2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
56 #include <linux/nodemask.h>
57 #include <linux/flex_array.h>
58 #include <linux/sched/signal.h>
60 #include <trace/events/block.h>
61 #include <linux/list_sort.h>
67 #include "raid5-log.h"
69 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
71 #define cpu_to_group(cpu) cpu_to_node(cpu)
72 #define ANY_GROUP NUMA_NO_NODE
74 static bool devices_handle_discard_safely
= false;
75 module_param(devices_handle_discard_safely
, bool, 0644);
76 MODULE_PARM_DESC(devices_handle_discard_safely
,
77 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
78 static struct workqueue_struct
*raid5_wq
;
80 static inline struct hlist_head
*stripe_hash(struct r5conf
*conf
, sector_t sect
)
82 int hash
= (sect
>> STRIPE_SHIFT
) & HASH_MASK
;
83 return &conf
->stripe_hashtbl
[hash
];
86 static inline int stripe_hash_locks_hash(sector_t sect
)
88 return (sect
>> STRIPE_SHIFT
) & STRIPE_HASH_LOCKS_MASK
;
91 static inline void lock_device_hash_lock(struct r5conf
*conf
, int hash
)
93 spin_lock_irq(conf
->hash_locks
+ hash
);
94 spin_lock(&conf
->device_lock
);
97 static inline void unlock_device_hash_lock(struct r5conf
*conf
, int hash
)
99 spin_unlock(&conf
->device_lock
);
100 spin_unlock_irq(conf
->hash_locks
+ hash
);
103 static inline void lock_all_device_hash_locks_irq(struct r5conf
*conf
)
106 spin_lock_irq(conf
->hash_locks
);
107 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
108 spin_lock_nest_lock(conf
->hash_locks
+ i
, conf
->hash_locks
);
109 spin_lock(&conf
->device_lock
);
112 static inline void unlock_all_device_hash_locks_irq(struct r5conf
*conf
)
115 spin_unlock(&conf
->device_lock
);
116 for (i
= NR_STRIPE_HASH_LOCKS
- 1; i
; i
--)
117 spin_unlock(conf
->hash_locks
+ i
);
118 spin_unlock_irq(conf
->hash_locks
);
121 /* Find first data disk in a raid6 stripe */
122 static inline int raid6_d0(struct stripe_head
*sh
)
125 /* ddf always start from first device */
127 /* md starts just after Q block */
128 if (sh
->qd_idx
== sh
->disks
- 1)
131 return sh
->qd_idx
+ 1;
133 static inline int raid6_next_disk(int disk
, int raid_disks
)
136 return (disk
< raid_disks
) ? disk
: 0;
139 /* When walking through the disks in a raid5, starting at raid6_d0,
140 * We need to map each disk to a 'slot', where the data disks are slot
141 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
142 * is raid_disks-1. This help does that mapping.
144 static int raid6_idx_to_slot(int idx
, struct stripe_head
*sh
,
145 int *count
, int syndrome_disks
)
151 if (idx
== sh
->pd_idx
)
152 return syndrome_disks
;
153 if (idx
== sh
->qd_idx
)
154 return syndrome_disks
+ 1;
160 static void print_raid5_conf (struct r5conf
*conf
);
162 static int stripe_operations_active(struct stripe_head
*sh
)
164 return sh
->check_state
|| sh
->reconstruct_state
||
165 test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
) ||
166 test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
169 static bool stripe_is_lowprio(struct stripe_head
*sh
)
171 return (test_bit(STRIPE_R5C_FULL_STRIPE
, &sh
->state
) ||
172 test_bit(STRIPE_R5C_PARTIAL_STRIPE
, &sh
->state
)) &&
173 !test_bit(STRIPE_R5C_CACHING
, &sh
->state
);
176 static void raid5_wakeup_stripe_thread(struct stripe_head
*sh
)
178 struct r5conf
*conf
= sh
->raid_conf
;
179 struct r5worker_group
*group
;
181 int i
, cpu
= sh
->cpu
;
183 if (!cpu_online(cpu
)) {
184 cpu
= cpumask_any(cpu_online_mask
);
188 if (list_empty(&sh
->lru
)) {
189 struct r5worker_group
*group
;
190 group
= conf
->worker_groups
+ cpu_to_group(cpu
);
191 if (stripe_is_lowprio(sh
))
192 list_add_tail(&sh
->lru
, &group
->loprio_list
);
194 list_add_tail(&sh
->lru
, &group
->handle_list
);
195 group
->stripes_cnt
++;
199 if (conf
->worker_cnt_per_group
== 0) {
200 md_wakeup_thread(conf
->mddev
->thread
);
204 group
= conf
->worker_groups
+ cpu_to_group(sh
->cpu
);
206 group
->workers
[0].working
= true;
207 /* at least one worker should run to avoid race */
208 queue_work_on(sh
->cpu
, raid5_wq
, &group
->workers
[0].work
);
210 thread_cnt
= group
->stripes_cnt
/ MAX_STRIPE_BATCH
- 1;
211 /* wakeup more workers */
212 for (i
= 1; i
< conf
->worker_cnt_per_group
&& thread_cnt
> 0; i
++) {
213 if (group
->workers
[i
].working
== false) {
214 group
->workers
[i
].working
= true;
215 queue_work_on(sh
->cpu
, raid5_wq
,
216 &group
->workers
[i
].work
);
222 static void do_release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
223 struct list_head
*temp_inactive_list
)
226 int injournal
= 0; /* number of date pages with R5_InJournal */
228 BUG_ON(!list_empty(&sh
->lru
));
229 BUG_ON(atomic_read(&conf
->active_stripes
)==0);
231 if (r5c_is_writeback(conf
->log
))
232 for (i
= sh
->disks
; i
--; )
233 if (test_bit(R5_InJournal
, &sh
->dev
[i
].flags
))
236 * In the following cases, the stripe cannot be released to cached
237 * lists. Therefore, we make the stripe write out and set
239 * 1. when quiesce in r5c write back;
240 * 2. when resync is requested fot the stripe.
242 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
) ||
243 (conf
->quiesce
&& r5c_is_writeback(conf
->log
) &&
244 !test_bit(STRIPE_HANDLE
, &sh
->state
) && injournal
!= 0)) {
245 if (test_bit(STRIPE_R5C_CACHING
, &sh
->state
))
246 r5c_make_stripe_write_out(sh
);
247 set_bit(STRIPE_HANDLE
, &sh
->state
);
250 if (test_bit(STRIPE_HANDLE
, &sh
->state
)) {
251 if (test_bit(STRIPE_DELAYED
, &sh
->state
) &&
252 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
253 list_add_tail(&sh
->lru
, &conf
->delayed_list
);
254 else if (test_bit(STRIPE_BIT_DELAY
, &sh
->state
) &&
255 sh
->bm_seq
- conf
->seq_write
> 0)
256 list_add_tail(&sh
->lru
, &conf
->bitmap_list
);
258 clear_bit(STRIPE_DELAYED
, &sh
->state
);
259 clear_bit(STRIPE_BIT_DELAY
, &sh
->state
);
260 if (conf
->worker_cnt_per_group
== 0) {
261 if (stripe_is_lowprio(sh
))
262 list_add_tail(&sh
->lru
,
265 list_add_tail(&sh
->lru
,
268 raid5_wakeup_stripe_thread(sh
);
272 md_wakeup_thread(conf
->mddev
->thread
);
274 BUG_ON(stripe_operations_active(sh
));
275 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
276 if (atomic_dec_return(&conf
->preread_active_stripes
)
278 md_wakeup_thread(conf
->mddev
->thread
);
279 atomic_dec(&conf
->active_stripes
);
280 if (!test_bit(STRIPE_EXPANDING
, &sh
->state
)) {
281 if (!r5c_is_writeback(conf
->log
))
282 list_add_tail(&sh
->lru
, temp_inactive_list
);
284 WARN_ON(test_bit(R5_InJournal
, &sh
->dev
[sh
->pd_idx
].flags
));
286 list_add_tail(&sh
->lru
, temp_inactive_list
);
287 else if (injournal
== conf
->raid_disks
- conf
->max_degraded
) {
289 if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE
, &sh
->state
))
290 atomic_inc(&conf
->r5c_cached_full_stripes
);
291 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE
, &sh
->state
))
292 atomic_dec(&conf
->r5c_cached_partial_stripes
);
293 list_add_tail(&sh
->lru
, &conf
->r5c_full_stripe_list
);
294 r5c_check_cached_full_stripe(conf
);
297 * STRIPE_R5C_PARTIAL_STRIPE is set in
298 * r5c_try_caching_write(). No need to
301 list_add_tail(&sh
->lru
, &conf
->r5c_partial_stripe_list
);
307 static void __release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
308 struct list_head
*temp_inactive_list
)
310 if (atomic_dec_and_test(&sh
->count
))
311 do_release_stripe(conf
, sh
, temp_inactive_list
);
315 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
317 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
318 * given time. Adding stripes only takes device lock, while deleting stripes
319 * only takes hash lock.
321 static void release_inactive_stripe_list(struct r5conf
*conf
,
322 struct list_head
*temp_inactive_list
,
326 bool do_wakeup
= false;
329 if (hash
== NR_STRIPE_HASH_LOCKS
) {
330 size
= NR_STRIPE_HASH_LOCKS
;
331 hash
= NR_STRIPE_HASH_LOCKS
- 1;
335 struct list_head
*list
= &temp_inactive_list
[size
- 1];
338 * We don't hold any lock here yet, raid5_get_active_stripe() might
339 * remove stripes from the list
341 if (!list_empty_careful(list
)) {
342 spin_lock_irqsave(conf
->hash_locks
+ hash
, flags
);
343 if (list_empty(conf
->inactive_list
+ hash
) &&
345 atomic_dec(&conf
->empty_inactive_list_nr
);
346 list_splice_tail_init(list
, conf
->inactive_list
+ hash
);
348 spin_unlock_irqrestore(conf
->hash_locks
+ hash
, flags
);
355 wake_up(&conf
->wait_for_stripe
);
356 if (atomic_read(&conf
->active_stripes
) == 0)
357 wake_up(&conf
->wait_for_quiescent
);
358 if (conf
->retry_read_aligned
)
359 md_wakeup_thread(conf
->mddev
->thread
);
363 /* should hold conf->device_lock already */
364 static int release_stripe_list(struct r5conf
*conf
,
365 struct list_head
*temp_inactive_list
)
367 struct stripe_head
*sh
, *t
;
369 struct llist_node
*head
;
371 head
= llist_del_all(&conf
->released_stripes
);
372 head
= llist_reverse_order(head
);
373 llist_for_each_entry_safe(sh
, t
, head
, release_list
) {
376 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
378 clear_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
);
380 * Don't worry the bit is set here, because if the bit is set
381 * again, the count is always > 1. This is true for
382 * STRIPE_ON_UNPLUG_LIST bit too.
384 hash
= sh
->hash_lock_index
;
385 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
392 void raid5_release_stripe(struct stripe_head
*sh
)
394 struct r5conf
*conf
= sh
->raid_conf
;
396 struct list_head list
;
400 /* Avoid release_list until the last reference.
402 if (atomic_add_unless(&sh
->count
, -1, 1))
405 if (unlikely(!conf
->mddev
->thread
) ||
406 test_and_set_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
))
408 wakeup
= llist_add(&sh
->release_list
, &conf
->released_stripes
);
410 md_wakeup_thread(conf
->mddev
->thread
);
413 local_irq_save(flags
);
414 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
415 if (atomic_dec_and_lock(&sh
->count
, &conf
->device_lock
)) {
416 INIT_LIST_HEAD(&list
);
417 hash
= sh
->hash_lock_index
;
418 do_release_stripe(conf
, sh
, &list
);
419 spin_unlock(&conf
->device_lock
);
420 release_inactive_stripe_list(conf
, &list
, hash
);
422 local_irq_restore(flags
);
425 static inline void remove_hash(struct stripe_head
*sh
)
427 pr_debug("remove_hash(), stripe %llu\n",
428 (unsigned long long)sh
->sector
);
430 hlist_del_init(&sh
->hash
);
433 static inline void insert_hash(struct r5conf
*conf
, struct stripe_head
*sh
)
435 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
437 pr_debug("insert_hash(), stripe %llu\n",
438 (unsigned long long)sh
->sector
);
440 hlist_add_head(&sh
->hash
, hp
);
443 /* find an idle stripe, make sure it is unhashed, and return it. */
444 static struct stripe_head
*get_free_stripe(struct r5conf
*conf
, int hash
)
446 struct stripe_head
*sh
= NULL
;
447 struct list_head
*first
;
449 if (list_empty(conf
->inactive_list
+ hash
))
451 first
= (conf
->inactive_list
+ hash
)->next
;
452 sh
= list_entry(first
, struct stripe_head
, lru
);
453 list_del_init(first
);
455 atomic_inc(&conf
->active_stripes
);
456 BUG_ON(hash
!= sh
->hash_lock_index
);
457 if (list_empty(conf
->inactive_list
+ hash
))
458 atomic_inc(&conf
->empty_inactive_list_nr
);
463 static void shrink_buffers(struct stripe_head
*sh
)
467 int num
= sh
->raid_conf
->pool_size
;
469 for (i
= 0; i
< num
; i
++) {
470 WARN_ON(sh
->dev
[i
].page
!= sh
->dev
[i
].orig_page
);
474 sh
->dev
[i
].page
= NULL
;
479 static int grow_buffers(struct stripe_head
*sh
, gfp_t gfp
)
482 int num
= sh
->raid_conf
->pool_size
;
484 for (i
= 0; i
< num
; i
++) {
487 if (!(page
= alloc_page(gfp
))) {
490 sh
->dev
[i
].page
= page
;
491 sh
->dev
[i
].orig_page
= page
;
497 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
498 struct stripe_head
*sh
);
500 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int previous
)
502 struct r5conf
*conf
= sh
->raid_conf
;
505 BUG_ON(atomic_read(&sh
->count
) != 0);
506 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
507 BUG_ON(stripe_operations_active(sh
));
508 BUG_ON(sh
->batch_head
);
510 pr_debug("init_stripe called, stripe %llu\n",
511 (unsigned long long)sector
);
513 seq
= read_seqcount_begin(&conf
->gen_lock
);
514 sh
->generation
= conf
->generation
- previous
;
515 sh
->disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
517 stripe_set_idx(sector
, conf
, previous
, sh
);
520 for (i
= sh
->disks
; i
--; ) {
521 struct r5dev
*dev
= &sh
->dev
[i
];
523 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
524 test_bit(R5_LOCKED
, &dev
->flags
)) {
525 pr_err("sector=%llx i=%d %p %p %p %p %d\n",
526 (unsigned long long)sh
->sector
, i
, dev
->toread
,
527 dev
->read
, dev
->towrite
, dev
->written
,
528 test_bit(R5_LOCKED
, &dev
->flags
));
532 dev
->sector
= raid5_compute_blocknr(sh
, i
, previous
);
534 if (read_seqcount_retry(&conf
->gen_lock
, seq
))
536 sh
->overwrite_disks
= 0;
537 insert_hash(conf
, sh
);
538 sh
->cpu
= smp_processor_id();
539 set_bit(STRIPE_BATCH_READY
, &sh
->state
);
542 static struct stripe_head
*__find_stripe(struct r5conf
*conf
, sector_t sector
,
545 struct stripe_head
*sh
;
547 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
548 hlist_for_each_entry(sh
, stripe_hash(conf
, sector
), hash
)
549 if (sh
->sector
== sector
&& sh
->generation
== generation
)
551 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
556 * Need to check if array has failed when deciding whether to:
558 * - remove non-faulty devices
561 * This determination is simple when no reshape is happening.
562 * However if there is a reshape, we need to carefully check
563 * both the before and after sections.
564 * This is because some failed devices may only affect one
565 * of the two sections, and some non-in_sync devices may
566 * be insync in the section most affected by failed devices.
568 int raid5_calc_degraded(struct r5conf
*conf
)
570 int degraded
, degraded2
;
575 for (i
= 0; i
< conf
->previous_raid_disks
; i
++) {
576 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
577 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
578 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
579 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
581 else if (test_bit(In_sync
, &rdev
->flags
))
584 /* not in-sync or faulty.
585 * If the reshape increases the number of devices,
586 * this is being recovered by the reshape, so
587 * this 'previous' section is not in_sync.
588 * If the number of devices is being reduced however,
589 * the device can only be part of the array if
590 * we are reverting a reshape, so this section will
593 if (conf
->raid_disks
>= conf
->previous_raid_disks
)
597 if (conf
->raid_disks
== conf
->previous_raid_disks
)
601 for (i
= 0; i
< conf
->raid_disks
; i
++) {
602 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
603 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
604 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
605 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
607 else if (test_bit(In_sync
, &rdev
->flags
))
610 /* not in-sync or faulty.
611 * If reshape increases the number of devices, this
612 * section has already been recovered, else it
613 * almost certainly hasn't.
615 if (conf
->raid_disks
<= conf
->previous_raid_disks
)
619 if (degraded2
> degraded
)
624 static int has_failed(struct r5conf
*conf
)
628 if (conf
->mddev
->reshape_position
== MaxSector
)
629 return conf
->mddev
->degraded
> conf
->max_degraded
;
631 degraded
= raid5_calc_degraded(conf
);
632 if (degraded
> conf
->max_degraded
)
638 raid5_get_active_stripe(struct r5conf
*conf
, sector_t sector
,
639 int previous
, int noblock
, int noquiesce
)
641 struct stripe_head
*sh
;
642 int hash
= stripe_hash_locks_hash(sector
);
643 int inc_empty_inactive_list_flag
;
645 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
647 spin_lock_irq(conf
->hash_locks
+ hash
);
650 wait_event_lock_irq(conf
->wait_for_quiescent
,
651 conf
->quiesce
== 0 || noquiesce
,
652 *(conf
->hash_locks
+ hash
));
653 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
655 if (!test_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
)) {
656 sh
= get_free_stripe(conf
, hash
);
657 if (!sh
&& !test_bit(R5_DID_ALLOC
,
659 set_bit(R5_ALLOC_MORE
,
662 if (noblock
&& sh
== NULL
)
665 r5c_check_stripe_cache_usage(conf
);
667 set_bit(R5_INACTIVE_BLOCKED
,
669 r5l_wake_reclaim(conf
->log
, 0);
671 conf
->wait_for_stripe
,
672 !list_empty(conf
->inactive_list
+ hash
) &&
673 (atomic_read(&conf
->active_stripes
)
674 < (conf
->max_nr_stripes
* 3 / 4)
675 || !test_bit(R5_INACTIVE_BLOCKED
,
676 &conf
->cache_state
)),
677 *(conf
->hash_locks
+ hash
));
678 clear_bit(R5_INACTIVE_BLOCKED
,
681 init_stripe(sh
, sector
, previous
);
682 atomic_inc(&sh
->count
);
684 } else if (!atomic_inc_not_zero(&sh
->count
)) {
685 spin_lock(&conf
->device_lock
);
686 if (!atomic_read(&sh
->count
)) {
687 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
688 atomic_inc(&conf
->active_stripes
);
689 BUG_ON(list_empty(&sh
->lru
) &&
690 !test_bit(STRIPE_EXPANDING
, &sh
->state
));
691 inc_empty_inactive_list_flag
= 0;
692 if (!list_empty(conf
->inactive_list
+ hash
))
693 inc_empty_inactive_list_flag
= 1;
694 list_del_init(&sh
->lru
);
695 if (list_empty(conf
->inactive_list
+ hash
) && inc_empty_inactive_list_flag
)
696 atomic_inc(&conf
->empty_inactive_list_nr
);
698 sh
->group
->stripes_cnt
--;
702 atomic_inc(&sh
->count
);
703 spin_unlock(&conf
->device_lock
);
705 } while (sh
== NULL
);
707 spin_unlock_irq(conf
->hash_locks
+ hash
);
711 static bool is_full_stripe_write(struct stripe_head
*sh
)
713 BUG_ON(sh
->overwrite_disks
> (sh
->disks
- sh
->raid_conf
->max_degraded
));
714 return sh
->overwrite_disks
== (sh
->disks
- sh
->raid_conf
->max_degraded
);
717 static void lock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
720 spin_lock_irq(&sh2
->stripe_lock
);
721 spin_lock_nested(&sh1
->stripe_lock
, 1);
723 spin_lock_irq(&sh1
->stripe_lock
);
724 spin_lock_nested(&sh2
->stripe_lock
, 1);
728 static void unlock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
730 spin_unlock(&sh1
->stripe_lock
);
731 spin_unlock_irq(&sh2
->stripe_lock
);
734 /* Only freshly new full stripe normal write stripe can be added to a batch list */
735 static bool stripe_can_batch(struct stripe_head
*sh
)
737 struct r5conf
*conf
= sh
->raid_conf
;
739 if (conf
->log
|| raid5_has_ppl(conf
))
741 return test_bit(STRIPE_BATCH_READY
, &sh
->state
) &&
742 !test_bit(STRIPE_BITMAP_PENDING
, &sh
->state
) &&
743 is_full_stripe_write(sh
);
746 /* we only do back search */
747 static void stripe_add_to_batch_list(struct r5conf
*conf
, struct stripe_head
*sh
)
749 struct stripe_head
*head
;
750 sector_t head_sector
, tmp_sec
;
753 int inc_empty_inactive_list_flag
;
755 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
756 tmp_sec
= sh
->sector
;
757 if (!sector_div(tmp_sec
, conf
->chunk_sectors
))
759 head_sector
= sh
->sector
- STRIPE_SECTORS
;
761 hash
= stripe_hash_locks_hash(head_sector
);
762 spin_lock_irq(conf
->hash_locks
+ hash
);
763 head
= __find_stripe(conf
, head_sector
, conf
->generation
);
764 if (head
&& !atomic_inc_not_zero(&head
->count
)) {
765 spin_lock(&conf
->device_lock
);
766 if (!atomic_read(&head
->count
)) {
767 if (!test_bit(STRIPE_HANDLE
, &head
->state
))
768 atomic_inc(&conf
->active_stripes
);
769 BUG_ON(list_empty(&head
->lru
) &&
770 !test_bit(STRIPE_EXPANDING
, &head
->state
));
771 inc_empty_inactive_list_flag
= 0;
772 if (!list_empty(conf
->inactive_list
+ hash
))
773 inc_empty_inactive_list_flag
= 1;
774 list_del_init(&head
->lru
);
775 if (list_empty(conf
->inactive_list
+ hash
) && inc_empty_inactive_list_flag
)
776 atomic_inc(&conf
->empty_inactive_list_nr
);
778 head
->group
->stripes_cnt
--;
782 atomic_inc(&head
->count
);
783 spin_unlock(&conf
->device_lock
);
785 spin_unlock_irq(conf
->hash_locks
+ hash
);
789 if (!stripe_can_batch(head
))
792 lock_two_stripes(head
, sh
);
793 /* clear_batch_ready clear the flag */
794 if (!stripe_can_batch(head
) || !stripe_can_batch(sh
))
801 while (dd_idx
== sh
->pd_idx
|| dd_idx
== sh
->qd_idx
)
803 if (head
->dev
[dd_idx
].towrite
->bi_opf
!= sh
->dev
[dd_idx
].towrite
->bi_opf
||
804 bio_op(head
->dev
[dd_idx
].towrite
) != bio_op(sh
->dev
[dd_idx
].towrite
))
807 if (head
->batch_head
) {
808 spin_lock(&head
->batch_head
->batch_lock
);
809 /* This batch list is already running */
810 if (!stripe_can_batch(head
)) {
811 spin_unlock(&head
->batch_head
->batch_lock
);
815 * We must assign batch_head of this stripe within the
816 * batch_lock, otherwise clear_batch_ready of batch head
817 * stripe could clear BATCH_READY bit of this stripe and
818 * this stripe->batch_head doesn't get assigned, which
819 * could confuse clear_batch_ready for this stripe
821 sh
->batch_head
= head
->batch_head
;
824 * at this point, head's BATCH_READY could be cleared, but we
825 * can still add the stripe to batch list
827 list_add(&sh
->batch_list
, &head
->batch_list
);
828 spin_unlock(&head
->batch_head
->batch_lock
);
830 head
->batch_head
= head
;
831 sh
->batch_head
= head
->batch_head
;
832 spin_lock(&head
->batch_lock
);
833 list_add_tail(&sh
->batch_list
, &head
->batch_list
);
834 spin_unlock(&head
->batch_lock
);
837 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
838 if (atomic_dec_return(&conf
->preread_active_stripes
)
840 md_wakeup_thread(conf
->mddev
->thread
);
842 if (test_and_clear_bit(STRIPE_BIT_DELAY
, &sh
->state
)) {
843 int seq
= sh
->bm_seq
;
844 if (test_bit(STRIPE_BIT_DELAY
, &sh
->batch_head
->state
) &&
845 sh
->batch_head
->bm_seq
> seq
)
846 seq
= sh
->batch_head
->bm_seq
;
847 set_bit(STRIPE_BIT_DELAY
, &sh
->batch_head
->state
);
848 sh
->batch_head
->bm_seq
= seq
;
851 atomic_inc(&sh
->count
);
853 unlock_two_stripes(head
, sh
);
855 raid5_release_stripe(head
);
858 /* Determine if 'data_offset' or 'new_data_offset' should be used
859 * in this stripe_head.
861 static int use_new_offset(struct r5conf
*conf
, struct stripe_head
*sh
)
863 sector_t progress
= conf
->reshape_progress
;
864 /* Need a memory barrier to make sure we see the value
865 * of conf->generation, or ->data_offset that was set before
866 * reshape_progress was updated.
869 if (progress
== MaxSector
)
871 if (sh
->generation
== conf
->generation
- 1)
873 /* We are in a reshape, and this is a new-generation stripe,
874 * so use new_data_offset.
879 static void dispatch_bio_list(struct bio_list
*tmp
)
883 while ((bio
= bio_list_pop(tmp
)))
884 generic_make_request(bio
);
887 static int cmp_stripe(void *priv
, struct list_head
*a
, struct list_head
*b
)
889 const struct r5pending_data
*da
= list_entry(a
,
890 struct r5pending_data
, sibling
);
891 const struct r5pending_data
*db
= list_entry(b
,
892 struct r5pending_data
, sibling
);
893 if (da
->sector
> db
->sector
)
895 if (da
->sector
< db
->sector
)
900 static void dispatch_defer_bios(struct r5conf
*conf
, int target
,
901 struct bio_list
*list
)
903 struct r5pending_data
*data
;
904 struct list_head
*first
, *next
= NULL
;
907 if (conf
->pending_data_cnt
== 0)
910 list_sort(NULL
, &conf
->pending_list
, cmp_stripe
);
912 first
= conf
->pending_list
.next
;
914 /* temporarily move the head */
915 if (conf
->next_pending_data
)
916 list_move_tail(&conf
->pending_list
,
917 &conf
->next_pending_data
->sibling
);
919 while (!list_empty(&conf
->pending_list
)) {
920 data
= list_first_entry(&conf
->pending_list
,
921 struct r5pending_data
, sibling
);
922 if (&data
->sibling
== first
)
923 first
= data
->sibling
.next
;
924 next
= data
->sibling
.next
;
926 bio_list_merge(list
, &data
->bios
);
927 list_move(&data
->sibling
, &conf
->free_list
);
932 conf
->pending_data_cnt
-= cnt
;
933 BUG_ON(conf
->pending_data_cnt
< 0 || cnt
< target
);
935 if (next
!= &conf
->pending_list
)
936 conf
->next_pending_data
= list_entry(next
,
937 struct r5pending_data
, sibling
);
939 conf
->next_pending_data
= NULL
;
940 /* list isn't empty */
941 if (first
!= &conf
->pending_list
)
942 list_move_tail(&conf
->pending_list
, first
);
945 static void flush_deferred_bios(struct r5conf
*conf
)
947 struct bio_list tmp
= BIO_EMPTY_LIST
;
949 if (conf
->pending_data_cnt
== 0)
952 spin_lock(&conf
->pending_bios_lock
);
953 dispatch_defer_bios(conf
, conf
->pending_data_cnt
, &tmp
);
954 BUG_ON(conf
->pending_data_cnt
!= 0);
955 spin_unlock(&conf
->pending_bios_lock
);
957 dispatch_bio_list(&tmp
);
960 static void defer_issue_bios(struct r5conf
*conf
, sector_t sector
,
961 struct bio_list
*bios
)
963 struct bio_list tmp
= BIO_EMPTY_LIST
;
964 struct r5pending_data
*ent
;
966 spin_lock(&conf
->pending_bios_lock
);
967 ent
= list_first_entry(&conf
->free_list
, struct r5pending_data
,
969 list_move_tail(&ent
->sibling
, &conf
->pending_list
);
970 ent
->sector
= sector
;
971 bio_list_init(&ent
->bios
);
972 bio_list_merge(&ent
->bios
, bios
);
973 conf
->pending_data_cnt
++;
974 if (conf
->pending_data_cnt
>= PENDING_IO_MAX
)
975 dispatch_defer_bios(conf
, PENDING_IO_ONE_FLUSH
, &tmp
);
977 spin_unlock(&conf
->pending_bios_lock
);
979 dispatch_bio_list(&tmp
);
983 raid5_end_read_request(struct bio
*bi
);
985 raid5_end_write_request(struct bio
*bi
);
987 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
989 struct r5conf
*conf
= sh
->raid_conf
;
990 int i
, disks
= sh
->disks
;
991 struct stripe_head
*head_sh
= sh
;
992 struct bio_list pending_bios
= BIO_EMPTY_LIST
;
997 if (log_stripe(sh
, s
) == 0)
1000 should_defer
= conf
->batch_bio_dispatch
&& conf
->group_cnt
;
1002 for (i
= disks
; i
--; ) {
1003 int op
, op_flags
= 0;
1004 int replace_only
= 0;
1005 struct bio
*bi
, *rbi
;
1006 struct md_rdev
*rdev
, *rrdev
= NULL
;
1009 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
1011 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
1013 if (test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1014 op
= REQ_OP_DISCARD
;
1015 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
1017 else if (test_and_clear_bit(R5_WantReplace
,
1018 &sh
->dev
[i
].flags
)) {
1023 if (test_and_clear_bit(R5_SyncIO
, &sh
->dev
[i
].flags
))
1024 op_flags
|= REQ_SYNC
;
1027 bi
= &sh
->dev
[i
].req
;
1028 rbi
= &sh
->dev
[i
].rreq
; /* For writing to replacement */
1031 rrdev
= rcu_dereference(conf
->disks
[i
].replacement
);
1032 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1033 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
1038 if (op_is_write(op
)) {
1042 /* We raced and saw duplicates */
1045 if (test_bit(R5_ReadRepl
, &head_sh
->dev
[i
].flags
) && rrdev
)
1050 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
1053 atomic_inc(&rdev
->nr_pending
);
1054 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
1057 atomic_inc(&rrdev
->nr_pending
);
1060 /* We have already checked bad blocks for reads. Now
1061 * need to check for writes. We never accept write errors
1062 * on the replacement, so we don't to check rrdev.
1064 while (op_is_write(op
) && rdev
&&
1065 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1068 int bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
1069 &first_bad
, &bad_sectors
);
1074 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1075 if (!conf
->mddev
->external
&&
1076 conf
->mddev
->sb_flags
) {
1077 /* It is very unlikely, but we might
1078 * still need to write out the
1079 * bad block log - better give it
1081 md_check_recovery(conf
->mddev
);
1084 * Because md_wait_for_blocked_rdev
1085 * will dec nr_pending, we must
1086 * increment it first.
1088 atomic_inc(&rdev
->nr_pending
);
1089 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
1091 /* Acknowledged bad block - skip the write */
1092 rdev_dec_pending(rdev
, conf
->mddev
);
1098 if (s
->syncing
|| s
->expanding
|| s
->expanded
1100 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
1102 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
1104 bio_set_dev(bi
, rdev
->bdev
);
1105 bio_set_op_attrs(bi
, op
, op_flags
);
1106 bi
->bi_end_io
= op_is_write(op
)
1107 ? raid5_end_write_request
1108 : raid5_end_read_request
;
1109 bi
->bi_private
= sh
;
1111 pr_debug("%s: for %llu schedule op %d on disc %d\n",
1112 __func__
, (unsigned long long)sh
->sector
,
1114 atomic_inc(&sh
->count
);
1116 atomic_inc(&head_sh
->count
);
1117 if (use_new_offset(conf
, sh
))
1118 bi
->bi_iter
.bi_sector
= (sh
->sector
1119 + rdev
->new_data_offset
);
1121 bi
->bi_iter
.bi_sector
= (sh
->sector
1122 + rdev
->data_offset
);
1123 if (test_bit(R5_ReadNoMerge
, &head_sh
->dev
[i
].flags
))
1124 bi
->bi_opf
|= REQ_NOMERGE
;
1126 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1127 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1129 if (!op_is_write(op
) &&
1130 test_bit(R5_InJournal
, &sh
->dev
[i
].flags
))
1132 * issuing read for a page in journal, this
1133 * must be preparing for prexor in rmw; read
1134 * the data into orig_page
1136 sh
->dev
[i
].vec
.bv_page
= sh
->dev
[i
].orig_page
;
1138 sh
->dev
[i
].vec
.bv_page
= sh
->dev
[i
].page
;
1140 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
1141 bi
->bi_io_vec
[0].bv_offset
= 0;
1142 bi
->bi_iter
.bi_size
= STRIPE_SIZE
;
1144 * If this is discard request, set bi_vcnt 0. We don't
1145 * want to confuse SCSI because SCSI will replace payload
1147 if (op
== REQ_OP_DISCARD
)
1150 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
1152 if (conf
->mddev
->gendisk
)
1153 trace_block_bio_remap(bi
->bi_disk
->queue
,
1154 bi
, disk_devt(conf
->mddev
->gendisk
),
1156 if (should_defer
&& op_is_write(op
))
1157 bio_list_add(&pending_bios
, bi
);
1159 generic_make_request(bi
);
1162 if (s
->syncing
|| s
->expanding
|| s
->expanded
1164 md_sync_acct(rrdev
->bdev
, STRIPE_SECTORS
);
1166 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
1168 bio_set_dev(rbi
, rrdev
->bdev
);
1169 bio_set_op_attrs(rbi
, op
, op_flags
);
1170 BUG_ON(!op_is_write(op
));
1171 rbi
->bi_end_io
= raid5_end_write_request
;
1172 rbi
->bi_private
= sh
;
1174 pr_debug("%s: for %llu schedule op %d on "
1175 "replacement disc %d\n",
1176 __func__
, (unsigned long long)sh
->sector
,
1178 atomic_inc(&sh
->count
);
1180 atomic_inc(&head_sh
->count
);
1181 if (use_new_offset(conf
, sh
))
1182 rbi
->bi_iter
.bi_sector
= (sh
->sector
1183 + rrdev
->new_data_offset
);
1185 rbi
->bi_iter
.bi_sector
= (sh
->sector
1186 + rrdev
->data_offset
);
1187 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1188 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1189 sh
->dev
[i
].rvec
.bv_page
= sh
->dev
[i
].page
;
1191 rbi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
1192 rbi
->bi_io_vec
[0].bv_offset
= 0;
1193 rbi
->bi_iter
.bi_size
= STRIPE_SIZE
;
1195 * If this is discard request, set bi_vcnt 0. We don't
1196 * want to confuse SCSI because SCSI will replace payload
1198 if (op
== REQ_OP_DISCARD
)
1200 if (conf
->mddev
->gendisk
)
1201 trace_block_bio_remap(rbi
->bi_disk
->queue
,
1202 rbi
, disk_devt(conf
->mddev
->gendisk
),
1204 if (should_defer
&& op_is_write(op
))
1205 bio_list_add(&pending_bios
, rbi
);
1207 generic_make_request(rbi
);
1209 if (!rdev
&& !rrdev
) {
1210 if (op_is_write(op
))
1211 set_bit(STRIPE_DEGRADED
, &sh
->state
);
1212 pr_debug("skip op %d on disc %d for sector %llu\n",
1213 bi
->bi_opf
, i
, (unsigned long long)sh
->sector
);
1214 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1215 set_bit(STRIPE_HANDLE
, &sh
->state
);
1218 if (!head_sh
->batch_head
)
1220 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1226 if (should_defer
&& !bio_list_empty(&pending_bios
))
1227 defer_issue_bios(conf
, head_sh
->sector
, &pending_bios
);
1230 static struct dma_async_tx_descriptor
*
1231 async_copy_data(int frombio
, struct bio
*bio
, struct page
**page
,
1232 sector_t sector
, struct dma_async_tx_descriptor
*tx
,
1233 struct stripe_head
*sh
, int no_skipcopy
)
1236 struct bvec_iter iter
;
1237 struct page
*bio_page
;
1239 struct async_submit_ctl submit
;
1240 enum async_tx_flags flags
= 0;
1242 if (bio
->bi_iter
.bi_sector
>= sector
)
1243 page_offset
= (signed)(bio
->bi_iter
.bi_sector
- sector
) * 512;
1245 page_offset
= (signed)(sector
- bio
->bi_iter
.bi_sector
) * -512;
1248 flags
|= ASYNC_TX_FENCE
;
1249 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
1251 bio_for_each_segment(bvl
, bio
, iter
) {
1252 int len
= bvl
.bv_len
;
1256 if (page_offset
< 0) {
1257 b_offset
= -page_offset
;
1258 page_offset
+= b_offset
;
1262 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
1263 clen
= STRIPE_SIZE
- page_offset
;
1268 b_offset
+= bvl
.bv_offset
;
1269 bio_page
= bvl
.bv_page
;
1271 if (sh
->raid_conf
->skip_copy
&&
1272 b_offset
== 0 && page_offset
== 0 &&
1273 clen
== STRIPE_SIZE
&&
1277 tx
= async_memcpy(*page
, bio_page
, page_offset
,
1278 b_offset
, clen
, &submit
);
1280 tx
= async_memcpy(bio_page
, *page
, b_offset
,
1281 page_offset
, clen
, &submit
);
1283 /* chain the operations */
1284 submit
.depend_tx
= tx
;
1286 if (clen
< len
) /* hit end of page */
1294 static void ops_complete_biofill(void *stripe_head_ref
)
1296 struct stripe_head
*sh
= stripe_head_ref
;
1299 pr_debug("%s: stripe %llu\n", __func__
,
1300 (unsigned long long)sh
->sector
);
1302 /* clear completed biofills */
1303 for (i
= sh
->disks
; i
--; ) {
1304 struct r5dev
*dev
= &sh
->dev
[i
];
1306 /* acknowledge completion of a biofill operation */
1307 /* and check if we need to reply to a read request,
1308 * new R5_Wantfill requests are held off until
1309 * !STRIPE_BIOFILL_RUN
1311 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
1312 struct bio
*rbi
, *rbi2
;
1317 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1318 dev
->sector
+ STRIPE_SECTORS
) {
1319 rbi2
= r5_next_bio(rbi
, dev
->sector
);
1325 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
1327 set_bit(STRIPE_HANDLE
, &sh
->state
);
1328 raid5_release_stripe(sh
);
1331 static void ops_run_biofill(struct stripe_head
*sh
)
1333 struct dma_async_tx_descriptor
*tx
= NULL
;
1334 struct async_submit_ctl submit
;
1337 BUG_ON(sh
->batch_head
);
1338 pr_debug("%s: stripe %llu\n", __func__
,
1339 (unsigned long long)sh
->sector
);
1341 for (i
= sh
->disks
; i
--; ) {
1342 struct r5dev
*dev
= &sh
->dev
[i
];
1343 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
1345 spin_lock_irq(&sh
->stripe_lock
);
1346 dev
->read
= rbi
= dev
->toread
;
1348 spin_unlock_irq(&sh
->stripe_lock
);
1349 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1350 dev
->sector
+ STRIPE_SECTORS
) {
1351 tx
= async_copy_data(0, rbi
, &dev
->page
,
1352 dev
->sector
, tx
, sh
, 0);
1353 rbi
= r5_next_bio(rbi
, dev
->sector
);
1358 atomic_inc(&sh
->count
);
1359 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
1360 async_trigger_callback(&submit
);
1363 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
1370 tgt
= &sh
->dev
[target
];
1371 set_bit(R5_UPTODATE
, &tgt
->flags
);
1372 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1373 clear_bit(R5_Wantcompute
, &tgt
->flags
);
1376 static void ops_complete_compute(void *stripe_head_ref
)
1378 struct stripe_head
*sh
= stripe_head_ref
;
1380 pr_debug("%s: stripe %llu\n", __func__
,
1381 (unsigned long long)sh
->sector
);
1383 /* mark the computed target(s) as uptodate */
1384 mark_target_uptodate(sh
, sh
->ops
.target
);
1385 mark_target_uptodate(sh
, sh
->ops
.target2
);
1387 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
1388 if (sh
->check_state
== check_state_compute_run
)
1389 sh
->check_state
= check_state_compute_result
;
1390 set_bit(STRIPE_HANDLE
, &sh
->state
);
1391 raid5_release_stripe(sh
);
1394 /* return a pointer to the address conversion region of the scribble buffer */
1395 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
1396 struct raid5_percpu
*percpu
, int i
)
1400 addr
= flex_array_get(percpu
->scribble
, i
);
1401 return addr
+ sizeof(struct page
*) * (sh
->disks
+ 2);
1404 /* return a pointer to the address conversion region of the scribble buffer */
1405 static struct page
**to_addr_page(struct raid5_percpu
*percpu
, int i
)
1409 addr
= flex_array_get(percpu
->scribble
, i
);
1413 static struct dma_async_tx_descriptor
*
1414 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1416 int disks
= sh
->disks
;
1417 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1418 int target
= sh
->ops
.target
;
1419 struct r5dev
*tgt
= &sh
->dev
[target
];
1420 struct page
*xor_dest
= tgt
->page
;
1422 struct dma_async_tx_descriptor
*tx
;
1423 struct async_submit_ctl submit
;
1426 BUG_ON(sh
->batch_head
);
1428 pr_debug("%s: stripe %llu block: %d\n",
1429 __func__
, (unsigned long long)sh
->sector
, target
);
1430 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1432 for (i
= disks
; i
--; )
1434 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1436 atomic_inc(&sh
->count
);
1438 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
1439 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
, 0));
1440 if (unlikely(count
== 1))
1441 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1443 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1448 /* set_syndrome_sources - populate source buffers for gen_syndrome
1449 * @srcs - (struct page *) array of size sh->disks
1450 * @sh - stripe_head to parse
1452 * Populates srcs in proper layout order for the stripe and returns the
1453 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1454 * destination buffer is recorded in srcs[count] and the Q destination
1455 * is recorded in srcs[count+1]].
1457 static int set_syndrome_sources(struct page
**srcs
,
1458 struct stripe_head
*sh
,
1461 int disks
= sh
->disks
;
1462 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
1463 int d0_idx
= raid6_d0(sh
);
1467 for (i
= 0; i
< disks
; i
++)
1473 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1474 struct r5dev
*dev
= &sh
->dev
[i
];
1476 if (i
== sh
->qd_idx
|| i
== sh
->pd_idx
||
1477 (srctype
== SYNDROME_SRC_ALL
) ||
1478 (srctype
== SYNDROME_SRC_WANT_DRAIN
&&
1479 (test_bit(R5_Wantdrain
, &dev
->flags
) ||
1480 test_bit(R5_InJournal
, &dev
->flags
))) ||
1481 (srctype
== SYNDROME_SRC_WRITTEN
&&
1483 test_bit(R5_InJournal
, &dev
->flags
)))) {
1484 if (test_bit(R5_InJournal
, &dev
->flags
))
1485 srcs
[slot
] = sh
->dev
[i
].orig_page
;
1487 srcs
[slot
] = sh
->dev
[i
].page
;
1489 i
= raid6_next_disk(i
, disks
);
1490 } while (i
!= d0_idx
);
1492 return syndrome_disks
;
1495 static struct dma_async_tx_descriptor
*
1496 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1498 int disks
= sh
->disks
;
1499 struct page
**blocks
= to_addr_page(percpu
, 0);
1501 int qd_idx
= sh
->qd_idx
;
1502 struct dma_async_tx_descriptor
*tx
;
1503 struct async_submit_ctl submit
;
1509 BUG_ON(sh
->batch_head
);
1510 if (sh
->ops
.target
< 0)
1511 target
= sh
->ops
.target2
;
1512 else if (sh
->ops
.target2
< 0)
1513 target
= sh
->ops
.target
;
1515 /* we should only have one valid target */
1518 pr_debug("%s: stripe %llu block: %d\n",
1519 __func__
, (unsigned long long)sh
->sector
, target
);
1521 tgt
= &sh
->dev
[target
];
1522 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1525 atomic_inc(&sh
->count
);
1527 if (target
== qd_idx
) {
1528 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1529 blocks
[count
] = NULL
; /* regenerating p is not necessary */
1530 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
1531 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1532 ops_complete_compute
, sh
,
1533 to_addr_conv(sh
, percpu
, 0));
1534 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1536 /* Compute any data- or p-drive using XOR */
1538 for (i
= disks
; i
-- ; ) {
1539 if (i
== target
|| i
== qd_idx
)
1541 blocks
[count
++] = sh
->dev
[i
].page
;
1544 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1545 NULL
, ops_complete_compute
, sh
,
1546 to_addr_conv(sh
, percpu
, 0));
1547 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
1553 static struct dma_async_tx_descriptor
*
1554 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1556 int i
, count
, disks
= sh
->disks
;
1557 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1558 int d0_idx
= raid6_d0(sh
);
1559 int faila
= -1, failb
= -1;
1560 int target
= sh
->ops
.target
;
1561 int target2
= sh
->ops
.target2
;
1562 struct r5dev
*tgt
= &sh
->dev
[target
];
1563 struct r5dev
*tgt2
= &sh
->dev
[target2
];
1564 struct dma_async_tx_descriptor
*tx
;
1565 struct page
**blocks
= to_addr_page(percpu
, 0);
1566 struct async_submit_ctl submit
;
1568 BUG_ON(sh
->batch_head
);
1569 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1570 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
1571 BUG_ON(target
< 0 || target2
< 0);
1572 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1573 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
1575 /* we need to open-code set_syndrome_sources to handle the
1576 * slot number conversion for 'faila' and 'failb'
1578 for (i
= 0; i
< disks
; i
++)
1583 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1585 blocks
[slot
] = sh
->dev
[i
].page
;
1591 i
= raid6_next_disk(i
, disks
);
1592 } while (i
!= d0_idx
);
1594 BUG_ON(faila
== failb
);
1597 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1598 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1600 atomic_inc(&sh
->count
);
1602 if (failb
== syndrome_disks
+1) {
1603 /* Q disk is one of the missing disks */
1604 if (faila
== syndrome_disks
) {
1605 /* Missing P+Q, just recompute */
1606 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1607 ops_complete_compute
, sh
,
1608 to_addr_conv(sh
, percpu
, 0));
1609 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
1610 STRIPE_SIZE
, &submit
);
1614 int qd_idx
= sh
->qd_idx
;
1616 /* Missing D+Q: recompute D from P, then recompute Q */
1617 if (target
== qd_idx
)
1618 data_target
= target2
;
1620 data_target
= target
;
1623 for (i
= disks
; i
-- ; ) {
1624 if (i
== data_target
|| i
== qd_idx
)
1626 blocks
[count
++] = sh
->dev
[i
].page
;
1628 dest
= sh
->dev
[data_target
].page
;
1629 init_async_submit(&submit
,
1630 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1632 to_addr_conv(sh
, percpu
, 0));
1633 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
1636 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1637 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1638 ops_complete_compute
, sh
,
1639 to_addr_conv(sh
, percpu
, 0));
1640 return async_gen_syndrome(blocks
, 0, count
+2,
1641 STRIPE_SIZE
, &submit
);
1644 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1645 ops_complete_compute
, sh
,
1646 to_addr_conv(sh
, percpu
, 0));
1647 if (failb
== syndrome_disks
) {
1648 /* We're missing D+P. */
1649 return async_raid6_datap_recov(syndrome_disks
+2,
1653 /* We're missing D+D. */
1654 return async_raid6_2data_recov(syndrome_disks
+2,
1655 STRIPE_SIZE
, faila
, failb
,
1661 static void ops_complete_prexor(void *stripe_head_ref
)
1663 struct stripe_head
*sh
= stripe_head_ref
;
1665 pr_debug("%s: stripe %llu\n", __func__
,
1666 (unsigned long long)sh
->sector
);
1668 if (r5c_is_writeback(sh
->raid_conf
->log
))
1670 * raid5-cache write back uses orig_page during prexor.
1671 * After prexor, it is time to free orig_page
1673 r5c_release_extra_page(sh
);
1676 static struct dma_async_tx_descriptor
*
1677 ops_run_prexor5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1678 struct dma_async_tx_descriptor
*tx
)
1680 int disks
= sh
->disks
;
1681 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1682 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1683 struct async_submit_ctl submit
;
1685 /* existing parity data subtracted */
1686 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1688 BUG_ON(sh
->batch_head
);
1689 pr_debug("%s: stripe %llu\n", __func__
,
1690 (unsigned long long)sh
->sector
);
1692 for (i
= disks
; i
--; ) {
1693 struct r5dev
*dev
= &sh
->dev
[i
];
1694 /* Only process blocks that are known to be uptodate */
1695 if (test_bit(R5_InJournal
, &dev
->flags
))
1696 xor_srcs
[count
++] = dev
->orig_page
;
1697 else if (test_bit(R5_Wantdrain
, &dev
->flags
))
1698 xor_srcs
[count
++] = dev
->page
;
1701 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1702 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1703 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1708 static struct dma_async_tx_descriptor
*
1709 ops_run_prexor6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1710 struct dma_async_tx_descriptor
*tx
)
1712 struct page
**blocks
= to_addr_page(percpu
, 0);
1714 struct async_submit_ctl submit
;
1716 pr_debug("%s: stripe %llu\n", __func__
,
1717 (unsigned long long)sh
->sector
);
1719 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_WANT_DRAIN
);
1721 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_PQ_XOR_DST
, tx
,
1722 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1723 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1728 static struct dma_async_tx_descriptor
*
1729 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1731 struct r5conf
*conf
= sh
->raid_conf
;
1732 int disks
= sh
->disks
;
1734 struct stripe_head
*head_sh
= sh
;
1736 pr_debug("%s: stripe %llu\n", __func__
,
1737 (unsigned long long)sh
->sector
);
1739 for (i
= disks
; i
--; ) {
1744 if (test_and_clear_bit(R5_Wantdrain
, &head_sh
->dev
[i
].flags
)) {
1750 * clear R5_InJournal, so when rewriting a page in
1751 * journal, it is not skipped by r5l_log_stripe()
1753 clear_bit(R5_InJournal
, &dev
->flags
);
1754 spin_lock_irq(&sh
->stripe_lock
);
1755 chosen
= dev
->towrite
;
1756 dev
->towrite
= NULL
;
1757 sh
->overwrite_disks
= 0;
1758 BUG_ON(dev
->written
);
1759 wbi
= dev
->written
= chosen
;
1760 spin_unlock_irq(&sh
->stripe_lock
);
1761 WARN_ON(dev
->page
!= dev
->orig_page
);
1763 while (wbi
&& wbi
->bi_iter
.bi_sector
<
1764 dev
->sector
+ STRIPE_SECTORS
) {
1765 if (wbi
->bi_opf
& REQ_FUA
)
1766 set_bit(R5_WantFUA
, &dev
->flags
);
1767 if (wbi
->bi_opf
& REQ_SYNC
)
1768 set_bit(R5_SyncIO
, &dev
->flags
);
1769 if (bio_op(wbi
) == REQ_OP_DISCARD
)
1770 set_bit(R5_Discard
, &dev
->flags
);
1772 tx
= async_copy_data(1, wbi
, &dev
->page
,
1773 dev
->sector
, tx
, sh
,
1774 r5c_is_writeback(conf
->log
));
1775 if (dev
->page
!= dev
->orig_page
&&
1776 !r5c_is_writeback(conf
->log
)) {
1777 set_bit(R5_SkipCopy
, &dev
->flags
);
1778 clear_bit(R5_UPTODATE
, &dev
->flags
);
1779 clear_bit(R5_OVERWRITE
, &dev
->flags
);
1782 wbi
= r5_next_bio(wbi
, dev
->sector
);
1785 if (head_sh
->batch_head
) {
1786 sh
= list_first_entry(&sh
->batch_list
,
1799 static void ops_complete_reconstruct(void *stripe_head_ref
)
1801 struct stripe_head
*sh
= stripe_head_ref
;
1802 int disks
= sh
->disks
;
1803 int pd_idx
= sh
->pd_idx
;
1804 int qd_idx
= sh
->qd_idx
;
1806 bool fua
= false, sync
= false, discard
= false;
1808 pr_debug("%s: stripe %llu\n", __func__
,
1809 (unsigned long long)sh
->sector
);
1811 for (i
= disks
; i
--; ) {
1812 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1813 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
1814 discard
|= test_bit(R5_Discard
, &sh
->dev
[i
].flags
);
1817 for (i
= disks
; i
--; ) {
1818 struct r5dev
*dev
= &sh
->dev
[i
];
1820 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1821 if (!discard
&& !test_bit(R5_SkipCopy
, &dev
->flags
))
1822 set_bit(R5_UPTODATE
, &dev
->flags
);
1824 set_bit(R5_WantFUA
, &dev
->flags
);
1826 set_bit(R5_SyncIO
, &dev
->flags
);
1830 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1831 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1832 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1833 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1835 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1836 sh
->reconstruct_state
= reconstruct_state_result
;
1839 set_bit(STRIPE_HANDLE
, &sh
->state
);
1840 raid5_release_stripe(sh
);
1844 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1845 struct dma_async_tx_descriptor
*tx
)
1847 int disks
= sh
->disks
;
1848 struct page
**xor_srcs
;
1849 struct async_submit_ctl submit
;
1850 int count
, pd_idx
= sh
->pd_idx
, i
;
1851 struct page
*xor_dest
;
1853 unsigned long flags
;
1855 struct stripe_head
*head_sh
= sh
;
1858 pr_debug("%s: stripe %llu\n", __func__
,
1859 (unsigned long long)sh
->sector
);
1861 for (i
= 0; i
< sh
->disks
; i
++) {
1864 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1867 if (i
>= sh
->disks
) {
1868 atomic_inc(&sh
->count
);
1869 set_bit(R5_Discard
, &sh
->dev
[pd_idx
].flags
);
1870 ops_complete_reconstruct(sh
);
1875 xor_srcs
= to_addr_page(percpu
, j
);
1876 /* check if prexor is active which means only process blocks
1877 * that are part of a read-modify-write (written)
1879 if (head_sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1881 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1882 for (i
= disks
; i
--; ) {
1883 struct r5dev
*dev
= &sh
->dev
[i
];
1884 if (head_sh
->dev
[i
].written
||
1885 test_bit(R5_InJournal
, &head_sh
->dev
[i
].flags
))
1886 xor_srcs
[count
++] = dev
->page
;
1889 xor_dest
= sh
->dev
[pd_idx
].page
;
1890 for (i
= disks
; i
--; ) {
1891 struct r5dev
*dev
= &sh
->dev
[i
];
1893 xor_srcs
[count
++] = dev
->page
;
1897 /* 1/ if we prexor'd then the dest is reused as a source
1898 * 2/ if we did not prexor then we are redoing the parity
1899 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1900 * for the synchronous xor case
1902 last_stripe
= !head_sh
->batch_head
||
1903 list_first_entry(&sh
->batch_list
,
1904 struct stripe_head
, batch_list
) == head_sh
;
1906 flags
= ASYNC_TX_ACK
|
1907 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1909 atomic_inc(&head_sh
->count
);
1910 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, head_sh
,
1911 to_addr_conv(sh
, percpu
, j
));
1913 flags
= prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
;
1914 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
,
1915 to_addr_conv(sh
, percpu
, j
));
1918 if (unlikely(count
== 1))
1919 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1921 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1924 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1931 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1932 struct dma_async_tx_descriptor
*tx
)
1934 struct async_submit_ctl submit
;
1935 struct page
**blocks
;
1936 int count
, i
, j
= 0;
1937 struct stripe_head
*head_sh
= sh
;
1940 unsigned long txflags
;
1942 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1944 for (i
= 0; i
< sh
->disks
; i
++) {
1945 if (sh
->pd_idx
== i
|| sh
->qd_idx
== i
)
1947 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1950 if (i
>= sh
->disks
) {
1951 atomic_inc(&sh
->count
);
1952 set_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
1953 set_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
1954 ops_complete_reconstruct(sh
);
1959 blocks
= to_addr_page(percpu
, j
);
1961 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1962 synflags
= SYNDROME_SRC_WRITTEN
;
1963 txflags
= ASYNC_TX_ACK
| ASYNC_TX_PQ_XOR_DST
;
1965 synflags
= SYNDROME_SRC_ALL
;
1966 txflags
= ASYNC_TX_ACK
;
1969 count
= set_syndrome_sources(blocks
, sh
, synflags
);
1970 last_stripe
= !head_sh
->batch_head
||
1971 list_first_entry(&sh
->batch_list
,
1972 struct stripe_head
, batch_list
) == head_sh
;
1975 atomic_inc(&head_sh
->count
);
1976 init_async_submit(&submit
, txflags
, tx
, ops_complete_reconstruct
,
1977 head_sh
, to_addr_conv(sh
, percpu
, j
));
1979 init_async_submit(&submit
, 0, tx
, NULL
, NULL
,
1980 to_addr_conv(sh
, percpu
, j
));
1981 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1984 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1990 static void ops_complete_check(void *stripe_head_ref
)
1992 struct stripe_head
*sh
= stripe_head_ref
;
1994 pr_debug("%s: stripe %llu\n", __func__
,
1995 (unsigned long long)sh
->sector
);
1997 sh
->check_state
= check_state_check_result
;
1998 set_bit(STRIPE_HANDLE
, &sh
->state
);
1999 raid5_release_stripe(sh
);
2002 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
2004 int disks
= sh
->disks
;
2005 int pd_idx
= sh
->pd_idx
;
2006 int qd_idx
= sh
->qd_idx
;
2007 struct page
*xor_dest
;
2008 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
2009 struct dma_async_tx_descriptor
*tx
;
2010 struct async_submit_ctl submit
;
2014 pr_debug("%s: stripe %llu\n", __func__
,
2015 (unsigned long long)sh
->sector
);
2017 BUG_ON(sh
->batch_head
);
2019 xor_dest
= sh
->dev
[pd_idx
].page
;
2020 xor_srcs
[count
++] = xor_dest
;
2021 for (i
= disks
; i
--; ) {
2022 if (i
== pd_idx
|| i
== qd_idx
)
2024 xor_srcs
[count
++] = sh
->dev
[i
].page
;
2027 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
2028 to_addr_conv(sh
, percpu
, 0));
2029 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
2030 &sh
->ops
.zero_sum_result
, &submit
);
2032 atomic_inc(&sh
->count
);
2033 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
2034 tx
= async_trigger_callback(&submit
);
2037 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
2039 struct page
**srcs
= to_addr_page(percpu
, 0);
2040 struct async_submit_ctl submit
;
2043 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
2044 (unsigned long long)sh
->sector
, checkp
);
2046 BUG_ON(sh
->batch_head
);
2047 count
= set_syndrome_sources(srcs
, sh
, SYNDROME_SRC_ALL
);
2051 atomic_inc(&sh
->count
);
2052 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
2053 sh
, to_addr_conv(sh
, percpu
, 0));
2054 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
2055 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
2058 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
2060 int overlap_clear
= 0, i
, disks
= sh
->disks
;
2061 struct dma_async_tx_descriptor
*tx
= NULL
;
2062 struct r5conf
*conf
= sh
->raid_conf
;
2063 int level
= conf
->level
;
2064 struct raid5_percpu
*percpu
;
2068 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
2069 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
2070 ops_run_biofill(sh
);
2074 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
2076 tx
= ops_run_compute5(sh
, percpu
);
2078 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
2079 tx
= ops_run_compute6_1(sh
, percpu
);
2081 tx
= ops_run_compute6_2(sh
, percpu
);
2083 /* terminate the chain if reconstruct is not set to be run */
2084 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
2088 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
)) {
2090 tx
= ops_run_prexor5(sh
, percpu
, tx
);
2092 tx
= ops_run_prexor6(sh
, percpu
, tx
);
2095 if (test_bit(STRIPE_OP_PARTIAL_PARITY
, &ops_request
))
2096 tx
= ops_run_partial_parity(sh
, percpu
, tx
);
2098 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
2099 tx
= ops_run_biodrain(sh
, tx
);
2103 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
2105 ops_run_reconstruct5(sh
, percpu
, tx
);
2107 ops_run_reconstruct6(sh
, percpu
, tx
);
2110 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
2111 if (sh
->check_state
== check_state_run
)
2112 ops_run_check_p(sh
, percpu
);
2113 else if (sh
->check_state
== check_state_run_q
)
2114 ops_run_check_pq(sh
, percpu
, 0);
2115 else if (sh
->check_state
== check_state_run_pq
)
2116 ops_run_check_pq(sh
, percpu
, 1);
2121 if (overlap_clear
&& !sh
->batch_head
)
2122 for (i
= disks
; i
--; ) {
2123 struct r5dev
*dev
= &sh
->dev
[i
];
2124 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
2125 wake_up(&sh
->raid_conf
->wait_for_overlap
);
2130 static void free_stripe(struct kmem_cache
*sc
, struct stripe_head
*sh
)
2133 __free_page(sh
->ppl_page
);
2134 kmem_cache_free(sc
, sh
);
2137 static struct stripe_head
*alloc_stripe(struct kmem_cache
*sc
, gfp_t gfp
,
2138 int disks
, struct r5conf
*conf
)
2140 struct stripe_head
*sh
;
2143 sh
= kmem_cache_zalloc(sc
, gfp
);
2145 spin_lock_init(&sh
->stripe_lock
);
2146 spin_lock_init(&sh
->batch_lock
);
2147 INIT_LIST_HEAD(&sh
->batch_list
);
2148 INIT_LIST_HEAD(&sh
->lru
);
2149 INIT_LIST_HEAD(&sh
->r5c
);
2150 INIT_LIST_HEAD(&sh
->log_list
);
2151 atomic_set(&sh
->count
, 1);
2152 sh
->raid_conf
= conf
;
2153 sh
->log_start
= MaxSector
;
2154 for (i
= 0; i
< disks
; i
++) {
2155 struct r5dev
*dev
= &sh
->dev
[i
];
2157 bio_init(&dev
->req
, &dev
->vec
, 1);
2158 bio_init(&dev
->rreq
, &dev
->rvec
, 1);
2161 if (raid5_has_ppl(conf
)) {
2162 sh
->ppl_page
= alloc_page(gfp
);
2163 if (!sh
->ppl_page
) {
2164 free_stripe(sc
, sh
);
2171 static int grow_one_stripe(struct r5conf
*conf
, gfp_t gfp
)
2173 struct stripe_head
*sh
;
2175 sh
= alloc_stripe(conf
->slab_cache
, gfp
, conf
->pool_size
, conf
);
2179 if (grow_buffers(sh
, gfp
)) {
2181 free_stripe(conf
->slab_cache
, sh
);
2184 sh
->hash_lock_index
=
2185 conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
;
2186 /* we just created an active stripe so... */
2187 atomic_inc(&conf
->active_stripes
);
2189 raid5_release_stripe(sh
);
2190 conf
->max_nr_stripes
++;
2194 static int grow_stripes(struct r5conf
*conf
, int num
)
2196 struct kmem_cache
*sc
;
2197 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
2199 if (conf
->mddev
->gendisk
)
2200 sprintf(conf
->cache_name
[0],
2201 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
2203 sprintf(conf
->cache_name
[0],
2204 "raid%d-%p", conf
->level
, conf
->mddev
);
2205 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
2207 conf
->active_name
= 0;
2208 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
2209 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
2213 conf
->slab_cache
= sc
;
2214 conf
->pool_size
= devs
;
2216 if (!grow_one_stripe(conf
, GFP_KERNEL
))
2223 * scribble_len - return the required size of the scribble region
2224 * @num - total number of disks in the array
2226 * The size must be enough to contain:
2227 * 1/ a struct page pointer for each device in the array +2
2228 * 2/ room to convert each entry in (1) to its corresponding dma
2229 * (dma_map_page()) or page (page_address()) address.
2231 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2232 * calculate over all devices (not just the data blocks), using zeros in place
2233 * of the P and Q blocks.
2235 static struct flex_array
*scribble_alloc(int num
, int cnt
, gfp_t flags
)
2237 struct flex_array
*ret
;
2240 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
2241 ret
= flex_array_alloc(len
, cnt
, flags
);
2244 /* always prealloc all elements, so no locking is required */
2245 if (flex_array_prealloc(ret
, 0, cnt
, flags
)) {
2246 flex_array_free(ret
);
2252 static int resize_chunks(struct r5conf
*conf
, int new_disks
, int new_sectors
)
2258 * Never shrink. And mddev_suspend() could deadlock if this is called
2259 * from raid5d. In that case, scribble_disks and scribble_sectors
2260 * should equal to new_disks and new_sectors
2262 if (conf
->scribble_disks
>= new_disks
&&
2263 conf
->scribble_sectors
>= new_sectors
)
2265 mddev_suspend(conf
->mddev
);
2267 for_each_present_cpu(cpu
) {
2268 struct raid5_percpu
*percpu
;
2269 struct flex_array
*scribble
;
2271 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
2272 scribble
= scribble_alloc(new_disks
,
2273 new_sectors
/ STRIPE_SECTORS
,
2277 flex_array_free(percpu
->scribble
);
2278 percpu
->scribble
= scribble
;
2285 mddev_resume(conf
->mddev
);
2287 conf
->scribble_disks
= new_disks
;
2288 conf
->scribble_sectors
= new_sectors
;
2293 static int resize_stripes(struct r5conf
*conf
, int newsize
)
2295 /* Make all the stripes able to hold 'newsize' devices.
2296 * New slots in each stripe get 'page' set to a new page.
2298 * This happens in stages:
2299 * 1/ create a new kmem_cache and allocate the required number of
2301 * 2/ gather all the old stripe_heads and transfer the pages across
2302 * to the new stripe_heads. This will have the side effect of
2303 * freezing the array as once all stripe_heads have been collected,
2304 * no IO will be possible. Old stripe heads are freed once their
2305 * pages have been transferred over, and the old kmem_cache is
2306 * freed when all stripes are done.
2307 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2308 * we simple return a failure status - no need to clean anything up.
2309 * 4/ allocate new pages for the new slots in the new stripe_heads.
2310 * If this fails, we don't bother trying the shrink the
2311 * stripe_heads down again, we just leave them as they are.
2312 * As each stripe_head is processed the new one is released into
2315 * Once step2 is started, we cannot afford to wait for a write,
2316 * so we use GFP_NOIO allocations.
2318 struct stripe_head
*osh
, *nsh
;
2319 LIST_HEAD(newstripes
);
2320 struct disk_info
*ndisks
;
2322 struct kmem_cache
*sc
;
2326 md_allow_write(conf
->mddev
);
2329 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
2330 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
2335 /* Need to ensure auto-resizing doesn't interfere */
2336 mutex_lock(&conf
->cache_size_mutex
);
2338 for (i
= conf
->max_nr_stripes
; i
; i
--) {
2339 nsh
= alloc_stripe(sc
, GFP_KERNEL
, newsize
, conf
);
2343 list_add(&nsh
->lru
, &newstripes
);
2346 /* didn't get enough, give up */
2347 while (!list_empty(&newstripes
)) {
2348 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2349 list_del(&nsh
->lru
);
2350 free_stripe(sc
, nsh
);
2352 kmem_cache_destroy(sc
);
2353 mutex_unlock(&conf
->cache_size_mutex
);
2356 /* Step 2 - Must use GFP_NOIO now.
2357 * OK, we have enough stripes, start collecting inactive
2358 * stripes and copying them over
2362 list_for_each_entry(nsh
, &newstripes
, lru
) {
2363 lock_device_hash_lock(conf
, hash
);
2364 wait_event_cmd(conf
->wait_for_stripe
,
2365 !list_empty(conf
->inactive_list
+ hash
),
2366 unlock_device_hash_lock(conf
, hash
),
2367 lock_device_hash_lock(conf
, hash
));
2368 osh
= get_free_stripe(conf
, hash
);
2369 unlock_device_hash_lock(conf
, hash
);
2371 for(i
=0; i
<conf
->pool_size
; i
++) {
2372 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
2373 nsh
->dev
[i
].orig_page
= osh
->dev
[i
].page
;
2375 nsh
->hash_lock_index
= hash
;
2376 free_stripe(conf
->slab_cache
, osh
);
2378 if (cnt
>= conf
->max_nr_stripes
/ NR_STRIPE_HASH_LOCKS
+
2379 !!((conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
) > hash
)) {
2384 kmem_cache_destroy(conf
->slab_cache
);
2387 * At this point, we are holding all the stripes so the array
2388 * is completely stalled, so now is a good time to resize
2389 * conf->disks and the scribble region
2391 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
2393 for (i
= 0; i
< conf
->pool_size
; i
++)
2394 ndisks
[i
] = conf
->disks
[i
];
2396 for (i
= conf
->pool_size
; i
< newsize
; i
++) {
2397 ndisks
[i
].extra_page
= alloc_page(GFP_NOIO
);
2398 if (!ndisks
[i
].extra_page
)
2403 for (i
= conf
->pool_size
; i
< newsize
; i
++)
2404 if (ndisks
[i
].extra_page
)
2405 put_page(ndisks
[i
].extra_page
);
2409 conf
->disks
= ndisks
;
2414 mutex_unlock(&conf
->cache_size_mutex
);
2416 conf
->slab_cache
= sc
;
2417 conf
->active_name
= 1-conf
->active_name
;
2419 /* Step 4, return new stripes to service */
2420 while(!list_empty(&newstripes
)) {
2421 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2422 list_del_init(&nsh
->lru
);
2424 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
2425 if (nsh
->dev
[i
].page
== NULL
) {
2426 struct page
*p
= alloc_page(GFP_NOIO
);
2427 nsh
->dev
[i
].page
= p
;
2428 nsh
->dev
[i
].orig_page
= p
;
2432 raid5_release_stripe(nsh
);
2434 /* critical section pass, GFP_NOIO no longer needed */
2437 conf
->pool_size
= newsize
;
2441 static int drop_one_stripe(struct r5conf
*conf
)
2443 struct stripe_head
*sh
;
2444 int hash
= (conf
->max_nr_stripes
- 1) & STRIPE_HASH_LOCKS_MASK
;
2446 spin_lock_irq(conf
->hash_locks
+ hash
);
2447 sh
= get_free_stripe(conf
, hash
);
2448 spin_unlock_irq(conf
->hash_locks
+ hash
);
2451 BUG_ON(atomic_read(&sh
->count
));
2453 free_stripe(conf
->slab_cache
, sh
);
2454 atomic_dec(&conf
->active_stripes
);
2455 conf
->max_nr_stripes
--;
2459 static void shrink_stripes(struct r5conf
*conf
)
2461 while (conf
->max_nr_stripes
&&
2462 drop_one_stripe(conf
))
2465 kmem_cache_destroy(conf
->slab_cache
);
2466 conf
->slab_cache
= NULL
;
2469 static void raid5_end_read_request(struct bio
* bi
)
2471 struct stripe_head
*sh
= bi
->bi_private
;
2472 struct r5conf
*conf
= sh
->raid_conf
;
2473 int disks
= sh
->disks
, i
;
2474 char b
[BDEVNAME_SIZE
];
2475 struct md_rdev
*rdev
= NULL
;
2478 for (i
=0 ; i
<disks
; i
++)
2479 if (bi
== &sh
->dev
[i
].req
)
2482 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2483 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2490 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2491 /* If replacement finished while this request was outstanding,
2492 * 'replacement' might be NULL already.
2493 * In that case it moved down to 'rdev'.
2494 * rdev is not removed until all requests are finished.
2496 rdev
= conf
->disks
[i
].replacement
;
2498 rdev
= conf
->disks
[i
].rdev
;
2500 if (use_new_offset(conf
, sh
))
2501 s
= sh
->sector
+ rdev
->new_data_offset
;
2503 s
= sh
->sector
+ rdev
->data_offset
;
2504 if (!bi
->bi_status
) {
2505 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2506 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2507 /* Note that this cannot happen on a
2508 * replacement device. We just fail those on
2511 pr_info_ratelimited(
2512 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2513 mdname(conf
->mddev
), STRIPE_SECTORS
,
2514 (unsigned long long)s
,
2515 bdevname(rdev
->bdev
, b
));
2516 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
2517 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2518 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2519 } else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2520 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2522 if (test_bit(R5_InJournal
, &sh
->dev
[i
].flags
))
2524 * end read for a page in journal, this
2525 * must be preparing for prexor in rmw
2527 set_bit(R5_OrigPageUPTDODATE
, &sh
->dev
[i
].flags
);
2529 if (atomic_read(&rdev
->read_errors
))
2530 atomic_set(&rdev
->read_errors
, 0);
2532 const char *bdn
= bdevname(rdev
->bdev
, b
);
2536 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2537 atomic_inc(&rdev
->read_errors
);
2538 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2539 pr_warn_ratelimited(
2540 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2541 mdname(conf
->mddev
),
2542 (unsigned long long)s
,
2544 else if (conf
->mddev
->degraded
>= conf
->max_degraded
) {
2546 pr_warn_ratelimited(
2547 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2548 mdname(conf
->mddev
),
2549 (unsigned long long)s
,
2551 } else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
)) {
2554 pr_warn_ratelimited(
2555 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2556 mdname(conf
->mddev
),
2557 (unsigned long long)s
,
2559 } else if (atomic_read(&rdev
->read_errors
)
2560 > conf
->max_nr_stripes
)
2561 pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2562 mdname(conf
->mddev
), bdn
);
2565 if (set_bad
&& test_bit(In_sync
, &rdev
->flags
)
2566 && !test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2569 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
)) {
2570 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2571 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2573 set_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2575 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2576 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2578 && test_bit(In_sync
, &rdev
->flags
)
2579 && rdev_set_badblocks(
2580 rdev
, sh
->sector
, STRIPE_SECTORS
, 0)))
2581 md_error(conf
->mddev
, rdev
);
2584 rdev_dec_pending(rdev
, conf
->mddev
);
2586 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2587 set_bit(STRIPE_HANDLE
, &sh
->state
);
2588 raid5_release_stripe(sh
);
2591 static void raid5_end_write_request(struct bio
*bi
)
2593 struct stripe_head
*sh
= bi
->bi_private
;
2594 struct r5conf
*conf
= sh
->raid_conf
;
2595 int disks
= sh
->disks
, i
;
2596 struct md_rdev
*uninitialized_var(rdev
);
2599 int replacement
= 0;
2601 for (i
= 0 ; i
< disks
; i
++) {
2602 if (bi
== &sh
->dev
[i
].req
) {
2603 rdev
= conf
->disks
[i
].rdev
;
2606 if (bi
== &sh
->dev
[i
].rreq
) {
2607 rdev
= conf
->disks
[i
].replacement
;
2611 /* rdev was removed and 'replacement'
2612 * replaced it. rdev is not removed
2613 * until all requests are finished.
2615 rdev
= conf
->disks
[i
].rdev
;
2619 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2620 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2630 md_error(conf
->mddev
, rdev
);
2631 else if (is_badblock(rdev
, sh
->sector
,
2633 &first_bad
, &bad_sectors
))
2634 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
2636 if (bi
->bi_status
) {
2637 set_bit(STRIPE_DEGRADED
, &sh
->state
);
2638 set_bit(WriteErrorSeen
, &rdev
->flags
);
2639 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
2640 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2641 set_bit(MD_RECOVERY_NEEDED
,
2642 &rdev
->mddev
->recovery
);
2643 } else if (is_badblock(rdev
, sh
->sector
,
2645 &first_bad
, &bad_sectors
)) {
2646 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
2647 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))
2648 /* That was a successful write so make
2649 * sure it looks like we already did
2652 set_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2655 rdev_dec_pending(rdev
, conf
->mddev
);
2657 if (sh
->batch_head
&& bi
->bi_status
&& !replacement
)
2658 set_bit(STRIPE_BATCH_ERR
, &sh
->batch_head
->state
);
2661 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
2662 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2663 set_bit(STRIPE_HANDLE
, &sh
->state
);
2664 raid5_release_stripe(sh
);
2666 if (sh
->batch_head
&& sh
!= sh
->batch_head
)
2667 raid5_release_stripe(sh
->batch_head
);
2670 static void raid5_error(struct mddev
*mddev
, struct md_rdev
*rdev
)
2672 char b
[BDEVNAME_SIZE
];
2673 struct r5conf
*conf
= mddev
->private;
2674 unsigned long flags
;
2675 pr_debug("raid456: error called\n");
2677 spin_lock_irqsave(&conf
->device_lock
, flags
);
2678 clear_bit(In_sync
, &rdev
->flags
);
2679 mddev
->degraded
= raid5_calc_degraded(conf
);
2680 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2681 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2683 set_bit(Blocked
, &rdev
->flags
);
2684 set_bit(Faulty
, &rdev
->flags
);
2685 set_mask_bits(&mddev
->sb_flags
, 0,
2686 BIT(MD_SB_CHANGE_DEVS
) | BIT(MD_SB_CHANGE_PENDING
));
2687 pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2688 "md/raid:%s: Operation continuing on %d devices.\n",
2690 bdevname(rdev
->bdev
, b
),
2692 conf
->raid_disks
- mddev
->degraded
);
2693 r5c_update_on_rdev_error(mddev
, rdev
);
2697 * Input: a 'big' sector number,
2698 * Output: index of the data and parity disk, and the sector # in them.
2700 sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
2701 int previous
, int *dd_idx
,
2702 struct stripe_head
*sh
)
2704 sector_t stripe
, stripe2
;
2705 sector_t chunk_number
;
2706 unsigned int chunk_offset
;
2709 sector_t new_sector
;
2710 int algorithm
= previous
? conf
->prev_algo
2712 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2713 : conf
->chunk_sectors
;
2714 int raid_disks
= previous
? conf
->previous_raid_disks
2716 int data_disks
= raid_disks
- conf
->max_degraded
;
2718 /* First compute the information on this sector */
2721 * Compute the chunk number and the sector offset inside the chunk
2723 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
2724 chunk_number
= r_sector
;
2727 * Compute the stripe number
2729 stripe
= chunk_number
;
2730 *dd_idx
= sector_div(stripe
, data_disks
);
2733 * Select the parity disk based on the user selected algorithm.
2735 pd_idx
= qd_idx
= -1;
2736 switch(conf
->level
) {
2738 pd_idx
= data_disks
;
2741 switch (algorithm
) {
2742 case ALGORITHM_LEFT_ASYMMETRIC
:
2743 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2744 if (*dd_idx
>= pd_idx
)
2747 case ALGORITHM_RIGHT_ASYMMETRIC
:
2748 pd_idx
= sector_div(stripe2
, raid_disks
);
2749 if (*dd_idx
>= pd_idx
)
2752 case ALGORITHM_LEFT_SYMMETRIC
:
2753 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2754 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2756 case ALGORITHM_RIGHT_SYMMETRIC
:
2757 pd_idx
= sector_div(stripe2
, raid_disks
);
2758 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2760 case ALGORITHM_PARITY_0
:
2764 case ALGORITHM_PARITY_N
:
2765 pd_idx
= data_disks
;
2773 switch (algorithm
) {
2774 case ALGORITHM_LEFT_ASYMMETRIC
:
2775 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2776 qd_idx
= pd_idx
+ 1;
2777 if (pd_idx
== raid_disks
-1) {
2778 (*dd_idx
)++; /* Q D D D P */
2780 } else if (*dd_idx
>= pd_idx
)
2781 (*dd_idx
) += 2; /* D D P Q D */
2783 case ALGORITHM_RIGHT_ASYMMETRIC
:
2784 pd_idx
= sector_div(stripe2
, raid_disks
);
2785 qd_idx
= pd_idx
+ 1;
2786 if (pd_idx
== raid_disks
-1) {
2787 (*dd_idx
)++; /* Q D D D P */
2789 } else if (*dd_idx
>= pd_idx
)
2790 (*dd_idx
) += 2; /* D D P Q D */
2792 case ALGORITHM_LEFT_SYMMETRIC
:
2793 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2794 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2795 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2797 case ALGORITHM_RIGHT_SYMMETRIC
:
2798 pd_idx
= sector_div(stripe2
, raid_disks
);
2799 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2800 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2803 case ALGORITHM_PARITY_0
:
2808 case ALGORITHM_PARITY_N
:
2809 pd_idx
= data_disks
;
2810 qd_idx
= data_disks
+ 1;
2813 case ALGORITHM_ROTATING_ZERO_RESTART
:
2814 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2815 * of blocks for computing Q is different.
2817 pd_idx
= sector_div(stripe2
, raid_disks
);
2818 qd_idx
= pd_idx
+ 1;
2819 if (pd_idx
== raid_disks
-1) {
2820 (*dd_idx
)++; /* Q D D D P */
2822 } else if (*dd_idx
>= pd_idx
)
2823 (*dd_idx
) += 2; /* D D P Q D */
2827 case ALGORITHM_ROTATING_N_RESTART
:
2828 /* Same a left_asymmetric, by first stripe is
2829 * D D D P Q rather than
2833 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2834 qd_idx
= pd_idx
+ 1;
2835 if (pd_idx
== raid_disks
-1) {
2836 (*dd_idx
)++; /* Q D D D P */
2838 } else if (*dd_idx
>= pd_idx
)
2839 (*dd_idx
) += 2; /* D D P Q D */
2843 case ALGORITHM_ROTATING_N_CONTINUE
:
2844 /* Same as left_symmetric but Q is before P */
2845 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2846 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
2847 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2851 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2852 /* RAID5 left_asymmetric, with Q on last device */
2853 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2854 if (*dd_idx
>= pd_idx
)
2856 qd_idx
= raid_disks
- 1;
2859 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2860 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2861 if (*dd_idx
>= pd_idx
)
2863 qd_idx
= raid_disks
- 1;
2866 case ALGORITHM_LEFT_SYMMETRIC_6
:
2867 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2868 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2869 qd_idx
= raid_disks
- 1;
2872 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2873 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2874 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2875 qd_idx
= raid_disks
- 1;
2878 case ALGORITHM_PARITY_0_6
:
2881 qd_idx
= raid_disks
- 1;
2891 sh
->pd_idx
= pd_idx
;
2892 sh
->qd_idx
= qd_idx
;
2893 sh
->ddf_layout
= ddf_layout
;
2896 * Finally, compute the new sector number
2898 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
2902 sector_t
raid5_compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
2904 struct r5conf
*conf
= sh
->raid_conf
;
2905 int raid_disks
= sh
->disks
;
2906 int data_disks
= raid_disks
- conf
->max_degraded
;
2907 sector_t new_sector
= sh
->sector
, check
;
2908 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2909 : conf
->chunk_sectors
;
2910 int algorithm
= previous
? conf
->prev_algo
2914 sector_t chunk_number
;
2915 int dummy1
, dd_idx
= i
;
2917 struct stripe_head sh2
;
2919 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
2920 stripe
= new_sector
;
2922 if (i
== sh
->pd_idx
)
2924 switch(conf
->level
) {
2927 switch (algorithm
) {
2928 case ALGORITHM_LEFT_ASYMMETRIC
:
2929 case ALGORITHM_RIGHT_ASYMMETRIC
:
2933 case ALGORITHM_LEFT_SYMMETRIC
:
2934 case ALGORITHM_RIGHT_SYMMETRIC
:
2937 i
-= (sh
->pd_idx
+ 1);
2939 case ALGORITHM_PARITY_0
:
2942 case ALGORITHM_PARITY_N
:
2949 if (i
== sh
->qd_idx
)
2950 return 0; /* It is the Q disk */
2951 switch (algorithm
) {
2952 case ALGORITHM_LEFT_ASYMMETRIC
:
2953 case ALGORITHM_RIGHT_ASYMMETRIC
:
2954 case ALGORITHM_ROTATING_ZERO_RESTART
:
2955 case ALGORITHM_ROTATING_N_RESTART
:
2956 if (sh
->pd_idx
== raid_disks
-1)
2957 i
--; /* Q D D D P */
2958 else if (i
> sh
->pd_idx
)
2959 i
-= 2; /* D D P Q D */
2961 case ALGORITHM_LEFT_SYMMETRIC
:
2962 case ALGORITHM_RIGHT_SYMMETRIC
:
2963 if (sh
->pd_idx
== raid_disks
-1)
2964 i
--; /* Q D D D P */
2969 i
-= (sh
->pd_idx
+ 2);
2972 case ALGORITHM_PARITY_0
:
2975 case ALGORITHM_PARITY_N
:
2977 case ALGORITHM_ROTATING_N_CONTINUE
:
2978 /* Like left_symmetric, but P is before Q */
2979 if (sh
->pd_idx
== 0)
2980 i
--; /* P D D D Q */
2985 i
-= (sh
->pd_idx
+ 1);
2988 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2989 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2993 case ALGORITHM_LEFT_SYMMETRIC_6
:
2994 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2996 i
+= data_disks
+ 1;
2997 i
-= (sh
->pd_idx
+ 1);
2999 case ALGORITHM_PARITY_0_6
:
3008 chunk_number
= stripe
* data_disks
+ i
;
3009 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
3011 check
= raid5_compute_sector(conf
, r_sector
,
3012 previous
, &dummy1
, &sh2
);
3013 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
3014 || sh2
.qd_idx
!= sh
->qd_idx
) {
3015 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3016 mdname(conf
->mddev
));
3023 * There are cases where we want handle_stripe_dirtying() and
3024 * schedule_reconstruction() to delay towrite to some dev of a stripe.
3026 * This function checks whether we want to delay the towrite. Specifically,
3027 * we delay the towrite when:
3029 * 1. degraded stripe has a non-overwrite to the missing dev, AND this
3030 * stripe has data in journal (for other devices).
3032 * In this case, when reading data for the non-overwrite dev, it is
3033 * necessary to handle complex rmw of write back cache (prexor with
3034 * orig_page, and xor with page). To keep read path simple, we would
3035 * like to flush data in journal to RAID disks first, so complex rmw
3036 * is handled in the write patch (handle_stripe_dirtying).
3038 * 2. when journal space is critical (R5C_LOG_CRITICAL=1)
3040 * It is important to be able to flush all stripes in raid5-cache.
3041 * Therefore, we need reserve some space on the journal device for
3042 * these flushes. If flush operation includes pending writes to the
3043 * stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3044 * for the flush out. If we exclude these pending writes from flush
3045 * operation, we only need (conf->max_degraded + 1) pages per stripe.
3046 * Therefore, excluding pending writes in these cases enables more
3047 * efficient use of the journal device.
3049 * Note: To make sure the stripe makes progress, we only delay
3050 * towrite for stripes with data already in journal (injournal > 0).
3051 * When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3052 * no_space_stripes list.
3054 * 3. during journal failure
3055 * In journal failure, we try to flush all cached data to raid disks
3056 * based on data in stripe cache. The array is read-only to upper
3057 * layers, so we would skip all pending writes.
3060 static inline bool delay_towrite(struct r5conf
*conf
,
3062 struct stripe_head_state
*s
)
3065 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3066 !test_bit(R5_Insync
, &dev
->flags
) && s
->injournal
)
3069 if (test_bit(R5C_LOG_CRITICAL
, &conf
->cache_state
) &&
3073 if (s
->log_failed
&& s
->injournal
)
3079 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3080 int rcw
, int expand
)
3082 int i
, pd_idx
= sh
->pd_idx
, qd_idx
= sh
->qd_idx
, disks
= sh
->disks
;
3083 struct r5conf
*conf
= sh
->raid_conf
;
3084 int level
= conf
->level
;
3088 * In some cases, handle_stripe_dirtying initially decided to
3089 * run rmw and allocates extra page for prexor. However, rcw is
3090 * cheaper later on. We need to free the extra page now,
3091 * because we won't be able to do that in ops_complete_prexor().
3093 r5c_release_extra_page(sh
);
3095 for (i
= disks
; i
--; ) {
3096 struct r5dev
*dev
= &sh
->dev
[i
];
3098 if (dev
->towrite
&& !delay_towrite(conf
, dev
, s
)) {
3099 set_bit(R5_LOCKED
, &dev
->flags
);
3100 set_bit(R5_Wantdrain
, &dev
->flags
);
3102 clear_bit(R5_UPTODATE
, &dev
->flags
);
3104 } else if (test_bit(R5_InJournal
, &dev
->flags
)) {
3105 set_bit(R5_LOCKED
, &dev
->flags
);
3109 /* if we are not expanding this is a proper write request, and
3110 * there will be bios with new data to be drained into the
3115 /* False alarm, nothing to do */
3117 sh
->reconstruct_state
= reconstruct_state_drain_run
;
3118 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
3120 sh
->reconstruct_state
= reconstruct_state_run
;
3122 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
3124 if (s
->locked
+ conf
->max_degraded
== disks
)
3125 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3126 atomic_inc(&conf
->pending_full_writes
);
3128 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
3129 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
3130 BUG_ON(level
== 6 &&
3131 (!(test_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
) ||
3132 test_bit(R5_Wantcompute
, &sh
->dev
[qd_idx
].flags
))));
3134 for (i
= disks
; i
--; ) {
3135 struct r5dev
*dev
= &sh
->dev
[i
];
3136 if (i
== pd_idx
|| i
== qd_idx
)
3140 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
3141 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3142 set_bit(R5_Wantdrain
, &dev
->flags
);
3143 set_bit(R5_LOCKED
, &dev
->flags
);
3144 clear_bit(R5_UPTODATE
, &dev
->flags
);
3146 } else if (test_bit(R5_InJournal
, &dev
->flags
)) {
3147 set_bit(R5_LOCKED
, &dev
->flags
);
3152 /* False alarm - nothing to do */
3154 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
3155 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
3156 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
3157 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
3160 /* keep the parity disk(s) locked while asynchronous operations
3163 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
3164 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
3168 int qd_idx
= sh
->qd_idx
;
3169 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
3171 set_bit(R5_LOCKED
, &dev
->flags
);
3172 clear_bit(R5_UPTODATE
, &dev
->flags
);
3176 if (raid5_has_ppl(sh
->raid_conf
) && sh
->ppl_page
&&
3177 test_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
) &&
3178 !test_bit(STRIPE_FULL_WRITE
, &sh
->state
) &&
3179 test_bit(R5_Insync
, &sh
->dev
[pd_idx
].flags
))
3180 set_bit(STRIPE_OP_PARTIAL_PARITY
, &s
->ops_request
);
3182 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3183 __func__
, (unsigned long long)sh
->sector
,
3184 s
->locked
, s
->ops_request
);
3188 * Each stripe/dev can have one or more bion attached.
3189 * toread/towrite point to the first in a chain.
3190 * The bi_next chain must be in order.
3192 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
,
3193 int forwrite
, int previous
)
3196 struct r5conf
*conf
= sh
->raid_conf
;
3199 pr_debug("adding bi b#%llu to stripe s#%llu\n",
3200 (unsigned long long)bi
->bi_iter
.bi_sector
,
3201 (unsigned long long)sh
->sector
);
3203 spin_lock_irq(&sh
->stripe_lock
);
3204 /* Don't allow new IO added to stripes in batch list */
3208 bip
= &sh
->dev
[dd_idx
].towrite
;
3212 bip
= &sh
->dev
[dd_idx
].toread
;
3213 while (*bip
&& (*bip
)->bi_iter
.bi_sector
< bi
->bi_iter
.bi_sector
) {
3214 if (bio_end_sector(*bip
) > bi
->bi_iter
.bi_sector
)
3216 bip
= & (*bip
)->bi_next
;
3218 if (*bip
&& (*bip
)->bi_iter
.bi_sector
< bio_end_sector(bi
))
3221 if (forwrite
&& raid5_has_ppl(conf
)) {
3223 * With PPL only writes to consecutive data chunks within a
3224 * stripe are allowed because for a single stripe_head we can
3225 * only have one PPL entry at a time, which describes one data
3226 * range. Not really an overlap, but wait_for_overlap can be
3227 * used to handle this.
3235 for (i
= 0; i
< sh
->disks
; i
++) {
3236 if (i
!= sh
->pd_idx
&&
3237 (i
== dd_idx
|| sh
->dev
[i
].towrite
)) {
3238 sector
= sh
->dev
[i
].sector
;
3239 if (count
== 0 || sector
< first
)
3247 if (first
+ conf
->chunk_sectors
* (count
- 1) != last
)
3251 if (!forwrite
|| previous
)
3252 clear_bit(STRIPE_BATCH_READY
, &sh
->state
);
3254 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
3258 bio_inc_remaining(bi
);
3259 md_write_inc(conf
->mddev
, bi
);
3262 /* check if page is covered */
3263 sector_t sector
= sh
->dev
[dd_idx
].sector
;
3264 for (bi
=sh
->dev
[dd_idx
].towrite
;
3265 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
3266 bi
&& bi
->bi_iter
.bi_sector
<= sector
;
3267 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
3268 if (bio_end_sector(bi
) >= sector
)
3269 sector
= bio_end_sector(bi
);
3271 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
3272 if (!test_and_set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
))
3273 sh
->overwrite_disks
++;
3276 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3277 (unsigned long long)(*bip
)->bi_iter
.bi_sector
,
3278 (unsigned long long)sh
->sector
, dd_idx
);
3280 if (conf
->mddev
->bitmap
&& firstwrite
) {
3281 /* Cannot hold spinlock over bitmap_startwrite,
3282 * but must ensure this isn't added to a batch until
3283 * we have added to the bitmap and set bm_seq.
3284 * So set STRIPE_BITMAP_PENDING to prevent
3286 * If multiple add_stripe_bio() calls race here they
3287 * much all set STRIPE_BITMAP_PENDING. So only the first one
3288 * to complete "bitmap_startwrite" gets to set
3289 * STRIPE_BIT_DELAY. This is important as once a stripe
3290 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3293 set_bit(STRIPE_BITMAP_PENDING
, &sh
->state
);
3294 spin_unlock_irq(&sh
->stripe_lock
);
3295 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
3297 spin_lock_irq(&sh
->stripe_lock
);
3298 clear_bit(STRIPE_BITMAP_PENDING
, &sh
->state
);
3299 if (!sh
->batch_head
) {
3300 sh
->bm_seq
= conf
->seq_flush
+1;
3301 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
3304 spin_unlock_irq(&sh
->stripe_lock
);
3306 if (stripe_can_batch(sh
))
3307 stripe_add_to_batch_list(conf
, sh
);
3311 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
3312 spin_unlock_irq(&sh
->stripe_lock
);
3316 static void end_reshape(struct r5conf
*conf
);
3318 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
3319 struct stripe_head
*sh
)
3321 int sectors_per_chunk
=
3322 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
3324 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
3325 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
3327 raid5_compute_sector(conf
,
3328 stripe
* (disks
- conf
->max_degraded
)
3329 *sectors_per_chunk
+ chunk_offset
,
3335 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
3336 struct stripe_head_state
*s
, int disks
)
3339 BUG_ON(sh
->batch_head
);
3340 for (i
= disks
; i
--; ) {
3344 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
3345 struct md_rdev
*rdev
;
3347 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3348 if (rdev
&& test_bit(In_sync
, &rdev
->flags
) &&
3349 !test_bit(Faulty
, &rdev
->flags
))
3350 atomic_inc(&rdev
->nr_pending
);
3355 if (!rdev_set_badblocks(
3359 md_error(conf
->mddev
, rdev
);
3360 rdev_dec_pending(rdev
, conf
->mddev
);
3363 spin_lock_irq(&sh
->stripe_lock
);
3364 /* fail all writes first */
3365 bi
= sh
->dev
[i
].towrite
;
3366 sh
->dev
[i
].towrite
= NULL
;
3367 sh
->overwrite_disks
= 0;
3368 spin_unlock_irq(&sh
->stripe_lock
);
3372 log_stripe_write_finished(sh
);
3374 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3375 wake_up(&conf
->wait_for_overlap
);
3377 while (bi
&& bi
->bi_iter
.bi_sector
<
3378 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3379 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3381 md_write_end(conf
->mddev
);
3386 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3387 STRIPE_SECTORS
, 0, 0);
3389 /* and fail all 'written' */
3390 bi
= sh
->dev
[i
].written
;
3391 sh
->dev
[i
].written
= NULL
;
3392 if (test_and_clear_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
)) {
3393 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
3394 sh
->dev
[i
].page
= sh
->dev
[i
].orig_page
;
3397 if (bi
) bitmap_end
= 1;
3398 while (bi
&& bi
->bi_iter
.bi_sector
<
3399 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3400 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3402 md_write_end(conf
->mddev
);
3407 /* fail any reads if this device is non-operational and
3408 * the data has not reached the cache yet.
3410 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
3411 s
->failed
> conf
->max_degraded
&&
3412 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
3413 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
3414 spin_lock_irq(&sh
->stripe_lock
);
3415 bi
= sh
->dev
[i
].toread
;
3416 sh
->dev
[i
].toread
= NULL
;
3417 spin_unlock_irq(&sh
->stripe_lock
);
3418 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3419 wake_up(&conf
->wait_for_overlap
);
3422 while (bi
&& bi
->bi_iter
.bi_sector
<
3423 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3424 struct bio
*nextbi
=
3425 r5_next_bio(bi
, sh
->dev
[i
].sector
);
3432 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3433 STRIPE_SECTORS
, 0, 0);
3434 /* If we were in the middle of a write the parity block might
3435 * still be locked - so just clear all R5_LOCKED flags
3437 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3442 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3443 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3444 md_wakeup_thread(conf
->mddev
->thread
);
3448 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
3449 struct stripe_head_state
*s
)
3454 BUG_ON(sh
->batch_head
);
3455 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3456 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
3457 wake_up(&conf
->wait_for_overlap
);
3460 /* There is nothing more to do for sync/check/repair.
3461 * Don't even need to abort as that is handled elsewhere
3462 * if needed, and not always wanted e.g. if there is a known
3464 * For recover/replace we need to record a bad block on all
3465 * non-sync devices, or abort the recovery
3467 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
3468 /* During recovery devices cannot be removed, so
3469 * locking and refcounting of rdevs is not needed
3472 for (i
= 0; i
< conf
->raid_disks
; i
++) {
3473 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3475 && !test_bit(Faulty
, &rdev
->flags
)
3476 && !test_bit(In_sync
, &rdev
->flags
)
3477 && !rdev_set_badblocks(rdev
, sh
->sector
,
3480 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
3482 && !test_bit(Faulty
, &rdev
->flags
)
3483 && !test_bit(In_sync
, &rdev
->flags
)
3484 && !rdev_set_badblocks(rdev
, sh
->sector
,
3490 conf
->recovery_disabled
=
3491 conf
->mddev
->recovery_disabled
;
3493 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, !abort
);
3496 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
3498 struct md_rdev
*rdev
;
3502 rdev
= rcu_dereference(sh
->raid_conf
->disks
[disk_idx
].replacement
);
3504 && !test_bit(Faulty
, &rdev
->flags
)
3505 && !test_bit(In_sync
, &rdev
->flags
)
3506 && (rdev
->recovery_offset
<= sh
->sector
3507 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
3513 static int need_this_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3514 int disk_idx
, int disks
)
3516 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3517 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
3518 &sh
->dev
[s
->failed_num
[1]] };
3522 if (test_bit(R5_LOCKED
, &dev
->flags
) ||
3523 test_bit(R5_UPTODATE
, &dev
->flags
))
3524 /* No point reading this as we already have it or have
3525 * decided to get it.
3530 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)))
3531 /* We need this block to directly satisfy a request */
3534 if (s
->syncing
|| s
->expanding
||
3535 (s
->replacing
&& want_replace(sh
, disk_idx
)))
3536 /* When syncing, or expanding we read everything.
3537 * When replacing, we need the replaced block.
3541 if ((s
->failed
>= 1 && fdev
[0]->toread
) ||
3542 (s
->failed
>= 2 && fdev
[1]->toread
))
3543 /* If we want to read from a failed device, then
3544 * we need to actually read every other device.
3548 /* Sometimes neither read-modify-write nor reconstruct-write
3549 * cycles can work. In those cases we read every block we
3550 * can. Then the parity-update is certain to have enough to
3552 * This can only be a problem when we need to write something,
3553 * and some device has failed. If either of those tests
3554 * fail we need look no further.
3556 if (!s
->failed
|| !s
->to_write
)
3559 if (test_bit(R5_Insync
, &dev
->flags
) &&
3560 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3561 /* Pre-reads at not permitted until after short delay
3562 * to gather multiple requests. However if this
3563 * device is no Insync, the block could only be computed
3564 * and there is no need to delay that.
3568 for (i
= 0; i
< s
->failed
&& i
< 2; i
++) {
3569 if (fdev
[i
]->towrite
&&
3570 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3571 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3572 /* If we have a partial write to a failed
3573 * device, then we will need to reconstruct
3574 * the content of that device, so all other
3575 * devices must be read.
3580 /* If we are forced to do a reconstruct-write, either because
3581 * the current RAID6 implementation only supports that, or
3582 * because parity cannot be trusted and we are currently
3583 * recovering it, there is extra need to be careful.
3584 * If one of the devices that we would need to read, because
3585 * it is not being overwritten (and maybe not written at all)
3586 * is missing/faulty, then we need to read everything we can.
3588 if (sh
->raid_conf
->level
!= 6 &&
3589 sh
->sector
< sh
->raid_conf
->mddev
->recovery_cp
)
3590 /* reconstruct-write isn't being forced */
3592 for (i
= 0; i
< s
->failed
&& i
< 2; i
++) {
3593 if (s
->failed_num
[i
] != sh
->pd_idx
&&
3594 s
->failed_num
[i
] != sh
->qd_idx
&&
3595 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3596 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3603 /* fetch_block - checks the given member device to see if its data needs
3604 * to be read or computed to satisfy a request.
3606 * Returns 1 when no more member devices need to be checked, otherwise returns
3607 * 0 to tell the loop in handle_stripe_fill to continue
3609 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3610 int disk_idx
, int disks
)
3612 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3614 /* is the data in this block needed, and can we get it? */
3615 if (need_this_block(sh
, s
, disk_idx
, disks
)) {
3616 /* we would like to get this block, possibly by computing it,
3617 * otherwise read it if the backing disk is insync
3619 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
3620 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
3621 BUG_ON(sh
->batch_head
);
3624 * In the raid6 case if the only non-uptodate disk is P
3625 * then we already trusted P to compute the other failed
3626 * drives. It is safe to compute rather than re-read P.
3627 * In other cases we only compute blocks from failed
3628 * devices, otherwise check/repair might fail to detect
3629 * a real inconsistency.
3632 if ((s
->uptodate
== disks
- 1) &&
3633 ((sh
->qd_idx
>= 0 && sh
->pd_idx
== disk_idx
) ||
3634 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
3635 disk_idx
== s
->failed_num
[1])))) {
3636 /* have disk failed, and we're requested to fetch it;
3639 pr_debug("Computing stripe %llu block %d\n",
3640 (unsigned long long)sh
->sector
, disk_idx
);
3641 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3642 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3643 set_bit(R5_Wantcompute
, &dev
->flags
);
3644 sh
->ops
.target
= disk_idx
;
3645 sh
->ops
.target2
= -1; /* no 2nd target */
3647 /* Careful: from this point on 'uptodate' is in the eye
3648 * of raid_run_ops which services 'compute' operations
3649 * before writes. R5_Wantcompute flags a block that will
3650 * be R5_UPTODATE by the time it is needed for a
3651 * subsequent operation.
3655 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
3656 /* Computing 2-failure is *very* expensive; only
3657 * do it if failed >= 2
3660 for (other
= disks
; other
--; ) {
3661 if (other
== disk_idx
)
3663 if (!test_bit(R5_UPTODATE
,
3664 &sh
->dev
[other
].flags
))
3668 pr_debug("Computing stripe %llu blocks %d,%d\n",
3669 (unsigned long long)sh
->sector
,
3671 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3672 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3673 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
3674 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
3675 sh
->ops
.target
= disk_idx
;
3676 sh
->ops
.target2
= other
;
3680 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
3681 set_bit(R5_LOCKED
, &dev
->flags
);
3682 set_bit(R5_Wantread
, &dev
->flags
);
3684 pr_debug("Reading block %d (sync=%d)\n",
3685 disk_idx
, s
->syncing
);
3693 * handle_stripe_fill - read or compute data to satisfy pending requests.
3695 static void handle_stripe_fill(struct stripe_head
*sh
,
3696 struct stripe_head_state
*s
,
3701 /* look for blocks to read/compute, skip this if a compute
3702 * is already in flight, or if the stripe contents are in the
3703 * midst of changing due to a write
3705 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
3706 !sh
->reconstruct_state
) {
3709 * For degraded stripe with data in journal, do not handle
3710 * read requests yet, instead, flush the stripe to raid
3711 * disks first, this avoids handling complex rmw of write
3712 * back cache (prexor with orig_page, and then xor with
3713 * page) in the read path
3715 if (s
->injournal
&& s
->failed
) {
3716 if (test_bit(STRIPE_R5C_CACHING
, &sh
->state
))
3717 r5c_make_stripe_write_out(sh
);
3721 for (i
= disks
; i
--; )
3722 if (fetch_block(sh
, s
, i
, disks
))
3726 set_bit(STRIPE_HANDLE
, &sh
->state
);
3729 static void break_stripe_batch_list(struct stripe_head
*head_sh
,
3730 unsigned long handle_flags
);
3731 /* handle_stripe_clean_event
3732 * any written block on an uptodate or failed drive can be returned.
3733 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3734 * never LOCKED, so we don't need to test 'failed' directly.
3736 static void handle_stripe_clean_event(struct r5conf
*conf
,
3737 struct stripe_head
*sh
, int disks
)
3741 int discard_pending
= 0;
3742 struct stripe_head
*head_sh
= sh
;
3743 bool do_endio
= false;
3745 for (i
= disks
; i
--; )
3746 if (sh
->dev
[i
].written
) {
3748 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
3749 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
3750 test_bit(R5_Discard
, &dev
->flags
) ||
3751 test_bit(R5_SkipCopy
, &dev
->flags
))) {
3752 /* We can return any write requests */
3753 struct bio
*wbi
, *wbi2
;
3754 pr_debug("Return write for disc %d\n", i
);
3755 if (test_and_clear_bit(R5_Discard
, &dev
->flags
))
3756 clear_bit(R5_UPTODATE
, &dev
->flags
);
3757 if (test_and_clear_bit(R5_SkipCopy
, &dev
->flags
)) {
3758 WARN_ON(test_bit(R5_UPTODATE
, &dev
->flags
));
3763 dev
->page
= dev
->orig_page
;
3765 dev
->written
= NULL
;
3766 while (wbi
&& wbi
->bi_iter
.bi_sector
<
3767 dev
->sector
+ STRIPE_SECTORS
) {
3768 wbi2
= r5_next_bio(wbi
, dev
->sector
);
3769 md_write_end(conf
->mddev
);
3773 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3775 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
3777 if (head_sh
->batch_head
) {
3778 sh
= list_first_entry(&sh
->batch_list
,
3781 if (sh
!= head_sh
) {
3788 } else if (test_bit(R5_Discard
, &dev
->flags
))
3789 discard_pending
= 1;
3792 log_stripe_write_finished(sh
);
3794 if (!discard_pending
&&
3795 test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
)) {
3797 clear_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
3798 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3799 if (sh
->qd_idx
>= 0) {
3800 clear_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
3801 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
);
3803 /* now that discard is done we can proceed with any sync */
3804 clear_bit(STRIPE_DISCARD
, &sh
->state
);
3806 * SCSI discard will change some bio fields and the stripe has
3807 * no updated data, so remove it from hash list and the stripe
3808 * will be reinitialized
3811 hash
= sh
->hash_lock_index
;
3812 spin_lock_irq(conf
->hash_locks
+ hash
);
3814 spin_unlock_irq(conf
->hash_locks
+ hash
);
3815 if (head_sh
->batch_head
) {
3816 sh
= list_first_entry(&sh
->batch_list
,
3817 struct stripe_head
, batch_list
);
3823 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
))
3824 set_bit(STRIPE_HANDLE
, &sh
->state
);
3828 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3829 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3830 md_wakeup_thread(conf
->mddev
->thread
);
3832 if (head_sh
->batch_head
&& do_endio
)
3833 break_stripe_batch_list(head_sh
, STRIPE_EXPAND_SYNC_FLAGS
);
3837 * For RMW in write back cache, we need extra page in prexor to store the
3838 * old data. This page is stored in dev->orig_page.
3840 * This function checks whether we have data for prexor. The exact logic
3842 * R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3844 static inline bool uptodate_for_rmw(struct r5dev
*dev
)
3846 return (test_bit(R5_UPTODATE
, &dev
->flags
)) &&
3847 (!test_bit(R5_InJournal
, &dev
->flags
) ||
3848 test_bit(R5_OrigPageUPTDODATE
, &dev
->flags
));
3851 static int handle_stripe_dirtying(struct r5conf
*conf
,
3852 struct stripe_head
*sh
,
3853 struct stripe_head_state
*s
,
3856 int rmw
= 0, rcw
= 0, i
;
3857 sector_t recovery_cp
= conf
->mddev
->recovery_cp
;
3859 /* Check whether resync is now happening or should start.
3860 * If yes, then the array is dirty (after unclean shutdown or
3861 * initial creation), so parity in some stripes might be inconsistent.
3862 * In this case, we need to always do reconstruct-write, to ensure
3863 * that in case of drive failure or read-error correction, we
3864 * generate correct data from the parity.
3866 if (conf
->rmw_level
== PARITY_DISABLE_RMW
||
3867 (recovery_cp
< MaxSector
&& sh
->sector
>= recovery_cp
&&
3869 /* Calculate the real rcw later - for now make it
3870 * look like rcw is cheaper
3873 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3874 conf
->rmw_level
, (unsigned long long)recovery_cp
,
3875 (unsigned long long)sh
->sector
);
3876 } else for (i
= disks
; i
--; ) {
3877 /* would I have to read this buffer for read_modify_write */
3878 struct r5dev
*dev
= &sh
->dev
[i
];
3879 if (((dev
->towrite
&& !delay_towrite(conf
, dev
, s
)) ||
3880 i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3881 test_bit(R5_InJournal
, &dev
->flags
)) &&
3882 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3883 !(uptodate_for_rmw(dev
) ||
3884 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3885 if (test_bit(R5_Insync
, &dev
->flags
))
3888 rmw
+= 2*disks
; /* cannot read it */
3890 /* Would I have to read this buffer for reconstruct_write */
3891 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3892 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3893 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3894 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3895 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3896 if (test_bit(R5_Insync
, &dev
->flags
))
3903 pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3904 (unsigned long long)sh
->sector
, sh
->state
, rmw
, rcw
);
3905 set_bit(STRIPE_HANDLE
, &sh
->state
);
3906 if ((rmw
< rcw
|| (rmw
== rcw
&& conf
->rmw_level
== PARITY_PREFER_RMW
)) && rmw
> 0) {
3907 /* prefer read-modify-write, but need to get some data */
3908 if (conf
->mddev
->queue
)
3909 blk_add_trace_msg(conf
->mddev
->queue
,
3910 "raid5 rmw %llu %d",
3911 (unsigned long long)sh
->sector
, rmw
);
3912 for (i
= disks
; i
--; ) {
3913 struct r5dev
*dev
= &sh
->dev
[i
];
3914 if (test_bit(R5_InJournal
, &dev
->flags
) &&
3915 dev
->page
== dev
->orig_page
&&
3916 !test_bit(R5_LOCKED
, &sh
->dev
[sh
->pd_idx
].flags
)) {
3917 /* alloc page for prexor */
3918 struct page
*p
= alloc_page(GFP_NOIO
);
3926 * alloc_page() failed, try use
3927 * disk_info->extra_page
3929 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE
,
3930 &conf
->cache_state
)) {
3931 r5c_use_extra_page(sh
);
3935 /* extra_page in use, add to delayed_list */
3936 set_bit(STRIPE_DELAYED
, &sh
->state
);
3937 s
->waiting_extra_page
= 1;
3942 for (i
= disks
; i
--; ) {
3943 struct r5dev
*dev
= &sh
->dev
[i
];
3944 if (((dev
->towrite
&& !delay_towrite(conf
, dev
, s
)) ||
3945 i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3946 test_bit(R5_InJournal
, &dev
->flags
)) &&
3947 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3948 !(uptodate_for_rmw(dev
) ||
3949 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
3950 test_bit(R5_Insync
, &dev
->flags
)) {
3951 if (test_bit(STRIPE_PREREAD_ACTIVE
,
3953 pr_debug("Read_old block %d for r-m-w\n",
3955 set_bit(R5_LOCKED
, &dev
->flags
);
3956 set_bit(R5_Wantread
, &dev
->flags
);
3959 set_bit(STRIPE_DELAYED
, &sh
->state
);
3960 set_bit(STRIPE_HANDLE
, &sh
->state
);
3965 if ((rcw
< rmw
|| (rcw
== rmw
&& conf
->rmw_level
!= PARITY_PREFER_RMW
)) && rcw
> 0) {
3966 /* want reconstruct write, but need to get some data */
3969 for (i
= disks
; i
--; ) {
3970 struct r5dev
*dev
= &sh
->dev
[i
];
3971 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3972 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3973 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3974 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3975 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3977 if (test_bit(R5_Insync
, &dev
->flags
) &&
3978 test_bit(STRIPE_PREREAD_ACTIVE
,
3980 pr_debug("Read_old block "
3981 "%d for Reconstruct\n", i
);
3982 set_bit(R5_LOCKED
, &dev
->flags
);
3983 set_bit(R5_Wantread
, &dev
->flags
);
3987 set_bit(STRIPE_DELAYED
, &sh
->state
);
3988 set_bit(STRIPE_HANDLE
, &sh
->state
);
3992 if (rcw
&& conf
->mddev
->queue
)
3993 blk_add_trace_msg(conf
->mddev
->queue
, "raid5 rcw %llu %d %d %d",
3994 (unsigned long long)sh
->sector
,
3995 rcw
, qread
, test_bit(STRIPE_DELAYED
, &sh
->state
));
3998 if (rcw
> disks
&& rmw
> disks
&&
3999 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4000 set_bit(STRIPE_DELAYED
, &sh
->state
);
4002 /* now if nothing is locked, and if we have enough data,
4003 * we can start a write request
4005 /* since handle_stripe can be called at any time we need to handle the
4006 * case where a compute block operation has been submitted and then a
4007 * subsequent call wants to start a write request. raid_run_ops only
4008 * handles the case where compute block and reconstruct are requested
4009 * simultaneously. If this is not the case then new writes need to be
4010 * held off until the compute completes.
4012 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
4013 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
4014 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
4015 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
4019 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
4020 struct stripe_head_state
*s
, int disks
)
4022 struct r5dev
*dev
= NULL
;
4024 BUG_ON(sh
->batch_head
);
4025 set_bit(STRIPE_HANDLE
, &sh
->state
);
4027 switch (sh
->check_state
) {
4028 case check_state_idle
:
4029 /* start a new check operation if there are no failures */
4030 if (s
->failed
== 0) {
4031 BUG_ON(s
->uptodate
!= disks
);
4032 sh
->check_state
= check_state_run
;
4033 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
4034 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
4038 dev
= &sh
->dev
[s
->failed_num
[0]];
4040 case check_state_compute_result
:
4041 sh
->check_state
= check_state_idle
;
4043 dev
= &sh
->dev
[sh
->pd_idx
];
4045 /* check that a write has not made the stripe insync */
4046 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
4049 /* either failed parity check, or recovery is happening */
4050 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
4051 BUG_ON(s
->uptodate
!= disks
);
4053 set_bit(R5_LOCKED
, &dev
->flags
);
4055 set_bit(R5_Wantwrite
, &dev
->flags
);
4057 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
4058 set_bit(STRIPE_INSYNC
, &sh
->state
);
4060 case check_state_run
:
4061 break; /* we will be called again upon completion */
4062 case check_state_check_result
:
4063 sh
->check_state
= check_state_idle
;
4065 /* if a failure occurred during the check operation, leave
4066 * STRIPE_INSYNC not set and let the stripe be handled again
4071 /* handle a successful check operation, if parity is correct
4072 * we are done. Otherwise update the mismatch count and repair
4073 * parity if !MD_RECOVERY_CHECK
4075 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
4076 /* parity is correct (on disc,
4077 * not in buffer any more)
4079 set_bit(STRIPE_INSYNC
, &sh
->state
);
4081 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
4082 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
)) {
4083 /* don't try to repair!! */
4084 set_bit(STRIPE_INSYNC
, &sh
->state
);
4085 pr_warn_ratelimited("%s: mismatch sector in range "
4086 "%llu-%llu\n", mdname(conf
->mddev
),
4087 (unsigned long long) sh
->sector
,
4088 (unsigned long long) sh
->sector
+
4091 sh
->check_state
= check_state_compute_run
;
4092 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
4093 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
4094 set_bit(R5_Wantcompute
,
4095 &sh
->dev
[sh
->pd_idx
].flags
);
4096 sh
->ops
.target
= sh
->pd_idx
;
4097 sh
->ops
.target2
= -1;
4102 case check_state_compute_run
:
4105 pr_err("%s: unknown check_state: %d sector: %llu\n",
4106 __func__
, sh
->check_state
,
4107 (unsigned long long) sh
->sector
);
4112 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
4113 struct stripe_head_state
*s
,
4116 int pd_idx
= sh
->pd_idx
;
4117 int qd_idx
= sh
->qd_idx
;
4120 BUG_ON(sh
->batch_head
);
4121 set_bit(STRIPE_HANDLE
, &sh
->state
);
4123 BUG_ON(s
->failed
> 2);
4125 /* Want to check and possibly repair P and Q.
4126 * However there could be one 'failed' device, in which
4127 * case we can only check one of them, possibly using the
4128 * other to generate missing data
4131 switch (sh
->check_state
) {
4132 case check_state_idle
:
4133 /* start a new check operation if there are < 2 failures */
4134 if (s
->failed
== s
->q_failed
) {
4135 /* The only possible failed device holds Q, so it
4136 * makes sense to check P (If anything else were failed,
4137 * we would have used P to recreate it).
4139 sh
->check_state
= check_state_run
;
4141 if (!s
->q_failed
&& s
->failed
< 2) {
4142 /* Q is not failed, and we didn't use it to generate
4143 * anything, so it makes sense to check it
4145 if (sh
->check_state
== check_state_run
)
4146 sh
->check_state
= check_state_run_pq
;
4148 sh
->check_state
= check_state_run_q
;
4151 /* discard potentially stale zero_sum_result */
4152 sh
->ops
.zero_sum_result
= 0;
4154 if (sh
->check_state
== check_state_run
) {
4155 /* async_xor_zero_sum destroys the contents of P */
4156 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
4159 if (sh
->check_state
>= check_state_run
&&
4160 sh
->check_state
<= check_state_run_pq
) {
4161 /* async_syndrome_zero_sum preserves P and Q, so
4162 * no need to mark them !uptodate here
4164 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
4168 /* we have 2-disk failure */
4169 BUG_ON(s
->failed
!= 2);
4171 case check_state_compute_result
:
4172 sh
->check_state
= check_state_idle
;
4174 /* check that a write has not made the stripe insync */
4175 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
4178 /* now write out any block on a failed drive,
4179 * or P or Q if they were recomputed
4181 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
4182 if (s
->failed
== 2) {
4183 dev
= &sh
->dev
[s
->failed_num
[1]];
4185 set_bit(R5_LOCKED
, &dev
->flags
);
4186 set_bit(R5_Wantwrite
, &dev
->flags
);
4188 if (s
->failed
>= 1) {
4189 dev
= &sh
->dev
[s
->failed_num
[0]];
4191 set_bit(R5_LOCKED
, &dev
->flags
);
4192 set_bit(R5_Wantwrite
, &dev
->flags
);
4194 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
4195 dev
= &sh
->dev
[pd_idx
];
4197 set_bit(R5_LOCKED
, &dev
->flags
);
4198 set_bit(R5_Wantwrite
, &dev
->flags
);
4200 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
4201 dev
= &sh
->dev
[qd_idx
];
4203 set_bit(R5_LOCKED
, &dev
->flags
);
4204 set_bit(R5_Wantwrite
, &dev
->flags
);
4206 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
4208 set_bit(STRIPE_INSYNC
, &sh
->state
);
4210 case check_state_run
:
4211 case check_state_run_q
:
4212 case check_state_run_pq
:
4213 break; /* we will be called again upon completion */
4214 case check_state_check_result
:
4215 sh
->check_state
= check_state_idle
;
4217 /* handle a successful check operation, if parity is correct
4218 * we are done. Otherwise update the mismatch count and repair
4219 * parity if !MD_RECOVERY_CHECK
4221 if (sh
->ops
.zero_sum_result
== 0) {
4222 /* both parities are correct */
4224 set_bit(STRIPE_INSYNC
, &sh
->state
);
4226 /* in contrast to the raid5 case we can validate
4227 * parity, but still have a failure to write
4230 sh
->check_state
= check_state_compute_result
;
4231 /* Returning at this point means that we may go
4232 * off and bring p and/or q uptodate again so
4233 * we make sure to check zero_sum_result again
4234 * to verify if p or q need writeback
4238 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
4239 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
)) {
4240 /* don't try to repair!! */
4241 set_bit(STRIPE_INSYNC
, &sh
->state
);
4242 pr_warn_ratelimited("%s: mismatch sector in range "
4243 "%llu-%llu\n", mdname(conf
->mddev
),
4244 (unsigned long long) sh
->sector
,
4245 (unsigned long long) sh
->sector
+
4248 int *target
= &sh
->ops
.target
;
4250 sh
->ops
.target
= -1;
4251 sh
->ops
.target2
= -1;
4252 sh
->check_state
= check_state_compute_run
;
4253 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
4254 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
4255 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
4256 set_bit(R5_Wantcompute
,
4257 &sh
->dev
[pd_idx
].flags
);
4259 target
= &sh
->ops
.target2
;
4262 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
4263 set_bit(R5_Wantcompute
,
4264 &sh
->dev
[qd_idx
].flags
);
4271 case check_state_compute_run
:
4274 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4275 __func__
, sh
->check_state
,
4276 (unsigned long long) sh
->sector
);
4281 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
4285 /* We have read all the blocks in this stripe and now we need to
4286 * copy some of them into a target stripe for expand.
4288 struct dma_async_tx_descriptor
*tx
= NULL
;
4289 BUG_ON(sh
->batch_head
);
4290 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
4291 for (i
= 0; i
< sh
->disks
; i
++)
4292 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
4294 struct stripe_head
*sh2
;
4295 struct async_submit_ctl submit
;
4297 sector_t bn
= raid5_compute_blocknr(sh
, i
, 1);
4298 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
4300 sh2
= raid5_get_active_stripe(conf
, s
, 0, 1, 1);
4302 /* so far only the early blocks of this stripe
4303 * have been requested. When later blocks
4304 * get requested, we will try again
4307 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
4308 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
4309 /* must have already done this block */
4310 raid5_release_stripe(sh2
);
4314 /* place all the copies on one channel */
4315 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
4316 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
4317 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
4320 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
4321 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
4322 for (j
= 0; j
< conf
->raid_disks
; j
++)
4323 if (j
!= sh2
->pd_idx
&&
4325 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
4327 if (j
== conf
->raid_disks
) {
4328 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
4329 set_bit(STRIPE_HANDLE
, &sh2
->state
);
4331 raid5_release_stripe(sh2
);
4334 /* done submitting copies, wait for them to complete */
4335 async_tx_quiesce(&tx
);
4339 * handle_stripe - do things to a stripe.
4341 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4342 * state of various bits to see what needs to be done.
4344 * return some read requests which now have data
4345 * return some write requests which are safely on storage
4346 * schedule a read on some buffers
4347 * schedule a write of some buffers
4348 * return confirmation of parity correctness
4352 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
4354 struct r5conf
*conf
= sh
->raid_conf
;
4355 int disks
= sh
->disks
;
4358 int do_recovery
= 0;
4360 memset(s
, 0, sizeof(*s
));
4362 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
) && !sh
->batch_head
;
4363 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
) && !sh
->batch_head
;
4364 s
->failed_num
[0] = -1;
4365 s
->failed_num
[1] = -1;
4366 s
->log_failed
= r5l_log_disk_error(conf
);
4368 /* Now to look around and see what can be done */
4370 for (i
=disks
; i
--; ) {
4371 struct md_rdev
*rdev
;
4378 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4380 dev
->toread
, dev
->towrite
, dev
->written
);
4381 /* maybe we can reply to a read
4383 * new wantfill requests are only permitted while
4384 * ops_complete_biofill is guaranteed to be inactive
4386 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
4387 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
4388 set_bit(R5_Wantfill
, &dev
->flags
);
4390 /* now count some things */
4391 if (test_bit(R5_LOCKED
, &dev
->flags
))
4393 if (test_bit(R5_UPTODATE
, &dev
->flags
))
4395 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
4397 BUG_ON(s
->compute
> 2);
4400 if (test_bit(R5_Wantfill
, &dev
->flags
))
4402 else if (dev
->toread
)
4406 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
4411 /* Prefer to use the replacement for reads, but only
4412 * if it is recovered enough and has no bad blocks.
4414 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
4415 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
4416 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
4417 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4418 &first_bad
, &bad_sectors
))
4419 set_bit(R5_ReadRepl
, &dev
->flags
);
4421 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4422 set_bit(R5_NeedReplace
, &dev
->flags
);
4424 clear_bit(R5_NeedReplace
, &dev
->flags
);
4425 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
4426 clear_bit(R5_ReadRepl
, &dev
->flags
);
4428 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
4431 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4432 &first_bad
, &bad_sectors
);
4433 if (s
->blocked_rdev
== NULL
4434 && (test_bit(Blocked
, &rdev
->flags
)
4437 set_bit(BlockedBadBlocks
,
4439 s
->blocked_rdev
= rdev
;
4440 atomic_inc(&rdev
->nr_pending
);
4443 clear_bit(R5_Insync
, &dev
->flags
);
4447 /* also not in-sync */
4448 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
4449 test_bit(R5_UPTODATE
, &dev
->flags
)) {
4450 /* treat as in-sync, but with a read error
4451 * which we can now try to correct
4453 set_bit(R5_Insync
, &dev
->flags
);
4454 set_bit(R5_ReadError
, &dev
->flags
);
4456 } else if (test_bit(In_sync
, &rdev
->flags
))
4457 set_bit(R5_Insync
, &dev
->flags
);
4458 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
4459 /* in sync if before recovery_offset */
4460 set_bit(R5_Insync
, &dev
->flags
);
4461 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
4462 test_bit(R5_Expanded
, &dev
->flags
))
4463 /* If we've reshaped into here, we assume it is Insync.
4464 * We will shortly update recovery_offset to make
4467 set_bit(R5_Insync
, &dev
->flags
);
4469 if (test_bit(R5_WriteError
, &dev
->flags
)) {
4470 /* This flag does not apply to '.replacement'
4471 * only to .rdev, so make sure to check that*/
4472 struct md_rdev
*rdev2
= rcu_dereference(
4473 conf
->disks
[i
].rdev
);
4475 clear_bit(R5_Insync
, &dev
->flags
);
4476 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4477 s
->handle_bad_blocks
= 1;
4478 atomic_inc(&rdev2
->nr_pending
);
4480 clear_bit(R5_WriteError
, &dev
->flags
);
4482 if (test_bit(R5_MadeGood
, &dev
->flags
)) {
4483 /* This flag does not apply to '.replacement'
4484 * only to .rdev, so make sure to check that*/
4485 struct md_rdev
*rdev2
= rcu_dereference(
4486 conf
->disks
[i
].rdev
);
4487 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4488 s
->handle_bad_blocks
= 1;
4489 atomic_inc(&rdev2
->nr_pending
);
4491 clear_bit(R5_MadeGood
, &dev
->flags
);
4493 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
4494 struct md_rdev
*rdev2
= rcu_dereference(
4495 conf
->disks
[i
].replacement
);
4496 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4497 s
->handle_bad_blocks
= 1;
4498 atomic_inc(&rdev2
->nr_pending
);
4500 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
4502 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4503 /* The ReadError flag will just be confusing now */
4504 clear_bit(R5_ReadError
, &dev
->flags
);
4505 clear_bit(R5_ReWrite
, &dev
->flags
);
4507 if (test_bit(R5_ReadError
, &dev
->flags
))
4508 clear_bit(R5_Insync
, &dev
->flags
);
4509 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4511 s
->failed_num
[s
->failed
] = i
;
4513 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4517 if (test_bit(R5_InJournal
, &dev
->flags
))
4519 if (test_bit(R5_InJournal
, &dev
->flags
) && dev
->written
)
4522 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
4523 /* If there is a failed device being replaced,
4524 * we must be recovering.
4525 * else if we are after recovery_cp, we must be syncing
4526 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4527 * else we can only be replacing
4528 * sync and recovery both need to read all devices, and so
4529 * use the same flag.
4532 sh
->sector
>= conf
->mddev
->recovery_cp
||
4533 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
4541 static int clear_batch_ready(struct stripe_head
*sh
)
4543 /* Return '1' if this is a member of batch, or
4544 * '0' if it is a lone stripe or a head which can now be
4547 struct stripe_head
*tmp
;
4548 if (!test_and_clear_bit(STRIPE_BATCH_READY
, &sh
->state
))
4549 return (sh
->batch_head
&& sh
->batch_head
!= sh
);
4550 spin_lock(&sh
->stripe_lock
);
4551 if (!sh
->batch_head
) {
4552 spin_unlock(&sh
->stripe_lock
);
4557 * this stripe could be added to a batch list before we check
4558 * BATCH_READY, skips it
4560 if (sh
->batch_head
!= sh
) {
4561 spin_unlock(&sh
->stripe_lock
);
4564 spin_lock(&sh
->batch_lock
);
4565 list_for_each_entry(tmp
, &sh
->batch_list
, batch_list
)
4566 clear_bit(STRIPE_BATCH_READY
, &tmp
->state
);
4567 spin_unlock(&sh
->batch_lock
);
4568 spin_unlock(&sh
->stripe_lock
);
4571 * BATCH_READY is cleared, no new stripes can be added.
4572 * batch_list can be accessed without lock
4577 static void break_stripe_batch_list(struct stripe_head
*head_sh
,
4578 unsigned long handle_flags
)
4580 struct stripe_head
*sh
, *next
;
4584 list_for_each_entry_safe(sh
, next
, &head_sh
->batch_list
, batch_list
) {
4586 list_del_init(&sh
->batch_list
);
4588 WARN_ONCE(sh
->state
& ((1 << STRIPE_ACTIVE
) |
4589 (1 << STRIPE_SYNCING
) |
4590 (1 << STRIPE_REPLACED
) |
4591 (1 << STRIPE_DELAYED
) |
4592 (1 << STRIPE_BIT_DELAY
) |
4593 (1 << STRIPE_FULL_WRITE
) |
4594 (1 << STRIPE_BIOFILL_RUN
) |
4595 (1 << STRIPE_COMPUTE_RUN
) |
4596 (1 << STRIPE_OPS_REQ_PENDING
) |
4597 (1 << STRIPE_DISCARD
) |
4598 (1 << STRIPE_BATCH_READY
) |
4599 (1 << STRIPE_BATCH_ERR
) |
4600 (1 << STRIPE_BITMAP_PENDING
)),
4601 "stripe state: %lx\n", sh
->state
);
4602 WARN_ONCE(head_sh
->state
& ((1 << STRIPE_DISCARD
) |
4603 (1 << STRIPE_REPLACED
)),
4604 "head stripe state: %lx\n", head_sh
->state
);
4606 set_mask_bits(&sh
->state
, ~(STRIPE_EXPAND_SYNC_FLAGS
|
4607 (1 << STRIPE_PREREAD_ACTIVE
) |
4608 (1 << STRIPE_DEGRADED
) |
4609 (1 << STRIPE_ON_UNPLUG_LIST
)),
4610 head_sh
->state
& (1 << STRIPE_INSYNC
));
4612 sh
->check_state
= head_sh
->check_state
;
4613 sh
->reconstruct_state
= head_sh
->reconstruct_state
;
4614 for (i
= 0; i
< sh
->disks
; i
++) {
4615 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
4617 sh
->dev
[i
].flags
= head_sh
->dev
[i
].flags
&
4618 (~((1 << R5_WriteError
) | (1 << R5_Overlap
)));
4620 spin_lock_irq(&sh
->stripe_lock
);
4621 sh
->batch_head
= NULL
;
4622 spin_unlock_irq(&sh
->stripe_lock
);
4623 if (handle_flags
== 0 ||
4624 sh
->state
& handle_flags
)
4625 set_bit(STRIPE_HANDLE
, &sh
->state
);
4626 raid5_release_stripe(sh
);
4628 spin_lock_irq(&head_sh
->stripe_lock
);
4629 head_sh
->batch_head
= NULL
;
4630 spin_unlock_irq(&head_sh
->stripe_lock
);
4631 for (i
= 0; i
< head_sh
->disks
; i
++)
4632 if (test_and_clear_bit(R5_Overlap
, &head_sh
->dev
[i
].flags
))
4634 if (head_sh
->state
& handle_flags
)
4635 set_bit(STRIPE_HANDLE
, &head_sh
->state
);
4638 wake_up(&head_sh
->raid_conf
->wait_for_overlap
);
4641 static void handle_stripe(struct stripe_head
*sh
)
4643 struct stripe_head_state s
;
4644 struct r5conf
*conf
= sh
->raid_conf
;
4647 int disks
= sh
->disks
;
4648 struct r5dev
*pdev
, *qdev
;
4650 clear_bit(STRIPE_HANDLE
, &sh
->state
);
4651 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
4652 /* already being handled, ensure it gets handled
4653 * again when current action finishes */
4654 set_bit(STRIPE_HANDLE
, &sh
->state
);
4658 if (clear_batch_ready(sh
) ) {
4659 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
4663 if (test_and_clear_bit(STRIPE_BATCH_ERR
, &sh
->state
))
4664 break_stripe_batch_list(sh
, 0);
4666 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
) && !sh
->batch_head
) {
4667 spin_lock(&sh
->stripe_lock
);
4669 * Cannot process 'sync' concurrently with 'discard'.
4670 * Flush data in r5cache before 'sync'.
4672 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE
, &sh
->state
) &&
4673 !test_bit(STRIPE_R5C_FULL_STRIPE
, &sh
->state
) &&
4674 !test_bit(STRIPE_DISCARD
, &sh
->state
) &&
4675 test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
4676 set_bit(STRIPE_SYNCING
, &sh
->state
);
4677 clear_bit(STRIPE_INSYNC
, &sh
->state
);
4678 clear_bit(STRIPE_REPLACED
, &sh
->state
);
4680 spin_unlock(&sh
->stripe_lock
);
4682 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4684 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4685 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4686 (unsigned long long)sh
->sector
, sh
->state
,
4687 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
4688 sh
->check_state
, sh
->reconstruct_state
);
4690 analyse_stripe(sh
, &s
);
4692 if (test_bit(STRIPE_LOG_TRAPPED
, &sh
->state
))
4695 if (s
.handle_bad_blocks
||
4696 test_bit(MD_SB_CHANGE_PENDING
, &conf
->mddev
->sb_flags
)) {
4697 set_bit(STRIPE_HANDLE
, &sh
->state
);
4701 if (unlikely(s
.blocked_rdev
)) {
4702 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
4703 s
.replacing
|| s
.to_write
|| s
.written
) {
4704 set_bit(STRIPE_HANDLE
, &sh
->state
);
4707 /* There is nothing for the blocked_rdev to block */
4708 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
4709 s
.blocked_rdev
= NULL
;
4712 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
4713 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
4714 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
4717 pr_debug("locked=%d uptodate=%d to_read=%d"
4718 " to_write=%d failed=%d failed_num=%d,%d\n",
4719 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
4720 s
.failed_num
[0], s
.failed_num
[1]);
4722 * check if the array has lost more than max_degraded devices and,
4723 * if so, some requests might need to be failed.
4725 * When journal device failed (log_failed), we will only process
4726 * the stripe if there is data need write to raid disks
4728 if (s
.failed
> conf
->max_degraded
||
4729 (s
.log_failed
&& s
.injournal
== 0)) {
4730 sh
->check_state
= 0;
4731 sh
->reconstruct_state
= 0;
4732 break_stripe_batch_list(sh
, 0);
4733 if (s
.to_read
+s
.to_write
+s
.written
)
4734 handle_failed_stripe(conf
, sh
, &s
, disks
);
4735 if (s
.syncing
+ s
.replacing
)
4736 handle_failed_sync(conf
, sh
, &s
);
4739 /* Now we check to see if any write operations have recently
4743 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
4745 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
4746 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
4747 sh
->reconstruct_state
= reconstruct_state_idle
;
4749 /* All the 'written' buffers and the parity block are ready to
4750 * be written back to disk
4752 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
) &&
4753 !test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
));
4754 BUG_ON(sh
->qd_idx
>= 0 &&
4755 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
) &&
4756 !test_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
));
4757 for (i
= disks
; i
--; ) {
4758 struct r5dev
*dev
= &sh
->dev
[i
];
4759 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
4760 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
4761 dev
->written
|| test_bit(R5_InJournal
,
4763 pr_debug("Writing block %d\n", i
);
4764 set_bit(R5_Wantwrite
, &dev
->flags
);
4769 if (!test_bit(R5_Insync
, &dev
->flags
) ||
4770 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
4772 set_bit(STRIPE_INSYNC
, &sh
->state
);
4775 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4776 s
.dec_preread_active
= 1;
4780 * might be able to return some write requests if the parity blocks
4781 * are safe, or on a failed drive
4783 pdev
= &sh
->dev
[sh
->pd_idx
];
4784 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
4785 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
4786 qdev
= &sh
->dev
[sh
->qd_idx
];
4787 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
4788 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
4792 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
4793 && !test_bit(R5_LOCKED
, &pdev
->flags
)
4794 && (test_bit(R5_UPTODATE
, &pdev
->flags
) ||
4795 test_bit(R5_Discard
, &pdev
->flags
))))) &&
4796 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
4797 && !test_bit(R5_LOCKED
, &qdev
->flags
)
4798 && (test_bit(R5_UPTODATE
, &qdev
->flags
) ||
4799 test_bit(R5_Discard
, &qdev
->flags
))))))
4800 handle_stripe_clean_event(conf
, sh
, disks
);
4803 r5c_handle_cached_data_endio(conf
, sh
, disks
);
4804 log_stripe_write_finished(sh
);
4806 /* Now we might consider reading some blocks, either to check/generate
4807 * parity, or to satisfy requests
4808 * or to load a block that is being partially written.
4810 if (s
.to_read
|| s
.non_overwrite
4811 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
4812 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
4815 handle_stripe_fill(sh
, &s
, disks
);
4818 * When the stripe finishes full journal write cycle (write to journal
4819 * and raid disk), this is the clean up procedure so it is ready for
4822 r5c_finish_stripe_write_out(conf
, sh
, &s
);
4825 * Now to consider new write requests, cache write back and what else,
4826 * if anything should be read. We do not handle new writes when:
4827 * 1/ A 'write' operation (copy+xor) is already in flight.
4828 * 2/ A 'check' operation is in flight, as it may clobber the parity
4830 * 3/ A r5c cache log write is in flight.
4833 if (!sh
->reconstruct_state
&& !sh
->check_state
&& !sh
->log_io
) {
4834 if (!r5c_is_writeback(conf
->log
)) {
4836 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
4837 } else { /* write back cache */
4840 /* First, try handle writes in caching phase */
4842 ret
= r5c_try_caching_write(conf
, sh
, &s
,
4845 * If caching phase failed: ret == -EAGAIN
4847 * stripe under reclaim: !caching && injournal
4849 * fall back to handle_stripe_dirtying()
4851 if (ret
== -EAGAIN
||
4852 /* stripe under reclaim: !caching && injournal */
4853 (!test_bit(STRIPE_R5C_CACHING
, &sh
->state
) &&
4855 ret
= handle_stripe_dirtying(conf
, sh
, &s
,
4863 /* maybe we need to check and possibly fix the parity for this stripe
4864 * Any reads will already have been scheduled, so we just see if enough
4865 * data is available. The parity check is held off while parity
4866 * dependent operations are in flight.
4868 if (sh
->check_state
||
4869 (s
.syncing
&& s
.locked
== 0 &&
4870 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4871 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
4872 if (conf
->level
== 6)
4873 handle_parity_checks6(conf
, sh
, &s
, disks
);
4875 handle_parity_checks5(conf
, sh
, &s
, disks
);
4878 if ((s
.replacing
|| s
.syncing
) && s
.locked
== 0
4879 && !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)
4880 && !test_bit(STRIPE_REPLACED
, &sh
->state
)) {
4881 /* Write out to replacement devices where possible */
4882 for (i
= 0; i
< conf
->raid_disks
; i
++)
4883 if (test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
4884 WARN_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
4885 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
4886 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4890 set_bit(STRIPE_INSYNC
, &sh
->state
);
4891 set_bit(STRIPE_REPLACED
, &sh
->state
);
4893 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
4894 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4895 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
4896 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4897 clear_bit(STRIPE_SYNCING
, &sh
->state
);
4898 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
4899 wake_up(&conf
->wait_for_overlap
);
4902 /* If the failed drives are just a ReadError, then we might need
4903 * to progress the repair/check process
4905 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
4906 for (i
= 0; i
< s
.failed
; i
++) {
4907 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
4908 if (test_bit(R5_ReadError
, &dev
->flags
)
4909 && !test_bit(R5_LOCKED
, &dev
->flags
)
4910 && test_bit(R5_UPTODATE
, &dev
->flags
)
4912 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
4913 set_bit(R5_Wantwrite
, &dev
->flags
);
4914 set_bit(R5_ReWrite
, &dev
->flags
);
4915 set_bit(R5_LOCKED
, &dev
->flags
);
4918 /* let's read it back */
4919 set_bit(R5_Wantread
, &dev
->flags
);
4920 set_bit(R5_LOCKED
, &dev
->flags
);
4926 /* Finish reconstruct operations initiated by the expansion process */
4927 if (sh
->reconstruct_state
== reconstruct_state_result
) {
4928 struct stripe_head
*sh_src
4929 = raid5_get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
4930 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
4931 /* sh cannot be written until sh_src has been read.
4932 * so arrange for sh to be delayed a little
4934 set_bit(STRIPE_DELAYED
, &sh
->state
);
4935 set_bit(STRIPE_HANDLE
, &sh
->state
);
4936 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
4938 atomic_inc(&conf
->preread_active_stripes
);
4939 raid5_release_stripe(sh_src
);
4943 raid5_release_stripe(sh_src
);
4945 sh
->reconstruct_state
= reconstruct_state_idle
;
4946 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
4947 for (i
= conf
->raid_disks
; i
--; ) {
4948 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
4949 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4954 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
4955 !sh
->reconstruct_state
) {
4956 /* Need to write out all blocks after computing parity */
4957 sh
->disks
= conf
->raid_disks
;
4958 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
4959 schedule_reconstruction(sh
, &s
, 1, 1);
4960 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
4961 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4962 atomic_dec(&conf
->reshape_stripes
);
4963 wake_up(&conf
->wait_for_overlap
);
4964 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4967 if (s
.expanding
&& s
.locked
== 0 &&
4968 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
4969 handle_stripe_expansion(conf
, sh
);
4972 /* wait for this device to become unblocked */
4973 if (unlikely(s
.blocked_rdev
)) {
4974 if (conf
->mddev
->external
)
4975 md_wait_for_blocked_rdev(s
.blocked_rdev
,
4978 /* Internal metadata will immediately
4979 * be written by raid5d, so we don't
4980 * need to wait here.
4982 rdev_dec_pending(s
.blocked_rdev
,
4986 if (s
.handle_bad_blocks
)
4987 for (i
= disks
; i
--; ) {
4988 struct md_rdev
*rdev
;
4989 struct r5dev
*dev
= &sh
->dev
[i
];
4990 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
4991 /* We own a safe reference to the rdev */
4992 rdev
= conf
->disks
[i
].rdev
;
4993 if (!rdev_set_badblocks(rdev
, sh
->sector
,
4995 md_error(conf
->mddev
, rdev
);
4996 rdev_dec_pending(rdev
, conf
->mddev
);
4998 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
4999 rdev
= conf
->disks
[i
].rdev
;
5000 rdev_clear_badblocks(rdev
, sh
->sector
,
5002 rdev_dec_pending(rdev
, conf
->mddev
);
5004 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
5005 rdev
= conf
->disks
[i
].replacement
;
5007 /* rdev have been moved down */
5008 rdev
= conf
->disks
[i
].rdev
;
5009 rdev_clear_badblocks(rdev
, sh
->sector
,
5011 rdev_dec_pending(rdev
, conf
->mddev
);
5016 raid_run_ops(sh
, s
.ops_request
);
5020 if (s
.dec_preread_active
) {
5021 /* We delay this until after ops_run_io so that if make_request
5022 * is waiting on a flush, it won't continue until the writes
5023 * have actually been submitted.
5025 atomic_dec(&conf
->preread_active_stripes
);
5026 if (atomic_read(&conf
->preread_active_stripes
) <
5028 md_wakeup_thread(conf
->mddev
->thread
);
5031 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
5034 static void raid5_activate_delayed(struct r5conf
*conf
)
5036 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
5037 while (!list_empty(&conf
->delayed_list
)) {
5038 struct list_head
*l
= conf
->delayed_list
.next
;
5039 struct stripe_head
*sh
;
5040 sh
= list_entry(l
, struct stripe_head
, lru
);
5042 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5043 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5044 atomic_inc(&conf
->preread_active_stripes
);
5045 list_add_tail(&sh
->lru
, &conf
->hold_list
);
5046 raid5_wakeup_stripe_thread(sh
);
5051 static void activate_bit_delay(struct r5conf
*conf
,
5052 struct list_head
*temp_inactive_list
)
5054 /* device_lock is held */
5055 struct list_head head
;
5056 list_add(&head
, &conf
->bitmap_list
);
5057 list_del_init(&conf
->bitmap_list
);
5058 while (!list_empty(&head
)) {
5059 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
5061 list_del_init(&sh
->lru
);
5062 atomic_inc(&sh
->count
);
5063 hash
= sh
->hash_lock_index
;
5064 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
5068 static int raid5_congested(struct mddev
*mddev
, int bits
)
5070 struct r5conf
*conf
= mddev
->private;
5072 /* No difference between reads and writes. Just check
5073 * how busy the stripe_cache is
5076 if (test_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
))
5079 /* Also checks whether there is pressure on r5cache log space */
5080 if (test_bit(R5C_LOG_TIGHT
, &conf
->cache_state
))
5084 if (atomic_read(&conf
->empty_inactive_list_nr
))
5090 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
5092 struct r5conf
*conf
= mddev
->private;
5093 sector_t sector
= bio
->bi_iter
.bi_sector
;
5094 unsigned int chunk_sectors
;
5095 unsigned int bio_sectors
= bio_sectors(bio
);
5097 WARN_ON_ONCE(bio
->bi_partno
);
5099 chunk_sectors
= min(conf
->chunk_sectors
, conf
->prev_chunk_sectors
);
5100 return chunk_sectors
>=
5101 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
5105 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
5106 * later sampled by raid5d.
5108 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
5110 unsigned long flags
;
5112 spin_lock_irqsave(&conf
->device_lock
, flags
);
5114 bi
->bi_next
= conf
->retry_read_aligned_list
;
5115 conf
->retry_read_aligned_list
= bi
;
5117 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5118 md_wakeup_thread(conf
->mddev
->thread
);
5121 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
,
5122 unsigned int *offset
)
5126 bi
= conf
->retry_read_aligned
;
5128 *offset
= conf
->retry_read_offset
;
5129 conf
->retry_read_aligned
= NULL
;
5132 bi
= conf
->retry_read_aligned_list
;
5134 conf
->retry_read_aligned_list
= bi
->bi_next
;
5143 * The "raid5_align_endio" should check if the read succeeded and if it
5144 * did, call bio_endio on the original bio (having bio_put the new bio
5146 * If the read failed..
5148 static void raid5_align_endio(struct bio
*bi
)
5150 struct bio
* raid_bi
= bi
->bi_private
;
5151 struct mddev
*mddev
;
5152 struct r5conf
*conf
;
5153 struct md_rdev
*rdev
;
5154 blk_status_t error
= bi
->bi_status
;
5158 rdev
= (void*)raid_bi
->bi_next
;
5159 raid_bi
->bi_next
= NULL
;
5160 mddev
= rdev
->mddev
;
5161 conf
= mddev
->private;
5163 rdev_dec_pending(rdev
, conf
->mddev
);
5167 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
5168 wake_up(&conf
->wait_for_quiescent
);
5172 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5174 add_bio_to_retry(raid_bi
, conf
);
5177 static int raid5_read_one_chunk(struct mddev
*mddev
, struct bio
*raid_bio
)
5179 struct r5conf
*conf
= mddev
->private;
5181 struct bio
* align_bi
;
5182 struct md_rdev
*rdev
;
5183 sector_t end_sector
;
5185 if (!in_chunk_boundary(mddev
, raid_bio
)) {
5186 pr_debug("%s: non aligned\n", __func__
);
5190 * use bio_clone_fast to make a copy of the bio
5192 align_bi
= bio_clone_fast(raid_bio
, GFP_NOIO
, mddev
->bio_set
);
5196 * set bi_end_io to a new function, and set bi_private to the
5199 align_bi
->bi_end_io
= raid5_align_endio
;
5200 align_bi
->bi_private
= raid_bio
;
5204 align_bi
->bi_iter
.bi_sector
=
5205 raid5_compute_sector(conf
, raid_bio
->bi_iter
.bi_sector
,
5208 end_sector
= bio_end_sector(align_bi
);
5210 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
5211 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
5212 rdev
->recovery_offset
< end_sector
) {
5213 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
5215 (test_bit(Faulty
, &rdev
->flags
) ||
5216 !(test_bit(In_sync
, &rdev
->flags
) ||
5217 rdev
->recovery_offset
>= end_sector
)))
5221 if (r5c_big_stripe_cached(conf
, align_bi
->bi_iter
.bi_sector
)) {
5231 atomic_inc(&rdev
->nr_pending
);
5233 raid_bio
->bi_next
= (void*)rdev
;
5234 bio_set_dev(align_bi
, rdev
->bdev
);
5235 bio_clear_flag(align_bi
, BIO_SEG_VALID
);
5237 if (is_badblock(rdev
, align_bi
->bi_iter
.bi_sector
,
5238 bio_sectors(align_bi
),
5239 &first_bad
, &bad_sectors
)) {
5241 rdev_dec_pending(rdev
, mddev
);
5245 /* No reshape active, so we can trust rdev->data_offset */
5246 align_bi
->bi_iter
.bi_sector
+= rdev
->data_offset
;
5248 spin_lock_irq(&conf
->device_lock
);
5249 wait_event_lock_irq(conf
->wait_for_quiescent
,
5252 atomic_inc(&conf
->active_aligned_reads
);
5253 spin_unlock_irq(&conf
->device_lock
);
5256 trace_block_bio_remap(align_bi
->bi_disk
->queue
,
5257 align_bi
, disk_devt(mddev
->gendisk
),
5258 raid_bio
->bi_iter
.bi_sector
);
5259 generic_make_request(align_bi
);
5268 static struct bio
*chunk_aligned_read(struct mddev
*mddev
, struct bio
*raid_bio
)
5271 sector_t sector
= raid_bio
->bi_iter
.bi_sector
;
5272 unsigned chunk_sects
= mddev
->chunk_sectors
;
5273 unsigned sectors
= chunk_sects
- (sector
& (chunk_sects
-1));
5275 if (sectors
< bio_sectors(raid_bio
)) {
5276 struct r5conf
*conf
= mddev
->private;
5277 split
= bio_split(raid_bio
, sectors
, GFP_NOIO
, conf
->bio_split
);
5278 bio_chain(split
, raid_bio
);
5279 generic_make_request(raid_bio
);
5283 if (!raid5_read_one_chunk(mddev
, raid_bio
))
5289 /* __get_priority_stripe - get the next stripe to process
5291 * Full stripe writes are allowed to pass preread active stripes up until
5292 * the bypass_threshold is exceeded. In general the bypass_count
5293 * increments when the handle_list is handled before the hold_list; however, it
5294 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5295 * stripe with in flight i/o. The bypass_count will be reset when the
5296 * head of the hold_list has changed, i.e. the head was promoted to the
5299 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
, int group
)
5301 struct stripe_head
*sh
, *tmp
;
5302 struct list_head
*handle_list
= NULL
;
5303 struct r5worker_group
*wg
;
5304 bool second_try
= !r5c_is_writeback(conf
->log
) &&
5305 !r5l_log_disk_error(conf
);
5306 bool try_loprio
= test_bit(R5C_LOG_TIGHT
, &conf
->cache_state
) ||
5307 r5l_log_disk_error(conf
);
5312 if (conf
->worker_cnt_per_group
== 0) {
5313 handle_list
= try_loprio
? &conf
->loprio_list
:
5315 } else if (group
!= ANY_GROUP
) {
5316 handle_list
= try_loprio
? &conf
->worker_groups
[group
].loprio_list
:
5317 &conf
->worker_groups
[group
].handle_list
;
5318 wg
= &conf
->worker_groups
[group
];
5321 for (i
= 0; i
< conf
->group_cnt
; i
++) {
5322 handle_list
= try_loprio
? &conf
->worker_groups
[i
].loprio_list
:
5323 &conf
->worker_groups
[i
].handle_list
;
5324 wg
= &conf
->worker_groups
[i
];
5325 if (!list_empty(handle_list
))
5330 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5332 list_empty(handle_list
) ? "empty" : "busy",
5333 list_empty(&conf
->hold_list
) ? "empty" : "busy",
5334 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
5336 if (!list_empty(handle_list
)) {
5337 sh
= list_entry(handle_list
->next
, typeof(*sh
), lru
);
5339 if (list_empty(&conf
->hold_list
))
5340 conf
->bypass_count
= 0;
5341 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
5342 if (conf
->hold_list
.next
== conf
->last_hold
)
5343 conf
->bypass_count
++;
5345 conf
->last_hold
= conf
->hold_list
.next
;
5346 conf
->bypass_count
-= conf
->bypass_threshold
;
5347 if (conf
->bypass_count
< 0)
5348 conf
->bypass_count
= 0;
5351 } else if (!list_empty(&conf
->hold_list
) &&
5352 ((conf
->bypass_threshold
&&
5353 conf
->bypass_count
> conf
->bypass_threshold
) ||
5354 atomic_read(&conf
->pending_full_writes
) == 0)) {
5356 list_for_each_entry(tmp
, &conf
->hold_list
, lru
) {
5357 if (conf
->worker_cnt_per_group
== 0 ||
5358 group
== ANY_GROUP
||
5359 !cpu_online(tmp
->cpu
) ||
5360 cpu_to_group(tmp
->cpu
) == group
) {
5367 conf
->bypass_count
-= conf
->bypass_threshold
;
5368 if (conf
->bypass_count
< 0)
5369 conf
->bypass_count
= 0;
5378 try_loprio
= !try_loprio
;
5386 list_del_init(&sh
->lru
);
5387 BUG_ON(atomic_inc_return(&sh
->count
) != 1);
5391 struct raid5_plug_cb
{
5392 struct blk_plug_cb cb
;
5393 struct list_head list
;
5394 struct list_head temp_inactive_list
[NR_STRIPE_HASH_LOCKS
];
5397 static void raid5_unplug(struct blk_plug_cb
*blk_cb
, bool from_schedule
)
5399 struct raid5_plug_cb
*cb
= container_of(
5400 blk_cb
, struct raid5_plug_cb
, cb
);
5401 struct stripe_head
*sh
;
5402 struct mddev
*mddev
= cb
->cb
.data
;
5403 struct r5conf
*conf
= mddev
->private;
5407 if (cb
->list
.next
&& !list_empty(&cb
->list
)) {
5408 spin_lock_irq(&conf
->device_lock
);
5409 while (!list_empty(&cb
->list
)) {
5410 sh
= list_first_entry(&cb
->list
, struct stripe_head
, lru
);
5411 list_del_init(&sh
->lru
);
5413 * avoid race release_stripe_plug() sees
5414 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5415 * is still in our list
5417 smp_mb__before_atomic();
5418 clear_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
);
5420 * STRIPE_ON_RELEASE_LIST could be set here. In that
5421 * case, the count is always > 1 here
5423 hash
= sh
->hash_lock_index
;
5424 __release_stripe(conf
, sh
, &cb
->temp_inactive_list
[hash
]);
5427 spin_unlock_irq(&conf
->device_lock
);
5429 release_inactive_stripe_list(conf
, cb
->temp_inactive_list
,
5430 NR_STRIPE_HASH_LOCKS
);
5432 trace_block_unplug(mddev
->queue
, cnt
, !from_schedule
);
5436 static void release_stripe_plug(struct mddev
*mddev
,
5437 struct stripe_head
*sh
)
5439 struct blk_plug_cb
*blk_cb
= blk_check_plugged(
5440 raid5_unplug
, mddev
,
5441 sizeof(struct raid5_plug_cb
));
5442 struct raid5_plug_cb
*cb
;
5445 raid5_release_stripe(sh
);
5449 cb
= container_of(blk_cb
, struct raid5_plug_cb
, cb
);
5451 if (cb
->list
.next
== NULL
) {
5453 INIT_LIST_HEAD(&cb
->list
);
5454 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5455 INIT_LIST_HEAD(cb
->temp_inactive_list
+ i
);
5458 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
))
5459 list_add_tail(&sh
->lru
, &cb
->list
);
5461 raid5_release_stripe(sh
);
5464 static void make_discard_request(struct mddev
*mddev
, struct bio
*bi
)
5466 struct r5conf
*conf
= mddev
->private;
5467 sector_t logical_sector
, last_sector
;
5468 struct stripe_head
*sh
;
5471 if (mddev
->reshape_position
!= MaxSector
)
5472 /* Skip discard while reshape is happening */
5475 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5476 last_sector
= bi
->bi_iter
.bi_sector
+ (bi
->bi_iter
.bi_size
>>9);
5480 stripe_sectors
= conf
->chunk_sectors
*
5481 (conf
->raid_disks
- conf
->max_degraded
);
5482 logical_sector
= DIV_ROUND_UP_SECTOR_T(logical_sector
,
5484 sector_div(last_sector
, stripe_sectors
);
5486 logical_sector
*= conf
->chunk_sectors
;
5487 last_sector
*= conf
->chunk_sectors
;
5489 for (; logical_sector
< last_sector
;
5490 logical_sector
+= STRIPE_SECTORS
) {
5494 sh
= raid5_get_active_stripe(conf
, logical_sector
, 0, 0, 0);
5495 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5496 TASK_UNINTERRUPTIBLE
);
5497 set_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5498 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
5499 raid5_release_stripe(sh
);
5503 clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5504 spin_lock_irq(&sh
->stripe_lock
);
5505 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5506 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5508 if (sh
->dev
[d
].towrite
|| sh
->dev
[d
].toread
) {
5509 set_bit(R5_Overlap
, &sh
->dev
[d
].flags
);
5510 spin_unlock_irq(&sh
->stripe_lock
);
5511 raid5_release_stripe(sh
);
5516 set_bit(STRIPE_DISCARD
, &sh
->state
);
5517 finish_wait(&conf
->wait_for_overlap
, &w
);
5518 sh
->overwrite_disks
= 0;
5519 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5520 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5522 sh
->dev
[d
].towrite
= bi
;
5523 set_bit(R5_OVERWRITE
, &sh
->dev
[d
].flags
);
5524 bio_inc_remaining(bi
);
5525 md_write_inc(mddev
, bi
);
5526 sh
->overwrite_disks
++;
5528 spin_unlock_irq(&sh
->stripe_lock
);
5529 if (conf
->mddev
->bitmap
) {
5531 d
< conf
->raid_disks
- conf
->max_degraded
;
5533 bitmap_startwrite(mddev
->bitmap
,
5537 sh
->bm_seq
= conf
->seq_flush
+ 1;
5538 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
5541 set_bit(STRIPE_HANDLE
, &sh
->state
);
5542 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5543 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5544 atomic_inc(&conf
->preread_active_stripes
);
5545 release_stripe_plug(mddev
, sh
);
5551 static bool raid5_make_request(struct mddev
*mddev
, struct bio
* bi
)
5553 struct r5conf
*conf
= mddev
->private;
5555 sector_t new_sector
;
5556 sector_t logical_sector
, last_sector
;
5557 struct stripe_head
*sh
;
5558 const int rw
= bio_data_dir(bi
);
5561 bool do_flush
= false;
5563 if (unlikely(bi
->bi_opf
& REQ_PREFLUSH
)) {
5564 int ret
= r5l_handle_flush_request(conf
->log
, bi
);
5568 if (ret
== -ENODEV
) {
5569 md_flush_request(mddev
, bi
);
5572 /* ret == -EAGAIN, fallback */
5574 * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5575 * we need to flush journal device
5577 do_flush
= bi
->bi_opf
& REQ_PREFLUSH
;
5580 if (!md_write_start(mddev
, bi
))
5583 * If array is degraded, better not do chunk aligned read because
5584 * later we might have to read it again in order to reconstruct
5585 * data on failed drives.
5587 if (rw
== READ
&& mddev
->degraded
== 0 &&
5588 mddev
->reshape_position
== MaxSector
) {
5589 bi
= chunk_aligned_read(mddev
, bi
);
5594 if (unlikely(bio_op(bi
) == REQ_OP_DISCARD
)) {
5595 make_discard_request(mddev
, bi
);
5596 md_write_end(mddev
);
5600 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5601 last_sector
= bio_end_sector(bi
);
5604 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
5605 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
5611 seq
= read_seqcount_begin(&conf
->gen_lock
);
5614 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5615 TASK_UNINTERRUPTIBLE
);
5616 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
5617 /* spinlock is needed as reshape_progress may be
5618 * 64bit on a 32bit platform, and so it might be
5619 * possible to see a half-updated value
5620 * Of course reshape_progress could change after
5621 * the lock is dropped, so once we get a reference
5622 * to the stripe that we think it is, we will have
5625 spin_lock_irq(&conf
->device_lock
);
5626 if (mddev
->reshape_backwards
5627 ? logical_sector
< conf
->reshape_progress
5628 : logical_sector
>= conf
->reshape_progress
) {
5631 if (mddev
->reshape_backwards
5632 ? logical_sector
< conf
->reshape_safe
5633 : logical_sector
>= conf
->reshape_safe
) {
5634 spin_unlock_irq(&conf
->device_lock
);
5640 spin_unlock_irq(&conf
->device_lock
);
5643 new_sector
= raid5_compute_sector(conf
, logical_sector
,
5646 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5647 (unsigned long long)new_sector
,
5648 (unsigned long long)logical_sector
);
5650 sh
= raid5_get_active_stripe(conf
, new_sector
, previous
,
5651 (bi
->bi_opf
& REQ_RAHEAD
), 0);
5653 if (unlikely(previous
)) {
5654 /* expansion might have moved on while waiting for a
5655 * stripe, so we must do the range check again.
5656 * Expansion could still move past after this
5657 * test, but as we are holding a reference to
5658 * 'sh', we know that if that happens,
5659 * STRIPE_EXPANDING will get set and the expansion
5660 * won't proceed until we finish with the stripe.
5663 spin_lock_irq(&conf
->device_lock
);
5664 if (mddev
->reshape_backwards
5665 ? logical_sector
>= conf
->reshape_progress
5666 : logical_sector
< conf
->reshape_progress
)
5667 /* mismatch, need to try again */
5669 spin_unlock_irq(&conf
->device_lock
);
5671 raid5_release_stripe(sh
);
5677 if (read_seqcount_retry(&conf
->gen_lock
, seq
)) {
5678 /* Might have got the wrong stripe_head
5681 raid5_release_stripe(sh
);
5686 logical_sector
>= mddev
->suspend_lo
&&
5687 logical_sector
< mddev
->suspend_hi
) {
5688 raid5_release_stripe(sh
);
5689 /* As the suspend_* range is controlled by
5690 * userspace, we want an interruptible
5693 prepare_to_wait(&conf
->wait_for_overlap
,
5694 &w
, TASK_INTERRUPTIBLE
);
5695 if (logical_sector
>= mddev
->suspend_lo
&&
5696 logical_sector
< mddev
->suspend_hi
) {
5699 sigprocmask(SIG_BLOCK
, &full
, &old
);
5701 sigprocmask(SIG_SETMASK
, &old
, NULL
);
5707 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
5708 !add_stripe_bio(sh
, bi
, dd_idx
, rw
, previous
)) {
5709 /* Stripe is busy expanding or
5710 * add failed due to overlap. Flush everything
5713 md_wakeup_thread(mddev
->thread
);
5714 raid5_release_stripe(sh
);
5720 set_bit(STRIPE_R5C_PREFLUSH
, &sh
->state
);
5721 /* we only need flush for one stripe */
5725 set_bit(STRIPE_HANDLE
, &sh
->state
);
5726 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5727 if ((!sh
->batch_head
|| sh
== sh
->batch_head
) &&
5728 (bi
->bi_opf
& REQ_SYNC
) &&
5729 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5730 atomic_inc(&conf
->preread_active_stripes
);
5731 release_stripe_plug(mddev
, sh
);
5733 /* cannot get stripe for read-ahead, just give-up */
5734 bi
->bi_status
= BLK_STS_IOERR
;
5738 finish_wait(&conf
->wait_for_overlap
, &w
);
5741 md_write_end(mddev
);
5746 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
5748 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
5750 /* reshaping is quite different to recovery/resync so it is
5751 * handled quite separately ... here.
5753 * On each call to sync_request, we gather one chunk worth of
5754 * destination stripes and flag them as expanding.
5755 * Then we find all the source stripes and request reads.
5756 * As the reads complete, handle_stripe will copy the data
5757 * into the destination stripe and release that stripe.
5759 struct r5conf
*conf
= mddev
->private;
5760 struct stripe_head
*sh
;
5761 sector_t first_sector
, last_sector
;
5762 int raid_disks
= conf
->previous_raid_disks
;
5763 int data_disks
= raid_disks
- conf
->max_degraded
;
5764 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5767 sector_t writepos
, readpos
, safepos
;
5768 sector_t stripe_addr
;
5769 int reshape_sectors
;
5770 struct list_head stripes
;
5773 if (sector_nr
== 0) {
5774 /* If restarting in the middle, skip the initial sectors */
5775 if (mddev
->reshape_backwards
&&
5776 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
5777 sector_nr
= raid5_size(mddev
, 0, 0)
5778 - conf
->reshape_progress
;
5779 } else if (mddev
->reshape_backwards
&&
5780 conf
->reshape_progress
== MaxSector
) {
5781 /* shouldn't happen, but just in case, finish up.*/
5782 sector_nr
= MaxSector
;
5783 } else if (!mddev
->reshape_backwards
&&
5784 conf
->reshape_progress
> 0)
5785 sector_nr
= conf
->reshape_progress
;
5786 sector_div(sector_nr
, new_data_disks
);
5788 mddev
->curr_resync_completed
= sector_nr
;
5789 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5796 /* We need to process a full chunk at a time.
5797 * If old and new chunk sizes differ, we need to process the
5801 reshape_sectors
= max(conf
->chunk_sectors
, conf
->prev_chunk_sectors
);
5803 /* We update the metadata at least every 10 seconds, or when
5804 * the data about to be copied would over-write the source of
5805 * the data at the front of the range. i.e. one new_stripe
5806 * along from reshape_progress new_maps to after where
5807 * reshape_safe old_maps to
5809 writepos
= conf
->reshape_progress
;
5810 sector_div(writepos
, new_data_disks
);
5811 readpos
= conf
->reshape_progress
;
5812 sector_div(readpos
, data_disks
);
5813 safepos
= conf
->reshape_safe
;
5814 sector_div(safepos
, data_disks
);
5815 if (mddev
->reshape_backwards
) {
5816 BUG_ON(writepos
< reshape_sectors
);
5817 writepos
-= reshape_sectors
;
5818 readpos
+= reshape_sectors
;
5819 safepos
+= reshape_sectors
;
5821 writepos
+= reshape_sectors
;
5822 /* readpos and safepos are worst-case calculations.
5823 * A negative number is overly pessimistic, and causes
5824 * obvious problems for unsigned storage. So clip to 0.
5826 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
5827 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
5830 /* Having calculated the 'writepos' possibly use it
5831 * to set 'stripe_addr' which is where we will write to.
5833 if (mddev
->reshape_backwards
) {
5834 BUG_ON(conf
->reshape_progress
== 0);
5835 stripe_addr
= writepos
;
5836 BUG_ON((mddev
->dev_sectors
&
5837 ~((sector_t
)reshape_sectors
- 1))
5838 - reshape_sectors
- stripe_addr
5841 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
5842 stripe_addr
= sector_nr
;
5845 /* 'writepos' is the most advanced device address we might write.
5846 * 'readpos' is the least advanced device address we might read.
5847 * 'safepos' is the least address recorded in the metadata as having
5849 * If there is a min_offset_diff, these are adjusted either by
5850 * increasing the safepos/readpos if diff is negative, or
5851 * increasing writepos if diff is positive.
5852 * If 'readpos' is then behind 'writepos', there is no way that we can
5853 * ensure safety in the face of a crash - that must be done by userspace
5854 * making a backup of the data. So in that case there is no particular
5855 * rush to update metadata.
5856 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5857 * update the metadata to advance 'safepos' to match 'readpos' so that
5858 * we can be safe in the event of a crash.
5859 * So we insist on updating metadata if safepos is behind writepos and
5860 * readpos is beyond writepos.
5861 * In any case, update the metadata every 10 seconds.
5862 * Maybe that number should be configurable, but I'm not sure it is
5863 * worth it.... maybe it could be a multiple of safemode_delay???
5865 if (conf
->min_offset_diff
< 0) {
5866 safepos
+= -conf
->min_offset_diff
;
5867 readpos
+= -conf
->min_offset_diff
;
5869 writepos
+= conf
->min_offset_diff
;
5871 if ((mddev
->reshape_backwards
5872 ? (safepos
> writepos
&& readpos
< writepos
)
5873 : (safepos
< writepos
&& readpos
> writepos
)) ||
5874 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
5875 /* Cannot proceed until we've updated the superblock... */
5876 wait_event(conf
->wait_for_overlap
,
5877 atomic_read(&conf
->reshape_stripes
)==0
5878 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5879 if (atomic_read(&conf
->reshape_stripes
) != 0)
5881 mddev
->reshape_position
= conf
->reshape_progress
;
5882 mddev
->curr_resync_completed
= sector_nr
;
5883 conf
->reshape_checkpoint
= jiffies
;
5884 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
5885 md_wakeup_thread(mddev
->thread
);
5886 wait_event(mddev
->sb_wait
, mddev
->sb_flags
== 0 ||
5887 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5888 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5890 spin_lock_irq(&conf
->device_lock
);
5891 conf
->reshape_safe
= mddev
->reshape_position
;
5892 spin_unlock_irq(&conf
->device_lock
);
5893 wake_up(&conf
->wait_for_overlap
);
5894 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5897 INIT_LIST_HEAD(&stripes
);
5898 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
5900 int skipped_disk
= 0;
5901 sh
= raid5_get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
5902 set_bit(STRIPE_EXPANDING
, &sh
->state
);
5903 atomic_inc(&conf
->reshape_stripes
);
5904 /* If any of this stripe is beyond the end of the old
5905 * array, then we need to zero those blocks
5907 for (j
=sh
->disks
; j
--;) {
5909 if (j
== sh
->pd_idx
)
5911 if (conf
->level
== 6 &&
5914 s
= raid5_compute_blocknr(sh
, j
, 0);
5915 if (s
< raid5_size(mddev
, 0, 0)) {
5919 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
5920 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
5921 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
5923 if (!skipped_disk
) {
5924 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
5925 set_bit(STRIPE_HANDLE
, &sh
->state
);
5927 list_add(&sh
->lru
, &stripes
);
5929 spin_lock_irq(&conf
->device_lock
);
5930 if (mddev
->reshape_backwards
)
5931 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
5933 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
5934 spin_unlock_irq(&conf
->device_lock
);
5935 /* Ok, those stripe are ready. We can start scheduling
5936 * reads on the source stripes.
5937 * The source stripes are determined by mapping the first and last
5938 * block on the destination stripes.
5941 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
5944 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
5945 * new_data_disks
- 1),
5947 if (last_sector
>= mddev
->dev_sectors
)
5948 last_sector
= mddev
->dev_sectors
- 1;
5949 while (first_sector
<= last_sector
) {
5950 sh
= raid5_get_active_stripe(conf
, first_sector
, 1, 0, 1);
5951 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
5952 set_bit(STRIPE_HANDLE
, &sh
->state
);
5953 raid5_release_stripe(sh
);
5954 first_sector
+= STRIPE_SECTORS
;
5956 /* Now that the sources are clearly marked, we can release
5957 * the destination stripes
5959 while (!list_empty(&stripes
)) {
5960 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
5961 list_del_init(&sh
->lru
);
5962 raid5_release_stripe(sh
);
5964 /* If this takes us to the resync_max point where we have to pause,
5965 * then we need to write out the superblock.
5967 sector_nr
+= reshape_sectors
;
5968 retn
= reshape_sectors
;
5970 if (mddev
->curr_resync_completed
> mddev
->resync_max
||
5971 (sector_nr
- mddev
->curr_resync_completed
) * 2
5972 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
5973 /* Cannot proceed until we've updated the superblock... */
5974 wait_event(conf
->wait_for_overlap
,
5975 atomic_read(&conf
->reshape_stripes
) == 0
5976 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5977 if (atomic_read(&conf
->reshape_stripes
) != 0)
5979 mddev
->reshape_position
= conf
->reshape_progress
;
5980 mddev
->curr_resync_completed
= sector_nr
;
5981 conf
->reshape_checkpoint
= jiffies
;
5982 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
5983 md_wakeup_thread(mddev
->thread
);
5984 wait_event(mddev
->sb_wait
,
5985 !test_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
)
5986 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5987 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5989 spin_lock_irq(&conf
->device_lock
);
5990 conf
->reshape_safe
= mddev
->reshape_position
;
5991 spin_unlock_irq(&conf
->device_lock
);
5992 wake_up(&conf
->wait_for_overlap
);
5993 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5999 static inline sector_t
raid5_sync_request(struct mddev
*mddev
, sector_t sector_nr
,
6002 struct r5conf
*conf
= mddev
->private;
6003 struct stripe_head
*sh
;
6004 sector_t max_sector
= mddev
->dev_sectors
;
6005 sector_t sync_blocks
;
6006 int still_degraded
= 0;
6009 if (sector_nr
>= max_sector
) {
6010 /* just being told to finish up .. nothing much to do */
6012 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
6017 if (mddev
->curr_resync
< max_sector
) /* aborted */
6018 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
6020 else /* completed sync */
6022 bitmap_close_sync(mddev
->bitmap
);
6027 /* Allow raid5_quiesce to complete */
6028 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
6030 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
6031 return reshape_request(mddev
, sector_nr
, skipped
);
6033 /* No need to check resync_max as we never do more than one
6034 * stripe, and as resync_max will always be on a chunk boundary,
6035 * if the check in md_do_sync didn't fire, there is no chance
6036 * of overstepping resync_max here
6039 /* if there is too many failed drives and we are trying
6040 * to resync, then assert that we are finished, because there is
6041 * nothing we can do.
6043 if (mddev
->degraded
>= conf
->max_degraded
&&
6044 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
6045 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
6049 if (!test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
6051 !bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
6052 sync_blocks
>= STRIPE_SECTORS
) {
6053 /* we can skip this block, and probably more */
6054 sync_blocks
/= STRIPE_SECTORS
;
6056 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
6059 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
, false);
6061 sh
= raid5_get_active_stripe(conf
, sector_nr
, 0, 1, 0);
6063 sh
= raid5_get_active_stripe(conf
, sector_nr
, 0, 0, 0);
6064 /* make sure we don't swamp the stripe cache if someone else
6065 * is trying to get access
6067 schedule_timeout_uninterruptible(1);
6069 /* Need to check if array will still be degraded after recovery/resync
6070 * Note in case of > 1 drive failures it's possible we're rebuilding
6071 * one drive while leaving another faulty drive in array.
6074 for (i
= 0; i
< conf
->raid_disks
; i
++) {
6075 struct md_rdev
*rdev
= READ_ONCE(conf
->disks
[i
].rdev
);
6077 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
))
6082 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
6084 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
6085 set_bit(STRIPE_HANDLE
, &sh
->state
);
6087 raid5_release_stripe(sh
);
6089 return STRIPE_SECTORS
;
6092 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
,
6093 unsigned int offset
)
6095 /* We may not be able to submit a whole bio at once as there
6096 * may not be enough stripe_heads available.
6097 * We cannot pre-allocate enough stripe_heads as we may need
6098 * more than exist in the cache (if we allow ever large chunks).
6099 * So we do one stripe head at a time and record in
6100 * ->bi_hw_segments how many have been done.
6102 * We *know* that this entire raid_bio is in one chunk, so
6103 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6105 struct stripe_head
*sh
;
6107 sector_t sector
, logical_sector
, last_sector
;
6111 logical_sector
= raid_bio
->bi_iter
.bi_sector
&
6112 ~((sector_t
)STRIPE_SECTORS
-1);
6113 sector
= raid5_compute_sector(conf
, logical_sector
,
6115 last_sector
= bio_end_sector(raid_bio
);
6117 for (; logical_sector
< last_sector
;
6118 logical_sector
+= STRIPE_SECTORS
,
6119 sector
+= STRIPE_SECTORS
,
6123 /* already done this stripe */
6126 sh
= raid5_get_active_stripe(conf
, sector
, 0, 1, 1);
6129 /* failed to get a stripe - must wait */
6130 conf
->retry_read_aligned
= raid_bio
;
6131 conf
->retry_read_offset
= scnt
;
6135 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0, 0)) {
6136 raid5_release_stripe(sh
);
6137 conf
->retry_read_aligned
= raid_bio
;
6138 conf
->retry_read_offset
= scnt
;
6142 set_bit(R5_ReadNoMerge
, &sh
->dev
[dd_idx
].flags
);
6144 raid5_release_stripe(sh
);
6148 bio_endio(raid_bio
);
6150 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
6151 wake_up(&conf
->wait_for_quiescent
);
6155 static int handle_active_stripes(struct r5conf
*conf
, int group
,
6156 struct r5worker
*worker
,
6157 struct list_head
*temp_inactive_list
)
6159 struct stripe_head
*batch
[MAX_STRIPE_BATCH
], *sh
;
6160 int i
, batch_size
= 0, hash
;
6161 bool release_inactive
= false;
6163 while (batch_size
< MAX_STRIPE_BATCH
&&
6164 (sh
= __get_priority_stripe(conf
, group
)) != NULL
)
6165 batch
[batch_size
++] = sh
;
6167 if (batch_size
== 0) {
6168 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6169 if (!list_empty(temp_inactive_list
+ i
))
6171 if (i
== NR_STRIPE_HASH_LOCKS
) {
6172 spin_unlock_irq(&conf
->device_lock
);
6173 r5l_flush_stripe_to_raid(conf
->log
);
6174 spin_lock_irq(&conf
->device_lock
);
6177 release_inactive
= true;
6179 spin_unlock_irq(&conf
->device_lock
);
6181 release_inactive_stripe_list(conf
, temp_inactive_list
,
6182 NR_STRIPE_HASH_LOCKS
);
6184 r5l_flush_stripe_to_raid(conf
->log
);
6185 if (release_inactive
) {
6186 spin_lock_irq(&conf
->device_lock
);
6190 for (i
= 0; i
< batch_size
; i
++)
6191 handle_stripe(batch
[i
]);
6192 log_write_stripe_run(conf
);
6196 spin_lock_irq(&conf
->device_lock
);
6197 for (i
= 0; i
< batch_size
; i
++) {
6198 hash
= batch
[i
]->hash_lock_index
;
6199 __release_stripe(conf
, batch
[i
], &temp_inactive_list
[hash
]);
6204 static void raid5_do_work(struct work_struct
*work
)
6206 struct r5worker
*worker
= container_of(work
, struct r5worker
, work
);
6207 struct r5worker_group
*group
= worker
->group
;
6208 struct r5conf
*conf
= group
->conf
;
6209 struct mddev
*mddev
= conf
->mddev
;
6210 int group_id
= group
- conf
->worker_groups
;
6212 struct blk_plug plug
;
6214 pr_debug("+++ raid5worker active\n");
6216 blk_start_plug(&plug
);
6218 spin_lock_irq(&conf
->device_lock
);
6220 int batch_size
, released
;
6222 released
= release_stripe_list(conf
, worker
->temp_inactive_list
);
6224 batch_size
= handle_active_stripes(conf
, group_id
, worker
,
6225 worker
->temp_inactive_list
);
6226 worker
->working
= false;
6227 if (!batch_size
&& !released
)
6229 handled
+= batch_size
;
6230 wait_event_lock_irq(mddev
->sb_wait
,
6231 !test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
),
6234 pr_debug("%d stripes handled\n", handled
);
6236 spin_unlock_irq(&conf
->device_lock
);
6238 flush_deferred_bios(conf
);
6240 r5l_flush_stripe_to_raid(conf
->log
);
6242 async_tx_issue_pending_all();
6243 blk_finish_plug(&plug
);
6245 pr_debug("--- raid5worker inactive\n");
6249 * This is our raid5 kernel thread.
6251 * We scan the hash table for stripes which can be handled now.
6252 * During the scan, completed stripes are saved for us by the interrupt
6253 * handler, so that they will not have to wait for our next wakeup.
6255 static void raid5d(struct md_thread
*thread
)
6257 struct mddev
*mddev
= thread
->mddev
;
6258 struct r5conf
*conf
= mddev
->private;
6260 struct blk_plug plug
;
6262 pr_debug("+++ raid5d active\n");
6264 md_check_recovery(mddev
);
6266 blk_start_plug(&plug
);
6268 spin_lock_irq(&conf
->device_lock
);
6271 int batch_size
, released
;
6272 unsigned int offset
;
6274 released
= release_stripe_list(conf
, conf
->temp_inactive_list
);
6276 clear_bit(R5_DID_ALLOC
, &conf
->cache_state
);
6279 !list_empty(&conf
->bitmap_list
)) {
6280 /* Now is a good time to flush some bitmap updates */
6282 spin_unlock_irq(&conf
->device_lock
);
6283 bitmap_unplug(mddev
->bitmap
);
6284 spin_lock_irq(&conf
->device_lock
);
6285 conf
->seq_write
= conf
->seq_flush
;
6286 activate_bit_delay(conf
, conf
->temp_inactive_list
);
6288 raid5_activate_delayed(conf
);
6290 while ((bio
= remove_bio_from_retry(conf
, &offset
))) {
6292 spin_unlock_irq(&conf
->device_lock
);
6293 ok
= retry_aligned_read(conf
, bio
, offset
);
6294 spin_lock_irq(&conf
->device_lock
);
6300 batch_size
= handle_active_stripes(conf
, ANY_GROUP
, NULL
,
6301 conf
->temp_inactive_list
);
6302 if (!batch_size
&& !released
)
6304 handled
+= batch_size
;
6306 if (mddev
->sb_flags
& ~(1 << MD_SB_CHANGE_PENDING
)) {
6307 spin_unlock_irq(&conf
->device_lock
);
6308 md_check_recovery(mddev
);
6309 spin_lock_irq(&conf
->device_lock
);
6312 pr_debug("%d stripes handled\n", handled
);
6314 spin_unlock_irq(&conf
->device_lock
);
6315 if (test_and_clear_bit(R5_ALLOC_MORE
, &conf
->cache_state
) &&
6316 mutex_trylock(&conf
->cache_size_mutex
)) {
6317 grow_one_stripe(conf
, __GFP_NOWARN
);
6318 /* Set flag even if allocation failed. This helps
6319 * slow down allocation requests when mem is short
6321 set_bit(R5_DID_ALLOC
, &conf
->cache_state
);
6322 mutex_unlock(&conf
->cache_size_mutex
);
6325 flush_deferred_bios(conf
);
6327 r5l_flush_stripe_to_raid(conf
->log
);
6329 async_tx_issue_pending_all();
6330 blk_finish_plug(&plug
);
6332 pr_debug("--- raid5d inactive\n");
6336 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
6338 struct r5conf
*conf
;
6340 spin_lock(&mddev
->lock
);
6341 conf
= mddev
->private;
6343 ret
= sprintf(page
, "%d\n", conf
->min_nr_stripes
);
6344 spin_unlock(&mddev
->lock
);
6349 raid5_set_cache_size(struct mddev
*mddev
, int size
)
6351 struct r5conf
*conf
= mddev
->private;
6353 if (size
<= 16 || size
> 32768)
6356 conf
->min_nr_stripes
= size
;
6357 mutex_lock(&conf
->cache_size_mutex
);
6358 while (size
< conf
->max_nr_stripes
&&
6359 drop_one_stripe(conf
))
6361 mutex_unlock(&conf
->cache_size_mutex
);
6363 md_allow_write(mddev
);
6365 mutex_lock(&conf
->cache_size_mutex
);
6366 while (size
> conf
->max_nr_stripes
)
6367 if (!grow_one_stripe(conf
, GFP_KERNEL
))
6369 mutex_unlock(&conf
->cache_size_mutex
);
6373 EXPORT_SYMBOL(raid5_set_cache_size
);
6376 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
6378 struct r5conf
*conf
;
6382 if (len
>= PAGE_SIZE
)
6384 if (kstrtoul(page
, 10, &new))
6386 err
= mddev_lock(mddev
);
6389 conf
= mddev
->private;
6393 err
= raid5_set_cache_size(mddev
, new);
6394 mddev_unlock(mddev
);
6399 static struct md_sysfs_entry
6400 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
6401 raid5_show_stripe_cache_size
,
6402 raid5_store_stripe_cache_size
);
6405 raid5_show_rmw_level(struct mddev
*mddev
, char *page
)
6407 struct r5conf
*conf
= mddev
->private;
6409 return sprintf(page
, "%d\n", conf
->rmw_level
);
6415 raid5_store_rmw_level(struct mddev
*mddev
, const char *page
, size_t len
)
6417 struct r5conf
*conf
= mddev
->private;
6423 if (len
>= PAGE_SIZE
)
6426 if (kstrtoul(page
, 10, &new))
6429 if (new != PARITY_DISABLE_RMW
&& !raid6_call
.xor_syndrome
)
6432 if (new != PARITY_DISABLE_RMW
&&
6433 new != PARITY_ENABLE_RMW
&&
6434 new != PARITY_PREFER_RMW
)
6437 conf
->rmw_level
= new;
6441 static struct md_sysfs_entry
6442 raid5_rmw_level
= __ATTR(rmw_level
, S_IRUGO
| S_IWUSR
,
6443 raid5_show_rmw_level
,
6444 raid5_store_rmw_level
);
6448 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
6450 struct r5conf
*conf
;
6452 spin_lock(&mddev
->lock
);
6453 conf
= mddev
->private;
6455 ret
= sprintf(page
, "%d\n", conf
->bypass_threshold
);
6456 spin_unlock(&mddev
->lock
);
6461 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
6463 struct r5conf
*conf
;
6467 if (len
>= PAGE_SIZE
)
6469 if (kstrtoul(page
, 10, &new))
6472 err
= mddev_lock(mddev
);
6475 conf
= mddev
->private;
6478 else if (new > conf
->min_nr_stripes
)
6481 conf
->bypass_threshold
= new;
6482 mddev_unlock(mddev
);
6486 static struct md_sysfs_entry
6487 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
6489 raid5_show_preread_threshold
,
6490 raid5_store_preread_threshold
);
6493 raid5_show_skip_copy(struct mddev
*mddev
, char *page
)
6495 struct r5conf
*conf
;
6497 spin_lock(&mddev
->lock
);
6498 conf
= mddev
->private;
6500 ret
= sprintf(page
, "%d\n", conf
->skip_copy
);
6501 spin_unlock(&mddev
->lock
);
6506 raid5_store_skip_copy(struct mddev
*mddev
, const char *page
, size_t len
)
6508 struct r5conf
*conf
;
6512 if (len
>= PAGE_SIZE
)
6514 if (kstrtoul(page
, 10, &new))
6518 err
= mddev_lock(mddev
);
6521 conf
= mddev
->private;
6524 else if (new != conf
->skip_copy
) {
6525 mddev_suspend(mddev
);
6526 conf
->skip_copy
= new;
6528 mddev
->queue
->backing_dev_info
->capabilities
|=
6529 BDI_CAP_STABLE_WRITES
;
6531 mddev
->queue
->backing_dev_info
->capabilities
&=
6532 ~BDI_CAP_STABLE_WRITES
;
6533 mddev_resume(mddev
);
6535 mddev_unlock(mddev
);
6539 static struct md_sysfs_entry
6540 raid5_skip_copy
= __ATTR(skip_copy
, S_IRUGO
| S_IWUSR
,
6541 raid5_show_skip_copy
,
6542 raid5_store_skip_copy
);
6545 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
6547 struct r5conf
*conf
= mddev
->private;
6549 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
6554 static struct md_sysfs_entry
6555 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
6558 raid5_show_group_thread_cnt(struct mddev
*mddev
, char *page
)
6560 struct r5conf
*conf
;
6562 spin_lock(&mddev
->lock
);
6563 conf
= mddev
->private;
6565 ret
= sprintf(page
, "%d\n", conf
->worker_cnt_per_group
);
6566 spin_unlock(&mddev
->lock
);
6570 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6572 int *worker_cnt_per_group
,
6573 struct r5worker_group
**worker_groups
);
6575 raid5_store_group_thread_cnt(struct mddev
*mddev
, const char *page
, size_t len
)
6577 struct r5conf
*conf
;
6580 struct r5worker_group
*new_groups
, *old_groups
;
6581 int group_cnt
, worker_cnt_per_group
;
6583 if (len
>= PAGE_SIZE
)
6585 if (kstrtouint(page
, 10, &new))
6587 /* 8192 should be big enough */
6591 err
= mddev_lock(mddev
);
6594 conf
= mddev
->private;
6597 else if (new != conf
->worker_cnt_per_group
) {
6598 mddev_suspend(mddev
);
6600 old_groups
= conf
->worker_groups
;
6602 flush_workqueue(raid5_wq
);
6604 err
= alloc_thread_groups(conf
, new,
6605 &group_cnt
, &worker_cnt_per_group
,
6608 spin_lock_irq(&conf
->device_lock
);
6609 conf
->group_cnt
= group_cnt
;
6610 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6611 conf
->worker_groups
= new_groups
;
6612 spin_unlock_irq(&conf
->device_lock
);
6615 kfree(old_groups
[0].workers
);
6618 mddev_resume(mddev
);
6620 mddev_unlock(mddev
);
6625 static struct md_sysfs_entry
6626 raid5_group_thread_cnt
= __ATTR(group_thread_cnt
, S_IRUGO
| S_IWUSR
,
6627 raid5_show_group_thread_cnt
,
6628 raid5_store_group_thread_cnt
);
6630 static struct attribute
*raid5_attrs
[] = {
6631 &raid5_stripecache_size
.attr
,
6632 &raid5_stripecache_active
.attr
,
6633 &raid5_preread_bypass_threshold
.attr
,
6634 &raid5_group_thread_cnt
.attr
,
6635 &raid5_skip_copy
.attr
,
6636 &raid5_rmw_level
.attr
,
6637 &r5c_journal_mode
.attr
,
6640 static struct attribute_group raid5_attrs_group
= {
6642 .attrs
= raid5_attrs
,
6645 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6647 int *worker_cnt_per_group
,
6648 struct r5worker_group
**worker_groups
)
6652 struct r5worker
*workers
;
6654 *worker_cnt_per_group
= cnt
;
6657 *worker_groups
= NULL
;
6660 *group_cnt
= num_possible_nodes();
6661 size
= sizeof(struct r5worker
) * cnt
;
6662 workers
= kzalloc(size
* *group_cnt
, GFP_NOIO
);
6663 *worker_groups
= kzalloc(sizeof(struct r5worker_group
) *
6664 *group_cnt
, GFP_NOIO
);
6665 if (!*worker_groups
|| !workers
) {
6667 kfree(*worker_groups
);
6671 for (i
= 0; i
< *group_cnt
; i
++) {
6672 struct r5worker_group
*group
;
6674 group
= &(*worker_groups
)[i
];
6675 INIT_LIST_HEAD(&group
->handle_list
);
6676 INIT_LIST_HEAD(&group
->loprio_list
);
6678 group
->workers
= workers
+ i
* cnt
;
6680 for (j
= 0; j
< cnt
; j
++) {
6681 struct r5worker
*worker
= group
->workers
+ j
;
6682 worker
->group
= group
;
6683 INIT_WORK(&worker
->work
, raid5_do_work
);
6685 for (k
= 0; k
< NR_STRIPE_HASH_LOCKS
; k
++)
6686 INIT_LIST_HEAD(worker
->temp_inactive_list
+ k
);
6693 static void free_thread_groups(struct r5conf
*conf
)
6695 if (conf
->worker_groups
)
6696 kfree(conf
->worker_groups
[0].workers
);
6697 kfree(conf
->worker_groups
);
6698 conf
->worker_groups
= NULL
;
6702 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
6704 struct r5conf
*conf
= mddev
->private;
6707 sectors
= mddev
->dev_sectors
;
6709 /* size is defined by the smallest of previous and new size */
6710 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
6712 sectors
&= ~((sector_t
)conf
->chunk_sectors
- 1);
6713 sectors
&= ~((sector_t
)conf
->prev_chunk_sectors
- 1);
6714 return sectors
* (raid_disks
- conf
->max_degraded
);
6717 static void free_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6719 safe_put_page(percpu
->spare_page
);
6720 if (percpu
->scribble
)
6721 flex_array_free(percpu
->scribble
);
6722 percpu
->spare_page
= NULL
;
6723 percpu
->scribble
= NULL
;
6726 static int alloc_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6728 if (conf
->level
== 6 && !percpu
->spare_page
)
6729 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
6730 if (!percpu
->scribble
)
6731 percpu
->scribble
= scribble_alloc(max(conf
->raid_disks
,
6732 conf
->previous_raid_disks
),
6733 max(conf
->chunk_sectors
,
6734 conf
->prev_chunk_sectors
)
6738 if (!percpu
->scribble
|| (conf
->level
== 6 && !percpu
->spare_page
)) {
6739 free_scratch_buffer(conf
, percpu
);
6746 static int raid456_cpu_dead(unsigned int cpu
, struct hlist_node
*node
)
6748 struct r5conf
*conf
= hlist_entry_safe(node
, struct r5conf
, node
);
6750 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6754 static void raid5_free_percpu(struct r5conf
*conf
)
6759 cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE
, &conf
->node
);
6760 free_percpu(conf
->percpu
);
6763 static void free_conf(struct r5conf
*conf
)
6769 if (conf
->shrinker
.nr_deferred
)
6770 unregister_shrinker(&conf
->shrinker
);
6772 free_thread_groups(conf
);
6773 shrink_stripes(conf
);
6774 raid5_free_percpu(conf
);
6775 for (i
= 0; i
< conf
->pool_size
; i
++)
6776 if (conf
->disks
[i
].extra_page
)
6777 put_page(conf
->disks
[i
].extra_page
);
6779 if (conf
->bio_split
)
6780 bioset_free(conf
->bio_split
);
6781 kfree(conf
->stripe_hashtbl
);
6782 kfree(conf
->pending_data
);
6786 static int raid456_cpu_up_prepare(unsigned int cpu
, struct hlist_node
*node
)
6788 struct r5conf
*conf
= hlist_entry_safe(node
, struct r5conf
, node
);
6789 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
6791 if (alloc_scratch_buffer(conf
, percpu
)) {
6792 pr_warn("%s: failed memory allocation for cpu%u\n",
6799 static int raid5_alloc_percpu(struct r5conf
*conf
)
6803 conf
->percpu
= alloc_percpu(struct raid5_percpu
);
6807 err
= cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE
, &conf
->node
);
6809 conf
->scribble_disks
= max(conf
->raid_disks
,
6810 conf
->previous_raid_disks
);
6811 conf
->scribble_sectors
= max(conf
->chunk_sectors
,
6812 conf
->prev_chunk_sectors
);
6817 static unsigned long raid5_cache_scan(struct shrinker
*shrink
,
6818 struct shrink_control
*sc
)
6820 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
6821 unsigned long ret
= SHRINK_STOP
;
6823 if (mutex_trylock(&conf
->cache_size_mutex
)) {
6825 while (ret
< sc
->nr_to_scan
&&
6826 conf
->max_nr_stripes
> conf
->min_nr_stripes
) {
6827 if (drop_one_stripe(conf
) == 0) {
6833 mutex_unlock(&conf
->cache_size_mutex
);
6838 static unsigned long raid5_cache_count(struct shrinker
*shrink
,
6839 struct shrink_control
*sc
)
6841 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
6843 if (conf
->max_nr_stripes
< conf
->min_nr_stripes
)
6844 /* unlikely, but not impossible */
6846 return conf
->max_nr_stripes
- conf
->min_nr_stripes
;
6849 static struct r5conf
*setup_conf(struct mddev
*mddev
)
6851 struct r5conf
*conf
;
6852 int raid_disk
, memory
, max_disks
;
6853 struct md_rdev
*rdev
;
6854 struct disk_info
*disk
;
6857 int group_cnt
, worker_cnt_per_group
;
6858 struct r5worker_group
*new_group
;
6860 if (mddev
->new_level
!= 5
6861 && mddev
->new_level
!= 4
6862 && mddev
->new_level
!= 6) {
6863 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6864 mdname(mddev
), mddev
->new_level
);
6865 return ERR_PTR(-EIO
);
6867 if ((mddev
->new_level
== 5
6868 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
6869 (mddev
->new_level
== 6
6870 && !algorithm_valid_raid6(mddev
->new_layout
))) {
6871 pr_warn("md/raid:%s: layout %d not supported\n",
6872 mdname(mddev
), mddev
->new_layout
);
6873 return ERR_PTR(-EIO
);
6875 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
6876 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6877 mdname(mddev
), mddev
->raid_disks
);
6878 return ERR_PTR(-EINVAL
);
6881 if (!mddev
->new_chunk_sectors
||
6882 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
6883 !is_power_of_2(mddev
->new_chunk_sectors
)) {
6884 pr_warn("md/raid:%s: invalid chunk size %d\n",
6885 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
6886 return ERR_PTR(-EINVAL
);
6889 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
6892 INIT_LIST_HEAD(&conf
->free_list
);
6893 INIT_LIST_HEAD(&conf
->pending_list
);
6894 conf
->pending_data
= kzalloc(sizeof(struct r5pending_data
) *
6895 PENDING_IO_MAX
, GFP_KERNEL
);
6896 if (!conf
->pending_data
)
6898 for (i
= 0; i
< PENDING_IO_MAX
; i
++)
6899 list_add(&conf
->pending_data
[i
].sibling
, &conf
->free_list
);
6900 /* Don't enable multi-threading by default*/
6901 if (!alloc_thread_groups(conf
, 0, &group_cnt
, &worker_cnt_per_group
,
6903 conf
->group_cnt
= group_cnt
;
6904 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6905 conf
->worker_groups
= new_group
;
6908 spin_lock_init(&conf
->device_lock
);
6909 seqcount_init(&conf
->gen_lock
);
6910 mutex_init(&conf
->cache_size_mutex
);
6911 init_waitqueue_head(&conf
->wait_for_quiescent
);
6912 init_waitqueue_head(&conf
->wait_for_stripe
);
6913 init_waitqueue_head(&conf
->wait_for_overlap
);
6914 INIT_LIST_HEAD(&conf
->handle_list
);
6915 INIT_LIST_HEAD(&conf
->loprio_list
);
6916 INIT_LIST_HEAD(&conf
->hold_list
);
6917 INIT_LIST_HEAD(&conf
->delayed_list
);
6918 INIT_LIST_HEAD(&conf
->bitmap_list
);
6919 init_llist_head(&conf
->released_stripes
);
6920 atomic_set(&conf
->active_stripes
, 0);
6921 atomic_set(&conf
->preread_active_stripes
, 0);
6922 atomic_set(&conf
->active_aligned_reads
, 0);
6923 spin_lock_init(&conf
->pending_bios_lock
);
6924 conf
->batch_bio_dispatch
= true;
6925 rdev_for_each(rdev
, mddev
) {
6926 if (test_bit(Journal
, &rdev
->flags
))
6928 if (blk_queue_nonrot(bdev_get_queue(rdev
->bdev
))) {
6929 conf
->batch_bio_dispatch
= false;
6934 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
6935 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
6937 conf
->raid_disks
= mddev
->raid_disks
;
6938 if (mddev
->reshape_position
== MaxSector
)
6939 conf
->previous_raid_disks
= mddev
->raid_disks
;
6941 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
6942 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
6944 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
6950 for (i
= 0; i
< max_disks
; i
++) {
6951 conf
->disks
[i
].extra_page
= alloc_page(GFP_KERNEL
);
6952 if (!conf
->disks
[i
].extra_page
)
6956 conf
->bio_split
= bioset_create(BIO_POOL_SIZE
, 0, 0);
6957 if (!conf
->bio_split
)
6959 conf
->mddev
= mddev
;
6961 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
6964 /* We init hash_locks[0] separately to that it can be used
6965 * as the reference lock in the spin_lock_nest_lock() call
6966 * in lock_all_device_hash_locks_irq in order to convince
6967 * lockdep that we know what we are doing.
6969 spin_lock_init(conf
->hash_locks
);
6970 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6971 spin_lock_init(conf
->hash_locks
+ i
);
6973 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6974 INIT_LIST_HEAD(conf
->inactive_list
+ i
);
6976 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6977 INIT_LIST_HEAD(conf
->temp_inactive_list
+ i
);
6979 atomic_set(&conf
->r5c_cached_full_stripes
, 0);
6980 INIT_LIST_HEAD(&conf
->r5c_full_stripe_list
);
6981 atomic_set(&conf
->r5c_cached_partial_stripes
, 0);
6982 INIT_LIST_HEAD(&conf
->r5c_partial_stripe_list
);
6983 atomic_set(&conf
->r5c_flushing_full_stripes
, 0);
6984 atomic_set(&conf
->r5c_flushing_partial_stripes
, 0);
6986 conf
->level
= mddev
->new_level
;
6987 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
6988 if (raid5_alloc_percpu(conf
) != 0)
6991 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
6993 rdev_for_each(rdev
, mddev
) {
6994 raid_disk
= rdev
->raid_disk
;
6995 if (raid_disk
>= max_disks
6996 || raid_disk
< 0 || test_bit(Journal
, &rdev
->flags
))
6998 disk
= conf
->disks
+ raid_disk
;
7000 if (test_bit(Replacement
, &rdev
->flags
)) {
7001 if (disk
->replacement
)
7003 disk
->replacement
= rdev
;
7010 if (test_bit(In_sync
, &rdev
->flags
)) {
7011 char b
[BDEVNAME_SIZE
];
7012 pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7013 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
7014 } else if (rdev
->saved_raid_disk
!= raid_disk
)
7015 /* Cannot rely on bitmap to complete recovery */
7019 conf
->level
= mddev
->new_level
;
7020 if (conf
->level
== 6) {
7021 conf
->max_degraded
= 2;
7022 if (raid6_call
.xor_syndrome
)
7023 conf
->rmw_level
= PARITY_ENABLE_RMW
;
7025 conf
->rmw_level
= PARITY_DISABLE_RMW
;
7027 conf
->max_degraded
= 1;
7028 conf
->rmw_level
= PARITY_ENABLE_RMW
;
7030 conf
->algorithm
= mddev
->new_layout
;
7031 conf
->reshape_progress
= mddev
->reshape_position
;
7032 if (conf
->reshape_progress
!= MaxSector
) {
7033 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
7034 conf
->prev_algo
= mddev
->layout
;
7036 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
7037 conf
->prev_algo
= conf
->algorithm
;
7040 conf
->min_nr_stripes
= NR_STRIPES
;
7041 if (mddev
->reshape_position
!= MaxSector
) {
7042 int stripes
= max_t(int,
7043 ((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4,
7044 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4);
7045 conf
->min_nr_stripes
= max(NR_STRIPES
, stripes
);
7046 if (conf
->min_nr_stripes
!= NR_STRIPES
)
7047 pr_info("md/raid:%s: force stripe size %d for reshape\n",
7048 mdname(mddev
), conf
->min_nr_stripes
);
7050 memory
= conf
->min_nr_stripes
* (sizeof(struct stripe_head
) +
7051 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
7052 atomic_set(&conf
->empty_inactive_list_nr
, NR_STRIPE_HASH_LOCKS
);
7053 if (grow_stripes(conf
, conf
->min_nr_stripes
)) {
7054 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7055 mdname(mddev
), memory
);
7058 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev
), memory
);
7060 * Losing a stripe head costs more than the time to refill it,
7061 * it reduces the queue depth and so can hurt throughput.
7062 * So set it rather large, scaled by number of devices.
7064 conf
->shrinker
.seeks
= DEFAULT_SEEKS
* conf
->raid_disks
* 4;
7065 conf
->shrinker
.scan_objects
= raid5_cache_scan
;
7066 conf
->shrinker
.count_objects
= raid5_cache_count
;
7067 conf
->shrinker
.batch
= 128;
7068 conf
->shrinker
.flags
= 0;
7069 if (register_shrinker(&conf
->shrinker
)) {
7070 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7075 sprintf(pers_name
, "raid%d", mddev
->new_level
);
7076 conf
->thread
= md_register_thread(raid5d
, mddev
, pers_name
);
7077 if (!conf
->thread
) {
7078 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7088 return ERR_PTR(-EIO
);
7090 return ERR_PTR(-ENOMEM
);
7093 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
7096 case ALGORITHM_PARITY_0
:
7097 if (raid_disk
< max_degraded
)
7100 case ALGORITHM_PARITY_N
:
7101 if (raid_disk
>= raid_disks
- max_degraded
)
7104 case ALGORITHM_PARITY_0_6
:
7105 if (raid_disk
== 0 ||
7106 raid_disk
== raid_disks
- 1)
7109 case ALGORITHM_LEFT_ASYMMETRIC_6
:
7110 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
7111 case ALGORITHM_LEFT_SYMMETRIC_6
:
7112 case ALGORITHM_RIGHT_SYMMETRIC_6
:
7113 if (raid_disk
== raid_disks
- 1)
7119 static int raid5_run(struct mddev
*mddev
)
7121 struct r5conf
*conf
;
7122 int working_disks
= 0;
7123 int dirty_parity_disks
= 0;
7124 struct md_rdev
*rdev
;
7125 struct md_rdev
*journal_dev
= NULL
;
7126 sector_t reshape_offset
= 0;
7128 long long min_offset_diff
= 0;
7131 if (mddev_init_writes_pending(mddev
) < 0)
7134 if (mddev
->recovery_cp
!= MaxSector
)
7135 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7138 rdev_for_each(rdev
, mddev
) {
7141 if (test_bit(Journal
, &rdev
->flags
)) {
7145 if (rdev
->raid_disk
< 0)
7147 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
7149 min_offset_diff
= diff
;
7151 } else if (mddev
->reshape_backwards
&&
7152 diff
< min_offset_diff
)
7153 min_offset_diff
= diff
;
7154 else if (!mddev
->reshape_backwards
&&
7155 diff
> min_offset_diff
)
7156 min_offset_diff
= diff
;
7159 if (mddev
->reshape_position
!= MaxSector
) {
7160 /* Check that we can continue the reshape.
7161 * Difficulties arise if the stripe we would write to
7162 * next is at or after the stripe we would read from next.
7163 * For a reshape that changes the number of devices, this
7164 * is only possible for a very short time, and mdadm makes
7165 * sure that time appears to have past before assembling
7166 * the array. So we fail if that time hasn't passed.
7167 * For a reshape that keeps the number of devices the same
7168 * mdadm must be monitoring the reshape can keeping the
7169 * critical areas read-only and backed up. It will start
7170 * the array in read-only mode, so we check for that.
7172 sector_t here_new
, here_old
;
7174 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
7179 pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7184 if (mddev
->new_level
!= mddev
->level
) {
7185 pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7189 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
7190 /* reshape_position must be on a new-stripe boundary, and one
7191 * further up in new geometry must map after here in old
7193 * If the chunk sizes are different, then as we perform reshape
7194 * in units of the largest of the two, reshape_position needs
7195 * be a multiple of the largest chunk size times new data disks.
7197 here_new
= mddev
->reshape_position
;
7198 chunk_sectors
= max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
);
7199 new_data_disks
= mddev
->raid_disks
- max_degraded
;
7200 if (sector_div(here_new
, chunk_sectors
* new_data_disks
)) {
7201 pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7205 reshape_offset
= here_new
* chunk_sectors
;
7206 /* here_new is the stripe we will write to */
7207 here_old
= mddev
->reshape_position
;
7208 sector_div(here_old
, chunk_sectors
* (old_disks
-max_degraded
));
7209 /* here_old is the first stripe that we might need to read
7211 if (mddev
->delta_disks
== 0) {
7212 /* We cannot be sure it is safe to start an in-place
7213 * reshape. It is only safe if user-space is monitoring
7214 * and taking constant backups.
7215 * mdadm always starts a situation like this in
7216 * readonly mode so it can take control before
7217 * allowing any writes. So just check for that.
7219 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
7220 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
7221 /* not really in-place - so OK */;
7222 else if (mddev
->ro
== 0) {
7223 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7227 } else if (mddev
->reshape_backwards
7228 ? (here_new
* chunk_sectors
+ min_offset_diff
<=
7229 here_old
* chunk_sectors
)
7230 : (here_new
* chunk_sectors
>=
7231 here_old
* chunk_sectors
+ (-min_offset_diff
))) {
7232 /* Reading from the same stripe as writing to - bad */
7233 pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7237 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev
));
7238 /* OK, we should be able to continue; */
7240 BUG_ON(mddev
->level
!= mddev
->new_level
);
7241 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
7242 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
7243 BUG_ON(mddev
->delta_disks
!= 0);
7246 if (test_bit(MD_HAS_JOURNAL
, &mddev
->flags
) &&
7247 test_bit(MD_HAS_PPL
, &mddev
->flags
)) {
7248 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7250 clear_bit(MD_HAS_PPL
, &mddev
->flags
);
7251 clear_bit(MD_HAS_MULTIPLE_PPLS
, &mddev
->flags
);
7254 if (mddev
->private == NULL
)
7255 conf
= setup_conf(mddev
);
7257 conf
= mddev
->private;
7260 return PTR_ERR(conf
);
7262 if (test_bit(MD_HAS_JOURNAL
, &mddev
->flags
)) {
7264 pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7267 set_disk_ro(mddev
->gendisk
, 1);
7268 } else if (mddev
->recovery_cp
== MaxSector
)
7269 set_bit(MD_JOURNAL_CLEAN
, &mddev
->flags
);
7272 conf
->min_offset_diff
= min_offset_diff
;
7273 mddev
->thread
= conf
->thread
;
7274 conf
->thread
= NULL
;
7275 mddev
->private = conf
;
7277 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
7279 rdev
= conf
->disks
[i
].rdev
;
7280 if (!rdev
&& conf
->disks
[i
].replacement
) {
7281 /* The replacement is all we have yet */
7282 rdev
= conf
->disks
[i
].replacement
;
7283 conf
->disks
[i
].replacement
= NULL
;
7284 clear_bit(Replacement
, &rdev
->flags
);
7285 conf
->disks
[i
].rdev
= rdev
;
7289 if (conf
->disks
[i
].replacement
&&
7290 conf
->reshape_progress
!= MaxSector
) {
7291 /* replacements and reshape simply do not mix. */
7292 pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7295 if (test_bit(In_sync
, &rdev
->flags
)) {
7299 /* This disc is not fully in-sync. However if it
7300 * just stored parity (beyond the recovery_offset),
7301 * when we don't need to be concerned about the
7302 * array being dirty.
7303 * When reshape goes 'backwards', we never have
7304 * partially completed devices, so we only need
7305 * to worry about reshape going forwards.
7307 /* Hack because v0.91 doesn't store recovery_offset properly. */
7308 if (mddev
->major_version
== 0 &&
7309 mddev
->minor_version
> 90)
7310 rdev
->recovery_offset
= reshape_offset
;
7312 if (rdev
->recovery_offset
< reshape_offset
) {
7313 /* We need to check old and new layout */
7314 if (!only_parity(rdev
->raid_disk
,
7317 conf
->max_degraded
))
7320 if (!only_parity(rdev
->raid_disk
,
7322 conf
->previous_raid_disks
,
7323 conf
->max_degraded
))
7325 dirty_parity_disks
++;
7329 * 0 for a fully functional array, 1 or 2 for a degraded array.
7331 mddev
->degraded
= raid5_calc_degraded(conf
);
7333 if (has_failed(conf
)) {
7334 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7335 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
7339 /* device size must be a multiple of chunk size */
7340 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
7341 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
7343 if (mddev
->degraded
> dirty_parity_disks
&&
7344 mddev
->recovery_cp
!= MaxSector
) {
7345 if (test_bit(MD_HAS_PPL
, &mddev
->flags
))
7346 pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7348 else if (mddev
->ok_start_degraded
)
7349 pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7352 pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7358 pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7359 mdname(mddev
), conf
->level
,
7360 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
7363 print_raid5_conf(conf
);
7365 if (conf
->reshape_progress
!= MaxSector
) {
7366 conf
->reshape_safe
= conf
->reshape_progress
;
7367 atomic_set(&conf
->reshape_stripes
, 0);
7368 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
7369 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
7370 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
7371 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
7372 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
7376 /* Ok, everything is just fine now */
7377 if (mddev
->to_remove
== &raid5_attrs_group
)
7378 mddev
->to_remove
= NULL
;
7379 else if (mddev
->kobj
.sd
&&
7380 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
7381 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7383 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
7387 /* read-ahead size must cover two whole stripes, which
7388 * is 2 * (datadisks) * chunksize where 'n' is the
7389 * number of raid devices
7391 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
7392 int stripe
= data_disks
*
7393 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
7394 if (mddev
->queue
->backing_dev_info
->ra_pages
< 2 * stripe
)
7395 mddev
->queue
->backing_dev_info
->ra_pages
= 2 * stripe
;
7397 chunk_size
= mddev
->chunk_sectors
<< 9;
7398 blk_queue_io_min(mddev
->queue
, chunk_size
);
7399 blk_queue_io_opt(mddev
->queue
, chunk_size
*
7400 (conf
->raid_disks
- conf
->max_degraded
));
7401 mddev
->queue
->limits
.raid_partial_stripes_expensive
= 1;
7403 * We can only discard a whole stripe. It doesn't make sense to
7404 * discard data disk but write parity disk
7406 stripe
= stripe
* PAGE_SIZE
;
7407 /* Round up to power of 2, as discard handling
7408 * currently assumes that */
7409 while ((stripe
-1) & stripe
)
7410 stripe
= (stripe
| (stripe
-1)) + 1;
7411 mddev
->queue
->limits
.discard_alignment
= stripe
;
7412 mddev
->queue
->limits
.discard_granularity
= stripe
;
7414 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
7415 blk_queue_max_write_zeroes_sectors(mddev
->queue
, 0);
7417 rdev_for_each(rdev
, mddev
) {
7418 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
7419 rdev
->data_offset
<< 9);
7420 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
7421 rdev
->new_data_offset
<< 9);
7425 * zeroing is required, otherwise data
7426 * could be lost. Consider a scenario: discard a stripe
7427 * (the stripe could be inconsistent if
7428 * discard_zeroes_data is 0); write one disk of the
7429 * stripe (the stripe could be inconsistent again
7430 * depending on which disks are used to calculate
7431 * parity); the disk is broken; The stripe data of this
7434 * We only allow DISCARD if the sysadmin has confirmed that
7435 * only safe devices are in use by setting a module parameter.
7436 * A better idea might be to turn DISCARD into WRITE_ZEROES
7437 * requests, as that is required to be safe.
7439 if (devices_handle_discard_safely
&&
7440 mddev
->queue
->limits
.max_discard_sectors
>= (stripe
>> 9) &&
7441 mddev
->queue
->limits
.discard_granularity
>= stripe
)
7442 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
7445 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
7448 blk_queue_max_hw_sectors(mddev
->queue
, UINT_MAX
);
7451 if (log_init(conf
, journal_dev
, raid5_has_ppl(conf
)))
7456 md_unregister_thread(&mddev
->thread
);
7457 print_raid5_conf(conf
);
7459 mddev
->private = NULL
;
7460 pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev
));
7464 static void raid5_free(struct mddev
*mddev
, void *priv
)
7466 struct r5conf
*conf
= priv
;
7469 mddev
->to_remove
= &raid5_attrs_group
;
7472 static void raid5_status(struct seq_file
*seq
, struct mddev
*mddev
)
7474 struct r5conf
*conf
= mddev
->private;
7477 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
7478 conf
->chunk_sectors
/ 2, mddev
->layout
);
7479 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
7481 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7482 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
7483 seq_printf (seq
, "%s", rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
7486 seq_printf (seq
, "]");
7489 static void print_raid5_conf (struct r5conf
*conf
)
7492 struct disk_info
*tmp
;
7494 pr_debug("RAID conf printout:\n");
7496 pr_debug("(conf==NULL)\n");
7499 pr_debug(" --- level:%d rd:%d wd:%d\n", conf
->level
,
7501 conf
->raid_disks
- conf
->mddev
->degraded
);
7503 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7504 char b
[BDEVNAME_SIZE
];
7505 tmp
= conf
->disks
+ i
;
7507 pr_debug(" disk %d, o:%d, dev:%s\n",
7508 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
7509 bdevname(tmp
->rdev
->bdev
, b
));
7513 static int raid5_spare_active(struct mddev
*mddev
)
7516 struct r5conf
*conf
= mddev
->private;
7517 struct disk_info
*tmp
;
7519 unsigned long flags
;
7521 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7522 tmp
= conf
->disks
+ i
;
7523 if (tmp
->replacement
7524 && tmp
->replacement
->recovery_offset
== MaxSector
7525 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
7526 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
7527 /* Replacement has just become active. */
7529 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
7532 /* Replaced device not technically faulty,
7533 * but we need to be sure it gets removed
7534 * and never re-added.
7536 set_bit(Faulty
, &tmp
->rdev
->flags
);
7537 sysfs_notify_dirent_safe(
7538 tmp
->rdev
->sysfs_state
);
7540 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
7541 } else if (tmp
->rdev
7542 && tmp
->rdev
->recovery_offset
== MaxSector
7543 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
7544 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
7546 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
7549 spin_lock_irqsave(&conf
->device_lock
, flags
);
7550 mddev
->degraded
= raid5_calc_degraded(conf
);
7551 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7552 print_raid5_conf(conf
);
7556 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7558 struct r5conf
*conf
= mddev
->private;
7560 int number
= rdev
->raid_disk
;
7561 struct md_rdev
**rdevp
;
7562 struct disk_info
*p
= conf
->disks
+ number
;
7564 print_raid5_conf(conf
);
7565 if (test_bit(Journal
, &rdev
->flags
) && conf
->log
) {
7567 * we can't wait pending write here, as this is called in
7568 * raid5d, wait will deadlock.
7569 * neilb: there is no locking about new writes here,
7570 * so this cannot be safe.
7572 if (atomic_read(&conf
->active_stripes
) ||
7573 atomic_read(&conf
->r5c_cached_full_stripes
) ||
7574 atomic_read(&conf
->r5c_cached_partial_stripes
)) {
7580 if (rdev
== p
->rdev
)
7582 else if (rdev
== p
->replacement
)
7583 rdevp
= &p
->replacement
;
7587 if (number
>= conf
->raid_disks
&&
7588 conf
->reshape_progress
== MaxSector
)
7589 clear_bit(In_sync
, &rdev
->flags
);
7591 if (test_bit(In_sync
, &rdev
->flags
) ||
7592 atomic_read(&rdev
->nr_pending
)) {
7596 /* Only remove non-faulty devices if recovery
7599 if (!test_bit(Faulty
, &rdev
->flags
) &&
7600 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
7601 !has_failed(conf
) &&
7602 (!p
->replacement
|| p
->replacement
== rdev
) &&
7603 number
< conf
->raid_disks
) {
7608 if (!test_bit(RemoveSynchronized
, &rdev
->flags
)) {
7610 if (atomic_read(&rdev
->nr_pending
)) {
7611 /* lost the race, try later */
7617 err
= log_modify(conf
, rdev
, false);
7621 if (p
->replacement
) {
7622 /* We must have just cleared 'rdev' */
7623 p
->rdev
= p
->replacement
;
7624 clear_bit(Replacement
, &p
->replacement
->flags
);
7625 smp_mb(); /* Make sure other CPUs may see both as identical
7626 * but will never see neither - if they are careful
7628 p
->replacement
= NULL
;
7631 err
= log_modify(conf
, p
->rdev
, true);
7634 clear_bit(WantReplacement
, &rdev
->flags
);
7637 print_raid5_conf(conf
);
7641 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7643 struct r5conf
*conf
= mddev
->private;
7646 struct disk_info
*p
;
7648 int last
= conf
->raid_disks
- 1;
7650 if (test_bit(Journal
, &rdev
->flags
)) {
7654 rdev
->raid_disk
= 0;
7656 * The array is in readonly mode if journal is missing, so no
7657 * write requests running. We should be safe
7659 log_init(conf
, rdev
, false);
7662 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
7665 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
7666 /* no point adding a device */
7669 if (rdev
->raid_disk
>= 0)
7670 first
= last
= rdev
->raid_disk
;
7673 * find the disk ... but prefer rdev->saved_raid_disk
7676 if (rdev
->saved_raid_disk
>= 0 &&
7677 rdev
->saved_raid_disk
>= first
&&
7678 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
7679 first
= rdev
->saved_raid_disk
;
7681 for (disk
= first
; disk
<= last
; disk
++) {
7682 p
= conf
->disks
+ disk
;
7683 if (p
->rdev
== NULL
) {
7684 clear_bit(In_sync
, &rdev
->flags
);
7685 rdev
->raid_disk
= disk
;
7686 if (rdev
->saved_raid_disk
!= disk
)
7688 rcu_assign_pointer(p
->rdev
, rdev
);
7690 err
= log_modify(conf
, rdev
, true);
7695 for (disk
= first
; disk
<= last
; disk
++) {
7696 p
= conf
->disks
+ disk
;
7697 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
7698 p
->replacement
== NULL
) {
7699 clear_bit(In_sync
, &rdev
->flags
);
7700 set_bit(Replacement
, &rdev
->flags
);
7701 rdev
->raid_disk
= disk
;
7704 rcu_assign_pointer(p
->replacement
, rdev
);
7709 print_raid5_conf(conf
);
7713 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
7715 /* no resync is happening, and there is enough space
7716 * on all devices, so we can resize.
7717 * We need to make sure resync covers any new space.
7718 * If the array is shrinking we should possibly wait until
7719 * any io in the removed space completes, but it hardly seems
7723 struct r5conf
*conf
= mddev
->private;
7725 if (conf
->log
|| raid5_has_ppl(conf
))
7727 sectors
&= ~((sector_t
)conf
->chunk_sectors
- 1);
7728 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
7729 if (mddev
->external_size
&&
7730 mddev
->array_sectors
> newsize
)
7732 if (mddev
->bitmap
) {
7733 int ret
= bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
7737 md_set_array_sectors(mddev
, newsize
);
7738 if (sectors
> mddev
->dev_sectors
&&
7739 mddev
->recovery_cp
> mddev
->dev_sectors
) {
7740 mddev
->recovery_cp
= mddev
->dev_sectors
;
7741 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
7743 mddev
->dev_sectors
= sectors
;
7744 mddev
->resync_max_sectors
= sectors
;
7748 static int check_stripe_cache(struct mddev
*mddev
)
7750 /* Can only proceed if there are plenty of stripe_heads.
7751 * We need a minimum of one full stripe,, and for sensible progress
7752 * it is best to have about 4 times that.
7753 * If we require 4 times, then the default 256 4K stripe_heads will
7754 * allow for chunk sizes up to 256K, which is probably OK.
7755 * If the chunk size is greater, user-space should request more
7756 * stripe_heads first.
7758 struct r5conf
*conf
= mddev
->private;
7759 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7760 > conf
->min_nr_stripes
||
7761 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7762 > conf
->min_nr_stripes
) {
7763 pr_warn("md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7765 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
7772 static int check_reshape(struct mddev
*mddev
)
7774 struct r5conf
*conf
= mddev
->private;
7776 if (conf
->log
|| raid5_has_ppl(conf
))
7778 if (mddev
->delta_disks
== 0 &&
7779 mddev
->new_layout
== mddev
->layout
&&
7780 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
7781 return 0; /* nothing to do */
7782 if (has_failed(conf
))
7784 if (mddev
->delta_disks
< 0 && mddev
->reshape_position
== MaxSector
) {
7785 /* We might be able to shrink, but the devices must
7786 * be made bigger first.
7787 * For raid6, 4 is the minimum size.
7788 * Otherwise 2 is the minimum
7791 if (mddev
->level
== 6)
7793 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
7797 if (!check_stripe_cache(mddev
))
7800 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
||
7801 mddev
->delta_disks
> 0)
7802 if (resize_chunks(conf
,
7803 conf
->previous_raid_disks
7804 + max(0, mddev
->delta_disks
),
7805 max(mddev
->new_chunk_sectors
,
7806 mddev
->chunk_sectors
)
7810 if (conf
->previous_raid_disks
+ mddev
->delta_disks
<= conf
->pool_size
)
7811 return 0; /* never bother to shrink */
7812 return resize_stripes(conf
, (conf
->previous_raid_disks
7813 + mddev
->delta_disks
));
7816 static int raid5_start_reshape(struct mddev
*mddev
)
7818 struct r5conf
*conf
= mddev
->private;
7819 struct md_rdev
*rdev
;
7821 unsigned long flags
;
7823 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
7826 if (!check_stripe_cache(mddev
))
7829 if (has_failed(conf
))
7832 rdev_for_each(rdev
, mddev
) {
7833 if (!test_bit(In_sync
, &rdev
->flags
)
7834 && !test_bit(Faulty
, &rdev
->flags
))
7838 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
7839 /* Not enough devices even to make a degraded array
7844 /* Refuse to reduce size of the array. Any reductions in
7845 * array size must be through explicit setting of array_size
7848 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
7849 < mddev
->array_sectors
) {
7850 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7855 atomic_set(&conf
->reshape_stripes
, 0);
7856 spin_lock_irq(&conf
->device_lock
);
7857 write_seqcount_begin(&conf
->gen_lock
);
7858 conf
->previous_raid_disks
= conf
->raid_disks
;
7859 conf
->raid_disks
+= mddev
->delta_disks
;
7860 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
7861 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
7862 conf
->prev_algo
= conf
->algorithm
;
7863 conf
->algorithm
= mddev
->new_layout
;
7865 /* Code that selects data_offset needs to see the generation update
7866 * if reshape_progress has been set - so a memory barrier needed.
7869 if (mddev
->reshape_backwards
)
7870 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
7872 conf
->reshape_progress
= 0;
7873 conf
->reshape_safe
= conf
->reshape_progress
;
7874 write_seqcount_end(&conf
->gen_lock
);
7875 spin_unlock_irq(&conf
->device_lock
);
7877 /* Now make sure any requests that proceeded on the assumption
7878 * the reshape wasn't running - like Discard or Read - have
7881 mddev_suspend(mddev
);
7882 mddev_resume(mddev
);
7884 /* Add some new drives, as many as will fit.
7885 * We know there are enough to make the newly sized array work.
7886 * Don't add devices if we are reducing the number of
7887 * devices in the array. This is because it is not possible
7888 * to correctly record the "partially reconstructed" state of
7889 * such devices during the reshape and confusion could result.
7891 if (mddev
->delta_disks
>= 0) {
7892 rdev_for_each(rdev
, mddev
)
7893 if (rdev
->raid_disk
< 0 &&
7894 !test_bit(Faulty
, &rdev
->flags
)) {
7895 if (raid5_add_disk(mddev
, rdev
) == 0) {
7897 >= conf
->previous_raid_disks
)
7898 set_bit(In_sync
, &rdev
->flags
);
7900 rdev
->recovery_offset
= 0;
7902 if (sysfs_link_rdev(mddev
, rdev
))
7903 /* Failure here is OK */;
7905 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
7906 && !test_bit(Faulty
, &rdev
->flags
)) {
7907 /* This is a spare that was manually added */
7908 set_bit(In_sync
, &rdev
->flags
);
7911 /* When a reshape changes the number of devices,
7912 * ->degraded is measured against the larger of the
7913 * pre and post number of devices.
7915 spin_lock_irqsave(&conf
->device_lock
, flags
);
7916 mddev
->degraded
= raid5_calc_degraded(conf
);
7917 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7919 mddev
->raid_disks
= conf
->raid_disks
;
7920 mddev
->reshape_position
= conf
->reshape_progress
;
7921 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
7923 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
7924 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
7925 clear_bit(MD_RECOVERY_DONE
, &mddev
->recovery
);
7926 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
7927 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
7928 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
7930 if (!mddev
->sync_thread
) {
7931 mddev
->recovery
= 0;
7932 spin_lock_irq(&conf
->device_lock
);
7933 write_seqcount_begin(&conf
->gen_lock
);
7934 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
7935 mddev
->new_chunk_sectors
=
7936 conf
->chunk_sectors
= conf
->prev_chunk_sectors
;
7937 mddev
->new_layout
= conf
->algorithm
= conf
->prev_algo
;
7938 rdev_for_each(rdev
, mddev
)
7939 rdev
->new_data_offset
= rdev
->data_offset
;
7941 conf
->generation
--;
7942 conf
->reshape_progress
= MaxSector
;
7943 mddev
->reshape_position
= MaxSector
;
7944 write_seqcount_end(&conf
->gen_lock
);
7945 spin_unlock_irq(&conf
->device_lock
);
7948 conf
->reshape_checkpoint
= jiffies
;
7949 md_wakeup_thread(mddev
->sync_thread
);
7950 md_new_event(mddev
);
7954 /* This is called from the reshape thread and should make any
7955 * changes needed in 'conf'
7957 static void end_reshape(struct r5conf
*conf
)
7960 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
7962 spin_lock_irq(&conf
->device_lock
);
7963 conf
->previous_raid_disks
= conf
->raid_disks
;
7964 md_finish_reshape(conf
->mddev
);
7966 conf
->reshape_progress
= MaxSector
;
7967 conf
->mddev
->reshape_position
= MaxSector
;
7968 spin_unlock_irq(&conf
->device_lock
);
7969 wake_up(&conf
->wait_for_overlap
);
7971 /* read-ahead size must cover two whole stripes, which is
7972 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7974 if (conf
->mddev
->queue
) {
7975 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
7976 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
7978 if (conf
->mddev
->queue
->backing_dev_info
->ra_pages
< 2 * stripe
)
7979 conf
->mddev
->queue
->backing_dev_info
->ra_pages
= 2 * stripe
;
7984 /* This is called from the raid5d thread with mddev_lock held.
7985 * It makes config changes to the device.
7987 static void raid5_finish_reshape(struct mddev
*mddev
)
7989 struct r5conf
*conf
= mddev
->private;
7991 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
7993 if (mddev
->delta_disks
> 0) {
7994 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
7996 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
7997 revalidate_disk(mddev
->gendisk
);
8001 spin_lock_irq(&conf
->device_lock
);
8002 mddev
->degraded
= raid5_calc_degraded(conf
);
8003 spin_unlock_irq(&conf
->device_lock
);
8004 for (d
= conf
->raid_disks
;
8005 d
< conf
->raid_disks
- mddev
->delta_disks
;
8007 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
8009 clear_bit(In_sync
, &rdev
->flags
);
8010 rdev
= conf
->disks
[d
].replacement
;
8012 clear_bit(In_sync
, &rdev
->flags
);
8015 mddev
->layout
= conf
->algorithm
;
8016 mddev
->chunk_sectors
= conf
->chunk_sectors
;
8017 mddev
->reshape_position
= MaxSector
;
8018 mddev
->delta_disks
= 0;
8019 mddev
->reshape_backwards
= 0;
8023 static void raid5_quiesce(struct mddev
*mddev
, int state
)
8025 struct r5conf
*conf
= mddev
->private;
8028 case 2: /* resume for a suspend */
8029 wake_up(&conf
->wait_for_overlap
);
8032 case 1: /* stop all writes */
8033 lock_all_device_hash_locks_irq(conf
);
8034 /* '2' tells resync/reshape to pause so that all
8035 * active stripes can drain
8037 r5c_flush_cache(conf
, INT_MAX
);
8039 wait_event_cmd(conf
->wait_for_quiescent
,
8040 atomic_read(&conf
->active_stripes
) == 0 &&
8041 atomic_read(&conf
->active_aligned_reads
) == 0,
8042 unlock_all_device_hash_locks_irq(conf
),
8043 lock_all_device_hash_locks_irq(conf
));
8045 unlock_all_device_hash_locks_irq(conf
);
8046 /* allow reshape to continue */
8047 wake_up(&conf
->wait_for_overlap
);
8050 case 0: /* re-enable writes */
8051 lock_all_device_hash_locks_irq(conf
);
8053 wake_up(&conf
->wait_for_quiescent
);
8054 wake_up(&conf
->wait_for_overlap
);
8055 unlock_all_device_hash_locks_irq(conf
);
8058 r5l_quiesce(conf
->log
, state
);
8061 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
8063 struct r0conf
*raid0_conf
= mddev
->private;
8066 /* for raid0 takeover only one zone is supported */
8067 if (raid0_conf
->nr_strip_zones
> 1) {
8068 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8070 return ERR_PTR(-EINVAL
);
8073 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
8074 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
8075 mddev
->dev_sectors
= sectors
;
8076 mddev
->new_level
= level
;
8077 mddev
->new_layout
= ALGORITHM_PARITY_N
;
8078 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
8079 mddev
->raid_disks
+= 1;
8080 mddev
->delta_disks
= 1;
8081 /* make sure it will be not marked as dirty */
8082 mddev
->recovery_cp
= MaxSector
;
8084 return setup_conf(mddev
);
8087 static void *raid5_takeover_raid1(struct mddev
*mddev
)
8092 if (mddev
->raid_disks
!= 2 ||
8093 mddev
->degraded
> 1)
8094 return ERR_PTR(-EINVAL
);
8096 /* Should check if there are write-behind devices? */
8098 chunksect
= 64*2; /* 64K by default */
8100 /* The array must be an exact multiple of chunksize */
8101 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
8104 if ((chunksect
<<9) < STRIPE_SIZE
)
8105 /* array size does not allow a suitable chunk size */
8106 return ERR_PTR(-EINVAL
);
8108 mddev
->new_level
= 5;
8109 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
8110 mddev
->new_chunk_sectors
= chunksect
;
8112 ret
= setup_conf(mddev
);
8114 mddev_clear_unsupported_flags(mddev
,
8115 UNSUPPORTED_MDDEV_FLAGS
);
8119 static void *raid5_takeover_raid6(struct mddev
*mddev
)
8123 switch (mddev
->layout
) {
8124 case ALGORITHM_LEFT_ASYMMETRIC_6
:
8125 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
8127 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
8128 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
8130 case ALGORITHM_LEFT_SYMMETRIC_6
:
8131 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
8133 case ALGORITHM_RIGHT_SYMMETRIC_6
:
8134 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
8136 case ALGORITHM_PARITY_0_6
:
8137 new_layout
= ALGORITHM_PARITY_0
;
8139 case ALGORITHM_PARITY_N
:
8140 new_layout
= ALGORITHM_PARITY_N
;
8143 return ERR_PTR(-EINVAL
);
8145 mddev
->new_level
= 5;
8146 mddev
->new_layout
= new_layout
;
8147 mddev
->delta_disks
= -1;
8148 mddev
->raid_disks
-= 1;
8149 return setup_conf(mddev
);
8152 static int raid5_check_reshape(struct mddev
*mddev
)
8154 /* For a 2-drive array, the layout and chunk size can be changed
8155 * immediately as not restriping is needed.
8156 * For larger arrays we record the new value - after validation
8157 * to be used by a reshape pass.
8159 struct r5conf
*conf
= mddev
->private;
8160 int new_chunk
= mddev
->new_chunk_sectors
;
8162 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
8164 if (new_chunk
> 0) {
8165 if (!is_power_of_2(new_chunk
))
8167 if (new_chunk
< (PAGE_SIZE
>>9))
8169 if (mddev
->array_sectors
& (new_chunk
-1))
8170 /* not factor of array size */
8174 /* They look valid */
8176 if (mddev
->raid_disks
== 2) {
8177 /* can make the change immediately */
8178 if (mddev
->new_layout
>= 0) {
8179 conf
->algorithm
= mddev
->new_layout
;
8180 mddev
->layout
= mddev
->new_layout
;
8182 if (new_chunk
> 0) {
8183 conf
->chunk_sectors
= new_chunk
;
8184 mddev
->chunk_sectors
= new_chunk
;
8186 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
8187 md_wakeup_thread(mddev
->thread
);
8189 return check_reshape(mddev
);
8192 static int raid6_check_reshape(struct mddev
*mddev
)
8194 int new_chunk
= mddev
->new_chunk_sectors
;
8196 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
8198 if (new_chunk
> 0) {
8199 if (!is_power_of_2(new_chunk
))
8201 if (new_chunk
< (PAGE_SIZE
>> 9))
8203 if (mddev
->array_sectors
& (new_chunk
-1))
8204 /* not factor of array size */
8208 /* They look valid */
8209 return check_reshape(mddev
);
8212 static void *raid5_takeover(struct mddev
*mddev
)
8214 /* raid5 can take over:
8215 * raid0 - if there is only one strip zone - make it a raid4 layout
8216 * raid1 - if there are two drives. We need to know the chunk size
8217 * raid4 - trivial - just use a raid4 layout.
8218 * raid6 - Providing it is a *_6 layout
8220 if (mddev
->level
== 0)
8221 return raid45_takeover_raid0(mddev
, 5);
8222 if (mddev
->level
== 1)
8223 return raid5_takeover_raid1(mddev
);
8224 if (mddev
->level
== 4) {
8225 mddev
->new_layout
= ALGORITHM_PARITY_N
;
8226 mddev
->new_level
= 5;
8227 return setup_conf(mddev
);
8229 if (mddev
->level
== 6)
8230 return raid5_takeover_raid6(mddev
);
8232 return ERR_PTR(-EINVAL
);
8235 static void *raid4_takeover(struct mddev
*mddev
)
8237 /* raid4 can take over:
8238 * raid0 - if there is only one strip zone
8239 * raid5 - if layout is right
8241 if (mddev
->level
== 0)
8242 return raid45_takeover_raid0(mddev
, 4);
8243 if (mddev
->level
== 5 &&
8244 mddev
->layout
== ALGORITHM_PARITY_N
) {
8245 mddev
->new_layout
= 0;
8246 mddev
->new_level
= 4;
8247 return setup_conf(mddev
);
8249 return ERR_PTR(-EINVAL
);
8252 static struct md_personality raid5_personality
;
8254 static void *raid6_takeover(struct mddev
*mddev
)
8256 /* Currently can only take over a raid5. We map the
8257 * personality to an equivalent raid6 personality
8258 * with the Q block at the end.
8262 if (mddev
->pers
!= &raid5_personality
)
8263 return ERR_PTR(-EINVAL
);
8264 if (mddev
->degraded
> 1)
8265 return ERR_PTR(-EINVAL
);
8266 if (mddev
->raid_disks
> 253)
8267 return ERR_PTR(-EINVAL
);
8268 if (mddev
->raid_disks
< 3)
8269 return ERR_PTR(-EINVAL
);
8271 switch (mddev
->layout
) {
8272 case ALGORITHM_LEFT_ASYMMETRIC
:
8273 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
8275 case ALGORITHM_RIGHT_ASYMMETRIC
:
8276 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
8278 case ALGORITHM_LEFT_SYMMETRIC
:
8279 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
8281 case ALGORITHM_RIGHT_SYMMETRIC
:
8282 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
8284 case ALGORITHM_PARITY_0
:
8285 new_layout
= ALGORITHM_PARITY_0_6
;
8287 case ALGORITHM_PARITY_N
:
8288 new_layout
= ALGORITHM_PARITY_N
;
8291 return ERR_PTR(-EINVAL
);
8293 mddev
->new_level
= 6;
8294 mddev
->new_layout
= new_layout
;
8295 mddev
->delta_disks
= 1;
8296 mddev
->raid_disks
+= 1;
8297 return setup_conf(mddev
);
8300 static int raid5_change_consistency_policy(struct mddev
*mddev
, const char *buf
)
8302 struct r5conf
*conf
;
8305 err
= mddev_lock(mddev
);
8308 conf
= mddev
->private;
8310 mddev_unlock(mddev
);
8314 if (strncmp(buf
, "ppl", 3) == 0) {
8315 /* ppl only works with RAID 5 */
8316 if (!raid5_has_ppl(conf
) && conf
->level
== 5) {
8317 err
= log_init(conf
, NULL
, true);
8319 err
= resize_stripes(conf
, conf
->pool_size
);
8325 } else if (strncmp(buf
, "resync", 6) == 0) {
8326 if (raid5_has_ppl(conf
)) {
8327 mddev_suspend(mddev
);
8329 mddev_resume(mddev
);
8330 err
= resize_stripes(conf
, conf
->pool_size
);
8331 } else if (test_bit(MD_HAS_JOURNAL
, &conf
->mddev
->flags
) &&
8332 r5l_log_disk_error(conf
)) {
8333 bool journal_dev_exists
= false;
8334 struct md_rdev
*rdev
;
8336 rdev_for_each(rdev
, mddev
)
8337 if (test_bit(Journal
, &rdev
->flags
)) {
8338 journal_dev_exists
= true;
8342 if (!journal_dev_exists
) {
8343 mddev_suspend(mddev
);
8344 clear_bit(MD_HAS_JOURNAL
, &mddev
->flags
);
8345 mddev_resume(mddev
);
8346 } else /* need remove journal device first */
8355 md_update_sb(mddev
, 1);
8357 mddev_unlock(mddev
);
8362 static struct md_personality raid6_personality
=
8366 .owner
= THIS_MODULE
,
8367 .make_request
= raid5_make_request
,
8370 .status
= raid5_status
,
8371 .error_handler
= raid5_error
,
8372 .hot_add_disk
= raid5_add_disk
,
8373 .hot_remove_disk
= raid5_remove_disk
,
8374 .spare_active
= raid5_spare_active
,
8375 .sync_request
= raid5_sync_request
,
8376 .resize
= raid5_resize
,
8378 .check_reshape
= raid6_check_reshape
,
8379 .start_reshape
= raid5_start_reshape
,
8380 .finish_reshape
= raid5_finish_reshape
,
8381 .quiesce
= raid5_quiesce
,
8382 .takeover
= raid6_takeover
,
8383 .congested
= raid5_congested
,
8384 .change_consistency_policy
= raid5_change_consistency_policy
,
8386 static struct md_personality raid5_personality
=
8390 .owner
= THIS_MODULE
,
8391 .make_request
= raid5_make_request
,
8394 .status
= raid5_status
,
8395 .error_handler
= raid5_error
,
8396 .hot_add_disk
= raid5_add_disk
,
8397 .hot_remove_disk
= raid5_remove_disk
,
8398 .spare_active
= raid5_spare_active
,
8399 .sync_request
= raid5_sync_request
,
8400 .resize
= raid5_resize
,
8402 .check_reshape
= raid5_check_reshape
,
8403 .start_reshape
= raid5_start_reshape
,
8404 .finish_reshape
= raid5_finish_reshape
,
8405 .quiesce
= raid5_quiesce
,
8406 .takeover
= raid5_takeover
,
8407 .congested
= raid5_congested
,
8408 .change_consistency_policy
= raid5_change_consistency_policy
,
8411 static struct md_personality raid4_personality
=
8415 .owner
= THIS_MODULE
,
8416 .make_request
= raid5_make_request
,
8419 .status
= raid5_status
,
8420 .error_handler
= raid5_error
,
8421 .hot_add_disk
= raid5_add_disk
,
8422 .hot_remove_disk
= raid5_remove_disk
,
8423 .spare_active
= raid5_spare_active
,
8424 .sync_request
= raid5_sync_request
,
8425 .resize
= raid5_resize
,
8427 .check_reshape
= raid5_check_reshape
,
8428 .start_reshape
= raid5_start_reshape
,
8429 .finish_reshape
= raid5_finish_reshape
,
8430 .quiesce
= raid5_quiesce
,
8431 .takeover
= raid4_takeover
,
8432 .congested
= raid5_congested
,
8433 .change_consistency_policy
= raid5_change_consistency_policy
,
8436 static int __init
raid5_init(void)
8440 raid5_wq
= alloc_workqueue("raid5wq",
8441 WQ_UNBOUND
|WQ_MEM_RECLAIM
|WQ_CPU_INTENSIVE
|WQ_SYSFS
, 0);
8445 ret
= cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE
,
8447 raid456_cpu_up_prepare
,
8450 destroy_workqueue(raid5_wq
);
8453 register_md_personality(&raid6_personality
);
8454 register_md_personality(&raid5_personality
);
8455 register_md_personality(&raid4_personality
);
8459 static void raid5_exit(void)
8461 unregister_md_personality(&raid6_personality
);
8462 unregister_md_personality(&raid5_personality
);
8463 unregister_md_personality(&raid4_personality
);
8464 cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE
);
8465 destroy_workqueue(raid5_wq
);
8468 module_init(raid5_init
);
8469 module_exit(raid5_exit
);
8470 MODULE_LICENSE("GPL");
8471 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8472 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8473 MODULE_ALIAS("md-raid5");
8474 MODULE_ALIAS("md-raid4");
8475 MODULE_ALIAS("md-level-5");
8476 MODULE_ALIAS("md-level-4");
8477 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8478 MODULE_ALIAS("md-raid6");
8479 MODULE_ALIAS("md-level-6");
8481 /* This used to be two separate modules, they were: */
8482 MODULE_ALIAS("raid5");
8483 MODULE_ALIAS("raid6");