_io(&io_req, 1, &io, NULL));
463 }
464 
465 static inline int region_in_sync(struct mirror_set *ms, region_t region,
466 				 int may_block)
467 {
468 	int state = dm_rh_get_state(ms->rh, region, may_block);
469 	return state == DM_RH_CLEAN || state == DM_RH_DIRTY;
470 }
471 
472 static void do_reads(struct mirror_set *ms, struct bio_list *reads)
473 {
474 	region_t region;
475 	struct bio *bio;
476 	struct mirror *m;
477 
478 	while ((bio = bio_list_pop(reads))) {
479 		region = dm_rh_bio_to_region(ms->rh, bio);
480 		m = get_default_mirror(ms);
481 
482 		/*
483 		 * We can only read balance if the region is in sync.
484 		 */
485 		if (likely(region_in_sync(ms, region, 1)))
486 			m = choose_mirror(ms, bio->bi_sector);
487 		else if (m && atomic_read(&m->error_count))
488 			m = NULL;
489 
490 		if (likely(m))
491 			read_async_bio(m, bio);
492 		else
493 			bio_endio(bio, -EIO);
494 	}
495 }
496 
497 /*-----------------------------------------------------------------
498  * Writes.
499  *
500  * We do different things with the write io depending on the
501  * state of the region that it's in:
502  *
503  * SYNC: 	increment pending, use kcopyd to write to *all* mirrors
504  * RECOVERING:	delay the io until recovery completes
505  * NOSYNC:	increment pending, just write to the default mirror
506  *---------------------------------------------------------------*/
507 
508 
509 static void write_callback(unsigned long error, void *context)
510 {
511 	unsigned i, ret = 0;
512 	struct bio *bio = (struct bio *) context;
513 	struct mirror_set *ms;
514 	int uptodate = 0;
515 	int should_wake = 0;
516 	unsigned long flags;
517 
518 	ms = bio_get_m(bio)->ms;
519 	bio_set_m(bio, NULL);
520 
521 	/*
522 	 * NOTE: We don't decrement the pending count here,
523 	 * instead it is done by the targets endio function.
524 	 * This way we handle both writes to SYNC and NOSYNC
525 	 * regions with the same code.
526 	 */
527 	if (likely(!error))
528 		goto out;
529 
530 	for (i = 0; i < ms->nr_mirrors; i++)
531 		if (test_bit(i, &error))
532 			fail_mirror(ms->mirror + i, DM_RAID1_WRITE_ERROR);
533 		else
534 			uptodate = 1;
535 
536 	if (unlikely(!uptodate)) {
537 		DMERR("All replicated volumes dead, failing I/O");
538 		/* None of the writes succeeded, fail the I/O. */
539 		ret = -EIO;
540 	} else if (errors_handled(ms)) {
541 		/*
542 		 * Need to raise event.  Since raising
543 		 * events can block, we need to do it in
544 		 * the main thread.
545 		 */
546 		spin_lock_irqsave(&ms->lock, flags);
547 		if (!ms->failures.head)
548 			should_wake = 1;
549 		bio_list_add(&ms->failures, bio);
550 		spin_unlock_irqrestore(&ms->lock, flags);
551 		if (should_wake)
552 			wakeup_mirrord(ms);
553 		return;
554 	}
555 out:
556 	bio_endio(bio, ret);
557 }
558 
559 static void do_write(struct mirror_set *ms, struct bio *bio)
560 {
561 	unsigned int i;
562 	struct dm_io_region io[ms->nr_mirrors], *dest = io;
563 	struct mirror *m;
564 	struct dm_io_request io_req = {
565 		.bi_rw = WRITE,
566 		.mem.type = DM_IO_BVEC,
567 		.mem.ptr.bvec = bio->bi_io_vec + bio->bi_idx,
568 		.notify.fn = write_callback,
569 		.notify.context = bio,
570 		.client = ms->io_client,
571 	};
572 
573 	for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++)
574 		map_region(dest++, m, bio);
575 
576 	/*
577 	 * Use default mirror because we only need it to retrieve the reference
578 	 * to the mirror set in write_callback().
579 	 */
580 	bio_set_m(bio, get_default_mirror(ms));
581 
582 	BUG_ON(dm_io(&io_req, ms->nr_mirrors, io, NULL));
583 }
584 
585 static void do_writes(struct mirror_set *ms, struct bio_list *writes)
586 {
587 	int state;
588 	struct bio *bio;
589 	struct bio_list sync, nosync, recover, *this_list = NULL;
590 	struct bio_list requeue;
591 	struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
592 	region_t region;
593 
594 	if (!writes->head)
595 		return;
596 
597 	/*
598 	 * Classify each write.
599 	 */
600 	bio_list_init(&sync);
601 	bio_list_init(&nosync);
602 	bio_list_init(&recover);
603 	bio_list_init(&requeue);
604 
605 	while ((bio = bio_list_pop(writes))) {
606 		region = dm_rh_bio_to_region(ms->rh, bio);
607 
608 		if (log->type->is_remote_recovering &&
609 		    log->type->is_remote_recovering(log, region)) {
610 			bio_list_add(&requeue, bio);
611 			continue;
612 		}
613 
614 		state = dm_rh_get_state(ms->rh, region, 1);
615 		switch (state) {
616 		case DM_RH_CLEAN:
617 		case DM_RH_DIRTY:
618 			this_list = &sync;
619 			break;
620 
621 		case DM_RH_NOSYNC:
622 			this_list = &nosync;
623 			break;
624 
625 		case DM_RH_RECOVERING:
626 			this_list = &recover;
627 			break;
628 		}
629 
630 		bio_list_add(this_list, bio);
631 	}
632 
633 	/*
634 	 * Add bios that are delayed due to remote recovery
635 	 * back on to the write queue
636 	 */
637 	if (unlikely(requeue.head)) {
638 		spin_lock_irq(&ms->lock);
639 		bio_list_merge(&ms->writes, &requeue);
640 		spin_unlock_irq(&ms->lock);
641 	}
642 
643 	/*
644 	 * Increment the pending counts for any regions that will
645 	 * be written to (writes to recover regions are going to
646 	 * be delayed).
647 	 */
648 	dm_rh_inc_pending(ms->rh, &sync);
649 	dm_rh_inc_pending(ms->rh, &nosync);
650 	ms->log_failure = dm_rh_flush(ms->rh) ? 1 : 0;
651 
652 	/*
653 	 * Dispatch io.
654 	 */
655 	if (unlikely(ms->log_failure)) {
656 		spin_lock_irq(&ms->lock);
657 		bio_list_merge(&ms->failures, &sync);
658 		spin_unlock_irq(&ms->lock);
659 		wakeup_mirrord(ms);
660 	} else
661 		while ((bio = bio_list_pop(&sync)))
662 			do_write(ms, bio);
663 
664 	while ((bio = bio_list_pop(&recover)))
665 		dm_rh_delay(ms->rh, bio);
666 
667 	while ((bio = bio_list_pop(&nosync))) {
668 		map_bio(get_default_mirror(ms), bio);
669 		generic_make_request(bio);
670 	}
671 }
672 
673 static void do_failures(struct mirror_set *ms, struct bio_list *failures)
674 {
675 	struct bio *bio;
676 
677 	if (!failures->head)
678 		return;
679 
680 	if (!ms->log_failure) {
681 		while ((bio = bio_list_pop(failures))) {
682 			ms->in_sync = 0;
683 			dm_rh_mark_nosync(ms->rh, bio, bio->bi_size, 0);
684 		}
685 		return;
686 	}
687 
688 	/*
689 	 * If the log has failed, unattempted writes are being
690 	 * put on the failures list.  We can't issue those writes
691 	 * until a log has been marked, so we must store them.
692 	 *
693 	 * If a 'noflush' suspend is in progress, we can requeue
694 	 * the I/O's to the core.  This give userspace a chance
695 	 * to reconfigure the mirror, at which point the core
696 	 * will reissue the writes.  If the 'noflush' flag is
697 	 * not set, we have no choice but to return errors.
698 	 *
699 	 * Some writes on the failures list may have been
700 	 * submitted before the log failure and represent a
701 	 * failure to write to one of the devices.  It is ok
702 	 * for us to treat them the same and requeue them
703 	 * as well.
704 	 */
705 	if (dm_noflush_suspending(ms->ti)) {
706 		while ((bio = bio_list_pop(failures)))
707 			bio_endio(bio, DM_ENDIO_REQUEUE);
708 		return;
709 	}
710 
711 	if (atomic_read(&ms->suspend)) {
712 		while ((bio = bio_list_pop(failures)))
713 			bio_endio(bio, -EIO);
714 		return;
715 	}
716 
717 	spin_lock_irq(&ms->lock);
718 	bio_list_merge(&ms->failures, failures);
719 	spin_unlock_irq(&ms->lock);
720 
721 	delayed_wake(ms);
722 }
723 
724 static void trigger_event(struct work_struct *work)
725 {
726 	struct mirror_set *ms =
727 		container_of(work, struct mirror_set, trigger_event);
728 
729 	dm_table_event(ms->ti->table);
730 }
731 
732 /*-----------------------------------------------------------------
733  * kmirrord
734  *---------------------------------------------------------------*/
735 static void do_mirror(struct work_struct *work)
736 {
737 	struct mirror_set *ms = container_of(work, struct mirror_set,
738 					     kmirrord_work);
739 	struct bio_list reads, writes, failures;
740 	unsigned long flags;
741 
742 	spin_lock_irqsave(&ms->lock, flags);
743 	reads = ms->reads;
744 	writes = ms->writes;
745 	failures = ms->failures;
746 	bio_list_init(&ms->reads);
747 	bio_list_init(&ms->writes);
748 	bio_list_init(&ms->failures);
749 	spin_unlock_irqrestore(&ms->lock, flags);
750 
751 	dm_rh_update_states(ms->rh, errors_handled(ms));
752 	do_recovery(ms);
753 	do_reads(ms, &reads);
754 	do_writes(ms, &writes);
755 	do_failures(ms, &failures);
756 
757 	dm_table_unplug_all(ms->ti->table);
758 }
759 
760 /*-----------------------------------------------------------------
761  * Target functions
762  *---------------------------------------------------------------*/
763 static struct mirror_set *alloc_context(unsigned int nr_mirrors,
764 					uint32_t region_size,
765 					struct dm_target *ti,
766 					struct dm_dirty_log *dl)
767 {
768 	size_t len;
769 	struct mirror_set *ms = NULL;
770 
771 	len = sizeof(*ms) + (sizeof(ms->mirror[0]) * nr_mirrors);
772 
773 	ms = kzalloc(len, GFP_KERNEL);
774 	if (!ms) {
775 		ti->error = "Cannot allocate mirror context";
776 		return NULL;
777 	}
778 
779 	spin_lock_init(&ms->lock);
780 
781 	ms->ti = ti;
782 	ms->nr_mirrors = nr_mirrors;
783 	ms->nr_regions = dm_sector_div_up(ti->len, region_size);
784 	ms->in_sync = 0;
785 	ms->log_failure = 0;
786 	atomic_set(&ms->suspend, 0);
787 	atomic_set(&ms->default_mirror, DEFAULT_MIRROR);
788 
789 	ms->read_record_pool = mempool_create_slab_pool(MIN_READ_RECORDS,
790 						_dm_raid1_read_record_cache);
791 
792 	if (!ms->read_record_pool) {
793 		ti->error = "Error creating mirror read_record_pool";
794 		kfree(ms);
795 		return NULL;
796 	}
797 
798 	ms->io_client = dm_io_client_create(DM_IO_PAGES);
799 	if (IS_ERR(ms->io_client)) {
800 		ti->error = "Error creating dm_io client";
801 		mempool_destroy(ms->read_record_pool);
802 		kfree(ms);
803  		return NULL;
804 	}
805 
806 	ms->rh = dm_region_hash_create(ms, dispatch_bios, wakeup_mirrord,
807 				       wakeup_all_recovery_waiters,
808 				       ms->ti->begin, MAX_RECOVERY,
809 				       dl, region_size, ms->nr_regions);
810 	if (IS_ERR(ms->rh)) {
811 		ti->error = "Error creating dirty region hash";
812 		dm_io_client_destroy(ms->io_client);
813 		mempool_destroy(ms->read_record_pool);
814 		kfree(ms);
815 		return NULL;
816 	}
817 
818 	return ms;
819 }
820 
821 static void free_context(struct mirror_set *ms, struct dm_target *ti,
822 			 unsigned int m)
823 {
824 	while (m--)
825 		dm_put_device(ti, ms->mirror[m].dev);
826 
827 	dm_io_client_destroy(ms->io_client);
828 	dm_region_hash_destroy(ms->rh);
829 	mempool_destroy(ms->read_record_pool);
830 	kfree(ms);
831 }
832 
833 static int get_mirror(struct mirror_set *ms, struct dm_target *ti,
834 		      unsigned int mirror, char **argv)
835 {
836 	unsigned long long offset;
837 
838 	if (sscanf(argv[1], "%llu", &offset) != 1) {
839 		ti->error = "Invalid offset";
840 		return -EINVAL;
841 	}
842 
843 	if (dm_get_device(ti, argv[0], offset, ti->len,
844 			  dm_table_get_mode(ti->table),
845 			  &ms->mirror[mirror].dev)) {
846 		ti->error = "Device lookup failure";
847 		return -ENXIO;
848 	}
849 
850 	ms->mirror[mirror].ms = ms;
851 	atomic_set(&(ms->mirror[mirror].error_count), 0);
852 	ms->mirror[mirror].error_type = 0;
853 	ms->mirror[mirror].offset = offset;
854 
855 	return 0;
856 }
857 
858 /*
859  * Create dirty log: log_type #log_params <log_params>
860  */
861 static struct dm_dirty_log *create_dirty_log(struct dm_target *ti,
862 					     unsigned argc, char **argv,
863 					     unsigned *args_used)
864 {
865 	unsigned param_count;
866 	struct dm_dirty_log *dl;
867 
868 	if (argc < 2) {
869 		ti->error = "Insufficient mirror log arguments";
870 		return NULL;
871 	}
872 
873 	if (sscanf(argv[1], "%u", &param_count) != 1) {
874 		ti->error = "Invalid mirror log argument count";
875 		return NULL;
876 	}
877 
878 	*args_used = 2 + param_count;
879 
880 	if (argc < *args_used) {
881 		ti->error = "Insufficient mirror log arguments";
882 		return NULL;
883 	}
884 
885 	dl = dm_dirty_log_create(argv[0], ti, param_count, argv + 2);
886 	if (!dl) {
887 		ti->error = "Error creating mirror dirty log";
888 		return NULL;
889 	}
890 
891 	return dl;
892 }
893 
894 static int parse_features(struct mirror_set *ms, unsigned argc, char **argv,
895 			  unsigned *args_used)
896 {
897 	unsigned num_features;
898 	struct dm_target *ti = ms->ti;
899 
900 	*args_used = 0;
901 
902 	if (!argc)
903 		return 0;
904 
905 	if (sscanf(argv[0], "%u", &num_features) != 1) {
906 		ti->error = "Invalid number of features";
907 		return -EINVAL;
908 	}
909 
910 	argc--;
911 	argv++;
912 	(*args_used)++;
913 
914 	if (num_features > argc) {
915 		ti->error = "Not enough arguments to support feature count";
916 		return -EINVAL;
917 	}
918 
919 	if (!strcmp("handle_errors", argv[0]))
920 		ms->features |= DM_RAID1_HANDLE_ERRORS;
921 	else {
922 		ti->error = "Unrecognised feature requested";
923 		return -EINVAL;
924 	}
925 
926 	(*args_used)++;
927 
928 	return 0;
929 }
930 
931 /*
932  * Construct a mirror mapping:
933  *
934  * log_type #log_params <log_params>
935  * #mirrors [mirror_path offset]{2,}
936  * [#features <features>]
937  *
938  * log_type is "core" or "disk"
939  * #log_params is between 1 and 3
940  *
941  * If present, features must be "handle_errors".
942  */
943 static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv)
944 {
945 	int r;
946 	unsigned int nr_mirrors, m, args_used;
947 	struct mirror_set *ms;
948 	struct dm_dirty_log *dl;
949 
950 	dl = create_dirty_log(ti, argc, argv, &args_used);
951 	if (!dl)
952 		return -EINVAL;
953 
954 	argv += args_used;
955 	argc -= args_used;
956 
957 	if (!argc || sscanf(argv[0], "%u", &nr_mirrors) != 1 ||
958 	    nr_mirrors < 2 || nr_mirrors > DM_KCOPYD_MAX_REGIONS + 1) {
959 		ti->error = "Invalid number of mirrors";
960 		dm_dirty_log_destroy(dl);
961 		return -EINVAL;
962 	}
963 
964 	argv++, argc--;
965 
966 	if (argc < nr_mirrors * 2) {
967 		ti->error = "Too few mirror arguments";
968 		dm_dirty_log_destroy(dl);
969 		return -EINVAL;
970 	}
971 
972 	ms = alloc_context(nr_mirrors, dl->type->get_region_size(dl), ti, dl);
973 	if (!ms) {
974 		dm_dirty_log_destroy(dl);
975 		return -ENOMEM;
976 	}
977 
978 	/* Get the mirror parameter sets */
979 	for (m = 0; m < nr_mirrors; m++) {
980 		r = get_mirror(ms, ti, m, argv);
981 		if (r) {
982 			free_context(ms, ti, m);
983 			return r;
984 		}
985 		argv += 2;
986 		argc -= 2;
987 	}
988 
989 	ti->private = ms;
990 	ti->split_io = dm_rh_get_region_size(ms->rh);
991 
992 	ms->kmirrord_wq = create_singlethread_workqueue("kmirrord");
993 	if (!ms->kmirrord_wq) {
994 		DMERR("couldn't start kmirrord");
995 		r = -ENOMEM;
996 		goto err_free_context;
997 	}
998 	INIT_WORK(&ms->kmirrord_work, do_mirror);
999 	init_timer(&ms->timer);
1000 	ms->timer_pending = 0;
1001 	INIT_WORK(&ms->trigger_event, trigger_event);
1002 
1003 	r = parse_features(ms, argc, argv, &args_used);
1004 	if (r)
1005 		goto err_destroy_wq;
1006 
1007 	argv += args_used;
1008 	argc -= args_used;
1009 
1010 	/*
1011 	 * Any read-balancing addition depends on the
1012 	 * DM_RAID1_HANDLE_ERRORS flag being present.
1013 	 * This is because the decision to balance depends
1014 	 * on the sync state of a region.  If the above
1015 	 * flag is not present, we ignore errors; and
1016 	 * the sync state may be inaccurate.
1017 	 */
1018 
1019 	if (argc) {
1020 		ti->error = "Too many mirror arguments";
1021 		r = -EINVAL;
1022 		goto err_destroy_wq;
1023 	}
1024 
1025 	r = dm_kcopyd_client_create(DM_KCOPYD_PAGES, &ms->kcopyd_client);
1026 	if (r)
1027 		goto err_destroy_wq;
1028 
1029 	wakeup_mirrord(ms);
1030 	return 0;
1031 
1032 err_destroy_wq:
1033 	destroy_workqueue(ms->kmirrord_wq);
1034 err_free_context:
1035 	free_context(ms, ti, ms->nr_mirrors);
1036 	return r;
1037 }
1038 
1039 static void mirror_dtr(struct dm_target *ti)
1040 {
1041 	struct mirror_set *ms = (struct mirror_set *) ti->private;
1042 
1043 	del_timer_sync(&ms->timer);
1044 	flush_workqueue(ms->kmirrord_wq);
1045 	flush_scheduled_work();
1046 	dm_kcopyd_client_destroy(ms->kcopyd_client);
1047 	destroy_workqueue(ms->kmirrord_wq);
1048 	free_context(ms, ti, ms->nr_mirrors);
1049 }
1050 
1051 /*
1052  * Mirror mapping function
1053  */
1054 static int mirror_map(struct dm_target *ti, struct bio *bio,
1055 		      union map_info *map_context)
1056 {
1057 	int r, rw = bio_rw(bio);
1058 	struct mirror *m;
1059 	struct mirror_set *ms = ti->private;
1060 	struct dm_raid1_read_record *read_record = NULL;
1061 	struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1062 
1063 	if (rw == WRITE) {
1064 		/* Save region for mirror_end_io() handler */
1065 		map_context->ll = dm_rh_bio_to_region(ms->rh, bio);
1066 		queue_bio(ms, bio, rw);
1067 		return DM_MAPIO_SUBMITTED;
1068 	}
1069 
1070 	r = log->type->in_sync(log, dm_rh_bio_to_region(ms->rh, bio), 0);
1071 	if (r < 0 && r != -EWOULDBLOCK)
1072 		return r;
1073 
1074 	/*
1075 	 * If region is not in-sync queue the bio.
1076 	 */
1077 	if (!r || (r == -EWOULDBLOCK)) {
1078 		if (rw == READA)
1079 			return -EWOULDBLOCK;
1080 
1081 		queue_bio(ms, bio, rw);
1082 		return DM_MAPIO_SUBMITTED;
1083 	}
1084 
1085 	/*
1086 	 * The region is in-sync and we can perform reads directly.
1087 	 * Store enough information so we can retry if it fails.
1088 	 */
1089 	m = choose_mirror(ms, bio->bi_sector);
1090 	if (unlikely(!m))
1091 		return -EIO;
1092 
1093 	read_record = mempool_alloc(ms->read_record_pool, GFP_NOIO);
1094 	if (likely(read_record)) {
1095 		dm_bio_record(&read_record->details, bio);
1096 		map_context->ptr = read_record;
1097 		read_record->m = m;
1098 	}
1099 
1100 	map_bio(m, bio);
1101 
1102 	return DM_MAPIO_REMAPPED;
1103 }
1104 
1105 static int mirror_end_io(struct dm_target *ti, struct bio *bio,
1106 			 int error, union map_info *map_context)
1107 {
1108 	int rw = bio_rw(bio);
1109 	struct mirror_set *ms = (struct mirror_set *) ti->private;
1110 	struct mirror *m = NULL;
1111 	struct dm_bio_details *bd = NULL;
1112 	struct dm_raid1_read_record *read_record = map_context->ptr;
1113 
1114 	/*
1115 	 * We need to dec pending if this was a write.
1116 	 */
1117 	if (rw == WRITE) {
1118 		dm_rh_dec(ms->rh, map_context->ll);
1119 		return error;
1120 	}
1121 
1122 	if (error == -EOPNOTSUPP)
1123 		goto out;
1124 
1125 	if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1126 		goto out;
1127 
1128 	if (unlikely(error)) {
1129 		if (!read_record) {
1130 			/*
1131 			 * There wasn't enough memory to record necessary
1132 			 * information for a retry or there was no other
1133 			 * mirror in-sync.
1134 			 */
1135 			DMERR_LIMIT("Mirror read failed.");
1136 			return -EIO;
1137 		}
1138 
1139 		m = read_record->m;
1140 
1141 		DMERR("Mirror read failed from %s. Trying alternative device.",
1142 		      m->dev->name);
1143 
1144 		fail_mirror(m, DM_RAID1_READ_ERROR);
1145 
1146 		/*
1147 		 * A failed read is requeued for another attempt using an intact
1148 		 * mirror.
1149 		 */
1150 		if (default_ok(m) || mirror_available(ms, bio)) {
1151 			bd = &read_record->details;
1152 
1153 			dm_bio_restore(bd, bio);
1154 			mempool_free(read_record, ms->read_record_pool);
1155 			map_context->ptr = NULL;
1156 			queue_bio(ms, bio, rw);
1157 			return 1;
1158 		}
1159 		DMERR("All replicated volumes dead, failing I/O");
1160 	}
1161 
1162 out:
1163 	if (read_record) {
1164 		mempool_free(read_record, ms->read_record_pool);
1165 		map_context->ptr = NULL;
1166 	}
1167 
1168 	return error;
1169 }
1170 
1171 static void mirror_presuspend(struct dm_target *ti)
1172 {
1173 	struct mirror_set *ms = (struct mirror_set *) ti->private;
1174 	struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1175 
1176 	atomic_set(&ms->suspend, 1);
1177 
1178 	/*
1179 	 * We must finish up all the work that we've
1180 	 * generated (i.e. recovery work).
1181 	 */
1182 	dm_rh_stop_recovery(ms->rh);
1183 
1184 	wait_event(_kmirrord_recovery_stopped,
1185 		   !dm_rh_recovery_in_flight(ms->rh));
1186 
1187 	if (log->type->presuspend && log->type->presuspend(log))
1188 		/* FIXME: need better error handling */
1189 		DMWARN("log presuspend failed");
1190 
1191 	/*
1192 	 * Now that recovery is complete/stopped and the
1193 	 * delayed bios are queued, we need to wait for
1194 	 * the worker thread to complete.  This way,
1195 	 * we know that all of our I/O has been pushed.
1196 	 */
1197 	flush_workqueue(ms->kmirrord_wq);
1198 }
1199 
1200 static void mirror_postsuspend(struct dm_target *ti)
1201 {
1202 	struct mirror_set *ms = ti->private;
1203 	struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1204 
1205 	if (log->type->postsuspend && log->type->postsuspend(log))
1206 		/* FIXME: need better error handling */
1207 		DMWARN("log postsuspend failed");
1208 }
1209 
1210 static void mirror_resume(struct dm_target *ti)
1211 {
1212 	struct mirror_set *ms = ti->private;
1213 	struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1214 
1215 	atomic_set(&ms->suspend, 0);
1216 	if (log->type->resume && log->type->resume(log))
1217 		/* FIXME: need better error handling */
1218 		DMWARN("log resume failed");
1219 	dm_rh_start_recovery(ms->rh);
1220 }
1221 
1222 /*
1223  * device_status_char
1224  * @m: mirror device/leg we want the status of
1225  *
1226  * We return one character representing the most severe error
1227  * we have encountered.
1228  *    A => Alive - No failures
1229  *    D => Dead - A write failure occurred leaving mirror out-of-sync
1230  *    S => Sync - A sychronization failure occurred, mirror out-of-sync
1231  *    R => Read - A read failure occurred, mirror data unaffected
1232  *
1233  * Returns: <char>
1234  */
1235 static char device_status_char(struct mirror *m)
1236 {
1237 	if (!atomic_read(&(m->error_count)))
1238 		return 'A';
1239 
1240 	return (test_bit(DM_RAID1_WRITE_ERROR, &(m->error_type))) ? 'D' :
1241 		(test_bit(DM_RAID1_SYNC_ERROR, &(m->error_type))) ? 'S' :
1242 		(test_bit(DM_RAID1_READ_ERROR, &(m->error_type))) ? 'R' : 'U';
1243 }
1244 
1245 
1246 static int mirror_status(struct dm_target *ti, status_type_t type,
1247 			 char *result, unsigned int maxlen)
1248 {
1249 	unsigned int m, sz = 0;
1250 	struct mirror_set *ms = (struct mirror_set *) ti->private;
1251 	struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1252 	char buffer[ms->nr_mirrors + 1];
1253 
1254 	switch (type) {
1255 	case STATUSTYPE_INFO:
1256 		DMEMIT("%d ", ms->nr_mirrors);
1257 		for (m = 0; m < ms->nr_mirrors; m++) {
1258 			DMEMIT("%s ", ms->mirror[m].dev->name);
1259 			buffer[m] = device_status_char(&(ms->mirror[m]));
1260 		}
1261 		buffer[m] = '\0';
1262 
1263 		DMEMIT("%llu/%llu 1 %s ",
1264 		      (unsigned long long)log->type->get_sync_count(log),
1265 		      (unsigned long long)ms->nr_regions, buffer);
1266 
1267 		sz += log->type->status(log, type, result+sz, maxlen-sz);
1268 
1269 		break;
1270 
1271 	case STATUSTYPE_TABLE:
1272 		sz = log->type->status(log, type, result, maxlen);
1273 
1274 		DMEMIT("%d", ms->nr_mirrors);
1275 		for (m = 0; m < ms->nr_mirrors; m++)
1276 			DMEMIT(" %s %llu", ms->mirror[m].dev->name,
1277 			       (unsigned long long)ms->mirror[m].offset);
1278 
1279 		if (ms->features & DM_RAID1_HANDLE_ERRORS)
1280 			DMEMIT(" 1 handle_errors");
1281 	}
1282 
1283 	return 0;
1284 }
1285 
1286 static struct target_type mirror_target = {
1287 	.name	 = "mirror",
1288 	.version = {1, 0, 20},
1289 	.module	 = THIS_MODULE,
1290 	.ctr	 = mirror_ctr,
1291 	.dtr	 = mirror_dtr,
1292 	.map	 = mirror_map,
1293 	.end_io	 = mirror_end_io,
1294 	.presuspend = mirror_presuspend,
1295 	.postsuspend = mirror_postsuspend,
1296 	.resume	 = mirror_resume,
1297 	.status	 = mirror_status,
1298 };
1299 
1300 static int __init dm_mirror_init(void)
1301 {
1302 	int r;
1303 
1304 	_dm_raid1_read_record_cache = KMEM_CACHE(dm_raid1_read_record, 0);
1305 	if (!_dm_raid1_read_record_cache) {
1306 		DMERR("Can't allocate dm_raid1_read_record cache");
1307 		r = -ENOMEM;
1308 		goto bad_cache;
1309 	}
1310 
1311 	r = dm_register_target(&mirror_target);
1312 	if (r < 0) {
1313 		DMERR("Failed to register mirror target");
1314 		goto bad_target;
1315 	}
1316 
1317 	return 0;
1318 
1319 bad_target:
1320 	kmem_cache_destroy(_dm_raid1_read_record_cache);
1321 bad_cache:
1322 	return r;
1323 }
1324 
1325 static void __exit dm_mirror_exit(void)
1326 {
1327 	dm_unregister_target(&mirror_target);
1328 	kmem_cache_destroy(_dm_raid1_read_record_cache);
1329 }
1330 
1331 /* Module hooks */
1332 module_init(dm_mirror_init);
1333 module_exit(dm_mirror_exit);
1334 
1335 MODULE_DESCRIPTION(DM_NAME " mirror target");
1336 MODULE_AUTHOR("Joe Thornber");
1337 MODULE_LICENSE("GPL");
1338