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
31static DECLARE_WAIT_QUEUE_HEAD(_kmirrord_recovery_stopped);
32
33/*-----------------------------------------------------------------
34 * Mirror set structures.
35 *---------------------------------------------------------------*/
36enum dm_raid1_error {
37	DM_RAID1_WRITE_ERROR,
38	DM_RAID1_FLUSH_ERROR,
39	DM_RAID1_SYNC_ERROR,
40	DM_RAID1_READ_ERROR
41};
42
43struct mirror {
44	struct mirror_set *ms;
45	atomic_t error_count;
46	unsigned long error_type;
47	struct dm_dev *dev;
48	sector_t offset;
49};
50
51struct mirror_set {
52	struct dm_target *ti;
53	struct list_head list;
54
55	uint64_t features;
56
57	spinlock_t lock;	/* protects the lists */
58	struct bio_list reads;
59	struct bio_list writes;
60	struct bio_list failures;
61	struct bio_list holds;	/* bios are waiting until suspend */
62
63	struct dm_region_hash *rh;
64	struct dm_kcopyd_client *kcopyd_client;
65	struct dm_io_client *io_client;
66	mempool_t *read_record_pool;
67
68	/* recovery */
69	region_t nr_regions;
70	int in_sync;
71	int log_failure;
72	int leg_failure;
73	atomic_t suspend;
74
75	atomic_t default_mirror;	/* Default mirror */
76
77	struct workqueue_struct *kmirrord_wq;
78	struct work_struct kmirrord_work;
79	struct timer_list timer;
80	unsigned long timer_pending;
81
82	struct work_struct trigger_event;
83
84	unsigned nr_mirrors;
85	struct mirror mirror[0];
86};
87
88static void wakeup_mirrord(void *context)
89{
90	struct mirror_set *ms = context;
91
92	queue_work(ms->kmirrord_wq, &ms->kmirrord_work);
93}
94
95static void delayed_wake_fn(unsigned long data)
96{
97	struct mirror_set *ms = (struct mirror_set *) data;
98
99	clear_bit(0, &ms->timer_pending);
100	wakeup_mirrord(ms);
101}
102
103static void delayed_wake(struct mirror_set *ms)
104{
105	if (test_and_set_bit(0, &ms->timer_pending))
106		return;
107
108	ms->timer.expires = jiffies + HZ / 5;
109	ms->timer.data = (unsigned long) ms;
110	ms->timer.function = delayed_wake_fn;
111	add_timer(&ms->timer);
112}
113
114static void wakeup_all_recovery_waiters(void *context)
115{
116	wake_up_all(&_kmirrord_recovery_stopped);
117}
118
119static void queue_bio(struct mirror_set *ms, struct bio *bio, int rw)
120{
121	unsigned long flags;
122	int should_wake = 0;
123	struct bio_list *bl;
124
125	bl = (rw == WRITE) ? &ms->writes : &ms->reads;
126	spin_lock_irqsave(&ms->lock, flags);
127	should_wake = !(bl->head);
128	bio_list_add(bl, bio);
129	spin_unlock_irqrestore(&ms->lock, flags);
130
131	if (should_wake)
132		wakeup_mirrord(ms);
133}
134
135static void dispatch_bios(void *context, struct bio_list *bio_list)
136{
137	struct mirror_set *ms = context;
138	struct bio *bio;
139
140	while ((bio = bio_list_pop(bio_list)))
141		queue_bio(ms, bio, WRITE);
142}
143
144#define MIN_READ_RECORDS 20
145struct dm_raid1_read_record {
146	struct mirror *m;
147	struct dm_bio_details details;
148};
149
150static struct kmem_cache *_dm_raid1_read_record_cache;
151
152/*
153 * Every mirror should look like this one.
154 */
155#define DEFAULT_MIRROR 0
156
157/*
158 * This is yucky.  We squirrel the mirror struct away inside
159 * bi_next for read/write buffers.  This is safe since the bh
160 * doesn't get submitted to the lower levels of block layer.
161 */
162static struct mirror *bio_get_m(struct bio *bio)
163{
164	return (struct mirror *) bio->bi_next;
165}
166
167static void bio_set_m(struct bio *bio, struct mirror *m)
168{
169	bio->bi_next = (struct bio *) m;
170}
171
172static struct mirror *get_default_mirror(struct mirror_set *ms)
173{
174	return &ms->mirror[atomic_read(&ms->default_mirror)];
175}
176
177static void set_default_mirror(struct mirror *m)
178{
179	struct mirror_set *ms = m->ms;
180	struct mirror *m0 = &(ms->mirror[0]);
181
182	atomic_set(&ms->default_mirror, m - m0);
183}
184
185static struct mirror *get_valid_mirror(struct mirror_set *ms)
186{
187	struct mirror *m;
188
189	for (m = ms->mirror; m < ms->mirror + ms->nr_mirrors; m++)
190		if (!atomic_read(&m->error_count))
191			return m;
192
193	return NULL;
194}
195
196/* fail_mirror
197 * @m: mirror device to fail
198 * @error_type: one of the enum's, DM_RAID1_*_ERROR
199 *
200 * If errors are being handled, record the type of
201 * error encountered for this device.  If this type
202 * of error has already been recorded, we can return;
203 * otherwise, we must signal userspace by triggering
204 * an event.  Additionally, if the device is the
205 * primary device, we must choose a new primary, but
206 * only if the mirror is in-sync.
207 *
208 * This function must not block.
209 */
210static void fail_mirror(struct mirror *m, enum dm_raid1_error error_type)
211{
212	struct mirror_set *ms = m->ms;
213	struct mirror *new;
214
215	ms->leg_failure = 1;
216
217	/*
218	 * error_count is used for nothing more than a
219	 * simple way to tell if a device has encountered
220	 * errors.
221	 */
222	atomic_inc(&m->error_count);
223
224	if (test_and_set_bit(error_type, &m->error_type))
225		return;
226
227	if (!errors_handled(ms))
228		return;
229
230	if (m != get_default_mirror(ms))
231		goto out;
232
233	if (!ms->in_sync) {
234		/*
235		 * Better to issue requests to same failing device
236		 * than to risk returning corrupt data.
237		 */
238		DMERR("Primary mirror (%s) failed while out-of-sync: "
239		      "Reads may fail.", m->dev->name);
240		goto out;
241	}
242
243	new = get_valid_mirror(ms);
244	if (new)
245		set_default_mirror(new);
246	else
247		DMWARN("All sides of mirror have failed.");
248
249out:
250	schedule_work(&ms->trigger_event);
251}
252
253static int mirror_flush(struct dm_target *ti)
254{
255	struct mirror_set *ms = ti->private;
256	unsigned long error_bits;
257
258	unsigned int i;
259	struct dm_io_region io[ms->nr_mirrors];
260	struct mirror *m;
261	struct dm_io_request io_req = {
262		.bi_rw = WRITE_BARRIER,
263		.mem.type = DM_IO_KMEM,
264		.mem.ptr.bvec = NULL,
265		.client = ms->io_client,
266	};
267
268	for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++) {
269		io[i].bdev = m->dev->bdev;
270		io[i].sector = 0;
271		io[i].count = 0;
272	}
273
274	error_bits = -1;
275	dm_io(&io_req, ms->nr_mirrors, io, &error_bits);
276	if (unlikely(error_bits != 0)) {
277		for (i = 0; i < ms->nr_mirrors; i++)
278			if (test_bit(i, &error_bits))
279				fail_mirror(ms->mirror + i,
280					    DM_RAID1_FLUSH_ERROR);
281		return -EIO;
282	}
283
284	return 0;
285}
286
287/*-----------------------------------------------------------------
288 * Recovery.
289 *
290 * When a mirror is first activated we may find that some regions
291 * are in the no-sync state.  We have to recover these by
292 * recopying from the default mirror to all the others.
293 *---------------------------------------------------------------*/
294static void recovery_complete(int read_err, unsigned long write_err,
295			      void *context)
296{
297	struct dm_region *reg = context;
298	struct mirror_set *ms = dm_rh_region_context(reg);
299	int m, bit = 0;
300
301	if (read_err) {
302		/* Read error means the failure of default mirror. */
303		DMERR_LIMIT("Unable to read primary mirror during recovery");
304		fail_mirror(get_default_mirror(ms), DM_RAID1_SYNC_ERROR);
305	}
306
307	if (write_err) {
308		DMERR_LIMIT("Write error during recovery (error = 0x%lx)",
309			    write_err);
310		/*
311		 * Bits correspond to devices (excluding default mirror).
312		 * The default mirror cannot change during recovery.
313		 */
314		for (m = 0; m < ms->nr_mirrors; m++) {
315			if (&ms->mirror[m] == get_default_mirror(ms))
316				continue;
317			if (test_bit(bit, &write_err))
318				fail_mirror(ms->mirror + m,
319					    DM_RAID1_SYNC_ERROR);
320			bit++;
321		}
322	}
323
324	dm_rh_recovery_end(reg, !(read_err || write_err));
325}
326
327static int recover(struct mirror_set *ms, struct dm_region *reg)
328{
329	int r;
330	unsigned i;
331	struct dm_io_region from, to[DM_KCOPYD_MAX_REGIONS], *dest;
332	struct mirror *m;
333	unsigned long flags = 0;
334	region_t key = dm_rh_get_region_key(reg);
335	sector_t region_size = dm_rh_get_region_size(ms->rh);
336
337	/* fill in the source */
338	m = get_default_mirror(ms);
339	from.bdev = m->dev->bdev;
340	from.sector = m->offset + dm_rh_region_to_sector(ms->rh, key);
341	if (key == (ms->nr_regions - 1)) {
342		/*
343		 * The final region may be smaller than
344		 * region_size.
345		 */
346		from.count = ms->ti->len & (region_size - 1);
347		if (!from.count)
348			from.count = region_size;
349	} else
350		from.count = region_size;
351
352	/* fill in the destinations */
353	for (i = 0, dest = to; i < ms->nr_mirrors; i++) {
354		if (&ms->mirror[i] == get_default_mirror(ms))
355			continue;
356
357		m = ms->mirror + i;
358		dest->bdev = m->dev->bdev;
359		dest->sector = m->offset + dm_rh_region_to_sector(ms->rh, key);
360		dest->count = from.count;
361		dest++;
362	}
363
364	/* hand to kcopyd */
365	if (!errors_handled(ms))
366		set_bit(DM_KCOPYD_IGNORE_ERROR, &flags);
367
368	r = dm_kcopyd_copy(ms->kcopyd_client, &from, ms->nr_mirrors - 1, to,
369			   flags, recovery_complete, reg);
370
371	return r;
372}
373
374static void do_recovery(struct mirror_set *ms)
375{
376	struct dm_region *reg;
377	struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
378	int r;
379
380	/*
381	 * Start quiescing some regions.
382	 */
383	dm_rh_recovery_prepare(ms->rh);
384
385	/*
386	 * Copy any already quiesced regions.
387	 */
388	while ((reg = dm_rh_recovery_start(ms->rh))) {
389		r = recover(ms, reg);
390		if (r)
391			dm_rh_recovery_end(reg, 0);
392	}
393
394	/*
395	 * Update the in sync flag.
396	 */
397	if (!ms->in_sync &&
398	    (log->type->get_sync_count(log) == ms->nr_regions)) {
399		/* the sync is complete */
400		dm_table_event(ms->ti->table);
401		ms->in_sync = 1;
402	}
403}
404
405/*-----------------------------------------------------------------
406 * Reads
407 *---------------------------------------------------------------*/
408static struct mirror *choose_mirror(struct mirror_set *ms, sector_t sector)
409{
410	struct mirror *m = get_default_mirror(ms);
411
412	do {
413		if (likely(!atomic_read(&m->error_count)))
414			return m;
415
416		if (m-- == ms->mirror)
417			m += ms->nr_mirrors;
418	} while (m != get_default_mirror(ms));
419
420	return NULL;
421}
422
423static int default_ok(struct mirror *m)
424{
425	struct mirror *default_mirror = get_default_mirror(m->ms);
426
427	return !atomic_read(&default_mirror->error_count);
428}
429
430static int mirror_available(struct mirror_set *ms, struct bio *bio)
431{
432	struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
433	region_t region = dm_rh_bio_to_region(ms->rh, bio);
434
435	if (log->type->in_sync(log, region, 0))
436		return choose_mirror(ms,  bio->bi_sector) ? 1 : 0;
437
438	return 0;
439}
440
441/*
442 * remap a buffer to a particular mirror.
443 */
444static sector_t map_sector(struct mirror *m, struct bio *bio)
445{
446	if (unlikely(!bio->bi_size))
447		return 0;
448	return m->offset + dm_target_offset(m->ms->ti, bio->bi_sector);
449}
450
451static void map_bio(struct mirror *m, struct bio *bio)
452{
453	bio->bi_bdev = m->dev->bdev;
454	bio->bi_sector = map_sector(m, bio);
455}
456
457static void map_region(struct dm_io_region *io, struct mirror *m,
458		       struct bio *bio)
459{
460	io->bdev = m->dev->bdev;
461	io->sector = map_sector(m, bio);
462	io->count = bio->bi_size >> 9;
463}
464
465static void hold_bio(struct mirror_set *ms, struct bio *bio)
466{
467	/*
468	 * Lock is required to avoid race condition during suspend
469	 * process.
470	 */
471	spin_lock_irq(&ms->lock);
472
473	if (atomic_read(&ms->suspend)) {
474		spin_unlock_irq(&ms->lock);
475
476		/*
477		 * If device is suspended, complete the bio.
478		 */
479		if (dm_noflush_suspending(ms->ti))
480			bio_endio(bio, DM_ENDIO_REQUEUE);
481		else
482			bio_endio(bio, -EIO);
483		return;
484	}
485
486	/*
487	 * Hold bio until the suspend is complete.
488	 */
489	bio_list_add(&ms->holds, bio);
490	spin_unlock_irq(&ms->lock);
491}
492
493/*-----------------------------------------------------------------
494 * Reads
495 *---------------------------------------------------------------*/
496static void read_callback(unsigned long error, void *context)
497{
498	struct bio *bio = context;
499	struct mirror *m;
500
501	m = bio_get_m(bio);
502	bio_set_m(bio, NULL);
503
504	if (likely(!error)) {
505		bio_endio(bio, 0);
506		return;
507	}
508
509	fail_mirror(m, DM_RAID1_READ_ERROR);
510
511	if (likely(default_ok(m)) || mirror_available(m->ms, bio)) {
512		DMWARN_LIMIT("Read failure on mirror device %s.  "
513			     "Trying alternative device.",
514			     m->dev->name);
515		queue_bio(m->ms, bio, bio_rw(bio));
516		return;
517	}
518
519	DMERR_LIMIT("Read failure on mirror device %s.  Failing I/O.",
520		    m->dev->name);
521	bio_endio(bio, -EIO);
522}
523
524/* Asynchronous read. */
525static void read_async_bio(struct mirror *m, struct bio *bio)
526{
527	struct dm_io_region io;
528	struct dm_io_request io_req = {
529		.bi_rw = READ,
530		.mem.type = DM_IO_BVEC,
531		.mem.ptr.bvec = bio->bi_io_vec + bio->bi_idx,
532		.notify.fn = read_callback,
533		.notify.context = bio,
534		.client = m->ms->io_client,
535	};
536
537	map_region(&io, m, bio);
538	bio_set_m(bio, m);
539	BUG_ON(dm_io(&io_req, 1, &io, NULL));
540}
541
542static inline int region_in_sync(struct mirror_set *ms, region_t region,
543				 int may_block)
544{
545	int state = dm_rh_get_state(ms->rh, region, may_block);
546	return state == DM_RH_CLEAN || state == DM_RH_DIRTY;
547}
548
549static void do_reads(struct mirror_set *ms, struct bio_list *reads)
550{
551	region_t region;
552	struct bio *bio;
553	struct mirror *m;
554
555	while ((bio = bio_list_pop(reads))) {
556		region = dm_rh_bio_to_region(ms->rh, bio);
557		m = get_default_mirror(ms);
558
559		/*
560		 * We can only read balance if the region is in sync.
561		 */
562		if (likely(region_in_sync(ms, region, 1)))
563			m = choose_mirror(ms, bio->bi_sector);
564		else if (m && atomic_read(&m->error_count))
565			m = NULL;
566
567		if (likely(m))
568			read_async_bio(m, bio);
569		else
570			bio_endio(bio, -EIO);
571	}
572}
573
574/*-----------------------------------------------------------------
575 * Writes.
576 *
577 * We do different things with the write io depending on the
578 * state of the region that it's in:
579 *
580 * SYNC: 	increment pending, use kcopyd to write to *all* mirrors
581 * RECOVERING:	delay the io until recovery completes
582 * NOSYNC:	increment pending, just write to the default mirror
583 *---------------------------------------------------------------*/
584
585
586static void write_callback(unsigned long error, void *context)
587{
588	unsigned i, ret = 0;
589	struct bio *bio = (struct bio *) context;
590	struct mirror_set *ms;
591	int should_wake = 0;
592	unsigned long flags;
593
594	ms = bio_get_m(bio)->ms;
595	bio_set_m(bio, NULL);
596
597	/*
598	 * NOTE: We don't decrement the pending count here,
599	 * instead it is done by the targets endio function.
600	 * This way we handle both writes to SYNC and NOSYNC
601	 * regions with the same code.
602	 */
603	if (likely(!error)) {
604		bio_endio(bio, ret);
605		return;
606	}
607
608	for (i = 0; i < ms->nr_mirrors; i++)
609		if (test_bit(i, &error))
610			fail_mirror(ms->mirror + i, DM_RAID1_WRITE_ERROR);
611
612	/*
613	 * Need to raise event.  Since raising
614	 * events can block, we need to do it in
615	 * the main thread.
616	 */
617	spin_lock_irqsave(&ms->lock, flags);
618	if (!ms->failures.head)
619		should_wake = 1;
620	bio_list_add(&ms->failures, bio);
621	spin_unlock_irqrestore(&ms->lock, flags);
622	if (should_wake)
623		wakeup_mirrord(ms);
624}
625
626static void do_write(struct mirror_set *ms, struct bio *bio)
627{
628	unsigned int i;
629	struct dm_io_region io[ms->nr_mirrors], *dest = io;
630	struct mirror *m;
631	struct dm_io_request io_req = {
632		.bi_rw = WRITE | (bio->bi_rw & WRITE_BARRIER),
633		.mem.type = DM_IO_BVEC,
634		.mem.ptr.bvec = bio->bi_io_vec + bio->bi_idx,
635		.notify.fn = write_callback,
636		.notify.context = bio,
637		.client = ms->io_client,
638	};
639
640	for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++)
641		map_region(dest++, m, bio);
642
643	/*
644	 * Use default mirror because we only need it to retrieve the reference
645	 * to the mirror set in write_callback().
646	 */
647	bio_set_m(bio, get_default_mirror(ms));
648
649	BUG_ON(dm_io(&io_req, ms->nr_mirrors, io, NULL));
650}
651
652static void do_writes(struct mirror_set *ms, struct bio_list *writes)
653{
654	int state;
655	struct bio *bio;
656	struct bio_list sync, nosync, recover, *this_list = NULL;
657	struct bio_list requeue;
658	struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
659	region_t region;
660
661	if (!writes->head)
662		return;
663
664	/*
665	 * Classify each write.
666	 */
667	bio_list_init(&sync);
668	bio_list_init(&nosync);
669	bio_list_init(&recover);
670	bio_list_init(&requeue);
671
672	while ((bio = bio_list_pop(writes))) {
673		if (unlikely(bio_empty_barrier(bio))) {
674			bio_list_add(&sync, bio);
675			continue;
676		}
677
678		region = dm_rh_bio_to_region(ms->rh, bio);
679
680		if (log->type->is_remote_recovering &&
681		    log->type->is_remote_recovering(log, region)) {
682			bio_list_add(&requeue, bio);
683			continue;
684		}
685
686		state = dm_rh_get_state(ms->rh, region, 1);
687		switch (state) {
688		case DM_RH_CLEAN:
689		case DM_RH_DIRTY:
690			this_list = &sync;
691			break;
692
693		case DM_RH_NOSYNC:
694			this_list = &nosync;
695			break;
696
697		case DM_RH_RECOVERING:
698			this_list = &recover;
699			break;
700		}
701
702		bio_list_add(this_list, bio);
703	}
704
705	/*
706	 * Add bios that are delayed due to remote recovery
707	 * back on to the write queue
708	 */
709	if (unlikely(requeue.head)) {
710		spin_lock_irq(&ms->lock);
711		bio_list_merge(&ms->writes, &requeue);
712		spin_unlock_irq(&ms->lock);
713		delayed_wake(ms);
714	}
715
716	/*
717	 * Increment the pending counts for any regions that will
718	 * be written to (writes to recover regions are going to
719	 * be delayed).
720	 */
721	dm_rh_inc_pending(ms->rh, &sync);
722	dm_rh_inc_pending(ms->rh, &nosync);
723
724	/*
725	 * If the flush fails on a previous call and succeeds here,
726	 * we must not reset the log_failure variable.  We need
727	 * userspace interaction to do that.
728	 */
729	ms->log_failure = dm_rh_flush(ms->rh) ? 1 : ms->log_failure;
730
731	/*
732	 * Dispatch io.
733	 */
734	if (unlikely(ms->log_failure) && errors_handled(ms)) {
735		spin_lock_irq(&ms->lock);
736		bio_list_merge(&ms->failures, &sync);
737		spin_unlock_irq(&ms->lock);
738		wakeup_mirrord(ms);
739	} else
740		while ((bio = bio_list_pop(&sync)))
741			do_write(ms, bio);
742
743	while ((bio = bio_list_pop(&recover)))
744		dm_rh_delay(ms->rh, bio);
745
746	while ((bio = bio_list_pop(&nosync))) {
747		if (unlikely(ms->leg_failure) && errors_handled(ms)) {
748			spin_lock_irq(&ms->lock);
749			bio_list_add(&ms->failures, bio);
750			spin_unlock_irq(&ms->lock);
751			wakeup_mirrord(ms);
752		} else {
753			map_bio(get_default_mirror(ms), bio);
754			generic_make_request(bio);
755		}
756	}
757}
758
759static void do_failures(struct mirror_set *ms, struct bio_list *failures)
760{
761	struct bio *bio;
762
763	if (likely(!failures->head))
764		return;
765
766	/*
767	 * If the log has failed, unattempted writes are being
768	 * put on the holds list.  We can't issue those writes
769	 * until a log has been marked, so we must store them.
770	 *
771	 * If a 'noflush' suspend is in progress, we can requeue
772	 * the I/O's to the core.  This give userspace a chance
773	 * to reconfigure the mirror, at which point the core
774	 * will reissue the writes.  If the 'noflush' flag is
775	 * not set, we have no choice but to return errors.
776	 *
777	 * Some writes on the failures list may have been
778	 * submitted before the log failure and represent a
779	 * failure to write to one of the devices.  It is ok
780	 * for us to treat them the same and requeue them
781	 * as well.
782	 */
783	while ((bio = bio_list_pop(failures))) {
784		if (!ms->log_failure) {
785			ms->in_sync = 0;
786			dm_rh_mark_nosync(ms->rh, bio);
787		}
788
789		/*
790		 * If all the legs are dead, fail the I/O.
791		 * If we have been told to handle errors, hold the bio
792		 * and wait for userspace to deal with the problem.
793		 * Otherwise pretend that the I/O succeeded. (This would
794		 * be wrong if the failed leg returned after reboot and
795		 * got replicated back to the good legs.)
796		 */
797		if (!get_valid_mirror(ms))
798			bio_endio(bio, -EIO);
799		else if (errors_handled(ms))
800			hold_bio(ms, bio);
801		else
802			bio_endio(bio, 0);
803	}
804}
805
806static void trigger_event(struct work_struct *work)
807{
808	struct mirror_set *ms =
809		container_of(work, struct mirror_set, trigger_event);
810
811	dm_table_event(ms->ti->table);
812}
813
814/*-----------------------------------------------------------------
815 * kmirrord
816 *---------------------------------------------------------------*/
817static void do_mirror(struct work_struct *work)
818{
819	struct mirror_set *ms = container_of(work, struct mirror_set,
820					     kmirrord_work);
821	struct bio_list reads, writes, failures;
822	unsigned long flags;
823
824	spin_lock_irqsave(&ms->lock, flags);
825	reads = ms->reads;
826	writes = ms->writes;
827	failures = ms->failures;
828	bio_list_init(&ms->reads);
829	bio_list_init(&ms->writes);
830	bio_list_init(&ms->failures);
831	spin_unlock_irqrestore(&ms->lock, flags);
832
833	dm_rh_update_states(ms->rh, errors_handled(ms));
834	do_recovery(ms);
835	do_reads(ms, &reads);
836	do_writes(ms, &writes);
837	do_failures(ms, &failures);
838
839	dm_table_unplug_all(ms->ti->table);
840}
841
842/*-----------------------------------------------------------------
843 * Target functions
844 *---------------------------------------------------------------*/
845static struct mirror_set *alloc_context(unsigned int nr_mirrors,
846					uint32_t region_size,
847					struct dm_target *ti,
848					struct dm_dirty_log *dl)
849{
850	size_t len;
851	struct mirror_set *ms = NULL;
852
853	len = sizeof(*ms) + (sizeof(ms->mirror[0]) * nr_mirrors);
854
855	ms = kzalloc(len, GFP_KERNEL);
856	if (!ms) {
857		ti->error = "Cannot allocate mirror context";
858		return NULL;
859	}
860
861	spin_lock_init(&ms->lock);
862	bio_list_init(&ms->reads);
863	bio_list_init(&ms->writes);
864	bio_list_init(&ms->failures);
865	bio_list_init(&ms->holds);
866
867	ms->ti = ti;
868	ms->nr_mirrors = nr_mirrors;
869	ms->nr_regions = dm_sector_div_up(ti->len, region_size);
870	ms->in_sync = 0;
871	ms->log_failure = 0;
872	ms->leg_failure = 0;
873	atomic_set(&ms->suspend, 0);
874	atomic_set(&ms->default_mirror, DEFAULT_MIRROR);
875
876	ms->read_record_pool = mempool_create_slab_pool(MIN_READ_RECORDS,
877						_dm_raid1_read_record_cache);
878
879	if (!ms->read_record_pool) {
880		ti->error = "Error creating mirror read_record_pool";
881		kfree(ms);
882		return NULL;
883	}
884
885	ms->io_client = dm_io_client_create(DM_IO_PAGES);
886	if (IS_ERR(ms->io_client)) {
887		ti->error = "Error creating dm_io client";
888		mempool_destroy(ms->read_record_pool);
889		kfree(ms);
890 		return NULL;
891	}
892
893	ms->rh = dm_region_hash_create(ms, dispatch_bios, wakeup_mirrord,
894				       wakeup_all_recovery_waiters,
895				       ms->ti->begin, MAX_RECOVERY,
896				       dl, region_size, ms->nr_regions);
897	if (IS_ERR(ms->rh)) {
898		ti->error = "Error creating dirty region hash";
899		dm_io_client_destroy(ms->io_client);
900		mempool_destroy(ms->read_record_pool);
901		kfree(ms);
902		return NULL;
903	}
904
905	return ms;
906}
907
908static void free_context(struct mirror_set *ms, struct dm_target *ti,
909			 unsigned int m)
910{
911	while (m--)
912		dm_put_device(ti, ms->mirror[m].dev);
913
914	dm_io_client_destroy(ms->io_client);
915	dm_region_hash_destroy(ms->rh);
916	mempool_destroy(ms->read_record_pool);
917	kfree(ms);
918}
919
920static int get_mirror(struct mirror_set *ms, struct dm_target *ti,
921		      unsigned int mirror, char **argv)
922{
923	unsigned long long offset;
924
925	if (sscanf(argv[1], "%llu", &offset) != 1) {
926		ti->error = "Invalid offset";
927		return -EINVAL;
928	}
929
930	if (dm_get_device(ti, argv[0], dm_table_get_mode(ti->table),
931			  &ms->mirror[mirror].dev)) {
932		ti->error = "Device lookup failure";
933		return -ENXIO;
934	}
935
936	ms->mirror[mirror].ms = ms;
937	atomic_set(&(ms->mirror[mirror].error_count), 0);
938	ms->mirror[mirror].error_type = 0;
939	ms->mirror[mirror].offset = offset;
940
941	return 0;
942}
943
944/*
945 * Create dirty log: log_type #log_params <log_params>
946 */
947static struct dm_dirty_log *create_dirty_log(struct dm_target *ti,
948					     unsigned argc, char **argv,
949					     unsigned *args_used)
950{
951	unsigned param_count;
952	struct dm_dirty_log *dl;
953
954	if (argc < 2) {
955		ti->error = "Insufficient mirror log arguments";
956		return NULL;
957	}
958
959	if (sscanf(argv[1], "%u", &param_count) != 1) {
960		ti->error = "Invalid mirror log argument count";
961		return NULL;
962	}
963
964	*args_used = 2 + param_count;
965
966	if (argc < *args_used) {
967		ti->error = "Insufficient mirror log arguments";
968		return NULL;
969	}
970
971	dl = dm_dirty_log_create(argv[0], ti, mirror_flush, param_count,
972				 argv + 2);
973	if (!dl) {
974		ti->error = "Error creating mirror dirty log";
975		return NULL;
976	}
977
978	return dl;
979}
980
981static int parse_features(struct mirror_set *ms, unsigned argc, char **argv,
982			  unsigned *args_used)
983{
984	unsigned num_features;
985	struct dm_target *ti = ms->ti;
986
987	*args_used = 0;
988
989	if (!argc)
990		return 0;
991
992	if (sscanf(argv[0], "%u", &num_features) != 1) {
993		ti->error = "Invalid number of features";
994		return -EINVAL;
995	}
996
997	argc--;
998	argv++;
999	(*args_used)++;
1000
1001	if (num_features > argc) {
1002		ti->error = "Not enough arguments to support feature count";
1003		return -EINVAL;
1004	}
1005
1006	if (!strcmp("handle_errors", argv[0]))
1007		ms->features |= DM_RAID1_HANDLE_ERRORS;
1008	else {
1009		ti->error = "Unrecognised feature requested";
1010		return -EINVAL;
1011	}
1012
1013	(*args_used)++;
1014
1015	return 0;
1016}
1017
1018/*
1019 * Construct a mirror mapping:
1020 *
1021 * log_type #log_params <log_params>
1022 * #mirrors [mirror_path offset]{2,}
1023 * [#features <features>]
1024 *
1025 * log_type is "core" or "disk"
1026 * #log_params is between 1 and 3
1027 *
1028 * If present, features must be "handle_errors".
1029 */
1030static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1031{
1032	int r;
1033	unsigned int nr_mirrors, m, args_used;
1034	struct mirror_set *ms;
1035	struct dm_dirty_log *dl;
1036
1037	dl = create_dirty_log(ti, argc, argv, &args_used);
1038	if (!dl)
1039		return -EINVAL;
1040
1041	argv += args_used;
1042	argc -= args_used;
1043
1044	if (!argc || sscanf(argv[0], "%u", &nr_mirrors) != 1 ||
1045	    nr_mirrors < 2 || nr_mirrors > DM_KCOPYD_MAX_REGIONS + 1) {
1046		ti->error = "Invalid number of mirrors";
1047		dm_dirty_log_destroy(dl);
1048		return -EINVAL;
1049	}
1050
1051	argv++, argc--;
1052
1053	if (argc < nr_mirrors * 2) {
1054		ti->error = "Too few mirror arguments";
1055		dm_dirty_log_destroy(dl);
1056		return -EINVAL;
1057	}
1058
1059	ms = alloc_context(nr_mirrors, dl->type->get_region_size(dl), ti, dl);
1060	if (!ms) {
1061		dm_dirty_log_destroy(dl);
1062		return -ENOMEM;
1063	}
1064
1065	/* Get the mirror parameter sets */
1066	for (m = 0; m < nr_mirrors; m++) {
1067		r = get_mirror(ms, ti, m, argv);
1068		if (r) {
1069			free_context(ms, ti, m);
1070			return r;
1071		}
1072		argv += 2;
1073		argc -= 2;
1074	}
1075
1076	ti->private = ms;
1077	ti->split_io = dm_rh_get_region_size(ms->rh);
1078	ti->num_flush_requests = 1;
1079
1080	ms->kmirrord_wq = create_singlethread_workqueue("kmirrord");
1081	if (!ms->kmirrord_wq) {
1082		DMERR("couldn't start kmirrord");
1083		r = -ENOMEM;
1084		goto err_free_context;
1085	}
1086	INIT_WORK(&ms->kmirrord_work, do_mirror);
1087	init_timer(&ms->timer);
1088	ms->timer_pending = 0;
1089	INIT_WORK(&ms->trigger_event, trigger_event);
1090
1091	r = parse_features(ms, argc, argv, &args_used);
1092	if (r)
1093		goto err_destroy_wq;
1094
1095	argv += args_used;
1096	argc -= args_used;
1097
1098	/*
1099	 * Any read-balancing addition depends on the
1100	 * DM_RAID1_HANDLE_ERRORS flag being present.
1101	 * This is because the decision to balance depends
1102	 * on the sync state of a region.  If the above
1103	 * flag is not present, we ignore errors; and
1104	 * the sync state may be inaccurate.
1105	 */
1106
1107	if (argc) {
1108		ti->error = "Too many mirror arguments";
1109		r = -EINVAL;
1110		goto err_destroy_wq;
1111	}
1112
1113	r = dm_kcopyd_client_create(DM_KCOPYD_PAGES, &ms->kcopyd_client);
1114	if (r)
1115		goto err_destroy_wq;
1116
1117	wakeup_mirrord(ms);
1118	return 0;
1119
1120err_destroy_wq:
1121	destroy_workqueue(ms->kmirrord_wq);
1122err_free_context:
1123	free_context(ms, ti, ms->nr_mirrors);
1124	return r;
1125}
1126
1127static void mirror_dtr(struct dm_target *ti)
1128{
1129	struct mirror_set *ms = (struct mirror_set *) ti->private;
1130
1131	del_timer_sync(&ms->timer);
1132	flush_workqueue(ms->kmirrord_wq);
1133	flush_scheduled_work();
1134	dm_kcopyd_client_destroy(ms->kcopyd_client);
1135	destroy_workqueue(ms->kmirrord_wq);
1136	free_context(ms, ti, ms->nr_mirrors);
1137}
1138
1139/*
1140 * Mirror mapping function
1141 */
1142static int mirror_map(struct dm_target *ti, struct bio *bio,
1143		      union map_info *map_context)
1144{
1145	int r, rw = bio_rw(bio);
1146	struct mirror *m;
1147	struct mirror_set *ms = ti->private;
1148	struct dm_raid1_read_record *read_record = NULL;
1149	struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1150
1151	if (rw == WRITE) {
1152		/* Save region for mirror_end_io() handler */
1153		map_context->ll = dm_rh_bio_to_region(ms->rh, bio);
1154		queue_bio(ms, bio, rw);
1155		return DM_MAPIO_SUBMITTED;
1156	}
1157
1158	r = log->type->in_sync(log, dm_rh_bio_to_region(ms->rh, bio), 0);
1159	if (r < 0 && r != -EWOULDBLOCK)
1160		return r;
1161
1162	/*
1163	 * If region is not in-sync queue the bio.
1164	 */
1165	if (!r || (r == -EWOULDBLOCK)) {
1166		if (rw == READA)
1167			return -EWOULDBLOCK;
1168
1169		queue_bio(ms, bio, rw);
1170		return DM_MAPIO_SUBMITTED;
1171	}
1172
1173	/*
1174	 * The region is in-sync and we can perform reads directly.
1175	 * Store enough information so we can retry if it fails.
1176	 */
1177	m = choose_mirror(ms, bio->bi_sector);
1178	if (unlikely(!m))
1179		return -EIO;
1180
1181	read_record = mempool_alloc(ms->read_record_pool, GFP_NOIO);
1182	if (likely(read_record)) {
1183		dm_bio_record(&read_record->details, bio);
1184		map_context->ptr = read_record;
1185		read_record->m = m;
1186	}
1187
1188	map_bio(m, bio);
1189
1190	return DM_MAPIO_REMAPPED;
1191}
1192
1193static int mirror_end_io(struct dm_target *ti, struct bio *bio,
1194			 int error, union map_info *map_context)
1195{
1196	int rw = bio_rw(bio);
1197	struct mirror_set *ms = (struct mirror_set *) ti->private;
1198	struct mirror *m = NULL;
1199	struct dm_bio_details *bd = NULL;
1200	struct dm_raid1_read_record *read_record = map_context->ptr;
1201
1202	/*
1203	 * We need to dec pending if this was a write.
1204	 */
1205	if (rw == WRITE) {
1206		if (likely(!bio_empty_barrier(bio)))
1207			dm_rh_dec(ms->rh, map_context->ll);
1208		return error;
1209	}
1210
1211	if (error == -EOPNOTSUPP)
1212		goto out;
1213
1214	if ((error == -EWOULDBLOCK) && (bio->bi_rw & REQ_RAHEAD))
1215		goto out;
1216
1217	if (unlikely(error)) {
1218		if (!read_record) {
1219			/*
1220			 * There wasn't enough memory to record necessary
1221			 * information for a retry or there was no other
1222			 * mirror in-sync.
1223			 */
1224			DMERR_LIMIT("Mirror read failed.");
1225			return -EIO;
1226		}
1227
1228		m = read_record->m;
1229
1230		DMERR("Mirror read failed from %s. Trying alternative device.",
1231		      m->dev->name);
1232
1233		fail_mirror(m, DM_RAID1_READ_ERROR);
1234
1235		/*
1236		 * A failed read is requeued for another attempt using an intact
1237		 * mirror.
1238		 */
1239		if (default_ok(m) || mirror_available(ms, bio)) {
1240			bd = &read_record->details;
1241
1242			dm_bio_restore(bd, bio);
1243			mempool_free(read_record, ms->read_record_pool);
1244			map_context->ptr = NULL;
1245			queue_bio(ms, bio, rw);
1246			return 1;
1247		}
1248		DMERR("All replicated volumes dead, failing I/O");
1249	}
1250
1251out:
1252	if (read_record) {
1253		mempool_free(read_record, ms->read_record_pool);
1254		map_context->ptr = NULL;
1255	}
1256
1257	return error;
1258}
1259
1260static void mirror_presuspend(struct dm_target *ti)
1261{
1262	struct mirror_set *ms = (struct mirror_set *) ti->private;
1263	struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1264
1265	struct bio_list holds;
1266	struct bio *bio;
1267
1268	atomic_set(&ms->suspend, 1);
1269
1270	/*
1271	 * Process bios in the hold list to start recovery waiting
1272	 * for bios in the hold list. After the process, no bio has
1273	 * a chance to be added in the hold list because ms->suspend
1274	 * is set.
1275	 */
1276	spin_lock_irq(&ms->lock);
1277	holds = ms->holds;
1278	bio_list_init(&ms->holds);
1279	spin_unlock_irq(&ms->lock);
1280
1281	while ((bio = bio_list_pop(&holds)))
1282		hold_bio(ms, bio);
1283
1284	/*
1285	 * We must finish up all the work that we've
1286	 * generated (i.e. recovery work).
1287	 */
1288	dm_rh_stop_recovery(ms->rh);
1289
1290	wait_event(_kmirrord_recovery_stopped,
1291		   !dm_rh_recovery_in_flight(ms->rh));
1292
1293	if (log->type->presuspend && log->type->presuspend(log))
1294		DMWARN("log presuspend failed");
1295
1296	/*
1297	 * Now that recovery is complete/stopped and the
1298	 * delayed bios are queued, we need to wait for
1299	 * the worker thread to complete.  This way,
1300	 * we know that all of our I/O has been pushed.
1301	 */
1302	flush_workqueue(ms->kmirrord_wq);
1303}
1304
1305static void mirror_postsuspend(struct dm_target *ti)
1306{
1307	struct mirror_set *ms = ti->private;
1308	struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1309
1310	if (log->type->postsuspend && log->type->postsuspend(log))
1311		DMWARN("log postsuspend failed");
1312}
1313
1314static void mirror_resume(struct dm_target *ti)
1315{
1316	struct mirror_set *ms = ti->private;
1317	struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1318
1319	atomic_set(&ms->suspend, 0);
1320	if (log->type->resume && log->type->resume(log))
1321		DMWARN("log resume failed");
1322	dm_rh_start_recovery(ms->rh);
1323}
1324
1325/*
1326 * device_status_char
1327 * @m: mirror device/leg we want the status of
1328 *
1329 * We return one character representing the most severe error
1330 * we have encountered.
1331 *    A => Alive - No failures
1332 *    D => Dead - A write failure occurred leaving mirror out-of-sync
1333 *    S => Sync - A sychronization failure occurred, mirror out-of-sync
1334 *    R => Read - A read failure occurred, mirror data unaffected
1335 *
1336 * Returns: <char>
1337 */
1338static char device_status_char(struct mirror *m)
1339{
1340	if (!atomic_read(&(m->error_count)))
1341		return 'A';
1342
1343	return (test_bit(DM_RAID1_FLUSH_ERROR, &(m->error_type))) ? 'F' :
1344		(test_bit(DM_RAID1_WRITE_ERROR, &(m->error_type))) ? 'D' :
1345		(test_bit(DM_RAID1_SYNC_ERROR, &(m->error_type))) ? 'S' :
1346		(test_bit(DM_RAID1_READ_ERROR, &(m->error_type))) ? 'R' : 'U';
1347}
1348
1349
1350static int mirror_status(struct dm_target *ti, status_type_t type,
1351			 char *result, unsigned int maxlen)
1352{
1353	unsigned int m, sz = 0;
1354	struct mirror_set *ms = (struct mirror_set *) ti->private;
1355	struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
1356	char buffer[ms->nr_mirrors + 1];
1357
1358	switch (type) {
1359	case STATUSTYPE_INFO:
1360		DMEMIT("%d ", ms->nr_mirrors);
1361		for (m = 0; m < ms->nr_mirrors; m++) {
1362			DMEMIT("%s ", ms->mirror[m].dev->name);
1363			buffer[m] = device_status_char(&(ms->mirror[m]));
1364		}
1365		buffer[m] = '\0';
1366
1367		DMEMIT("%llu/%llu 1 %s ",
1368		      (unsigned long long)log->type->get_sync_count(log),
1369		      (unsigned long long)ms->nr_regions, buffer);
1370
1371		sz += log->type->status(log, type, result+sz, maxlen-sz);
1372
1373		break;
1374
1375	case STATUSTYPE_TABLE:
1376		sz = log->type->status(log, type, result, maxlen);
1377
1378		DMEMIT("%d", ms->nr_mirrors);
1379		for (m = 0; m < ms->nr_mirrors; m++)
1380			DMEMIT(" %s %llu", ms->mirror[m].dev->name,
1381			       (unsigned long long)ms->mirror[m].offset);
1382
1383		if (ms->features & DM_RAID1_HANDLE_ERRORS)
1384			DMEMIT(" 1 handle_errors");
1385	}
1386
1387	return 0;
1388}
1389
1390static int mirror_iterate_devices(struct dm_target *ti,
1391				  iterate_devices_callout_fn fn, void *data)
1392{
1393	struct mirror_set *ms = ti->private;
1394	int ret = 0;
1395	unsigned i;
1396
1397	for (i = 0; !ret && i < ms->nr_mirrors; i++)
1398		ret = fn(ti, ms->mirror[i].dev,
1399			 ms->mirror[i].offset, ti->len, data);
1400
1401	return ret;
1402}
1403
1404static struct target_type mirror_target = {
1405	.name	 = "mirror",
1406	.version = {1, 12, 0},
1407	.module	 = THIS_MODULE,
1408	.ctr	 = mirror_ctr,
1409	.dtr	 = mirror_dtr,
1410	.map	 = mirror_map,
1411	.end_io	 = mirror_end_io,
1412	.presuspend = mirror_presuspend,
1413	.postsuspend = mirror_postsuspend,
1414	.resume	 = mirror_resume,
1415	.status	 = mirror_status,
1416	.iterate_devices = mirror_iterate_devices,
1417};
1418
1419static int __init dm_mirror_init(void)
1420{
1421	int r;
1422
1423	_dm_raid1_read_record_cache = KMEM_CACHE(dm_raid1_read_record, 0);
1424	if (!_dm_raid1_read_record_cache) {
1425		DMERR("Can't allocate dm_raid1_read_record cache");
1426		r = -ENOMEM;
1427		goto bad_cache;
1428	}
1429
1430	r = dm_register_target(&mirror_target);
1431	if (r < 0) {
1432		DMERR("Failed to register mirror target");
1433		goto bad_target;
1434	}
1435
1436	return 0;
1437
1438bad_target:
1439	kmem_cache_destroy(_dm_raid1_read_record_cache);
1440bad_cache:
1441	return r;
1442}
1443
1444static void __exit dm_mirror_exit(void)
1445{
1446	dm_unregister_target(&mirror_target);
1447	kmem_cache_destroy(_dm_raid1_read_record_cache);
1448}
1449
1450/* Module hooks */
1451module_init(dm_mirror_init);
1452module_exit(dm_mirror_exit);
1453
1454MODULE_DESCRIPTION(DM_NAME " mirror target");
1455MODULE_AUTHOR("Joe Thornber");
1456MODULE_LICENSE("GPL");
1457