blob: 762cb086bb7f6e864fca4998f407644e8add96cc [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Copyright (C) 2003 Sistina Software Limited.
3 *
4 * This file is released under the GPL.
5 */
6
7#include "dm.h"
8#include "dm-bio-list.h"
Jonathan Brassow06386bb2008-02-08 02:11:37 +00009#include "dm-bio-record.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include "dm-io.h"
11#include "dm-log.h"
12#include "kcopyd.h"
13
14#include <linux/ctype.h>
15#include <linux/init.h>
16#include <linux/mempool.h>
17#include <linux/module.h>
18#include <linux/pagemap.h>
19#include <linux/slab.h>
20#include <linux/time.h>
21#include <linux/vmalloc.h>
22#include <linux/workqueue.h>
vignesh babu6f3c3f02007-10-19 22:38:44 +010023#include <linux/log2.h>
Jonathan Brassow72f4b312008-02-08 02:11:29 +000024#include <linux/hardirq.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070025
Alasdair G Kergon72d94862006-06-26 00:27:35 -070026#define DM_MSG_PREFIX "raid1"
Milan Broz88be1632007-05-09 02:33:04 -070027#define DM_IO_PAGES 64
Alasdair G Kergon72d94862006-06-26 00:27:35 -070028
Jonathan E Brassowa8e6afa2007-05-09 02:32:59 -070029#define DM_RAID1_HANDLE_ERRORS 0x01
Jonathan Brassowf44db672007-07-12 17:29:04 +010030#define errors_handled(p) ((p)->features & DM_RAID1_HANDLE_ERRORS)
Jonathan E Brassowa8e6afa2007-05-09 02:32:59 -070031
Jonathan E Brassow33184042006-11-08 17:44:44 -080032static DECLARE_WAIT_QUEUE_HEAD(_kmirrord_recovery_stopped);
Linus Torvalds1da177e2005-04-16 15:20:36 -070033
Linus Torvalds1da177e2005-04-16 15:20:36 -070034/*-----------------------------------------------------------------
35 * Region hash
36 *
37 * The mirror splits itself up into discrete regions. Each
38 * region can be in one of three states: clean, dirty,
39 * nosync. There is no need to put clean regions in the hash.
40 *
41 * In addition to being present in the hash table a region _may_
42 * be present on one of three lists.
43 *
44 * clean_regions: Regions on this list have no io pending to
45 * them, they are in sync, we are no longer interested in them,
46 * they are dull. rh_update_states() will remove them from the
47 * hash table.
48 *
49 * quiesced_regions: These regions have been spun down, ready
50 * for recovery. rh_recovery_start() will remove regions from
51 * this list and hand them to kmirrord, which will schedule the
52 * recovery io with kcopyd.
53 *
54 * recovered_regions: Regions that kcopyd has successfully
55 * recovered. rh_update_states() will now schedule any delayed
56 * io, up the recovery_count, and remove the region from the
57 * hash.
58 *
59 * There are 2 locks:
60 * A rw spin lock 'hash_lock' protects just the hash table,
61 * this is never held in write mode from interrupt context,
62 * which I believe means that we only have to disable irqs when
63 * doing a write lock.
64 *
65 * An ordinary spin lock 'region_lock' that protects the three
66 * lists in the region_hash, with the 'state', 'list' and
67 * 'bhs_delayed' fields of the regions. This is used from irq
68 * context, so all other uses will have to suspend local irqs.
69 *---------------------------------------------------------------*/
70struct mirror_set;
71struct region_hash {
72 struct mirror_set *ms;
73 uint32_t region_size;
74 unsigned region_shift;
75
76 /* holds persistent region state */
77 struct dirty_log *log;
78
79 /* hash table */
80 rwlock_t hash_lock;
81 mempool_t *region_pool;
82 unsigned int mask;
83 unsigned int nr_buckets;
84 struct list_head *buckets;
85
86 spinlock_t region_lock;
Jonathan E Brassow33184042006-11-08 17:44:44 -080087 atomic_t recovery_in_flight;
Linus Torvalds1da177e2005-04-16 15:20:36 -070088 struct semaphore recovery_count;
89 struct list_head clean_regions;
90 struct list_head quiesced_regions;
91 struct list_head recovered_regions;
Jonathan Brassowf44db672007-07-12 17:29:04 +010092 struct list_head failed_recovered_regions;
Linus Torvalds1da177e2005-04-16 15:20:36 -070093};
94
95enum {
96 RH_CLEAN,
97 RH_DIRTY,
98 RH_NOSYNC,
99 RH_RECOVERING
100};
101
102struct region {
103 struct region_hash *rh; /* FIXME: can we get rid of this ? */
104 region_t key;
105 int state;
106
107 struct list_head hash_list;
108 struct list_head list;
109
110 atomic_t pending;
111 struct bio_list delayed_bios;
112};
113
Neil Browne4c8b3b2006-06-26 00:27:26 -0700114
115/*-----------------------------------------------------------------
116 * Mirror set structures.
117 *---------------------------------------------------------------*/
Jonathan Brassow72f4b312008-02-08 02:11:29 +0000118enum dm_raid1_error {
119 DM_RAID1_WRITE_ERROR,
120 DM_RAID1_SYNC_ERROR,
121 DM_RAID1_READ_ERROR
122};
123
Neil Browne4c8b3b2006-06-26 00:27:26 -0700124struct mirror {
Jonathan Brassowaa5617c2007-10-19 22:47:58 +0100125 struct mirror_set *ms;
Neil Browne4c8b3b2006-06-26 00:27:26 -0700126 atomic_t error_count;
Al Viro39ed7ad2008-02-13 03:53:00 +0000127 unsigned long error_type;
Neil Browne4c8b3b2006-06-26 00:27:26 -0700128 struct dm_dev *dev;
129 sector_t offset;
130};
131
132struct mirror_set {
133 struct dm_target *ti;
134 struct list_head list;
135 struct region_hash rh;
136 struct kcopyd_client *kcopyd_client;
Jonathan E Brassowa8e6afa2007-05-09 02:32:59 -0700137 uint64_t features;
Neil Browne4c8b3b2006-06-26 00:27:26 -0700138
Jonathan Brassow72f4b312008-02-08 02:11:29 +0000139 spinlock_t lock; /* protects the lists */
Neil Browne4c8b3b2006-06-26 00:27:26 -0700140 struct bio_list reads;
141 struct bio_list writes;
Jonathan Brassow72f4b312008-02-08 02:11:29 +0000142 struct bio_list failures;
Neil Browne4c8b3b2006-06-26 00:27:26 -0700143
Milan Broz88be1632007-05-09 02:33:04 -0700144 struct dm_io_client *io_client;
Jonathan Brassow06386bb2008-02-08 02:11:37 +0000145 mempool_t *read_record_pool;
Milan Broz88be1632007-05-09 02:33:04 -0700146
Neil Browne4c8b3b2006-06-26 00:27:26 -0700147 /* recovery */
148 region_t nr_regions;
149 int in_sync;
Jonathan Brassowfc1ff952007-07-12 17:29:15 +0100150 int log_failure;
Jonathan Brassowb80aa7a2008-02-08 02:11:35 +0000151 atomic_t suspend;
Neil Browne4c8b3b2006-06-26 00:27:26 -0700152
Jonathan Brassow72f4b312008-02-08 02:11:29 +0000153 atomic_t default_mirror; /* Default mirror */
Neil Browne4c8b3b2006-06-26 00:27:26 -0700154
Holger Smolinski6ad36fe2007-05-09 02:32:50 -0700155 struct workqueue_struct *kmirrord_wq;
156 struct work_struct kmirrord_work;
Jonathan Brassow72f4b312008-02-08 02:11:29 +0000157 struct work_struct trigger_event;
Holger Smolinski6ad36fe2007-05-09 02:32:50 -0700158
Neil Browne4c8b3b2006-06-26 00:27:26 -0700159 unsigned int nr_mirrors;
160 struct mirror mirror[0];
161};
162
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163/*
164 * Conversion fns
165 */
166static inline region_t bio_to_region(struct region_hash *rh, struct bio *bio)
167{
Neil Browne4c8b3b2006-06-26 00:27:26 -0700168 return (bio->bi_sector - rh->ms->ti->begin) >> rh->region_shift;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169}
170
171static inline sector_t region_to_sector(struct region_hash *rh, region_t region)
172{
173 return region << rh->region_shift;
174}
175
Holger Smolinski6ad36fe2007-05-09 02:32:50 -0700176static void wake(struct mirror_set *ms)
177{
178 queue_work(ms->kmirrord_wq, &ms->kmirrord_work);
179}
180
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181/* FIXME move this */
182static void queue_bio(struct mirror_set *ms, struct bio *bio, int rw);
183
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184#define MIN_REGIONS 64
185#define MAX_RECOVERY 1
186static int rh_init(struct region_hash *rh, struct mirror_set *ms,
187 struct dirty_log *log, uint32_t region_size,
188 region_t nr_regions)
189{
190 unsigned int nr_buckets, max_buckets;
191 size_t i;
192
193 /*
194 * Calculate a suitable number of buckets for our hash
195 * table.
196 */
197 max_buckets = nr_regions >> 6;
198 for (nr_buckets = 128u; nr_buckets < max_buckets; nr_buckets <<= 1)
199 ;
200 nr_buckets >>= 1;
201
202 rh->ms = ms;
203 rh->log = log;
204 rh->region_size = region_size;
205 rh->region_shift = ffs(region_size) - 1;
206 rwlock_init(&rh->hash_lock);
207 rh->mask = nr_buckets - 1;
208 rh->nr_buckets = nr_buckets;
209
210 rh->buckets = vmalloc(nr_buckets * sizeof(*rh->buckets));
211 if (!rh->buckets) {
212 DMERR("unable to allocate region hash memory");
213 return -ENOMEM;
214 }
215
216 for (i = 0; i < nr_buckets; i++)
217 INIT_LIST_HEAD(rh->buckets + i);
218
219 spin_lock_init(&rh->region_lock);
220 sema_init(&rh->recovery_count, 0);
Jonathan E Brassow33184042006-11-08 17:44:44 -0800221 atomic_set(&rh->recovery_in_flight, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222 INIT_LIST_HEAD(&rh->clean_regions);
223 INIT_LIST_HEAD(&rh->quiesced_regions);
224 INIT_LIST_HEAD(&rh->recovered_regions);
Jonathan Brassowf44db672007-07-12 17:29:04 +0100225 INIT_LIST_HEAD(&rh->failed_recovered_regions);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226
Matthew Dobson0eaae62a2006-03-26 01:37:47 -0800227 rh->region_pool = mempool_create_kmalloc_pool(MIN_REGIONS,
228 sizeof(struct region));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229 if (!rh->region_pool) {
230 vfree(rh->buckets);
231 rh->buckets = NULL;
232 return -ENOMEM;
233 }
234
235 return 0;
236}
237
238static void rh_exit(struct region_hash *rh)
239{
240 unsigned int h;
241 struct region *reg, *nreg;
242
243 BUG_ON(!list_empty(&rh->quiesced_regions));
244 for (h = 0; h < rh->nr_buckets; h++) {
245 list_for_each_entry_safe(reg, nreg, rh->buckets + h, hash_list) {
246 BUG_ON(atomic_read(&reg->pending));
247 mempool_free(reg, rh->region_pool);
248 }
249 }
250
251 if (rh->log)
252 dm_destroy_dirty_log(rh->log);
253 if (rh->region_pool)
254 mempool_destroy(rh->region_pool);
255 vfree(rh->buckets);
256}
257
258#define RH_HASH_MULT 2654435387U
259
260static inline unsigned int rh_hash(struct region_hash *rh, region_t region)
261{
262 return (unsigned int) ((region * RH_HASH_MULT) >> 12) & rh->mask;
263}
264
265static struct region *__rh_lookup(struct region_hash *rh, region_t region)
266{
267 struct region *reg;
268
269 list_for_each_entry (reg, rh->buckets + rh_hash(rh, region), hash_list)
270 if (reg->key == region)
271 return reg;
272
273 return NULL;
274}
275
276static void __rh_insert(struct region_hash *rh, struct region *reg)
277{
278 unsigned int h = rh_hash(rh, reg->key);
279 list_add(&reg->hash_list, rh->buckets + h);
280}
281
282static struct region *__rh_alloc(struct region_hash *rh, region_t region)
283{
284 struct region *reg, *nreg;
285
286 read_unlock(&rh->hash_lock);
Daniel Kobrasc06aad82006-08-27 01:23:24 -0700287 nreg = mempool_alloc(rh->region_pool, GFP_ATOMIC);
288 if (unlikely(!nreg))
289 nreg = kmalloc(sizeof(struct region), GFP_NOIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290 nreg->state = rh->log->type->in_sync(rh->log, region, 1) ?
291 RH_CLEAN : RH_NOSYNC;
292 nreg->rh = rh;
293 nreg->key = region;
294
295 INIT_LIST_HEAD(&nreg->list);
296
297 atomic_set(&nreg->pending, 0);
298 bio_list_init(&nreg->delayed_bios);
299 write_lock_irq(&rh->hash_lock);
300
301 reg = __rh_lookup(rh, region);
302 if (reg)
303 /* we lost the race */
304 mempool_free(nreg, rh->region_pool);
305
306 else {
307 __rh_insert(rh, nreg);
308 if (nreg->state == RH_CLEAN) {
309 spin_lock(&rh->region_lock);
310 list_add(&nreg->list, &rh->clean_regions);
311 spin_unlock(&rh->region_lock);
312 }
313 reg = nreg;
314 }
315 write_unlock_irq(&rh->hash_lock);
316 read_lock(&rh->hash_lock);
317
318 return reg;
319}
320
321static inline struct region *__rh_find(struct region_hash *rh, region_t region)
322{
323 struct region *reg;
324
325 reg = __rh_lookup(rh, region);
326 if (!reg)
327 reg = __rh_alloc(rh, region);
328
329 return reg;
330}
331
332static int rh_state(struct region_hash *rh, region_t region, int may_block)
333{
334 int r;
335 struct region *reg;
336
337 read_lock(&rh->hash_lock);
338 reg = __rh_lookup(rh, region);
339 read_unlock(&rh->hash_lock);
340
341 if (reg)
342 return reg->state;
343
344 /*
345 * The region wasn't in the hash, so we fall back to the
346 * dirty log.
347 */
348 r = rh->log->type->in_sync(rh->log, region, may_block);
349
350 /*
351 * Any error from the dirty log (eg. -EWOULDBLOCK) gets
352 * taken as a RH_NOSYNC
353 */
354 return r == 1 ? RH_CLEAN : RH_NOSYNC;
355}
356
357static inline int rh_in_sync(struct region_hash *rh,
358 region_t region, int may_block)
359{
360 int state = rh_state(rh, region, may_block);
361 return state == RH_CLEAN || state == RH_DIRTY;
362}
363
364static void dispatch_bios(struct mirror_set *ms, struct bio_list *bio_list)
365{
366 struct bio *bio;
367
368 while ((bio = bio_list_pop(bio_list))) {
369 queue_bio(ms, bio, WRITE);
370 }
371}
372
Jonathan E Brassowf3ee6b22006-12-08 02:41:11 -0800373static void complete_resync_work(struct region *reg, int success)
374{
375 struct region_hash *rh = reg->rh;
376
377 rh->log->type->set_region_sync(rh->log, reg->key, success);
Jonathan Brassowb80aa7a2008-02-08 02:11:35 +0000378
379 /*
380 * Dispatch the bios before we call 'wake_up_all'.
381 * This is important because if we are suspending,
382 * we want to know that recovery is complete and
383 * the work queue is flushed. If we wake_up_all
384 * before we dispatch_bios (queue bios and call wake()),
385 * then we risk suspending before the work queue
386 * has been properly flushed.
387 */
Jonathan E Brassowf3ee6b22006-12-08 02:41:11 -0800388 dispatch_bios(rh->ms, &reg->delayed_bios);
389 if (atomic_dec_and_test(&rh->recovery_in_flight))
390 wake_up_all(&_kmirrord_recovery_stopped);
391 up(&rh->recovery_count);
392}
393
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394static void rh_update_states(struct region_hash *rh)
395{
396 struct region *reg, *next;
397
398 LIST_HEAD(clean);
399 LIST_HEAD(recovered);
Jonathan Brassowf44db672007-07-12 17:29:04 +0100400 LIST_HEAD(failed_recovered);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401
402 /*
403 * Quickly grab the lists.
404 */
405 write_lock_irq(&rh->hash_lock);
406 spin_lock(&rh->region_lock);
407 if (!list_empty(&rh->clean_regions)) {
408 list_splice(&rh->clean_regions, &clean);
409 INIT_LIST_HEAD(&rh->clean_regions);
410
Jonathan Brassow943317e2007-07-12 17:28:25 +0100411 list_for_each_entry(reg, &clean, list)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412 list_del(&reg->hash_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413 }
414
415 if (!list_empty(&rh->recovered_regions)) {
416 list_splice(&rh->recovered_regions, &recovered);
417 INIT_LIST_HEAD(&rh->recovered_regions);
418
419 list_for_each_entry (reg, &recovered, list)
420 list_del(&reg->hash_list);
421 }
Jonathan Brassowf44db672007-07-12 17:29:04 +0100422
423 if (!list_empty(&rh->failed_recovered_regions)) {
424 list_splice(&rh->failed_recovered_regions, &failed_recovered);
425 INIT_LIST_HEAD(&rh->failed_recovered_regions);
426
427 list_for_each_entry(reg, &failed_recovered, list)
428 list_del(&reg->hash_list);
429 }
430
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431 spin_unlock(&rh->region_lock);
432 write_unlock_irq(&rh->hash_lock);
433
434 /*
435 * All the regions on the recovered and clean lists have
436 * now been pulled out of the system, so no need to do
437 * any more locking.
438 */
439 list_for_each_entry_safe (reg, next, &recovered, list) {
440 rh->log->type->clear_region(rh->log, reg->key);
Jonathan E Brassowf3ee6b22006-12-08 02:41:11 -0800441 complete_resync_work(reg, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442 mempool_free(reg, rh->region_pool);
443 }
444
Jonathan Brassowf44db672007-07-12 17:29:04 +0100445 list_for_each_entry_safe(reg, next, &failed_recovered, list) {
446 complete_resync_work(reg, errors_handled(rh->ms) ? 0 : 1);
447 mempool_free(reg, rh->region_pool);
448 }
449
Jonathan Brassow943317e2007-07-12 17:28:25 +0100450 list_for_each_entry_safe(reg, next, &clean, list) {
451 rh->log->type->clear_region(rh->log, reg->key);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452 mempool_free(reg, rh->region_pool);
Jonathan Brassow943317e2007-07-12 17:28:25 +0100453 }
454
455 rh->log->type->flush(rh->log);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456}
457
458static void rh_inc(struct region_hash *rh, region_t region)
459{
460 struct region *reg;
461
462 read_lock(&rh->hash_lock);
463 reg = __rh_find(rh, region);
Jun'ichi Nomura844e8d92005-09-09 16:23:42 -0700464
Jonathan E Brassow7692c5d2005-11-21 21:32:37 -0800465 spin_lock_irq(&rh->region_lock);
Jun'ichi Nomura844e8d92005-09-09 16:23:42 -0700466 atomic_inc(&reg->pending);
467
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468 if (reg->state == RH_CLEAN) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469 reg->state = RH_DIRTY;
470 list_del_init(&reg->list); /* take off the clean list */
Jonathan E Brassow7692c5d2005-11-21 21:32:37 -0800471 spin_unlock_irq(&rh->region_lock);
472
473 rh->log->type->mark_region(rh->log, reg->key);
474 } else
475 spin_unlock_irq(&rh->region_lock);
476
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478 read_unlock(&rh->hash_lock);
479}
480
481static void rh_inc_pending(struct region_hash *rh, struct bio_list *bios)
482{
483 struct bio *bio;
484
485 for (bio = bios->head; bio; bio = bio->bi_next)
486 rh_inc(rh, bio_to_region(rh, bio));
487}
488
489static void rh_dec(struct region_hash *rh, region_t region)
490{
491 unsigned long flags;
492 struct region *reg;
493 int should_wake = 0;
494
495 read_lock(&rh->hash_lock);
496 reg = __rh_lookup(rh, region);
497 read_unlock(&rh->hash_lock);
498
Jonathan E Brassow7692c5d2005-11-21 21:32:37 -0800499 spin_lock_irqsave(&rh->region_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500 if (atomic_dec_and_test(&reg->pending)) {
Jun'ichi Nomura930d3322006-03-27 01:17:47 -0800501 /*
502 * There is no pending I/O for this region.
503 * We can move the region to corresponding list for next action.
504 * At this point, the region is not yet connected to any list.
505 *
506 * If the state is RH_NOSYNC, the region should be kept off
507 * from clean list.
508 * The hash entry for RH_NOSYNC will remain in memory
509 * until the region is recovered or the map is reloaded.
510 */
511
512 /* do nothing for RH_NOSYNC */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513 if (reg->state == RH_RECOVERING) {
514 list_add_tail(&reg->list, &rh->quiesced_regions);
Jun'ichi Nomura930d3322006-03-27 01:17:47 -0800515 } else if (reg->state == RH_DIRTY) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516 reg->state = RH_CLEAN;
517 list_add(&reg->list, &rh->clean_regions);
518 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 should_wake = 1;
520 }
Jonathan E Brassow7692c5d2005-11-21 21:32:37 -0800521 spin_unlock_irqrestore(&rh->region_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522
523 if (should_wake)
Holger Smolinski6ad36fe2007-05-09 02:32:50 -0700524 wake(rh->ms);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525}
526
527/*
528 * Starts quiescing a region in preparation for recovery.
529 */
530static int __rh_recovery_prepare(struct region_hash *rh)
531{
532 int r;
533 struct region *reg;
534 region_t region;
535
536 /*
537 * Ask the dirty log what's next.
538 */
539 r = rh->log->type->get_resync_work(rh->log, &region);
540 if (r <= 0)
541 return r;
542
543 /*
544 * Get this region, and start it quiescing by setting the
545 * recovering flag.
546 */
547 read_lock(&rh->hash_lock);
548 reg = __rh_find(rh, region);
549 read_unlock(&rh->hash_lock);
550
551 spin_lock_irq(&rh->region_lock);
552 reg->state = RH_RECOVERING;
553
554 /* Already quiesced ? */
555 if (atomic_read(&reg->pending))
556 list_del_init(&reg->list);
Akinobu Mita179e0912006-06-26 00:24:41 -0700557 else
558 list_move(&reg->list, &rh->quiesced_regions);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560 spin_unlock_irq(&rh->region_lock);
561
562 return 1;
563}
564
565static void rh_recovery_prepare(struct region_hash *rh)
566{
Jonathan E Brassow33184042006-11-08 17:44:44 -0800567 /* Extra reference to avoid race with rh_stop_recovery */
568 atomic_inc(&rh->recovery_in_flight);
569
570 while (!down_trylock(&rh->recovery_count)) {
571 atomic_inc(&rh->recovery_in_flight);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572 if (__rh_recovery_prepare(rh) <= 0) {
Jonathan E Brassow33184042006-11-08 17:44:44 -0800573 atomic_dec(&rh->recovery_in_flight);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574 up(&rh->recovery_count);
575 break;
576 }
Jonathan E Brassow33184042006-11-08 17:44:44 -0800577 }
578
579 /* Drop the extra reference */
580 if (atomic_dec_and_test(&rh->recovery_in_flight))
581 wake_up_all(&_kmirrord_recovery_stopped);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582}
583
584/*
585 * Returns any quiesced regions.
586 */
587static struct region *rh_recovery_start(struct region_hash *rh)
588{
589 struct region *reg = NULL;
590
591 spin_lock_irq(&rh->region_lock);
592 if (!list_empty(&rh->quiesced_regions)) {
593 reg = list_entry(rh->quiesced_regions.next,
594 struct region, list);
595 list_del_init(&reg->list); /* remove from the quiesced list */
596 }
597 spin_unlock_irq(&rh->region_lock);
598
599 return reg;
600}
601
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602static void rh_recovery_end(struct region *reg, int success)
603{
604 struct region_hash *rh = reg->rh;
605
606 spin_lock_irq(&rh->region_lock);
Jonathan Brassowf44db672007-07-12 17:29:04 +0100607 if (success)
608 list_add(&reg->list, &reg->rh->recovered_regions);
609 else {
610 reg->state = RH_NOSYNC;
611 list_add(&reg->list, &reg->rh->failed_recovered_regions);
612 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613 spin_unlock_irq(&rh->region_lock);
614
Holger Smolinski6ad36fe2007-05-09 02:32:50 -0700615 wake(rh->ms);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616}
617
Jonathan Brassowfc1ff952007-07-12 17:29:15 +0100618static int rh_flush(struct region_hash *rh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619{
Jonathan Brassowfc1ff952007-07-12 17:29:15 +0100620 return rh->log->type->flush(rh->log);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621}
622
623static void rh_delay(struct region_hash *rh, struct bio *bio)
624{
625 struct region *reg;
626
627 read_lock(&rh->hash_lock);
628 reg = __rh_find(rh, bio_to_region(rh, bio));
629 bio_list_add(&reg->delayed_bios, bio);
630 read_unlock(&rh->hash_lock);
631}
632
633static void rh_stop_recovery(struct region_hash *rh)
634{
635 int i;
636
637 /* wait for any recovering regions */
638 for (i = 0; i < MAX_RECOVERY; i++)
639 down(&rh->recovery_count);
640}
641
642static void rh_start_recovery(struct region_hash *rh)
643{
644 int i;
645
646 for (i = 0; i < MAX_RECOVERY; i++)
647 up(&rh->recovery_count);
648
Holger Smolinski6ad36fe2007-05-09 02:32:50 -0700649 wake(rh->ms);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650}
651
Jonathan Brassow06386bb2008-02-08 02:11:37 +0000652#define MIN_READ_RECORDS 20
653struct dm_raid1_read_record {
654 struct mirror *m;
655 struct dm_bio_details details;
656};
657
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658/*
659 * Every mirror should look like this one.
660 */
661#define DEFAULT_MIRROR 0
662
663/*
Jonathan Brassow06386bb2008-02-08 02:11:37 +0000664 * This is yucky. We squirrel the mirror struct away inside
665 * bi_next for read/write buffers. This is safe since the bh
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666 * doesn't get submitted to the lower levels of block layer.
667 */
Jonathan Brassow06386bb2008-02-08 02:11:37 +0000668static struct mirror *bio_get_m(struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669{
Jonathan Brassow06386bb2008-02-08 02:11:37 +0000670 return (struct mirror *) bio->bi_next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671}
672
Jonathan Brassow06386bb2008-02-08 02:11:37 +0000673static void bio_set_m(struct bio *bio, struct mirror *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674{
Jonathan Brassow06386bb2008-02-08 02:11:37 +0000675 bio->bi_next = (struct bio *) m;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676}
677
Jonathan Brassow72f4b312008-02-08 02:11:29 +0000678static struct mirror *get_default_mirror(struct mirror_set *ms)
679{
680 return &ms->mirror[atomic_read(&ms->default_mirror)];
681}
682
683static void set_default_mirror(struct mirror *m)
684{
685 struct mirror_set *ms = m->ms;
686 struct mirror *m0 = &(ms->mirror[0]);
687
688 atomic_set(&ms->default_mirror, m - m0);
689}
690
691/* fail_mirror
692 * @m: mirror device to fail
693 * @error_type: one of the enum's, DM_RAID1_*_ERROR
694 *
695 * If errors are being handled, record the type of
696 * error encountered for this device. If this type
697 * of error has already been recorded, we can return;
698 * otherwise, we must signal userspace by triggering
699 * an event. Additionally, if the device is the
700 * primary device, we must choose a new primary, but
701 * only if the mirror is in-sync.
702 *
703 * This function must not block.
704 */
705static void fail_mirror(struct mirror *m, enum dm_raid1_error error_type)
706{
707 struct mirror_set *ms = m->ms;
708 struct mirror *new;
709
710 if (!errors_handled(ms))
711 return;
712
713 /*
714 * error_count is used for nothing more than a
715 * simple way to tell if a device has encountered
716 * errors.
717 */
718 atomic_inc(&m->error_count);
719
720 if (test_and_set_bit(error_type, &m->error_type))
721 return;
722
723 if (m != get_default_mirror(ms))
724 goto out;
725
726 if (!ms->in_sync) {
727 /*
728 * Better to issue requests to same failing device
729 * than to risk returning corrupt data.
730 */
731 DMERR("Primary mirror (%s) failed while out-of-sync: "
732 "Reads may fail.", m->dev->name);
733 goto out;
734 }
735
736 for (new = ms->mirror; new < ms->mirror + ms->nr_mirrors; new++)
737 if (!atomic_read(&new->error_count)) {
738 set_default_mirror(new);
739 break;
740 }
741
742 if (unlikely(new == ms->mirror + ms->nr_mirrors))
743 DMWARN("All sides of mirror have failed.");
744
745out:
746 schedule_work(&ms->trigger_event);
747}
748
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749/*-----------------------------------------------------------------
750 * Recovery.
751 *
752 * When a mirror is first activated we may find that some regions
753 * are in the no-sync state. We have to recover these by
754 * recopying from the default mirror to all the others.
755 *---------------------------------------------------------------*/
Alasdair G Kergon4cdc1d12008-03-28 14:16:10 -0700756static void recovery_complete(int read_err, unsigned long write_err,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757 void *context)
758{
Jonathan Brassow8f0205b2008-02-08 02:11:32 +0000759 struct region *reg = (struct region *)context;
760 struct mirror_set *ms = reg->rh->ms;
761 int m, bit = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762
Jonathan Brassow8f0205b2008-02-08 02:11:32 +0000763 if (read_err) {
Jonathan Brassowf44db672007-07-12 17:29:04 +0100764 /* Read error means the failure of default mirror. */
765 DMERR_LIMIT("Unable to read primary mirror during recovery");
Jonathan Brassow8f0205b2008-02-08 02:11:32 +0000766 fail_mirror(get_default_mirror(ms), DM_RAID1_SYNC_ERROR);
767 }
Jonathan Brassowf44db672007-07-12 17:29:04 +0100768
Jonathan Brassow8f0205b2008-02-08 02:11:32 +0000769 if (write_err) {
Alasdair G Kergon4cdc1d12008-03-28 14:16:10 -0700770 DMERR_LIMIT("Write error during recovery (error = 0x%lx)",
Jonathan Brassowf44db672007-07-12 17:29:04 +0100771 write_err);
Jonathan Brassow8f0205b2008-02-08 02:11:32 +0000772 /*
773 * Bits correspond to devices (excluding default mirror).
774 * The default mirror cannot change during recovery.
775 */
776 for (m = 0; m < ms->nr_mirrors; m++) {
777 if (&ms->mirror[m] == get_default_mirror(ms))
778 continue;
779 if (test_bit(bit, &write_err))
780 fail_mirror(ms->mirror + m,
781 DM_RAID1_SYNC_ERROR);
782 bit++;
783 }
784 }
Jonathan Brassowf44db672007-07-12 17:29:04 +0100785
Jonathan Brassowce503f52006-06-26 00:27:30 -0700786 rh_recovery_end(reg, !(read_err || write_err));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787}
788
789static int recover(struct mirror_set *ms, struct region *reg)
790{
791 int r;
792 unsigned int i;
793 struct io_region from, to[KCOPYD_MAX_REGIONS], *dest;
794 struct mirror *m;
795 unsigned long flags = 0;
796
797 /* fill in the source */
Jonathan Brassow72f4b312008-02-08 02:11:29 +0000798 m = get_default_mirror(ms);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799 from.bdev = m->dev->bdev;
800 from.sector = m->offset + region_to_sector(reg->rh, reg->key);
801 if (reg->key == (ms->nr_regions - 1)) {
802 /*
803 * The final region may be smaller than
804 * region_size.
805 */
806 from.count = ms->ti->len & (reg->rh->region_size - 1);
807 if (!from.count)
808 from.count = reg->rh->region_size;
809 } else
810 from.count = reg->rh->region_size;
811
812 /* fill in the destinations */
813 for (i = 0, dest = to; i < ms->nr_mirrors; i++) {
Jonathan Brassow72f4b312008-02-08 02:11:29 +0000814 if (&ms->mirror[i] == get_default_mirror(ms))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815 continue;
816
817 m = ms->mirror + i;
818 dest->bdev = m->dev->bdev;
819 dest->sector = m->offset + region_to_sector(reg->rh, reg->key);
820 dest->count = from.count;
821 dest++;
822 }
823
824 /* hand to kcopyd */
825 set_bit(KCOPYD_IGNORE_ERROR, &flags);
826 r = kcopyd_copy(ms->kcopyd_client, &from, ms->nr_mirrors - 1, to, flags,
827 recovery_complete, reg);
828
829 return r;
830}
831
832static void do_recovery(struct mirror_set *ms)
833{
834 int r;
835 struct region *reg;
836 struct dirty_log *log = ms->rh.log;
837
838 /*
839 * Start quiescing some regions.
840 */
841 rh_recovery_prepare(&ms->rh);
842
843 /*
844 * Copy any already quiesced regions.
845 */
846 while ((reg = rh_recovery_start(&ms->rh))) {
847 r = recover(ms, reg);
848 if (r)
849 rh_recovery_end(reg, 0);
850 }
851
852 /*
853 * Update the in sync flag.
854 */
855 if (!ms->in_sync &&
856 (log->type->get_sync_count(log) == ms->nr_regions)) {
857 /* the sync is complete */
858 dm_table_event(ms->ti->table);
859 ms->in_sync = 1;
860 }
861}
862
863/*-----------------------------------------------------------------
864 * Reads
865 *---------------------------------------------------------------*/
866static struct mirror *choose_mirror(struct mirror_set *ms, sector_t sector)
867{
Jonathan Brassow06386bb2008-02-08 02:11:37 +0000868 struct mirror *m = get_default_mirror(ms);
869
870 do {
871 if (likely(!atomic_read(&m->error_count)))
872 return m;
873
874 if (m-- == ms->mirror)
875 m += ms->nr_mirrors;
876 } while (m != get_default_mirror(ms));
877
878 return NULL;
879}
880
881static int default_ok(struct mirror *m)
882{
883 struct mirror *default_mirror = get_default_mirror(m->ms);
884
885 return !atomic_read(&default_mirror->error_count);
886}
887
888static int mirror_available(struct mirror_set *ms, struct bio *bio)
889{
890 region_t region = bio_to_region(&ms->rh, bio);
891
892 if (ms->rh.log->type->in_sync(ms->rh.log, region, 0))
893 return choose_mirror(ms, bio->bi_sector) ? 1 : 0;
894
895 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896}
897
898/*
899 * remap a buffer to a particular mirror.
900 */
Jonathan Brassow06386bb2008-02-08 02:11:37 +0000901static sector_t map_sector(struct mirror *m, struct bio *bio)
902{
903 return m->offset + (bio->bi_sector - m->ms->ti->begin);
904}
905
906static void map_bio(struct mirror *m, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700907{
908 bio->bi_bdev = m->dev->bdev;
Jonathan Brassow06386bb2008-02-08 02:11:37 +0000909 bio->bi_sector = map_sector(m, bio);
910}
911
912static void map_region(struct io_region *io, struct mirror *m,
913 struct bio *bio)
914{
915 io->bdev = m->dev->bdev;
916 io->sector = map_sector(m, bio);
917 io->count = bio->bi_size >> 9;
918}
919
920/*-----------------------------------------------------------------
921 * Reads
922 *---------------------------------------------------------------*/
923static void read_callback(unsigned long error, void *context)
924{
925 struct bio *bio = context;
926 struct mirror *m;
927
928 m = bio_get_m(bio);
929 bio_set_m(bio, NULL);
930
931 if (likely(!error)) {
932 bio_endio(bio, 0);
933 return;
934 }
935
936 fail_mirror(m, DM_RAID1_READ_ERROR);
937
938 if (likely(default_ok(m)) || mirror_available(m->ms, bio)) {
939 DMWARN_LIMIT("Read failure on mirror device %s. "
940 "Trying alternative device.",
941 m->dev->name);
942 queue_bio(m->ms, bio, bio_rw(bio));
943 return;
944 }
945
946 DMERR_LIMIT("Read failure on mirror device %s. Failing I/O.",
947 m->dev->name);
948 bio_endio(bio, -EIO);
949}
950
951/* Asynchronous read. */
952static void read_async_bio(struct mirror *m, struct bio *bio)
953{
954 struct io_region io;
955 struct dm_io_request io_req = {
956 .bi_rw = READ,
957 .mem.type = DM_IO_BVEC,
958 .mem.ptr.bvec = bio->bi_io_vec + bio->bi_idx,
959 .notify.fn = read_callback,
960 .notify.context = bio,
961 .client = m->ms->io_client,
962 };
963
964 map_region(&io, m, bio);
965 bio_set_m(bio, m);
966 (void) dm_io(&io_req, 1, &io, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967}
968
969static void do_reads(struct mirror_set *ms, struct bio_list *reads)
970{
971 region_t region;
972 struct bio *bio;
973 struct mirror *m;
974
975 while ((bio = bio_list_pop(reads))) {
976 region = bio_to_region(&ms->rh, bio);
Jonathan Brassow06386bb2008-02-08 02:11:37 +0000977 m = get_default_mirror(ms);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700978
979 /*
980 * We can only read balance if the region is in sync.
981 */
Jonathan Brassow06386bb2008-02-08 02:11:37 +0000982 if (likely(rh_in_sync(&ms->rh, region, 1)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983 m = choose_mirror(ms, bio->bi_sector);
Jonathan Brassow06386bb2008-02-08 02:11:37 +0000984 else if (m && atomic_read(&m->error_count))
985 m = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986
Jonathan Brassow06386bb2008-02-08 02:11:37 +0000987 if (likely(m))
988 read_async_bio(m, bio);
989 else
990 bio_endio(bio, -EIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991 }
992}
993
994/*-----------------------------------------------------------------
995 * Writes.
996 *
997 * We do different things with the write io depending on the
998 * state of the region that it's in:
999 *
1000 * SYNC: increment pending, use kcopyd to write to *all* mirrors
1001 * RECOVERING: delay the io until recovery completes
1002 * NOSYNC: increment pending, just write to the default mirror
1003 *---------------------------------------------------------------*/
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001004
1005/* __bio_mark_nosync
1006 * @ms
1007 * @bio
1008 * @done
1009 * @error
1010 *
1011 * The bio was written on some mirror(s) but failed on other mirror(s).
1012 * We can successfully endio the bio but should avoid the region being
1013 * marked clean by setting the state RH_NOSYNC.
1014 *
1015 * This function is _not_ safe in interrupt context!
1016 */
1017static void __bio_mark_nosync(struct mirror_set *ms,
1018 struct bio *bio, unsigned done, int error)
1019{
1020 unsigned long flags;
1021 struct region_hash *rh = &ms->rh;
1022 struct dirty_log *log = ms->rh.log;
1023 struct region *reg;
1024 region_t region = bio_to_region(rh, bio);
1025 int recovering = 0;
1026
1027 /* We must inform the log that the sync count has changed. */
1028 log->type->set_region_sync(log, region, 0);
1029 ms->in_sync = 0;
1030
1031 read_lock(&rh->hash_lock);
1032 reg = __rh_find(rh, region);
1033 read_unlock(&rh->hash_lock);
1034
1035 /* region hash entry should exist because write was in-flight */
1036 BUG_ON(!reg);
1037 BUG_ON(!list_empty(&reg->list));
1038
1039 spin_lock_irqsave(&rh->region_lock, flags);
1040 /*
1041 * Possible cases:
1042 * 1) RH_DIRTY
1043 * 2) RH_NOSYNC: was dirty, other preceeding writes failed
1044 * 3) RH_RECOVERING: flushing pending writes
1045 * Either case, the region should have not been connected to list.
1046 */
1047 recovering = (reg->state == RH_RECOVERING);
1048 reg->state = RH_NOSYNC;
1049 BUG_ON(!list_empty(&reg->list));
1050 spin_unlock_irqrestore(&rh->region_lock, flags);
1051
1052 bio_endio(bio, error);
1053 if (recovering)
1054 complete_resync_work(reg, 0);
1055}
1056
Linus Torvalds1da177e2005-04-16 15:20:36 -07001057static void write_callback(unsigned long error, void *context)
1058{
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001059 unsigned i, ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001060 struct bio *bio = (struct bio *) context;
1061 struct mirror_set *ms;
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001062 int uptodate = 0;
1063 int should_wake = 0;
1064 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001065
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001066 ms = bio_get_m(bio)->ms;
1067 bio_set_m(bio, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068
1069 /*
1070 * NOTE: We don't decrement the pending count here,
1071 * instead it is done by the targets endio function.
1072 * This way we handle both writes to SYNC and NOSYNC
1073 * regions with the same code.
1074 */
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001075 if (likely(!error))
1076 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001077
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001078 for (i = 0; i < ms->nr_mirrors; i++)
1079 if (test_bit(i, &error))
1080 fail_mirror(ms->mirror + i, DM_RAID1_WRITE_ERROR);
1081 else
1082 uptodate = 1;
1083
1084 if (unlikely(!uptodate)) {
1085 DMERR("All replicated volumes dead, failing I/O");
1086 /* None of the writes succeeded, fail the I/O. */
1087 ret = -EIO;
1088 } else if (errors_handled(ms)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089 /*
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001090 * Need to raise event. Since raising
1091 * events can block, we need to do it in
1092 * the main thread.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093 */
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001094 spin_lock_irqsave(&ms->lock, flags);
1095 if (!ms->failures.head)
1096 should_wake = 1;
1097 bio_list_add(&ms->failures, bio);
1098 spin_unlock_irqrestore(&ms->lock, flags);
1099 if (should_wake)
1100 wake(ms);
1101 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102 }
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001103out:
1104 bio_endio(bio, ret);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001105}
1106
1107static void do_write(struct mirror_set *ms, struct bio *bio)
1108{
1109 unsigned int i;
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001110 struct io_region io[ms->nr_mirrors], *dest = io;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111 struct mirror *m;
Milan Broz88be1632007-05-09 02:33:04 -07001112 struct dm_io_request io_req = {
1113 .bi_rw = WRITE,
1114 .mem.type = DM_IO_BVEC,
1115 .mem.ptr.bvec = bio->bi_io_vec + bio->bi_idx,
1116 .notify.fn = write_callback,
1117 .notify.context = bio,
1118 .client = ms->io_client,
1119 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001121 for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++)
1122 map_region(dest++, m, bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001123
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001124 /*
1125 * Use default mirror because we only need it to retrieve the reference
1126 * to the mirror set in write_callback().
1127 */
1128 bio_set_m(bio, get_default_mirror(ms));
Milan Broz88be1632007-05-09 02:33:04 -07001129
1130 (void) dm_io(&io_req, ms->nr_mirrors, io, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131}
1132
1133static void do_writes(struct mirror_set *ms, struct bio_list *writes)
1134{
1135 int state;
1136 struct bio *bio;
1137 struct bio_list sync, nosync, recover, *this_list = NULL;
1138
1139 if (!writes->head)
1140 return;
1141
1142 /*
1143 * Classify each write.
1144 */
1145 bio_list_init(&sync);
1146 bio_list_init(&nosync);
1147 bio_list_init(&recover);
1148
1149 while ((bio = bio_list_pop(writes))) {
1150 state = rh_state(&ms->rh, bio_to_region(&ms->rh, bio), 1);
1151 switch (state) {
1152 case RH_CLEAN:
1153 case RH_DIRTY:
1154 this_list = &sync;
1155 break;
1156
1157 case RH_NOSYNC:
1158 this_list = &nosync;
1159 break;
1160
1161 case RH_RECOVERING:
1162 this_list = &recover;
1163 break;
1164 }
1165
1166 bio_list_add(this_list, bio);
1167 }
1168
1169 /*
1170 * Increment the pending counts for any regions that will
1171 * be written to (writes to recover regions are going to
1172 * be delayed).
1173 */
1174 rh_inc_pending(&ms->rh, &sync);
1175 rh_inc_pending(&ms->rh, &nosync);
Jonathan Brassowfc1ff952007-07-12 17:29:15 +01001176 ms->log_failure = rh_flush(&ms->rh) ? 1 : 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177
1178 /*
1179 * Dispatch io.
1180 */
Jonathan Brassowb80aa7a2008-02-08 02:11:35 +00001181 if (unlikely(ms->log_failure)) {
1182 spin_lock_irq(&ms->lock);
1183 bio_list_merge(&ms->failures, &sync);
1184 spin_unlock_irq(&ms->lock);
1185 } else
Jonathan Brassowfc1ff952007-07-12 17:29:15 +01001186 while ((bio = bio_list_pop(&sync)))
Jonathan Brassowb80aa7a2008-02-08 02:11:35 +00001187 do_write(ms, bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188
1189 while ((bio = bio_list_pop(&recover)))
1190 rh_delay(&ms->rh, bio);
1191
1192 while ((bio = bio_list_pop(&nosync))) {
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001193 map_bio(get_default_mirror(ms), bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001194 generic_make_request(bio);
1195 }
1196}
1197
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001198static void do_failures(struct mirror_set *ms, struct bio_list *failures)
1199{
1200 struct bio *bio;
1201
1202 if (!failures->head)
1203 return;
1204
Jonathan Brassowb80aa7a2008-02-08 02:11:35 +00001205 if (!ms->log_failure) {
1206 while ((bio = bio_list_pop(failures)))
1207 __bio_mark_nosync(ms, bio, bio->bi_size, 0);
1208 return;
1209 }
1210
1211 /*
1212 * If the log has failed, unattempted writes are being
1213 * put on the failures list. We can't issue those writes
1214 * until a log has been marked, so we must store them.
1215 *
1216 * If a 'noflush' suspend is in progress, we can requeue
1217 * the I/O's to the core. This give userspace a chance
1218 * to reconfigure the mirror, at which point the core
1219 * will reissue the writes. If the 'noflush' flag is
1220 * not set, we have no choice but to return errors.
1221 *
1222 * Some writes on the failures list may have been
1223 * submitted before the log failure and represent a
1224 * failure to write to one of the devices. It is ok
1225 * for us to treat them the same and requeue them
1226 * as well.
1227 */
1228 if (dm_noflush_suspending(ms->ti)) {
1229 while ((bio = bio_list_pop(failures)))
1230 bio_endio(bio, DM_ENDIO_REQUEUE);
1231 return;
1232 }
1233
1234 if (atomic_read(&ms->suspend)) {
1235 while ((bio = bio_list_pop(failures)))
1236 bio_endio(bio, -EIO);
1237 return;
1238 }
1239
1240 spin_lock_irq(&ms->lock);
1241 bio_list_merge(&ms->failures, failures);
1242 spin_unlock_irq(&ms->lock);
1243
1244 wake(ms);
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001245}
1246
1247static void trigger_event(struct work_struct *work)
1248{
1249 struct mirror_set *ms =
1250 container_of(work, struct mirror_set, trigger_event);
1251
1252 dm_table_event(ms->ti->table);
1253}
1254
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255/*-----------------------------------------------------------------
1256 * kmirrord
1257 *---------------------------------------------------------------*/
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001258static int _do_mirror(struct work_struct *work)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259{
Holger Smolinski6ad36fe2007-05-09 02:32:50 -07001260 struct mirror_set *ms =container_of(work, struct mirror_set,
1261 kmirrord_work);
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001262 struct bio_list reads, writes, failures;
1263 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001265 spin_lock_irqsave(&ms->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266 reads = ms->reads;
1267 writes = ms->writes;
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001268 failures = ms->failures;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001269 bio_list_init(&ms->reads);
1270 bio_list_init(&ms->writes);
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001271 bio_list_init(&ms->failures);
1272 spin_unlock_irqrestore(&ms->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001273
1274 rh_update_states(&ms->rh);
1275 do_recovery(ms);
1276 do_reads(ms, &reads);
1277 do_writes(ms, &writes);
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001278 do_failures(ms, &failures);
1279
1280 return (ms->failures.head) ? 1 : 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001281}
1282
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001283static void do_mirror(struct work_struct *work)
1284{
1285 /*
1286 * If _do_mirror returns 1, we give it
1287 * another shot. This helps for cases like
1288 * 'suspend' where we call flush_workqueue
1289 * and expect all work to be finished. If
1290 * a failure happens during a suspend, we
1291 * couldn't issue a 'wake' because it would
1292 * not be honored. Therefore, we return '1'
1293 * from _do_mirror, and retry here.
1294 */
1295 while (_do_mirror(work))
1296 schedule();
1297}
1298
1299
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300/*-----------------------------------------------------------------
1301 * Target functions
1302 *---------------------------------------------------------------*/
1303static struct mirror_set *alloc_context(unsigned int nr_mirrors,
1304 uint32_t region_size,
1305 struct dm_target *ti,
1306 struct dirty_log *dl)
1307{
1308 size_t len;
1309 struct mirror_set *ms = NULL;
1310
1311 if (array_too_big(sizeof(*ms), sizeof(ms->mirror[0]), nr_mirrors))
1312 return NULL;
1313
1314 len = sizeof(*ms) + (sizeof(ms->mirror[0]) * nr_mirrors);
1315
Yoann Padioleaudd00cc42007-07-19 01:49:03 -07001316 ms = kzalloc(len, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001317 if (!ms) {
Alasdair G Kergon72d94862006-06-26 00:27:35 -07001318 ti->error = "Cannot allocate mirror context";
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319 return NULL;
1320 }
1321
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322 spin_lock_init(&ms->lock);
1323
1324 ms->ti = ti;
1325 ms->nr_mirrors = nr_mirrors;
1326 ms->nr_regions = dm_sector_div_up(ti->len, region_size);
1327 ms->in_sync = 0;
Jonathan Brassowb80aa7a2008-02-08 02:11:35 +00001328 ms->log_failure = 0;
1329 atomic_set(&ms->suspend, 0);
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001330 atomic_set(&ms->default_mirror, DEFAULT_MIRROR);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001331
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001332 len = sizeof(struct dm_raid1_read_record);
1333 ms->read_record_pool = mempool_create_kmalloc_pool(MIN_READ_RECORDS,
1334 len);
1335 if (!ms->read_record_pool) {
1336 ti->error = "Error creating mirror read_record_pool";
1337 kfree(ms);
1338 return NULL;
1339 }
1340
Milan Broz88be1632007-05-09 02:33:04 -07001341 ms->io_client = dm_io_client_create(DM_IO_PAGES);
1342 if (IS_ERR(ms->io_client)) {
1343 ti->error = "Error creating dm_io client";
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001344 mempool_destroy(ms->read_record_pool);
Milan Broz88be1632007-05-09 02:33:04 -07001345 kfree(ms);
1346 return NULL;
1347 }
1348
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349 if (rh_init(&ms->rh, ms, dl, region_size, ms->nr_regions)) {
Alasdair G Kergon72d94862006-06-26 00:27:35 -07001350 ti->error = "Error creating dirty region hash";
Dmitry Monakhova72cf732007-10-19 22:38:39 +01001351 dm_io_client_destroy(ms->io_client);
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001352 mempool_destroy(ms->read_record_pool);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353 kfree(ms);
1354 return NULL;
1355 }
1356
1357 return ms;
1358}
1359
1360static void free_context(struct mirror_set *ms, struct dm_target *ti,
1361 unsigned int m)
1362{
1363 while (m--)
1364 dm_put_device(ti, ms->mirror[m].dev);
1365
Milan Broz88be1632007-05-09 02:33:04 -07001366 dm_io_client_destroy(ms->io_client);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367 rh_exit(&ms->rh);
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001368 mempool_destroy(ms->read_record_pool);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369 kfree(ms);
1370}
1371
1372static inline int _check_region_size(struct dm_target *ti, uint32_t size)
1373{
vignesh babu6f3c3f02007-10-19 22:38:44 +01001374 return !(size % (PAGE_SIZE >> 9) || !is_power_of_2(size) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375 size > ti->len);
1376}
1377
1378static int get_mirror(struct mirror_set *ms, struct dm_target *ti,
1379 unsigned int mirror, char **argv)
1380{
Andrew Morton4ee218c2006-03-27 01:17:48 -08001381 unsigned long long offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382
Andrew Morton4ee218c2006-03-27 01:17:48 -08001383 if (sscanf(argv[1], "%llu", &offset) != 1) {
Alasdair G Kergon72d94862006-06-26 00:27:35 -07001384 ti->error = "Invalid offset";
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385 return -EINVAL;
1386 }
1387
1388 if (dm_get_device(ti, argv[0], offset, ti->len,
1389 dm_table_get_mode(ti->table),
1390 &ms->mirror[mirror].dev)) {
Alasdair G Kergon72d94862006-06-26 00:27:35 -07001391 ti->error = "Device lookup failure";
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392 return -ENXIO;
1393 }
1394
Jonathan Brassowaa5617c2007-10-19 22:47:58 +01001395 ms->mirror[mirror].ms = ms;
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001396 atomic_set(&(ms->mirror[mirror].error_count), 0);
1397 ms->mirror[mirror].error_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001398 ms->mirror[mirror].offset = offset;
1399
1400 return 0;
1401}
1402
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403/*
1404 * Create dirty log: log_type #log_params <log_params>
1405 */
1406static struct dirty_log *create_dirty_log(struct dm_target *ti,
1407 unsigned int argc, char **argv,
1408 unsigned int *args_used)
1409{
1410 unsigned int param_count;
1411 struct dirty_log *dl;
1412
1413 if (argc < 2) {
Alasdair G Kergon72d94862006-06-26 00:27:35 -07001414 ti->error = "Insufficient mirror log arguments";
Linus Torvalds1da177e2005-04-16 15:20:36 -07001415 return NULL;
1416 }
1417
1418 if (sscanf(argv[1], "%u", &param_count) != 1) {
Alasdair G Kergon72d94862006-06-26 00:27:35 -07001419 ti->error = "Invalid mirror log argument count";
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420 return NULL;
1421 }
1422
1423 *args_used = 2 + param_count;
1424
1425 if (argc < *args_used) {
Alasdair G Kergon72d94862006-06-26 00:27:35 -07001426 ti->error = "Insufficient mirror log arguments";
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427 return NULL;
1428 }
1429
1430 dl = dm_create_dirty_log(argv[0], ti, param_count, argv + 2);
1431 if (!dl) {
Alasdair G Kergon72d94862006-06-26 00:27:35 -07001432 ti->error = "Error creating mirror dirty log";
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433 return NULL;
1434 }
1435
1436 if (!_check_region_size(ti, dl->type->get_region_size(dl))) {
Alasdair G Kergon72d94862006-06-26 00:27:35 -07001437 ti->error = "Invalid region size";
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438 dm_destroy_dirty_log(dl);
1439 return NULL;
1440 }
1441
1442 return dl;
1443}
1444
Jonathan E Brassowa8e6afa2007-05-09 02:32:59 -07001445static int parse_features(struct mirror_set *ms, unsigned argc, char **argv,
1446 unsigned *args_used)
1447{
1448 unsigned num_features;
1449 struct dm_target *ti = ms->ti;
1450
1451 *args_used = 0;
1452
1453 if (!argc)
1454 return 0;
1455
1456 if (sscanf(argv[0], "%u", &num_features) != 1) {
1457 ti->error = "Invalid number of features";
1458 return -EINVAL;
1459 }
1460
1461 argc--;
1462 argv++;
1463 (*args_used)++;
1464
1465 if (num_features > argc) {
1466 ti->error = "Not enough arguments to support feature count";
1467 return -EINVAL;
1468 }
1469
1470 if (!strcmp("handle_errors", argv[0]))
1471 ms->features |= DM_RAID1_HANDLE_ERRORS;
1472 else {
1473 ti->error = "Unrecognised feature requested";
1474 return -EINVAL;
1475 }
1476
1477 (*args_used)++;
1478
1479 return 0;
1480}
1481
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482/*
1483 * Construct a mirror mapping:
1484 *
1485 * log_type #log_params <log_params>
1486 * #mirrors [mirror_path offset]{2,}
Jonathan E Brassowa8e6afa2007-05-09 02:32:59 -07001487 * [#features <features>]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001488 *
1489 * log_type is "core" or "disk"
1490 * #log_params is between 1 and 3
Jonathan E Brassowa8e6afa2007-05-09 02:32:59 -07001491 *
1492 * If present, features must be "handle_errors".
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1495{
1496 int r;
1497 unsigned int nr_mirrors, m, args_used;
1498 struct mirror_set *ms;
1499 struct dirty_log *dl;
1500
1501 dl = create_dirty_log(ti, argc, argv, &args_used);
1502 if (!dl)
1503 return -EINVAL;
1504
1505 argv += args_used;
1506 argc -= args_used;
1507
1508 if (!argc || sscanf(argv[0], "%u", &nr_mirrors) != 1 ||
1509 nr_mirrors < 2 || nr_mirrors > KCOPYD_MAX_REGIONS + 1) {
Alasdair G Kergon72d94862006-06-26 00:27:35 -07001510 ti->error = "Invalid number of mirrors";
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511 dm_destroy_dirty_log(dl);
1512 return -EINVAL;
1513 }
1514
1515 argv++, argc--;
1516
Jonathan E Brassowa8e6afa2007-05-09 02:32:59 -07001517 if (argc < nr_mirrors * 2) {
1518 ti->error = "Too few mirror arguments";
Linus Torvalds1da177e2005-04-16 15:20:36 -07001519 dm_destroy_dirty_log(dl);
1520 return -EINVAL;
1521 }
1522
1523 ms = alloc_context(nr_mirrors, dl->type->get_region_size(dl), ti, dl);
1524 if (!ms) {
1525 dm_destroy_dirty_log(dl);
1526 return -ENOMEM;
1527 }
1528
1529 /* Get the mirror parameter sets */
1530 for (m = 0; m < nr_mirrors; m++) {
1531 r = get_mirror(ms, ti, m, argv);
1532 if (r) {
1533 free_context(ms, ti, m);
1534 return r;
1535 }
1536 argv += 2;
1537 argc -= 2;
1538 }
1539
1540 ti->private = ms;
Alasdair G Kergond88854f2005-07-07 17:59:34 -07001541 ti->split_io = ms->rh.region_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542
Holger Smolinski6ad36fe2007-05-09 02:32:50 -07001543 ms->kmirrord_wq = create_singlethread_workqueue("kmirrord");
1544 if (!ms->kmirrord_wq) {
1545 DMERR("couldn't start kmirrord");
Dmitry Monakhova72cf732007-10-19 22:38:39 +01001546 r = -ENOMEM;
1547 goto err_free_context;
Holger Smolinski6ad36fe2007-05-09 02:32:50 -07001548 }
1549 INIT_WORK(&ms->kmirrord_work, do_mirror);
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001550 INIT_WORK(&ms->trigger_event, trigger_event);
Holger Smolinski6ad36fe2007-05-09 02:32:50 -07001551
Jonathan E Brassowa8e6afa2007-05-09 02:32:59 -07001552 r = parse_features(ms, argc, argv, &args_used);
Dmitry Monakhova72cf732007-10-19 22:38:39 +01001553 if (r)
1554 goto err_destroy_wq;
Jonathan E Brassowa8e6afa2007-05-09 02:32:59 -07001555
1556 argv += args_used;
1557 argc -= args_used;
1558
Jonathan Brassowf44db672007-07-12 17:29:04 +01001559 /*
1560 * Any read-balancing addition depends on the
1561 * DM_RAID1_HANDLE_ERRORS flag being present.
1562 * This is because the decision to balance depends
1563 * on the sync state of a region. If the above
1564 * flag is not present, we ignore errors; and
1565 * the sync state may be inaccurate.
1566 */
1567
Jonathan E Brassowa8e6afa2007-05-09 02:32:59 -07001568 if (argc) {
1569 ti->error = "Too many mirror arguments";
Dmitry Monakhova72cf732007-10-19 22:38:39 +01001570 r = -EINVAL;
1571 goto err_destroy_wq;
Jonathan E Brassowa8e6afa2007-05-09 02:32:59 -07001572 }
1573
Linus Torvalds1da177e2005-04-16 15:20:36 -07001574 r = kcopyd_client_create(DM_IO_PAGES, &ms->kcopyd_client);
Dmitry Monakhova72cf732007-10-19 22:38:39 +01001575 if (r)
1576 goto err_destroy_wq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001577
Holger Smolinski6ad36fe2007-05-09 02:32:50 -07001578 wake(ms);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001579 return 0;
Dmitry Monakhova72cf732007-10-19 22:38:39 +01001580
1581err_destroy_wq:
1582 destroy_workqueue(ms->kmirrord_wq);
1583err_free_context:
1584 free_context(ms, ti, ms->nr_mirrors);
1585 return r;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001586}
1587
1588static void mirror_dtr(struct dm_target *ti)
1589{
1590 struct mirror_set *ms = (struct mirror_set *) ti->private;
1591
Holger Smolinski6ad36fe2007-05-09 02:32:50 -07001592 flush_workqueue(ms->kmirrord_wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593 kcopyd_client_destroy(ms->kcopyd_client);
Holger Smolinski6ad36fe2007-05-09 02:32:50 -07001594 destroy_workqueue(ms->kmirrord_wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001595 free_context(ms, ti, ms->nr_mirrors);
1596}
1597
1598static void queue_bio(struct mirror_set *ms, struct bio *bio, int rw)
1599{
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001600 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601 int should_wake = 0;
1602 struct bio_list *bl;
1603
1604 bl = (rw == WRITE) ? &ms->writes : &ms->reads;
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001605 spin_lock_irqsave(&ms->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001606 should_wake = !(bl->head);
1607 bio_list_add(bl, bio);
Jonathan Brassow72f4b312008-02-08 02:11:29 +00001608 spin_unlock_irqrestore(&ms->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001609
1610 if (should_wake)
Holger Smolinski6ad36fe2007-05-09 02:32:50 -07001611 wake(ms);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001612}
1613
1614/*
1615 * Mirror mapping function
1616 */
1617static int mirror_map(struct dm_target *ti, struct bio *bio,
1618 union map_info *map_context)
1619{
1620 int r, rw = bio_rw(bio);
1621 struct mirror *m;
1622 struct mirror_set *ms = ti->private;
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001623 struct dm_raid1_read_record *read_record = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624
1625 if (rw == WRITE) {
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001626 /* Save region for mirror_end_io() handler */
1627 map_context->ll = bio_to_region(&ms->rh, bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001628 queue_bio(ms, bio, rw);
Kiyoshi Uedad2a7ad22006-12-08 02:41:06 -08001629 return DM_MAPIO_SUBMITTED;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630 }
1631
1632 r = ms->rh.log->type->in_sync(ms->rh.log,
1633 bio_to_region(&ms->rh, bio), 0);
1634 if (r < 0 && r != -EWOULDBLOCK)
1635 return r;
1636
Linus Torvalds1da177e2005-04-16 15:20:36 -07001637 /*
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001638 * If region is not in-sync queue the bio.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001639 */
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001640 if (!r || (r == -EWOULDBLOCK)) {
1641 if (rw == READA)
1642 return -EWOULDBLOCK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644 queue_bio(ms, bio, rw);
Kiyoshi Uedad2a7ad22006-12-08 02:41:06 -08001645 return DM_MAPIO_SUBMITTED;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646 }
1647
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001648 /*
1649 * The region is in-sync and we can perform reads directly.
1650 * Store enough information so we can retry if it fails.
1651 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001652 m = choose_mirror(ms, bio->bi_sector);
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001653 if (unlikely(!m))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654 return -EIO;
1655
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001656 read_record = mempool_alloc(ms->read_record_pool, GFP_NOIO);
1657 if (likely(read_record)) {
1658 dm_bio_record(&read_record->details, bio);
1659 map_context->ptr = read_record;
1660 read_record->m = m;
1661 }
1662
1663 map_bio(m, bio);
1664
Kiyoshi Uedad2a7ad22006-12-08 02:41:06 -08001665 return DM_MAPIO_REMAPPED;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666}
1667
1668static int mirror_end_io(struct dm_target *ti, struct bio *bio,
1669 int error, union map_info *map_context)
1670{
1671 int rw = bio_rw(bio);
1672 struct mirror_set *ms = (struct mirror_set *) ti->private;
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001673 struct mirror *m = NULL;
1674 struct dm_bio_details *bd = NULL;
1675 struct dm_raid1_read_record *read_record = map_context->ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001676
1677 /*
1678 * We need to dec pending if this was a write.
1679 */
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001680 if (rw == WRITE) {
1681 rh_dec(&ms->rh, map_context->ll);
1682 return error;
1683 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001685 if (error == -EOPNOTSUPP)
1686 goto out;
1687
1688 if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1689 goto out;
1690
1691 if (unlikely(error)) {
1692 if (!read_record) {
1693 /*
1694 * There wasn't enough memory to record necessary
1695 * information for a retry or there was no other
1696 * mirror in-sync.
1697 */
Adrian Bunke03f1a82008-02-19 19:44:19 +00001698 DMERR_LIMIT("Mirror read failed.");
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001699 return -EIO;
1700 }
Adrian Bunke03f1a82008-02-19 19:44:19 +00001701
1702 m = read_record->m;
1703
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001704 DMERR("Mirror read failed from %s. Trying alternative device.",
1705 m->dev->name);
1706
Jonathan Brassow06386bb2008-02-08 02:11:37 +00001707 fail_mirror(m, DM_RAID1_READ_ERROR);
1708
1709 /*
1710 * A failed read is requeued for another attempt using an intact
1711 * mirror.
1712 */
1713 if (default_ok(m) || mirror_available(ms, bio)) {
1714 bd = &read_record->details;
1715
1716 dm_bio_restore(bd, bio);
1717 mempool_free(read_record, ms->read_record_pool);
1718 map_context->ptr = NULL;
1719 queue_bio(ms, bio, rw);
1720 return 1;
1721 }
1722 DMERR("All replicated volumes dead, failing I/O");
1723 }
1724
1725out:
1726 if (read_record) {
1727 mempool_free(read_record, ms->read_record_pool);
1728 map_context->ptr = NULL;
1729 }
1730
1731 return error;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732}
1733
Jonathan Brassowb80aa7a2008-02-08 02:11:35 +00001734static void mirror_presuspend(struct dm_target *ti)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001735{
1736 struct mirror_set *ms = (struct mirror_set *) ti->private;
1737 struct dirty_log *log = ms->rh.log;
1738
Jonathan Brassowb80aa7a2008-02-08 02:11:35 +00001739 atomic_set(&ms->suspend, 1);
1740
1741 /*
1742 * We must finish up all the work that we've
1743 * generated (i.e. recovery work).
1744 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745 rh_stop_recovery(&ms->rh);
Jonathan E Brassow33184042006-11-08 17:44:44 -08001746
Jonathan E Brassow33184042006-11-08 17:44:44 -08001747 wait_event(_kmirrord_recovery_stopped,
1748 !atomic_read(&ms->rh.recovery_in_flight));
1749
Jonathan Brassowb80aa7a2008-02-08 02:11:35 +00001750 if (log->type->presuspend && log->type->presuspend(log))
1751 /* FIXME: need better error handling */
1752 DMWARN("log presuspend failed");
1753
1754 /*
1755 * Now that recovery is complete/stopped and the
1756 * delayed bios are queued, we need to wait for
1757 * the worker thread to complete. This way,
1758 * we know that all of our I/O has been pushed.
1759 */
1760 flush_workqueue(ms->kmirrord_wq);
1761}
1762
1763static void mirror_postsuspend(struct dm_target *ti)
1764{
1765 struct mirror_set *ms = ti->private;
1766 struct dirty_log *log = ms->rh.log;
1767
Jonathan Brassow6b3df0d2007-10-19 22:47:57 +01001768 if (log->type->postsuspend && log->type->postsuspend(log))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001769 /* FIXME: need better error handling */
Jonathan Brassowb80aa7a2008-02-08 02:11:35 +00001770 DMWARN("log postsuspend failed");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001771}
1772
1773static void mirror_resume(struct dm_target *ti)
1774{
Jonathan Brassowb80aa7a2008-02-08 02:11:35 +00001775 struct mirror_set *ms = ti->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001776 struct dirty_log *log = ms->rh.log;
Jonathan Brassowb80aa7a2008-02-08 02:11:35 +00001777
1778 atomic_set(&ms->suspend, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001779 if (log->type->resume && log->type->resume(log))
1780 /* FIXME: need better error handling */
1781 DMWARN("log resume failed");
1782 rh_start_recovery(&ms->rh);
1783}
1784
Jonathan Brassowaf195ac2008-02-08 02:11:39 +00001785/*
1786 * device_status_char
1787 * @m: mirror device/leg we want the status of
1788 *
1789 * We return one character representing the most severe error
1790 * we have encountered.
1791 * A => Alive - No failures
1792 * D => Dead - A write failure occurred leaving mirror out-of-sync
1793 * S => Sync - A sychronization failure occurred, mirror out-of-sync
1794 * R => Read - A read failure occurred, mirror data unaffected
1795 *
1796 * Returns: <char>
1797 */
1798static char device_status_char(struct mirror *m)
1799{
1800 if (!atomic_read(&(m->error_count)))
1801 return 'A';
1802
1803 return (test_bit(DM_RAID1_WRITE_ERROR, &(m->error_type))) ? 'D' :
1804 (test_bit(DM_RAID1_SYNC_ERROR, &(m->error_type))) ? 'S' :
1805 (test_bit(DM_RAID1_READ_ERROR, &(m->error_type))) ? 'R' : 'U';
1806}
1807
1808
Linus Torvalds1da177e2005-04-16 15:20:36 -07001809static int mirror_status(struct dm_target *ti, status_type_t type,
1810 char *result, unsigned int maxlen)
1811{
Jonathan E Brassow315dcc22007-05-09 02:32:58 -07001812 unsigned int m, sz = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813 struct mirror_set *ms = (struct mirror_set *) ti->private;
Jonathan Brassowaf195ac2008-02-08 02:11:39 +00001814 struct dirty_log *log = ms->rh.log;
1815 char buffer[ms->nr_mirrors + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001816
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817 switch (type) {
1818 case STATUSTYPE_INFO:
1819 DMEMIT("%d ", ms->nr_mirrors);
Jonathan Brassowaf195ac2008-02-08 02:11:39 +00001820 for (m = 0; m < ms->nr_mirrors; m++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001821 DMEMIT("%s ", ms->mirror[m].dev->name);
Jonathan Brassowaf195ac2008-02-08 02:11:39 +00001822 buffer[m] = device_status_char(&(ms->mirror[m]));
1823 }
1824 buffer[m] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825
Jonathan Brassowaf195ac2008-02-08 02:11:39 +00001826 DMEMIT("%llu/%llu 1 %s ",
1827 (unsigned long long)log->type->get_sync_count(ms->rh.log),
1828 (unsigned long long)ms->nr_regions, buffer);
Jonathan E Brassow315dcc22007-05-09 02:32:58 -07001829
Jonathan Brassowaf195ac2008-02-08 02:11:39 +00001830 sz += log->type->status(ms->rh.log, type, result+sz, maxlen-sz);
Jonathan E Brassow315dcc22007-05-09 02:32:58 -07001831
Linus Torvalds1da177e2005-04-16 15:20:36 -07001832 break;
1833
1834 case STATUSTYPE_TABLE:
Jonathan Brassowaf195ac2008-02-08 02:11:39 +00001835 sz = log->type->status(ms->rh.log, type, result, maxlen);
Jonathan E Brassow315dcc22007-05-09 02:32:58 -07001836
Jonathan Brassowe52b8f62006-10-03 01:15:32 -07001837 DMEMIT("%d", ms->nr_mirrors);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838 for (m = 0; m < ms->nr_mirrors; m++)
Jonathan Brassowe52b8f62006-10-03 01:15:32 -07001839 DMEMIT(" %s %llu", ms->mirror[m].dev->name,
Jonathan Brassowb80aa7a2008-02-08 02:11:35 +00001840 (unsigned long long)ms->mirror[m].offset);
Jonathan E Brassowa8e6afa2007-05-09 02:32:59 -07001841
1842 if (ms->features & DM_RAID1_HANDLE_ERRORS)
1843 DMEMIT(" 1 handle_errors");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001844 }
1845
1846 return 0;
1847}
1848
1849static struct target_type mirror_target = {
1850 .name = "mirror",
Jonathan Brassowaf195ac2008-02-08 02:11:39 +00001851 .version = {1, 0, 20},
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 .module = THIS_MODULE,
1853 .ctr = mirror_ctr,
1854 .dtr = mirror_dtr,
1855 .map = mirror_map,
1856 .end_io = mirror_end_io,
Jonathan Brassowb80aa7a2008-02-08 02:11:35 +00001857 .presuspend = mirror_presuspend,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001858 .postsuspend = mirror_postsuspend,
1859 .resume = mirror_resume,
1860 .status = mirror_status,
1861};
1862
1863static int __init dm_mirror_init(void)
1864{
1865 int r;
1866
1867 r = dm_dirty_log_init();
1868 if (r)
1869 return r;
1870
Linus Torvalds1da177e2005-04-16 15:20:36 -07001871 r = dm_register_target(&mirror_target);
1872 if (r < 0) {
Alasdair G Kergon0cd33122007-07-12 17:27:01 +01001873 DMERR("Failed to register mirror target");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001874 dm_dirty_log_exit();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875 }
1876
1877 return r;
1878}
1879
1880static void __exit dm_mirror_exit(void)
1881{
1882 int r;
1883
1884 r = dm_unregister_target(&mirror_target);
1885 if (r < 0)
Alasdair G Kergon0cd33122007-07-12 17:27:01 +01001886 DMERR("unregister failed %d", r);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001887
Linus Torvalds1da177e2005-04-16 15:20:36 -07001888 dm_dirty_log_exit();
1889}
1890
1891/* Module hooks */
1892module_init(dm_mirror_init);
1893module_exit(dm_mirror_exit);
1894
1895MODULE_DESCRIPTION(DM_NAME " mirror target");
1896MODULE_AUTHOR("Joe Thornber");
1897MODULE_LICENSE("GPL");