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>
59 #include <trace/events/block.h>
60 #include <linux/list_sort.h>
65 #include "md-bitmap.h"
66 #include "raid5-log.h"
68 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
70 #define cpu_to_group(cpu) cpu_to_node(cpu)
71 #define ANY_GROUP NUMA_NO_NODE
73 static bool devices_handle_discard_safely
= false;
74 module_param(devices_handle_discard_safely
, bool, 0644);
75 MODULE_PARM_DESC(devices_handle_discard_safely
,
76 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
77 static struct workqueue_struct
*raid5_wq
;
79 static inline struct hlist_head
*stripe_hash(struct r5conf
*conf
, sector_t sect
)
81 int hash
= (sect
>> STRIPE_SHIFT
) & HASH_MASK
;
82 return &conf
->stripe_hashtbl
[hash
];
85 static inline int stripe_hash_locks_hash(sector_t sect
)
87 return (sect
>> STRIPE_SHIFT
) & STRIPE_HASH_LOCKS_MASK
;
90 static inline void lock_device_hash_lock(struct r5conf
*conf
, int hash
)
92 spin_lock_irq(conf
->hash_locks
+ hash
);
93 spin_lock(&conf
->device_lock
);
96 static inline void unlock_device_hash_lock(struct r5conf
*conf
, int hash
)
98 spin_unlock(&conf
->device_lock
);
99 spin_unlock_irq(conf
->hash_locks
+ hash
);
102 static inline void lock_all_device_hash_locks_irq(struct r5conf
*conf
)
105 spin_lock_irq(conf
->hash_locks
);
106 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
107 spin_lock_nest_lock(conf
->hash_locks
+ i
, conf
->hash_locks
);
108 spin_lock(&conf
->device_lock
);
111 static inline void unlock_all_device_hash_locks_irq(struct r5conf
*conf
)
114 spin_unlock(&conf
->device_lock
);
115 for (i
= NR_STRIPE_HASH_LOCKS
- 1; i
; i
--)
116 spin_unlock(conf
->hash_locks
+ i
);
117 spin_unlock_irq(conf
->hash_locks
);
120 /* Find first data disk in a raid6 stripe */
121 static inline int raid6_d0(struct stripe_head
*sh
)
124 /* ddf always start from first device */
126 /* md starts just after Q block */
127 if (sh
->qd_idx
== sh
->disks
- 1)
130 return sh
->qd_idx
+ 1;
132 static inline int raid6_next_disk(int disk
, int raid_disks
)
135 return (disk
< raid_disks
) ? disk
: 0;
138 /* When walking through the disks in a raid5, starting at raid6_d0,
139 * We need to map each disk to a 'slot', where the data disks are slot
140 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
141 * is raid_disks-1. This help does that mapping.
143 static int raid6_idx_to_slot(int idx
, struct stripe_head
*sh
,
144 int *count
, int syndrome_disks
)
150 if (idx
== sh
->pd_idx
)
151 return syndrome_disks
;
152 if (idx
== sh
->qd_idx
)
153 return syndrome_disks
+ 1;
159 static void print_raid5_conf (struct r5conf
*conf
);
161 static int stripe_operations_active(struct stripe_head
*sh
)
163 return sh
->check_state
|| sh
->reconstruct_state
||
164 test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
) ||
165 test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
168 static bool stripe_is_lowprio(struct stripe_head
*sh
)
170 return (test_bit(STRIPE_R5C_FULL_STRIPE
, &sh
->state
) ||
171 test_bit(STRIPE_R5C_PARTIAL_STRIPE
, &sh
->state
)) &&
172 !test_bit(STRIPE_R5C_CACHING
, &sh
->state
);
175 static void raid5_wakeup_stripe_thread(struct stripe_head
*sh
)
177 struct r5conf
*conf
= sh
->raid_conf
;
178 struct r5worker_group
*group
;
180 int i
, cpu
= sh
->cpu
;
182 if (!cpu_online(cpu
)) {
183 cpu
= cpumask_any(cpu_online_mask
);
187 if (list_empty(&sh
->lru
)) {
188 struct r5worker_group
*group
;
189 group
= conf
->worker_groups
+ cpu_to_group(cpu
);
190 if (stripe_is_lowprio(sh
))
191 list_add_tail(&sh
->lru
, &group
->loprio_list
);
193 list_add_tail(&sh
->lru
, &group
->handle_list
);
194 group
->stripes_cnt
++;
198 if (conf
->worker_cnt_per_group
== 0) {
199 md_wakeup_thread(conf
->mddev
->thread
);
203 group
= conf
->worker_groups
+ cpu_to_group(sh
->cpu
);
205 group
->workers
[0].working
= true;
206 /* at least one worker should run to avoid race */
207 queue_work_on(sh
->cpu
, raid5_wq
, &group
->workers
[0].work
);
209 thread_cnt
= group
->stripes_cnt
/ MAX_STRIPE_BATCH
- 1;
210 /* wakeup more workers */
211 for (i
= 1; i
< conf
->worker_cnt_per_group
&& thread_cnt
> 0; i
++) {
212 if (group
->workers
[i
].working
== false) {
213 group
->workers
[i
].working
= true;
214 queue_work_on(sh
->cpu
, raid5_wq
,
215 &group
->workers
[i
].work
);
221 static void do_release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
222 struct list_head
*temp_inactive_list
)
225 int injournal
= 0; /* number of date pages with R5_InJournal */
227 BUG_ON(!list_empty(&sh
->lru
));
228 BUG_ON(atomic_read(&conf
->active_stripes
)==0);
230 if (r5c_is_writeback(conf
->log
))
231 for (i
= sh
->disks
; i
--; )
232 if (test_bit(R5_InJournal
, &sh
->dev
[i
].flags
))
235 * In the following cases, the stripe cannot be released to cached
236 * lists. Therefore, we make the stripe write out and set
238 * 1. when quiesce in r5c write back;
239 * 2. when resync is requested fot the stripe.
241 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
) ||
242 (conf
->quiesce
&& r5c_is_writeback(conf
->log
) &&
243 !test_bit(STRIPE_HANDLE
, &sh
->state
) && injournal
!= 0)) {
244 if (test_bit(STRIPE_R5C_CACHING
, &sh
->state
))
245 r5c_make_stripe_write_out(sh
);
246 set_bit(STRIPE_HANDLE
, &sh
->state
);
249 if (test_bit(STRIPE_HANDLE
, &sh
->state
)) {
250 if (test_bit(STRIPE_DELAYED
, &sh
->state
) &&
251 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
252 list_add_tail(&sh
->lru
, &conf
->delayed_list
);
253 else if (test_bit(STRIPE_BIT_DELAY
, &sh
->state
) &&
254 sh
->bm_seq
- conf
->seq_write
> 0)
255 list_add_tail(&sh
->lru
, &conf
->bitmap_list
);
257 clear_bit(STRIPE_DELAYED
, &sh
->state
);
258 clear_bit(STRIPE_BIT_DELAY
, &sh
->state
);
259 if (conf
->worker_cnt_per_group
== 0) {
260 if (stripe_is_lowprio(sh
))
261 list_add_tail(&sh
->lru
,
264 list_add_tail(&sh
->lru
,
267 raid5_wakeup_stripe_thread(sh
);
271 md_wakeup_thread(conf
->mddev
->thread
);
273 BUG_ON(stripe_operations_active(sh
));
274 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
275 if (atomic_dec_return(&conf
->preread_active_stripes
)
277 md_wakeup_thread(conf
->mddev
->thread
);
278 atomic_dec(&conf
->active_stripes
);
279 if (!test_bit(STRIPE_EXPANDING
, &sh
->state
)) {
280 if (!r5c_is_writeback(conf
->log
))
281 list_add_tail(&sh
->lru
, temp_inactive_list
);
283 WARN_ON(test_bit(R5_InJournal
, &sh
->dev
[sh
->pd_idx
].flags
));
285 list_add_tail(&sh
->lru
, temp_inactive_list
);
286 else if (injournal
== conf
->raid_disks
- conf
->max_degraded
) {
288 if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE
, &sh
->state
))
289 atomic_inc(&conf
->r5c_cached_full_stripes
);
290 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE
, &sh
->state
))
291 atomic_dec(&conf
->r5c_cached_partial_stripes
);
292 list_add_tail(&sh
->lru
, &conf
->r5c_full_stripe_list
);
293 r5c_check_cached_full_stripe(conf
);
296 * STRIPE_R5C_PARTIAL_STRIPE is set in
297 * r5c_try_caching_write(). No need to
300 list_add_tail(&sh
->lru
, &conf
->r5c_partial_stripe_list
);
306 static void __release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
307 struct list_head
*temp_inactive_list
)
309 if (atomic_dec_and_test(&sh
->count
))
310 do_release_stripe(conf
, sh
, temp_inactive_list
);
314 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
316 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
317 * given time. Adding stripes only takes device lock, while deleting stripes
318 * only takes hash lock.
320 static void release_inactive_stripe_list(struct r5conf
*conf
,
321 struct list_head
*temp_inactive_list
,
325 bool do_wakeup
= false;
328 if (hash
== NR_STRIPE_HASH_LOCKS
) {
329 size
= NR_STRIPE_HASH_LOCKS
;
330 hash
= NR_STRIPE_HASH_LOCKS
- 1;
334 struct list_head
*list
= &temp_inactive_list
[size
- 1];
337 * We don't hold any lock here yet, raid5_get_active_stripe() might
338 * remove stripes from the list
340 if (!list_empty_careful(list
)) {
341 spin_lock_irqsave(conf
->hash_locks
+ hash
, flags
);
342 if (list_empty(conf
->inactive_list
+ hash
) &&
344 atomic_dec(&conf
->empty_inactive_list_nr
);
345 list_splice_tail_init(list
, conf
->inactive_list
+ hash
);
347 spin_unlock_irqrestore(conf
->hash_locks
+ hash
, flags
);
354 wake_up(&conf
->wait_for_stripe
);
355 if (atomic_read(&conf
->active_stripes
) == 0)
356 wake_up(&conf
->wait_for_quiescent
);
357 if (conf
->retry_read_aligned
)
358 md_wakeup_thread(conf
->mddev
->thread
);
362 /* should hold conf->device_lock already */
363 static int release_stripe_list(struct r5conf
*conf
,
364 struct list_head
*temp_inactive_list
)
366 struct stripe_head
*sh
, *t
;
368 struct llist_node
*head
;
370 head
= llist_del_all(&conf
->released_stripes
);
371 head
= llist_reverse_order(head
);
372 llist_for_each_entry_safe(sh
, t
, head
, release_list
) {
375 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
377 clear_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
);
379 * Don't worry the bit is set here, because if the bit is set
380 * again, the count is always > 1. This is true for
381 * STRIPE_ON_UNPLUG_LIST bit too.
383 hash
= sh
->hash_lock_index
;
384 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
391 void raid5_release_stripe(struct stripe_head
*sh
)
393 struct r5conf
*conf
= sh
->raid_conf
;
395 struct list_head list
;
399 /* Avoid release_list until the last reference.
401 if (atomic_add_unless(&sh
->count
, -1, 1))
404 if (unlikely(!conf
->mddev
->thread
) ||
405 test_and_set_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
))
407 wakeup
= llist_add(&sh
->release_list
, &conf
->released_stripes
);
409 md_wakeup_thread(conf
->mddev
->thread
);
412 local_irq_save(flags
);
413 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
414 if (atomic_dec_and_lock(&sh
->count
, &conf
->device_lock
)) {
415 INIT_LIST_HEAD(&list
);
416 hash
= sh
->hash_lock_index
;
417 do_release_stripe(conf
, sh
, &list
);
418 spin_unlock(&conf
->device_lock
);
419 release_inactive_stripe_list(conf
, &list
, hash
);
421 local_irq_restore(flags
);
424 static inline void remove_hash(struct stripe_head
*sh
)
426 pr_debug("remove_hash(), stripe %llu\n",
427 (unsigned long long)sh
->sector
);
429 hlist_del_init(&sh
->hash
);
432 static inline void insert_hash(struct r5conf
*conf
, struct stripe_head
*sh
)
434 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
436 pr_debug("insert_hash(), stripe %llu\n",
437 (unsigned long long)sh
->sector
);
439 hlist_add_head(&sh
->hash
, hp
);
442 /* find an idle stripe, make sure it is unhashed, and return it. */
443 static struct stripe_head
*get_free_stripe(struct r5conf
*conf
, int hash
)
445 struct stripe_head
*sh
= NULL
;
446 struct list_head
*first
;
448 if (list_empty(conf
->inactive_list
+ hash
))
450 first
= (conf
->inactive_list
+ hash
)->next
;
451 sh
= list_entry(first
, struct stripe_head
, lru
);
452 list_del_init(first
);
454 atomic_inc(&conf
->active_stripes
);
455 BUG_ON(hash
!= sh
->hash_lock_index
);
456 if (list_empty(conf
->inactive_list
+ hash
))
457 atomic_inc(&conf
->empty_inactive_list_nr
);
462 static void shrink_buffers(struct stripe_head
*sh
)
466 int num
= sh
->raid_conf
->pool_size
;
468 for (i
= 0; i
< num
; i
++) {
469 WARN_ON(sh
->dev
[i
].page
!= sh
->dev
[i
].orig_page
);
473 sh
->dev
[i
].page
= NULL
;
478 static int grow_buffers(struct stripe_head
*sh
, gfp_t gfp
)
481 int num
= sh
->raid_conf
->pool_size
;
483 for (i
= 0; i
< num
; i
++) {
486 if (!(page
= alloc_page(gfp
))) {
489 sh
->dev
[i
].page
= page
;
490 sh
->dev
[i
].orig_page
= page
;
496 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
497 struct stripe_head
*sh
);
499 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int previous
)
501 struct r5conf
*conf
= sh
->raid_conf
;
504 BUG_ON(atomic_read(&sh
->count
) != 0);
505 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
506 BUG_ON(stripe_operations_active(sh
));
507 BUG_ON(sh
->batch_head
);
509 pr_debug("init_stripe called, stripe %llu\n",
510 (unsigned long long)sector
);
512 seq
= read_seqcount_begin(&conf
->gen_lock
);
513 sh
->generation
= conf
->generation
- previous
;
514 sh
->disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
516 stripe_set_idx(sector
, conf
, previous
, sh
);
519 for (i
= sh
->disks
; i
--; ) {
520 struct r5dev
*dev
= &sh
->dev
[i
];
522 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
523 test_bit(R5_LOCKED
, &dev
->flags
)) {
524 pr_err("sector=%llx i=%d %p %p %p %p %d\n",
525 (unsigned long long)sh
->sector
, i
, dev
->toread
,
526 dev
->read
, dev
->towrite
, dev
->written
,
527 test_bit(R5_LOCKED
, &dev
->flags
));
531 dev
->sector
= raid5_compute_blocknr(sh
, i
, previous
);
533 if (read_seqcount_retry(&conf
->gen_lock
, seq
))
535 sh
->overwrite_disks
= 0;
536 insert_hash(conf
, sh
);
537 sh
->cpu
= smp_processor_id();
538 set_bit(STRIPE_BATCH_READY
, &sh
->state
);
541 static struct stripe_head
*__find_stripe(struct r5conf
*conf
, sector_t sector
,
544 struct stripe_head
*sh
;
546 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
547 hlist_for_each_entry(sh
, stripe_hash(conf
, sector
), hash
)
548 if (sh
->sector
== sector
&& sh
->generation
== generation
)
550 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
555 * Need to check if array has failed when deciding whether to:
557 * - remove non-faulty devices
560 * This determination is simple when no reshape is happening.
561 * However if there is a reshape, we need to carefully check
562 * both the before and after sections.
563 * This is because some failed devices may only affect one
564 * of the two sections, and some non-in_sync devices may
565 * be insync in the section most affected by failed devices.
567 int raid5_calc_degraded(struct r5conf
*conf
)
569 int degraded
, degraded2
;
574 for (i
= 0; i
< conf
->previous_raid_disks
; i
++) {
575 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
576 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
577 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
578 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
580 else if (test_bit(In_sync
, &rdev
->flags
))
583 /* not in-sync or faulty.
584 * If the reshape increases the number of devices,
585 * this is being recovered by the reshape, so
586 * this 'previous' section is not in_sync.
587 * If the number of devices is being reduced however,
588 * the device can only be part of the array if
589 * we are reverting a reshape, so this section will
592 if (conf
->raid_disks
>= conf
->previous_raid_disks
)
596 if (conf
->raid_disks
== conf
->previous_raid_disks
)
600 for (i
= 0; i
< conf
->raid_disks
; i
++) {
601 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
602 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
603 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
604 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
606 else if (test_bit(In_sync
, &rdev
->flags
))
609 /* not in-sync or faulty.
610 * If reshape increases the number of devices, this
611 * section has already been recovered, else it
612 * almost certainly hasn't.
614 if (conf
->raid_disks
<= conf
->previous_raid_disks
)
618 if (degraded2
> degraded
)
623 static int has_failed(struct r5conf
*conf
)
627 if (conf
->mddev
->reshape_position
== MaxSector
)
628 return conf
->mddev
->degraded
> conf
->max_degraded
;
630 degraded
= raid5_calc_degraded(conf
);
631 if (degraded
> conf
->max_degraded
)
637 raid5_get_active_stripe(struct r5conf
*conf
, sector_t sector
,
638 int previous
, int noblock
, int noquiesce
)
640 struct stripe_head
*sh
;
641 int hash
= stripe_hash_locks_hash(sector
);
642 int inc_empty_inactive_list_flag
;
644 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
646 spin_lock_irq(conf
->hash_locks
+ hash
);
649 wait_event_lock_irq(conf
->wait_for_quiescent
,
650 conf
->quiesce
== 0 || noquiesce
,
651 *(conf
->hash_locks
+ hash
));
652 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
654 if (!test_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
)) {
655 sh
= get_free_stripe(conf
, hash
);
656 if (!sh
&& !test_bit(R5_DID_ALLOC
,
658 set_bit(R5_ALLOC_MORE
,
661 if (noblock
&& sh
== NULL
)
664 r5c_check_stripe_cache_usage(conf
);
666 set_bit(R5_INACTIVE_BLOCKED
,
668 r5l_wake_reclaim(conf
->log
, 0);
670 conf
->wait_for_stripe
,
671 !list_empty(conf
->inactive_list
+ hash
) &&
672 (atomic_read(&conf
->active_stripes
)
673 < (conf
->max_nr_stripes
* 3 / 4)
674 || !test_bit(R5_INACTIVE_BLOCKED
,
675 &conf
->cache_state
)),
676 *(conf
->hash_locks
+ hash
));
677 clear_bit(R5_INACTIVE_BLOCKED
,
680 init_stripe(sh
, sector
, previous
);
681 atomic_inc(&sh
->count
);
683 } else if (!atomic_inc_not_zero(&sh
->count
)) {
684 spin_lock(&conf
->device_lock
);
685 if (!atomic_read(&sh
->count
)) {
686 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
687 atomic_inc(&conf
->active_stripes
);
688 BUG_ON(list_empty(&sh
->lru
) &&
689 !test_bit(STRIPE_EXPANDING
, &sh
->state
));
690 inc_empty_inactive_list_flag
= 0;
691 if (!list_empty(conf
->inactive_list
+ hash
))
692 inc_empty_inactive_list_flag
= 1;
693 list_del_init(&sh
->lru
);
694 if (list_empty(conf
->inactive_list
+ hash
) && inc_empty_inactive_list_flag
)
695 atomic_inc(&conf
->empty_inactive_list_nr
);
697 sh
->group
->stripes_cnt
--;
701 atomic_inc(&sh
->count
);
702 spin_unlock(&conf
->device_lock
);
704 } while (sh
== NULL
);
706 spin_unlock_irq(conf
->hash_locks
+ hash
);
710 static bool is_full_stripe_write(struct stripe_head
*sh
)
712 BUG_ON(sh
->overwrite_disks
> (sh
->disks
- sh
->raid_conf
->max_degraded
));
713 return sh
->overwrite_disks
== (sh
->disks
- sh
->raid_conf
->max_degraded
);
716 static void lock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
719 spin_lock_irq(&sh2
->stripe_lock
);
720 spin_lock_nested(&sh1
->stripe_lock
, 1);
722 spin_lock_irq(&sh1
->stripe_lock
);
723 spin_lock_nested(&sh2
->stripe_lock
, 1);
727 static void unlock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
729 spin_unlock(&sh1
->stripe_lock
);
730 spin_unlock_irq(&sh2
->stripe_lock
);
733 /* Only freshly new full stripe normal write stripe can be added to a batch list */
734 static bool stripe_can_batch(struct stripe_head
*sh
)
736 struct r5conf
*conf
= sh
->raid_conf
;
738 if (conf
->log
|| raid5_has_ppl(conf
))
740 return test_bit(STRIPE_BATCH_READY
, &sh
->state
) &&
741 !test_bit(STRIPE_BITMAP_PENDING
, &sh
->state
) &&
742 is_full_stripe_write(sh
);
745 /* we only do back search */
746 static void stripe_add_to_batch_list(struct r5conf
*conf
, struct stripe_head
*sh
)
748 struct stripe_head
*head
;
749 sector_t head_sector
, tmp_sec
;
752 int inc_empty_inactive_list_flag
;
754 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
755 tmp_sec
= sh
->sector
;
756 if (!sector_div(tmp_sec
, conf
->chunk_sectors
))
758 head_sector
= sh
->sector
- STRIPE_SECTORS
;
760 hash
= stripe_hash_locks_hash(head_sector
);
761 spin_lock_irq(conf
->hash_locks
+ hash
);
762 head
= __find_stripe(conf
, head_sector
, conf
->generation
);
763 if (head
&& !atomic_inc_not_zero(&head
->count
)) {
764 spin_lock(&conf
->device_lock
);
765 if (!atomic_read(&head
->count
)) {
766 if (!test_bit(STRIPE_HANDLE
, &head
->state
))
767 atomic_inc(&conf
->active_stripes
);
768 BUG_ON(list_empty(&head
->lru
) &&
769 !test_bit(STRIPE_EXPANDING
, &head
->state
));
770 inc_empty_inactive_list_flag
= 0;
771 if (!list_empty(conf
->inactive_list
+ hash
))
772 inc_empty_inactive_list_flag
= 1;
773 list_del_init(&head
->lru
);
774 if (list_empty(conf
->inactive_list
+ hash
) && inc_empty_inactive_list_flag
)
775 atomic_inc(&conf
->empty_inactive_list_nr
);
777 head
->group
->stripes_cnt
--;
781 atomic_inc(&head
->count
);
782 spin_unlock(&conf
->device_lock
);
784 spin_unlock_irq(conf
->hash_locks
+ hash
);
788 if (!stripe_can_batch(head
))
791 lock_two_stripes(head
, sh
);
792 /* clear_batch_ready clear the flag */
793 if (!stripe_can_batch(head
) || !stripe_can_batch(sh
))
800 while (dd_idx
== sh
->pd_idx
|| dd_idx
== sh
->qd_idx
)
802 if (head
->dev
[dd_idx
].towrite
->bi_opf
!= sh
->dev
[dd_idx
].towrite
->bi_opf
||
803 bio_op(head
->dev
[dd_idx
].towrite
) != bio_op(sh
->dev
[dd_idx
].towrite
))
806 if (head
->batch_head
) {
807 spin_lock(&head
->batch_head
->batch_lock
);
808 /* This batch list is already running */
809 if (!stripe_can_batch(head
)) {
810 spin_unlock(&head
->batch_head
->batch_lock
);
814 * We must assign batch_head of this stripe within the
815 * batch_lock, otherwise clear_batch_ready of batch head
816 * stripe could clear BATCH_READY bit of this stripe and
817 * this stripe->batch_head doesn't get assigned, which
818 * could confuse clear_batch_ready for this stripe
820 sh
->batch_head
= head
->batch_head
;
823 * at this point, head's BATCH_READY could be cleared, but we
824 * can still add the stripe to batch list
826 list_add(&sh
->batch_list
, &head
->batch_list
);
827 spin_unlock(&head
->batch_head
->batch_lock
);
829 head
->batch_head
= head
;
830 sh
->batch_head
= head
->batch_head
;
831 spin_lock(&head
->batch_lock
);
832 list_add_tail(&sh
->batch_list
, &head
->batch_list
);
833 spin_unlock(&head
->batch_lock
);
836 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
837 if (atomic_dec_return(&conf
->preread_active_stripes
)
839 md_wakeup_thread(conf
->mddev
->thread
);
841 if (test_and_clear_bit(STRIPE_BIT_DELAY
, &sh
->state
)) {
842 int seq
= sh
->bm_seq
;
843 if (test_bit(STRIPE_BIT_DELAY
, &sh
->batch_head
->state
) &&
844 sh
->batch_head
->bm_seq
> seq
)
845 seq
= sh
->batch_head
->bm_seq
;
846 set_bit(STRIPE_BIT_DELAY
, &sh
->batch_head
->state
);
847 sh
->batch_head
->bm_seq
= seq
;
850 atomic_inc(&sh
->count
);
852 unlock_two_stripes(head
, sh
);
854 raid5_release_stripe(head
);
857 /* Determine if 'data_offset' or 'new_data_offset' should be used
858 * in this stripe_head.
860 static int use_new_offset(struct r5conf
*conf
, struct stripe_head
*sh
)
862 sector_t progress
= conf
->reshape_progress
;
863 /* Need a memory barrier to make sure we see the value
864 * of conf->generation, or ->data_offset that was set before
865 * reshape_progress was updated.
868 if (progress
== MaxSector
)
870 if (sh
->generation
== conf
->generation
- 1)
872 /* We are in a reshape, and this is a new-generation stripe,
873 * so use new_data_offset.
878 static void dispatch_bio_list(struct bio_list
*tmp
)
882 while ((bio
= bio_list_pop(tmp
)))
883 generic_make_request(bio
);
886 static int cmp_stripe(void *priv
, struct list_head
*a
, struct list_head
*b
)
888 const struct r5pending_data
*da
= list_entry(a
,
889 struct r5pending_data
, sibling
);
890 const struct r5pending_data
*db
= list_entry(b
,
891 struct r5pending_data
, sibling
);
892 if (da
->sector
> db
->sector
)
894 if (da
->sector
< db
->sector
)
899 static void dispatch_defer_bios(struct r5conf
*conf
, int target
,
900 struct bio_list
*list
)
902 struct r5pending_data
*data
;
903 struct list_head
*first
, *next
= NULL
;
906 if (conf
->pending_data_cnt
== 0)
909 list_sort(NULL
, &conf
->pending_list
, cmp_stripe
);
911 first
= conf
->pending_list
.next
;
913 /* temporarily move the head */
914 if (conf
->next_pending_data
)
915 list_move_tail(&conf
->pending_list
,
916 &conf
->next_pending_data
->sibling
);
918 while (!list_empty(&conf
->pending_list
)) {
919 data
= list_first_entry(&conf
->pending_list
,
920 struct r5pending_data
, sibling
);
921 if (&data
->sibling
== first
)
922 first
= data
->sibling
.next
;
923 next
= data
->sibling
.next
;
925 bio_list_merge(list
, &data
->bios
);
926 list_move(&data
->sibling
, &conf
->free_list
);
931 conf
->pending_data_cnt
-= cnt
;
932 BUG_ON(conf
->pending_data_cnt
< 0 || cnt
< target
);
934 if (next
!= &conf
->pending_list
)
935 conf
->next_pending_data
= list_entry(next
,
936 struct r5pending_data
, sibling
);
938 conf
->next_pending_data
= NULL
;
939 /* list isn't empty */
940 if (first
!= &conf
->pending_list
)
941 list_move_tail(&conf
->pending_list
, first
);
944 static void flush_deferred_bios(struct r5conf
*conf
)
946 struct bio_list tmp
= BIO_EMPTY_LIST
;
948 if (conf
->pending_data_cnt
== 0)
951 spin_lock(&conf
->pending_bios_lock
);
952 dispatch_defer_bios(conf
, conf
->pending_data_cnt
, &tmp
);
953 BUG_ON(conf
->pending_data_cnt
!= 0);
954 spin_unlock(&conf
->pending_bios_lock
);
956 dispatch_bio_list(&tmp
);
959 static void defer_issue_bios(struct r5conf
*conf
, sector_t sector
,
960 struct bio_list
*bios
)
962 struct bio_list tmp
= BIO_EMPTY_LIST
;
963 struct r5pending_data
*ent
;
965 spin_lock(&conf
->pending_bios_lock
);
966 ent
= list_first_entry(&conf
->free_list
, struct r5pending_data
,
968 list_move_tail(&ent
->sibling
, &conf
->pending_list
);
969 ent
->sector
= sector
;
970 bio_list_init(&ent
->bios
);
971 bio_list_merge(&ent
->bios
, bios
);
972 conf
->pending_data_cnt
++;
973 if (conf
->pending_data_cnt
>= PENDING_IO_MAX
)
974 dispatch_defer_bios(conf
, PENDING_IO_ONE_FLUSH
, &tmp
);
976 spin_unlock(&conf
->pending_bios_lock
);
978 dispatch_bio_list(&tmp
);
982 raid5_end_read_request(struct bio
*bi
);
984 raid5_end_write_request(struct bio
*bi
);
986 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
988 struct r5conf
*conf
= sh
->raid_conf
;
989 int i
, disks
= sh
->disks
;
990 struct stripe_head
*head_sh
= sh
;
991 struct bio_list pending_bios
= BIO_EMPTY_LIST
;
996 if (log_stripe(sh
, s
) == 0)
999 should_defer
= conf
->batch_bio_dispatch
&& conf
->group_cnt
;
1001 for (i
= disks
; i
--; ) {
1002 int op
, op_flags
= 0;
1003 int replace_only
= 0;
1004 struct bio
*bi
, *rbi
;
1005 struct md_rdev
*rdev
, *rrdev
= NULL
;
1008 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
1010 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
1012 if (test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1013 op
= REQ_OP_DISCARD
;
1014 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
1016 else if (test_and_clear_bit(R5_WantReplace
,
1017 &sh
->dev
[i
].flags
)) {
1022 if (test_and_clear_bit(R5_SyncIO
, &sh
->dev
[i
].flags
))
1023 op_flags
|= REQ_SYNC
;
1026 bi
= &sh
->dev
[i
].req
;
1027 rbi
= &sh
->dev
[i
].rreq
; /* For writing to replacement */
1030 rrdev
= rcu_dereference(conf
->disks
[i
].replacement
);
1031 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1032 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
1037 if (op_is_write(op
)) {
1041 /* We raced and saw duplicates */
1044 if (test_bit(R5_ReadRepl
, &head_sh
->dev
[i
].flags
) && rrdev
)
1049 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
1052 atomic_inc(&rdev
->nr_pending
);
1053 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
1056 atomic_inc(&rrdev
->nr_pending
);
1059 /* We have already checked bad blocks for reads. Now
1060 * need to check for writes. We never accept write errors
1061 * on the replacement, so we don't to check rrdev.
1063 while (op_is_write(op
) && rdev
&&
1064 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1067 int bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
1068 &first_bad
, &bad_sectors
);
1073 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1074 if (!conf
->mddev
->external
&&
1075 conf
->mddev
->sb_flags
) {
1076 /* It is very unlikely, but we might
1077 * still need to write out the
1078 * bad block log - better give it
1080 md_check_recovery(conf
->mddev
);
1083 * Because md_wait_for_blocked_rdev
1084 * will dec nr_pending, we must
1085 * increment it first.
1087 atomic_inc(&rdev
->nr_pending
);
1088 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
1090 /* Acknowledged bad block - skip the write */
1091 rdev_dec_pending(rdev
, conf
->mddev
);
1097 if (s
->syncing
|| s
->expanding
|| s
->expanded
1099 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
1101 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
1103 bio_set_dev(bi
, rdev
->bdev
);
1104 bio_set_op_attrs(bi
, op
, op_flags
);
1105 bi
->bi_end_io
= op_is_write(op
)
1106 ? raid5_end_write_request
1107 : raid5_end_read_request
;
1108 bi
->bi_private
= sh
;
1110 pr_debug("%s: for %llu schedule op %d on disc %d\n",
1111 __func__
, (unsigned long long)sh
->sector
,
1113 atomic_inc(&sh
->count
);
1115 atomic_inc(&head_sh
->count
);
1116 if (use_new_offset(conf
, sh
))
1117 bi
->bi_iter
.bi_sector
= (sh
->sector
1118 + rdev
->new_data_offset
);
1120 bi
->bi_iter
.bi_sector
= (sh
->sector
1121 + rdev
->data_offset
);
1122 if (test_bit(R5_ReadNoMerge
, &head_sh
->dev
[i
].flags
))
1123 bi
->bi_opf
|= REQ_NOMERGE
;
1125 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1126 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1128 if (!op_is_write(op
) &&
1129 test_bit(R5_InJournal
, &sh
->dev
[i
].flags
))
1131 * issuing read for a page in journal, this
1132 * must be preparing for prexor in rmw; read
1133 * the data into orig_page
1135 sh
->dev
[i
].vec
.bv_page
= sh
->dev
[i
].orig_page
;
1137 sh
->dev
[i
].vec
.bv_page
= sh
->dev
[i
].page
;
1139 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
1140 bi
->bi_io_vec
[0].bv_offset
= 0;
1141 bi
->bi_iter
.bi_size
= STRIPE_SIZE
;
1142 bi
->bi_write_hint
= sh
->dev
[i
].write_hint
;
1144 sh
->dev
[i
].write_hint
= RWF_WRITE_LIFE_NOT_SET
;
1146 * If this is discard request, set bi_vcnt 0. We don't
1147 * want to confuse SCSI because SCSI will replace payload
1149 if (op
== REQ_OP_DISCARD
)
1152 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
1154 if (conf
->mddev
->gendisk
)
1155 trace_block_bio_remap(bi
->bi_disk
->queue
,
1156 bi
, disk_devt(conf
->mddev
->gendisk
),
1158 if (should_defer
&& op_is_write(op
))
1159 bio_list_add(&pending_bios
, bi
);
1161 generic_make_request(bi
);
1164 if (s
->syncing
|| s
->expanding
|| s
->expanded
1166 md_sync_acct(rrdev
->bdev
, STRIPE_SECTORS
);
1168 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
1170 bio_set_dev(rbi
, rrdev
->bdev
);
1171 bio_set_op_attrs(rbi
, op
, op_flags
);
1172 BUG_ON(!op_is_write(op
));
1173 rbi
->bi_end_io
= raid5_end_write_request
;
1174 rbi
->bi_private
= sh
;
1176 pr_debug("%s: for %llu schedule op %d on "
1177 "replacement disc %d\n",
1178 __func__
, (unsigned long long)sh
->sector
,
1180 atomic_inc(&sh
->count
);
1182 atomic_inc(&head_sh
->count
);
1183 if (use_new_offset(conf
, sh
))
1184 rbi
->bi_iter
.bi_sector
= (sh
->sector
1185 + rrdev
->new_data_offset
);
1187 rbi
->bi_iter
.bi_sector
= (sh
->sector
1188 + rrdev
->data_offset
);
1189 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1190 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1191 sh
->dev
[i
].rvec
.bv_page
= sh
->dev
[i
].page
;
1193 rbi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
1194 rbi
->bi_io_vec
[0].bv_offset
= 0;
1195 rbi
->bi_iter
.bi_size
= STRIPE_SIZE
;
1196 rbi
->bi_write_hint
= sh
->dev
[i
].write_hint
;
1197 sh
->dev
[i
].write_hint
= RWF_WRITE_LIFE_NOT_SET
;
1199 * If this is discard request, set bi_vcnt 0. We don't
1200 * want to confuse SCSI because SCSI will replace payload
1202 if (op
== REQ_OP_DISCARD
)
1204 if (conf
->mddev
->gendisk
)
1205 trace_block_bio_remap(rbi
->bi_disk
->queue
,
1206 rbi
, disk_devt(conf
->mddev
->gendisk
),
1208 if (should_defer
&& op_is_write(op
))
1209 bio_list_add(&pending_bios
, rbi
);
1211 generic_make_request(rbi
);
1213 if (!rdev
&& !rrdev
) {
1214 if (op_is_write(op
))
1215 set_bit(STRIPE_DEGRADED
, &sh
->state
);
1216 pr_debug("skip op %d on disc %d for sector %llu\n",
1217 bi
->bi_opf
, i
, (unsigned long long)sh
->sector
);
1218 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1219 set_bit(STRIPE_HANDLE
, &sh
->state
);
1222 if (!head_sh
->batch_head
)
1224 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1230 if (should_defer
&& !bio_list_empty(&pending_bios
))
1231 defer_issue_bios(conf
, head_sh
->sector
, &pending_bios
);
1234 static struct dma_async_tx_descriptor
*
1235 async_copy_data(int frombio
, struct bio
*bio
, struct page
**page
,
1236 sector_t sector
, struct dma_async_tx_descriptor
*tx
,
1237 struct stripe_head
*sh
, int no_skipcopy
)
1240 struct bvec_iter iter
;
1241 struct page
*bio_page
;
1243 struct async_submit_ctl submit
;
1244 enum async_tx_flags flags
= 0;
1246 if (bio
->bi_iter
.bi_sector
>= sector
)
1247 page_offset
= (signed)(bio
->bi_iter
.bi_sector
- sector
) * 512;
1249 page_offset
= (signed)(sector
- bio
->bi_iter
.bi_sector
) * -512;
1252 flags
|= ASYNC_TX_FENCE
;
1253 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
1255 bio_for_each_segment(bvl
, bio
, iter
) {
1256 int len
= bvl
.bv_len
;
1260 if (page_offset
< 0) {
1261 b_offset
= -page_offset
;
1262 page_offset
+= b_offset
;
1266 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
1267 clen
= STRIPE_SIZE
- page_offset
;
1272 b_offset
+= bvl
.bv_offset
;
1273 bio_page
= bvl
.bv_page
;
1275 if (sh
->raid_conf
->skip_copy
&&
1276 b_offset
== 0 && page_offset
== 0 &&
1277 clen
== STRIPE_SIZE
&&
1281 tx
= async_memcpy(*page
, bio_page
, page_offset
,
1282 b_offset
, clen
, &submit
);
1284 tx
= async_memcpy(bio_page
, *page
, b_offset
,
1285 page_offset
, clen
, &submit
);
1287 /* chain the operations */
1288 submit
.depend_tx
= tx
;
1290 if (clen
< len
) /* hit end of page */
1298 static void ops_complete_biofill(void *stripe_head_ref
)
1300 struct stripe_head
*sh
= stripe_head_ref
;
1303 pr_debug("%s: stripe %llu\n", __func__
,
1304 (unsigned long long)sh
->sector
);
1306 /* clear completed biofills */
1307 for (i
= sh
->disks
; i
--; ) {
1308 struct r5dev
*dev
= &sh
->dev
[i
];
1310 /* acknowledge completion of a biofill operation */
1311 /* and check if we need to reply to a read request,
1312 * new R5_Wantfill requests are held off until
1313 * !STRIPE_BIOFILL_RUN
1315 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
1316 struct bio
*rbi
, *rbi2
;
1321 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1322 dev
->sector
+ STRIPE_SECTORS
) {
1323 rbi2
= r5_next_bio(rbi
, dev
->sector
);
1329 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
1331 set_bit(STRIPE_HANDLE
, &sh
->state
);
1332 raid5_release_stripe(sh
);
1335 static void ops_run_biofill(struct stripe_head
*sh
)
1337 struct dma_async_tx_descriptor
*tx
= NULL
;
1338 struct async_submit_ctl submit
;
1341 BUG_ON(sh
->batch_head
);
1342 pr_debug("%s: stripe %llu\n", __func__
,
1343 (unsigned long long)sh
->sector
);
1345 for (i
= sh
->disks
; i
--; ) {
1346 struct r5dev
*dev
= &sh
->dev
[i
];
1347 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
1349 spin_lock_irq(&sh
->stripe_lock
);
1350 dev
->read
= rbi
= dev
->toread
;
1352 spin_unlock_irq(&sh
->stripe_lock
);
1353 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1354 dev
->sector
+ STRIPE_SECTORS
) {
1355 tx
= async_copy_data(0, rbi
, &dev
->page
,
1356 dev
->sector
, tx
, sh
, 0);
1357 rbi
= r5_next_bio(rbi
, dev
->sector
);
1362 atomic_inc(&sh
->count
);
1363 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
1364 async_trigger_callback(&submit
);
1367 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
1374 tgt
= &sh
->dev
[target
];
1375 set_bit(R5_UPTODATE
, &tgt
->flags
);
1376 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1377 clear_bit(R5_Wantcompute
, &tgt
->flags
);
1380 static void ops_complete_compute(void *stripe_head_ref
)
1382 struct stripe_head
*sh
= stripe_head_ref
;
1384 pr_debug("%s: stripe %llu\n", __func__
,
1385 (unsigned long long)sh
->sector
);
1387 /* mark the computed target(s) as uptodate */
1388 mark_target_uptodate(sh
, sh
->ops
.target
);
1389 mark_target_uptodate(sh
, sh
->ops
.target2
);
1391 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
1392 if (sh
->check_state
== check_state_compute_run
)
1393 sh
->check_state
= check_state_compute_result
;
1394 set_bit(STRIPE_HANDLE
, &sh
->state
);
1395 raid5_release_stripe(sh
);
1398 /* return a pointer to the address conversion region of the scribble buffer */
1399 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
1400 struct raid5_percpu
*percpu
, int i
)
1404 addr
= flex_array_get(percpu
->scribble
, i
);
1405 return addr
+ sizeof(struct page
*) * (sh
->disks
+ 2);
1408 /* return a pointer to the address conversion region of the scribble buffer */
1409 static struct page
**to_addr_page(struct raid5_percpu
*percpu
, int i
)
1413 addr
= flex_array_get(percpu
->scribble
, i
);
1417 static struct dma_async_tx_descriptor
*
1418 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1420 int disks
= sh
->disks
;
1421 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1422 int target
= sh
->ops
.target
;
1423 struct r5dev
*tgt
= &sh
->dev
[target
];
1424 struct page
*xor_dest
= tgt
->page
;
1426 struct dma_async_tx_descriptor
*tx
;
1427 struct async_submit_ctl submit
;
1430 BUG_ON(sh
->batch_head
);
1432 pr_debug("%s: stripe %llu block: %d\n",
1433 __func__
, (unsigned long long)sh
->sector
, target
);
1434 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1436 for (i
= disks
; i
--; )
1438 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1440 atomic_inc(&sh
->count
);
1442 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
1443 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
, 0));
1444 if (unlikely(count
== 1))
1445 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1447 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1452 /* set_syndrome_sources - populate source buffers for gen_syndrome
1453 * @srcs - (struct page *) array of size sh->disks
1454 * @sh - stripe_head to parse
1456 * Populates srcs in proper layout order for the stripe and returns the
1457 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1458 * destination buffer is recorded in srcs[count] and the Q destination
1459 * is recorded in srcs[count+1]].
1461 static int set_syndrome_sources(struct page
**srcs
,
1462 struct stripe_head
*sh
,
1465 int disks
= sh
->disks
;
1466 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
1467 int d0_idx
= raid6_d0(sh
);
1471 for (i
= 0; i
< disks
; i
++)
1477 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1478 struct r5dev
*dev
= &sh
->dev
[i
];
1480 if (i
== sh
->qd_idx
|| i
== sh
->pd_idx
||
1481 (srctype
== SYNDROME_SRC_ALL
) ||
1482 (srctype
== SYNDROME_SRC_WANT_DRAIN
&&
1483 (test_bit(R5_Wantdrain
, &dev
->flags
) ||
1484 test_bit(R5_InJournal
, &dev
->flags
))) ||
1485 (srctype
== SYNDROME_SRC_WRITTEN
&&
1487 test_bit(R5_InJournal
, &dev
->flags
)))) {
1488 if (test_bit(R5_InJournal
, &dev
->flags
))
1489 srcs
[slot
] = sh
->dev
[i
].orig_page
;
1491 srcs
[slot
] = sh
->dev
[i
].page
;
1493 i
= raid6_next_disk(i
, disks
);
1494 } while (i
!= d0_idx
);
1496 return syndrome_disks
;
1499 static struct dma_async_tx_descriptor
*
1500 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1502 int disks
= sh
->disks
;
1503 struct page
**blocks
= to_addr_page(percpu
, 0);
1505 int qd_idx
= sh
->qd_idx
;
1506 struct dma_async_tx_descriptor
*tx
;
1507 struct async_submit_ctl submit
;
1513 BUG_ON(sh
->batch_head
);
1514 if (sh
->ops
.target
< 0)
1515 target
= sh
->ops
.target2
;
1516 else if (sh
->ops
.target2
< 0)
1517 target
= sh
->ops
.target
;
1519 /* we should only have one valid target */
1522 pr_debug("%s: stripe %llu block: %d\n",
1523 __func__
, (unsigned long long)sh
->sector
, target
);
1525 tgt
= &sh
->dev
[target
];
1526 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1529 atomic_inc(&sh
->count
);
1531 if (target
== qd_idx
) {
1532 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1533 blocks
[count
] = NULL
; /* regenerating p is not necessary */
1534 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
1535 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1536 ops_complete_compute
, sh
,
1537 to_addr_conv(sh
, percpu
, 0));
1538 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1540 /* Compute any data- or p-drive using XOR */
1542 for (i
= disks
; i
-- ; ) {
1543 if (i
== target
|| i
== qd_idx
)
1545 blocks
[count
++] = sh
->dev
[i
].page
;
1548 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1549 NULL
, ops_complete_compute
, sh
,
1550 to_addr_conv(sh
, percpu
, 0));
1551 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
1557 static struct dma_async_tx_descriptor
*
1558 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1560 int i
, count
, disks
= sh
->disks
;
1561 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1562 int d0_idx
= raid6_d0(sh
);
1563 int faila
= -1, failb
= -1;
1564 int target
= sh
->ops
.target
;
1565 int target2
= sh
->ops
.target2
;
1566 struct r5dev
*tgt
= &sh
->dev
[target
];
1567 struct r5dev
*tgt2
= &sh
->dev
[target2
];
1568 struct dma_async_tx_descriptor
*tx
;
1569 struct page
**blocks
= to_addr_page(percpu
, 0);
1570 struct async_submit_ctl submit
;
1572 BUG_ON(sh
->batch_head
);
1573 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1574 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
1575 BUG_ON(target
< 0 || target2
< 0);
1576 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1577 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
1579 /* we need to open-code set_syndrome_sources to handle the
1580 * slot number conversion for 'faila' and 'failb'
1582 for (i
= 0; i
< disks
; i
++)
1587 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1589 blocks
[slot
] = sh
->dev
[i
].page
;
1595 i
= raid6_next_disk(i
, disks
);
1596 } while (i
!= d0_idx
);
1598 BUG_ON(faila
== failb
);
1601 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1602 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1604 atomic_inc(&sh
->count
);
1606 if (failb
== syndrome_disks
+1) {
1607 /* Q disk is one of the missing disks */
1608 if (faila
== syndrome_disks
) {
1609 /* Missing P+Q, just recompute */
1610 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1611 ops_complete_compute
, sh
,
1612 to_addr_conv(sh
, percpu
, 0));
1613 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
1614 STRIPE_SIZE
, &submit
);
1618 int qd_idx
= sh
->qd_idx
;
1620 /* Missing D+Q: recompute D from P, then recompute Q */
1621 if (target
== qd_idx
)
1622 data_target
= target2
;
1624 data_target
= target
;
1627 for (i
= disks
; i
-- ; ) {
1628 if (i
== data_target
|| i
== qd_idx
)
1630 blocks
[count
++] = sh
->dev
[i
].page
;
1632 dest
= sh
->dev
[data_target
].page
;
1633 init_async_submit(&submit
,
1634 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1636 to_addr_conv(sh
, percpu
, 0));
1637 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
1640 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1641 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1642 ops_complete_compute
, sh
,
1643 to_addr_conv(sh
, percpu
, 0));
1644 return async_gen_syndrome(blocks
, 0, count
+2,
1645 STRIPE_SIZE
, &submit
);
1648 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1649 ops_complete_compute
, sh
,
1650 to_addr_conv(sh
, percpu
, 0));
1651 if (failb
== syndrome_disks
) {
1652 /* We're missing D+P. */
1653 return async_raid6_datap_recov(syndrome_disks
+2,
1657 /* We're missing D+D. */
1658 return async_raid6_2data_recov(syndrome_disks
+2,
1659 STRIPE_SIZE
, faila
, failb
,
1665 static void ops_complete_prexor(void *stripe_head_ref
)
1667 struct stripe_head
*sh
= stripe_head_ref
;
1669 pr_debug("%s: stripe %llu\n", __func__
,
1670 (unsigned long long)sh
->sector
);
1672 if (r5c_is_writeback(sh
->raid_conf
->log
))
1674 * raid5-cache write back uses orig_page during prexor.
1675 * After prexor, it is time to free orig_page
1677 r5c_release_extra_page(sh
);
1680 static struct dma_async_tx_descriptor
*
1681 ops_run_prexor5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1682 struct dma_async_tx_descriptor
*tx
)
1684 int disks
= sh
->disks
;
1685 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1686 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1687 struct async_submit_ctl submit
;
1689 /* existing parity data subtracted */
1690 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1692 BUG_ON(sh
->batch_head
);
1693 pr_debug("%s: stripe %llu\n", __func__
,
1694 (unsigned long long)sh
->sector
);
1696 for (i
= disks
; i
--; ) {
1697 struct r5dev
*dev
= &sh
->dev
[i
];
1698 /* Only process blocks that are known to be uptodate */
1699 if (test_bit(R5_InJournal
, &dev
->flags
))
1700 xor_srcs
[count
++] = dev
->orig_page
;
1701 else if (test_bit(R5_Wantdrain
, &dev
->flags
))
1702 xor_srcs
[count
++] = dev
->page
;
1705 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1706 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1707 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1712 static struct dma_async_tx_descriptor
*
1713 ops_run_prexor6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1714 struct dma_async_tx_descriptor
*tx
)
1716 struct page
**blocks
= to_addr_page(percpu
, 0);
1718 struct async_submit_ctl submit
;
1720 pr_debug("%s: stripe %llu\n", __func__
,
1721 (unsigned long long)sh
->sector
);
1723 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_WANT_DRAIN
);
1725 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_PQ_XOR_DST
, tx
,
1726 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1727 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1732 static struct dma_async_tx_descriptor
*
1733 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1735 struct r5conf
*conf
= sh
->raid_conf
;
1736 int disks
= sh
->disks
;
1738 struct stripe_head
*head_sh
= sh
;
1740 pr_debug("%s: stripe %llu\n", __func__
,
1741 (unsigned long long)sh
->sector
);
1743 for (i
= disks
; i
--; ) {
1748 if (test_and_clear_bit(R5_Wantdrain
, &head_sh
->dev
[i
].flags
)) {
1754 * clear R5_InJournal, so when rewriting a page in
1755 * journal, it is not skipped by r5l_log_stripe()
1757 clear_bit(R5_InJournal
, &dev
->flags
);
1758 spin_lock_irq(&sh
->stripe_lock
);
1759 chosen
= dev
->towrite
;
1760 dev
->towrite
= NULL
;
1761 sh
->overwrite_disks
= 0;
1762 BUG_ON(dev
->written
);
1763 wbi
= dev
->written
= chosen
;
1764 spin_unlock_irq(&sh
->stripe_lock
);
1765 WARN_ON(dev
->page
!= dev
->orig_page
);
1767 while (wbi
&& wbi
->bi_iter
.bi_sector
<
1768 dev
->sector
+ STRIPE_SECTORS
) {
1769 if (wbi
->bi_opf
& REQ_FUA
)
1770 set_bit(R5_WantFUA
, &dev
->flags
);
1771 if (wbi
->bi_opf
& REQ_SYNC
)
1772 set_bit(R5_SyncIO
, &dev
->flags
);
1773 if (bio_op(wbi
) == REQ_OP_DISCARD
)
1774 set_bit(R5_Discard
, &dev
->flags
);
1776 tx
= async_copy_data(1, wbi
, &dev
->page
,
1777 dev
->sector
, tx
, sh
,
1778 r5c_is_writeback(conf
->log
));
1779 if (dev
->page
!= dev
->orig_page
&&
1780 !r5c_is_writeback(conf
->log
)) {
1781 set_bit(R5_SkipCopy
, &dev
->flags
);
1782 clear_bit(R5_UPTODATE
, &dev
->flags
);
1783 clear_bit(R5_OVERWRITE
, &dev
->flags
);
1786 wbi
= r5_next_bio(wbi
, dev
->sector
);
1789 if (head_sh
->batch_head
) {
1790 sh
= list_first_entry(&sh
->batch_list
,
1803 static void ops_complete_reconstruct(void *stripe_head_ref
)
1805 struct stripe_head
*sh
= stripe_head_ref
;
1806 int disks
= sh
->disks
;
1807 int pd_idx
= sh
->pd_idx
;
1808 int qd_idx
= sh
->qd_idx
;
1810 bool fua
= false, sync
= false, discard
= false;
1812 pr_debug("%s: stripe %llu\n", __func__
,
1813 (unsigned long long)sh
->sector
);
1815 for (i
= disks
; i
--; ) {
1816 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1817 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
1818 discard
|= test_bit(R5_Discard
, &sh
->dev
[i
].flags
);
1821 for (i
= disks
; i
--; ) {
1822 struct r5dev
*dev
= &sh
->dev
[i
];
1824 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1825 if (!discard
&& !test_bit(R5_SkipCopy
, &dev
->flags
)) {
1826 set_bit(R5_UPTODATE
, &dev
->flags
);
1827 if (test_bit(STRIPE_EXPAND_READY
, &sh
->state
))
1828 set_bit(R5_Expanded
, &dev
->flags
);
1831 set_bit(R5_WantFUA
, &dev
->flags
);
1833 set_bit(R5_SyncIO
, &dev
->flags
);
1837 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1838 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1839 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1840 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1842 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1843 sh
->reconstruct_state
= reconstruct_state_result
;
1846 set_bit(STRIPE_HANDLE
, &sh
->state
);
1847 raid5_release_stripe(sh
);
1851 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1852 struct dma_async_tx_descriptor
*tx
)
1854 int disks
= sh
->disks
;
1855 struct page
**xor_srcs
;
1856 struct async_submit_ctl submit
;
1857 int count
, pd_idx
= sh
->pd_idx
, i
;
1858 struct page
*xor_dest
;
1860 unsigned long flags
;
1862 struct stripe_head
*head_sh
= sh
;
1865 pr_debug("%s: stripe %llu\n", __func__
,
1866 (unsigned long long)sh
->sector
);
1868 for (i
= 0; i
< sh
->disks
; i
++) {
1871 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1874 if (i
>= sh
->disks
) {
1875 atomic_inc(&sh
->count
);
1876 set_bit(R5_Discard
, &sh
->dev
[pd_idx
].flags
);
1877 ops_complete_reconstruct(sh
);
1882 xor_srcs
= to_addr_page(percpu
, j
);
1883 /* check if prexor is active which means only process blocks
1884 * that are part of a read-modify-write (written)
1886 if (head_sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1888 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1889 for (i
= disks
; i
--; ) {
1890 struct r5dev
*dev
= &sh
->dev
[i
];
1891 if (head_sh
->dev
[i
].written
||
1892 test_bit(R5_InJournal
, &head_sh
->dev
[i
].flags
))
1893 xor_srcs
[count
++] = dev
->page
;
1896 xor_dest
= sh
->dev
[pd_idx
].page
;
1897 for (i
= disks
; i
--; ) {
1898 struct r5dev
*dev
= &sh
->dev
[i
];
1900 xor_srcs
[count
++] = dev
->page
;
1904 /* 1/ if we prexor'd then the dest is reused as a source
1905 * 2/ if we did not prexor then we are redoing the parity
1906 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1907 * for the synchronous xor case
1909 last_stripe
= !head_sh
->batch_head
||
1910 list_first_entry(&sh
->batch_list
,
1911 struct stripe_head
, batch_list
) == head_sh
;
1913 flags
= ASYNC_TX_ACK
|
1914 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1916 atomic_inc(&head_sh
->count
);
1917 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, head_sh
,
1918 to_addr_conv(sh
, percpu
, j
));
1920 flags
= prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
;
1921 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
,
1922 to_addr_conv(sh
, percpu
, j
));
1925 if (unlikely(count
== 1))
1926 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1928 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1931 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1938 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1939 struct dma_async_tx_descriptor
*tx
)
1941 struct async_submit_ctl submit
;
1942 struct page
**blocks
;
1943 int count
, i
, j
= 0;
1944 struct stripe_head
*head_sh
= sh
;
1947 unsigned long txflags
;
1949 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1951 for (i
= 0; i
< sh
->disks
; i
++) {
1952 if (sh
->pd_idx
== i
|| sh
->qd_idx
== i
)
1954 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1957 if (i
>= sh
->disks
) {
1958 atomic_inc(&sh
->count
);
1959 set_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
1960 set_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
1961 ops_complete_reconstruct(sh
);
1966 blocks
= to_addr_page(percpu
, j
);
1968 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1969 synflags
= SYNDROME_SRC_WRITTEN
;
1970 txflags
= ASYNC_TX_ACK
| ASYNC_TX_PQ_XOR_DST
;
1972 synflags
= SYNDROME_SRC_ALL
;
1973 txflags
= ASYNC_TX_ACK
;
1976 count
= set_syndrome_sources(blocks
, sh
, synflags
);
1977 last_stripe
= !head_sh
->batch_head
||
1978 list_first_entry(&sh
->batch_list
,
1979 struct stripe_head
, batch_list
) == head_sh
;
1982 atomic_inc(&head_sh
->count
);
1983 init_async_submit(&submit
, txflags
, tx
, ops_complete_reconstruct
,
1984 head_sh
, to_addr_conv(sh
, percpu
, j
));
1986 init_async_submit(&submit
, 0, tx
, NULL
, NULL
,
1987 to_addr_conv(sh
, percpu
, j
));
1988 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1991 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1997 static void ops_complete_check(void *stripe_head_ref
)
1999 struct stripe_head
*sh
= stripe_head_ref
;
2001 pr_debug("%s: stripe %llu\n", __func__
,
2002 (unsigned long long)sh
->sector
);
2004 sh
->check_state
= check_state_check_result
;
2005 set_bit(STRIPE_HANDLE
, &sh
->state
);
2006 raid5_release_stripe(sh
);
2009 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
2011 int disks
= sh
->disks
;
2012 int pd_idx
= sh
->pd_idx
;
2013 int qd_idx
= sh
->qd_idx
;
2014 struct page
*xor_dest
;
2015 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
2016 struct dma_async_tx_descriptor
*tx
;
2017 struct async_submit_ctl submit
;
2021 pr_debug("%s: stripe %llu\n", __func__
,
2022 (unsigned long long)sh
->sector
);
2024 BUG_ON(sh
->batch_head
);
2026 xor_dest
= sh
->dev
[pd_idx
].page
;
2027 xor_srcs
[count
++] = xor_dest
;
2028 for (i
= disks
; i
--; ) {
2029 if (i
== pd_idx
|| i
== qd_idx
)
2031 xor_srcs
[count
++] = sh
->dev
[i
].page
;
2034 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
2035 to_addr_conv(sh
, percpu
, 0));
2036 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
2037 &sh
->ops
.zero_sum_result
, &submit
);
2039 atomic_inc(&sh
->count
);
2040 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
2041 tx
= async_trigger_callback(&submit
);
2044 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
2046 struct page
**srcs
= to_addr_page(percpu
, 0);
2047 struct async_submit_ctl submit
;
2050 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
2051 (unsigned long long)sh
->sector
, checkp
);
2053 BUG_ON(sh
->batch_head
);
2054 count
= set_syndrome_sources(srcs
, sh
, SYNDROME_SRC_ALL
);
2058 atomic_inc(&sh
->count
);
2059 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
2060 sh
, to_addr_conv(sh
, percpu
, 0));
2061 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
2062 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
2065 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
2067 int overlap_clear
= 0, i
, disks
= sh
->disks
;
2068 struct dma_async_tx_descriptor
*tx
= NULL
;
2069 struct r5conf
*conf
= sh
->raid_conf
;
2070 int level
= conf
->level
;
2071 struct raid5_percpu
*percpu
;
2075 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
2076 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
2077 ops_run_biofill(sh
);
2081 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
2083 tx
= ops_run_compute5(sh
, percpu
);
2085 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
2086 tx
= ops_run_compute6_1(sh
, percpu
);
2088 tx
= ops_run_compute6_2(sh
, percpu
);
2090 /* terminate the chain if reconstruct is not set to be run */
2091 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
2095 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
)) {
2097 tx
= ops_run_prexor5(sh
, percpu
, tx
);
2099 tx
= ops_run_prexor6(sh
, percpu
, tx
);
2102 if (test_bit(STRIPE_OP_PARTIAL_PARITY
, &ops_request
))
2103 tx
= ops_run_partial_parity(sh
, percpu
, tx
);
2105 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
2106 tx
= ops_run_biodrain(sh
, tx
);
2110 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
2112 ops_run_reconstruct5(sh
, percpu
, tx
);
2114 ops_run_reconstruct6(sh
, percpu
, tx
);
2117 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
2118 if (sh
->check_state
== check_state_run
)
2119 ops_run_check_p(sh
, percpu
);
2120 else if (sh
->check_state
== check_state_run_q
)
2121 ops_run_check_pq(sh
, percpu
, 0);
2122 else if (sh
->check_state
== check_state_run_pq
)
2123 ops_run_check_pq(sh
, percpu
, 1);
2128 if (overlap_clear
&& !sh
->batch_head
)
2129 for (i
= disks
; i
--; ) {
2130 struct r5dev
*dev
= &sh
->dev
[i
];
2131 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
2132 wake_up(&sh
->raid_conf
->wait_for_overlap
);
2137 static void free_stripe(struct kmem_cache
*sc
, struct stripe_head
*sh
)
2140 __free_page(sh
->ppl_page
);
2141 kmem_cache_free(sc
, sh
);
2144 static struct stripe_head
*alloc_stripe(struct kmem_cache
*sc
, gfp_t gfp
,
2145 int disks
, struct r5conf
*conf
)
2147 struct stripe_head
*sh
;
2150 sh
= kmem_cache_zalloc(sc
, gfp
);
2152 spin_lock_init(&sh
->stripe_lock
);
2153 spin_lock_init(&sh
->batch_lock
);
2154 INIT_LIST_HEAD(&sh
->batch_list
);
2155 INIT_LIST_HEAD(&sh
->lru
);
2156 INIT_LIST_HEAD(&sh
->r5c
);
2157 INIT_LIST_HEAD(&sh
->log_list
);
2158 atomic_set(&sh
->count
, 1);
2159 sh
->raid_conf
= conf
;
2160 sh
->log_start
= MaxSector
;
2161 for (i
= 0; i
< disks
; i
++) {
2162 struct r5dev
*dev
= &sh
->dev
[i
];
2164 bio_init(&dev
->req
, &dev
->vec
, 1);
2165 bio_init(&dev
->rreq
, &dev
->rvec
, 1);
2168 if (raid5_has_ppl(conf
)) {
2169 sh
->ppl_page
= alloc_page(gfp
);
2170 if (!sh
->ppl_page
) {
2171 free_stripe(sc
, sh
);
2178 static int grow_one_stripe(struct r5conf
*conf
, gfp_t gfp
)
2180 struct stripe_head
*sh
;
2182 sh
= alloc_stripe(conf
->slab_cache
, gfp
, conf
->pool_size
, conf
);
2186 if (grow_buffers(sh
, gfp
)) {
2188 free_stripe(conf
->slab_cache
, sh
);
2191 sh
->hash_lock_index
=
2192 conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
;
2193 /* we just created an active stripe so... */
2194 atomic_inc(&conf
->active_stripes
);
2196 raid5_release_stripe(sh
);
2197 conf
->max_nr_stripes
++;
2201 static int grow_stripes(struct r5conf
*conf
, int num
)
2203 struct kmem_cache
*sc
;
2204 size_t namelen
= sizeof(conf
->cache_name
[0]);
2205 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
2207 if (conf
->mddev
->gendisk
)
2208 snprintf(conf
->cache_name
[0], namelen
,
2209 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
2211 snprintf(conf
->cache_name
[0], namelen
,
2212 "raid%d-%p", conf
->level
, conf
->mddev
);
2213 snprintf(conf
->cache_name
[1], namelen
, "%.27s-alt", conf
->cache_name
[0]);
2215 conf
->active_name
= 0;
2216 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
2217 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
2221 conf
->slab_cache
= sc
;
2222 conf
->pool_size
= devs
;
2224 if (!grow_one_stripe(conf
, GFP_KERNEL
))
2231 * scribble_len - return the required size of the scribble region
2232 * @num - total number of disks in the array
2234 * The size must be enough to contain:
2235 * 1/ a struct page pointer for each device in the array +2
2236 * 2/ room to convert each entry in (1) to its corresponding dma
2237 * (dma_map_page()) or page (page_address()) address.
2239 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2240 * calculate over all devices (not just the data blocks), using zeros in place
2241 * of the P and Q blocks.
2243 static struct flex_array
*scribble_alloc(int num
, int cnt
, gfp_t flags
)
2245 struct flex_array
*ret
;
2248 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
2249 ret
= flex_array_alloc(len
, cnt
, flags
);
2252 /* always prealloc all elements, so no locking is required */
2253 if (flex_array_prealloc(ret
, 0, cnt
, flags
)) {
2254 flex_array_free(ret
);
2260 static int resize_chunks(struct r5conf
*conf
, int new_disks
, int new_sectors
)
2266 * Never shrink. And mddev_suspend() could deadlock if this is called
2267 * from raid5d. In that case, scribble_disks and scribble_sectors
2268 * should equal to new_disks and new_sectors
2270 if (conf
->scribble_disks
>= new_disks
&&
2271 conf
->scribble_sectors
>= new_sectors
)
2273 mddev_suspend(conf
->mddev
);
2275 for_each_present_cpu(cpu
) {
2276 struct raid5_percpu
*percpu
;
2277 struct flex_array
*scribble
;
2279 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
2280 scribble
= scribble_alloc(new_disks
,
2281 new_sectors
/ STRIPE_SECTORS
,
2285 flex_array_free(percpu
->scribble
);
2286 percpu
->scribble
= scribble
;
2293 mddev_resume(conf
->mddev
);
2295 conf
->scribble_disks
= new_disks
;
2296 conf
->scribble_sectors
= new_sectors
;
2301 static int resize_stripes(struct r5conf
*conf
, int newsize
)
2303 /* Make all the stripes able to hold 'newsize' devices.
2304 * New slots in each stripe get 'page' set to a new page.
2306 * This happens in stages:
2307 * 1/ create a new kmem_cache and allocate the required number of
2309 * 2/ gather all the old stripe_heads and transfer the pages across
2310 * to the new stripe_heads. This will have the side effect of
2311 * freezing the array as once all stripe_heads have been collected,
2312 * no IO will be possible. Old stripe heads are freed once their
2313 * pages have been transferred over, and the old kmem_cache is
2314 * freed when all stripes are done.
2315 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2316 * we simple return a failure status - no need to clean anything up.
2317 * 4/ allocate new pages for the new slots in the new stripe_heads.
2318 * If this fails, we don't bother trying the shrink the
2319 * stripe_heads down again, we just leave them as they are.
2320 * As each stripe_head is processed the new one is released into
2323 * Once step2 is started, we cannot afford to wait for a write,
2324 * so we use GFP_NOIO allocations.
2326 struct stripe_head
*osh
, *nsh
;
2327 LIST_HEAD(newstripes
);
2328 struct disk_info
*ndisks
;
2330 struct kmem_cache
*sc
;
2334 md_allow_write(conf
->mddev
);
2337 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
2338 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
2343 /* Need to ensure auto-resizing doesn't interfere */
2344 mutex_lock(&conf
->cache_size_mutex
);
2346 for (i
= conf
->max_nr_stripes
; i
; i
--) {
2347 nsh
= alloc_stripe(sc
, GFP_KERNEL
, newsize
, conf
);
2351 list_add(&nsh
->lru
, &newstripes
);
2354 /* didn't get enough, give up */
2355 while (!list_empty(&newstripes
)) {
2356 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2357 list_del(&nsh
->lru
);
2358 free_stripe(sc
, nsh
);
2360 kmem_cache_destroy(sc
);
2361 mutex_unlock(&conf
->cache_size_mutex
);
2364 /* Step 2 - Must use GFP_NOIO now.
2365 * OK, we have enough stripes, start collecting inactive
2366 * stripes and copying them over
2370 list_for_each_entry(nsh
, &newstripes
, lru
) {
2371 lock_device_hash_lock(conf
, hash
);
2372 wait_event_cmd(conf
->wait_for_stripe
,
2373 !list_empty(conf
->inactive_list
+ hash
),
2374 unlock_device_hash_lock(conf
, hash
),
2375 lock_device_hash_lock(conf
, hash
));
2376 osh
= get_free_stripe(conf
, hash
);
2377 unlock_device_hash_lock(conf
, hash
);
2379 for(i
=0; i
<conf
->pool_size
; i
++) {
2380 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
2381 nsh
->dev
[i
].orig_page
= osh
->dev
[i
].page
;
2383 nsh
->hash_lock_index
= hash
;
2384 free_stripe(conf
->slab_cache
, osh
);
2386 if (cnt
>= conf
->max_nr_stripes
/ NR_STRIPE_HASH_LOCKS
+
2387 !!((conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
) > hash
)) {
2392 kmem_cache_destroy(conf
->slab_cache
);
2395 * At this point, we are holding all the stripes so the array
2396 * is completely stalled, so now is a good time to resize
2397 * conf->disks and the scribble region
2399 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
2401 for (i
= 0; i
< conf
->pool_size
; i
++)
2402 ndisks
[i
] = conf
->disks
[i
];
2404 for (i
= conf
->pool_size
; i
< newsize
; i
++) {
2405 ndisks
[i
].extra_page
= alloc_page(GFP_NOIO
);
2406 if (!ndisks
[i
].extra_page
)
2411 for (i
= conf
->pool_size
; i
< newsize
; i
++)
2412 if (ndisks
[i
].extra_page
)
2413 put_page(ndisks
[i
].extra_page
);
2417 conf
->disks
= ndisks
;
2422 mutex_unlock(&conf
->cache_size_mutex
);
2424 conf
->slab_cache
= sc
;
2425 conf
->active_name
= 1-conf
->active_name
;
2427 /* Step 4, return new stripes to service */
2428 while(!list_empty(&newstripes
)) {
2429 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2430 list_del_init(&nsh
->lru
);
2432 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
2433 if (nsh
->dev
[i
].page
== NULL
) {
2434 struct page
*p
= alloc_page(GFP_NOIO
);
2435 nsh
->dev
[i
].page
= p
;
2436 nsh
->dev
[i
].orig_page
= p
;
2440 raid5_release_stripe(nsh
);
2442 /* critical section pass, GFP_NOIO no longer needed */
2445 conf
->pool_size
= newsize
;
2449 static int drop_one_stripe(struct r5conf
*conf
)
2451 struct stripe_head
*sh
;
2452 int hash
= (conf
->max_nr_stripes
- 1) & STRIPE_HASH_LOCKS_MASK
;
2454 spin_lock_irq(conf
->hash_locks
+ hash
);
2455 sh
= get_free_stripe(conf
, hash
);
2456 spin_unlock_irq(conf
->hash_locks
+ hash
);
2459 BUG_ON(atomic_read(&sh
->count
));
2461 free_stripe(conf
->slab_cache
, sh
);
2462 atomic_dec(&conf
->active_stripes
);
2463 conf
->max_nr_stripes
--;
2467 static void shrink_stripes(struct r5conf
*conf
)
2469 while (conf
->max_nr_stripes
&&
2470 drop_one_stripe(conf
))
2473 kmem_cache_destroy(conf
->slab_cache
);
2474 conf
->slab_cache
= NULL
;
2477 static void raid5_end_read_request(struct bio
* bi
)
2479 struct stripe_head
*sh
= bi
->bi_private
;
2480 struct r5conf
*conf
= sh
->raid_conf
;
2481 int disks
= sh
->disks
, i
;
2482 char b
[BDEVNAME_SIZE
];
2483 struct md_rdev
*rdev
= NULL
;
2486 for (i
=0 ; i
<disks
; i
++)
2487 if (bi
== &sh
->dev
[i
].req
)
2490 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2491 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2498 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2499 /* If replacement finished while this request was outstanding,
2500 * 'replacement' might be NULL already.
2501 * In that case it moved down to 'rdev'.
2502 * rdev is not removed until all requests are finished.
2504 rdev
= conf
->disks
[i
].replacement
;
2506 rdev
= conf
->disks
[i
].rdev
;
2508 if (use_new_offset(conf
, sh
))
2509 s
= sh
->sector
+ rdev
->new_data_offset
;
2511 s
= sh
->sector
+ rdev
->data_offset
;
2512 if (!bi
->bi_status
) {
2513 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2514 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2515 /* Note that this cannot happen on a
2516 * replacement device. We just fail those on
2519 pr_info_ratelimited(
2520 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2521 mdname(conf
->mddev
), STRIPE_SECTORS
,
2522 (unsigned long long)s
,
2523 bdevname(rdev
->bdev
, b
));
2524 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
2525 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2526 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2527 } else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2528 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2530 if (test_bit(R5_InJournal
, &sh
->dev
[i
].flags
))
2532 * end read for a page in journal, this
2533 * must be preparing for prexor in rmw
2535 set_bit(R5_OrigPageUPTDODATE
, &sh
->dev
[i
].flags
);
2537 if (atomic_read(&rdev
->read_errors
))
2538 atomic_set(&rdev
->read_errors
, 0);
2540 const char *bdn
= bdevname(rdev
->bdev
, b
);
2544 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2545 atomic_inc(&rdev
->read_errors
);
2546 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2547 pr_warn_ratelimited(
2548 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2549 mdname(conf
->mddev
),
2550 (unsigned long long)s
,
2552 else if (conf
->mddev
->degraded
>= conf
->max_degraded
) {
2554 pr_warn_ratelimited(
2555 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2556 mdname(conf
->mddev
),
2557 (unsigned long long)s
,
2559 } else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
)) {
2562 pr_warn_ratelimited(
2563 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2564 mdname(conf
->mddev
),
2565 (unsigned long long)s
,
2567 } else if (atomic_read(&rdev
->read_errors
)
2568 > conf
->max_nr_stripes
)
2569 pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2570 mdname(conf
->mddev
), bdn
);
2573 if (set_bad
&& test_bit(In_sync
, &rdev
->flags
)
2574 && !test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2577 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
)) {
2578 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2579 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2581 set_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2583 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2584 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2586 && test_bit(In_sync
, &rdev
->flags
)
2587 && rdev_set_badblocks(
2588 rdev
, sh
->sector
, STRIPE_SECTORS
, 0)))
2589 md_error(conf
->mddev
, rdev
);
2592 rdev_dec_pending(rdev
, conf
->mddev
);
2594 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2595 set_bit(STRIPE_HANDLE
, &sh
->state
);
2596 raid5_release_stripe(sh
);
2599 static void raid5_end_write_request(struct bio
*bi
)
2601 struct stripe_head
*sh
= bi
->bi_private
;
2602 struct r5conf
*conf
= sh
->raid_conf
;
2603 int disks
= sh
->disks
, i
;
2604 struct md_rdev
*uninitialized_var(rdev
);
2607 int replacement
= 0;
2609 for (i
= 0 ; i
< disks
; i
++) {
2610 if (bi
== &sh
->dev
[i
].req
) {
2611 rdev
= conf
->disks
[i
].rdev
;
2614 if (bi
== &sh
->dev
[i
].rreq
) {
2615 rdev
= conf
->disks
[i
].replacement
;
2619 /* rdev was removed and 'replacement'
2620 * replaced it. rdev is not removed
2621 * until all requests are finished.
2623 rdev
= conf
->disks
[i
].rdev
;
2627 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2628 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2638 md_error(conf
->mddev
, rdev
);
2639 else if (is_badblock(rdev
, sh
->sector
,
2641 &first_bad
, &bad_sectors
))
2642 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
2644 if (bi
->bi_status
) {
2645 set_bit(STRIPE_DEGRADED
, &sh
->state
);
2646 set_bit(WriteErrorSeen
, &rdev
->flags
);
2647 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
2648 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2649 set_bit(MD_RECOVERY_NEEDED
,
2650 &rdev
->mddev
->recovery
);
2651 } else if (is_badblock(rdev
, sh
->sector
,
2653 &first_bad
, &bad_sectors
)) {
2654 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
2655 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))
2656 /* That was a successful write so make
2657 * sure it looks like we already did
2660 set_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2663 rdev_dec_pending(rdev
, conf
->mddev
);
2665 if (sh
->batch_head
&& bi
->bi_status
&& !replacement
)
2666 set_bit(STRIPE_BATCH_ERR
, &sh
->batch_head
->state
);
2669 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
2670 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2671 set_bit(STRIPE_HANDLE
, &sh
->state
);
2672 raid5_release_stripe(sh
);
2674 if (sh
->batch_head
&& sh
!= sh
->batch_head
)
2675 raid5_release_stripe(sh
->batch_head
);
2678 static void raid5_error(struct mddev
*mddev
, struct md_rdev
*rdev
)
2680 char b
[BDEVNAME_SIZE
];
2681 struct r5conf
*conf
= mddev
->private;
2682 unsigned long flags
;
2683 pr_debug("raid456: error called\n");
2685 spin_lock_irqsave(&conf
->device_lock
, flags
);
2686 set_bit(Faulty
, &rdev
->flags
);
2687 clear_bit(In_sync
, &rdev
->flags
);
2688 mddev
->degraded
= raid5_calc_degraded(conf
);
2689 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2690 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2692 set_bit(Blocked
, &rdev
->flags
);
2693 set_mask_bits(&mddev
->sb_flags
, 0,
2694 BIT(MD_SB_CHANGE_DEVS
) | BIT(MD_SB_CHANGE_PENDING
));
2695 pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2696 "md/raid:%s: Operation continuing on %d devices.\n",
2698 bdevname(rdev
->bdev
, b
),
2700 conf
->raid_disks
- mddev
->degraded
);
2701 r5c_update_on_rdev_error(mddev
, rdev
);
2705 * Input: a 'big' sector number,
2706 * Output: index of the data and parity disk, and the sector # in them.
2708 sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
2709 int previous
, int *dd_idx
,
2710 struct stripe_head
*sh
)
2712 sector_t stripe
, stripe2
;
2713 sector_t chunk_number
;
2714 unsigned int chunk_offset
;
2717 sector_t new_sector
;
2718 int algorithm
= previous
? conf
->prev_algo
2720 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2721 : conf
->chunk_sectors
;
2722 int raid_disks
= previous
? conf
->previous_raid_disks
2724 int data_disks
= raid_disks
- conf
->max_degraded
;
2726 /* First compute the information on this sector */
2729 * Compute the chunk number and the sector offset inside the chunk
2731 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
2732 chunk_number
= r_sector
;
2735 * Compute the stripe number
2737 stripe
= chunk_number
;
2738 *dd_idx
= sector_div(stripe
, data_disks
);
2741 * Select the parity disk based on the user selected algorithm.
2743 pd_idx
= qd_idx
= -1;
2744 switch(conf
->level
) {
2746 pd_idx
= data_disks
;
2749 switch (algorithm
) {
2750 case ALGORITHM_LEFT_ASYMMETRIC
:
2751 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2752 if (*dd_idx
>= pd_idx
)
2755 case ALGORITHM_RIGHT_ASYMMETRIC
:
2756 pd_idx
= sector_div(stripe2
, raid_disks
);
2757 if (*dd_idx
>= pd_idx
)
2760 case ALGORITHM_LEFT_SYMMETRIC
:
2761 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2762 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2764 case ALGORITHM_RIGHT_SYMMETRIC
:
2765 pd_idx
= sector_div(stripe2
, raid_disks
);
2766 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2768 case ALGORITHM_PARITY_0
:
2772 case ALGORITHM_PARITY_N
:
2773 pd_idx
= data_disks
;
2781 switch (algorithm
) {
2782 case ALGORITHM_LEFT_ASYMMETRIC
:
2783 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2784 qd_idx
= pd_idx
+ 1;
2785 if (pd_idx
== raid_disks
-1) {
2786 (*dd_idx
)++; /* Q D D D P */
2788 } else if (*dd_idx
>= pd_idx
)
2789 (*dd_idx
) += 2; /* D D P Q D */
2791 case ALGORITHM_RIGHT_ASYMMETRIC
:
2792 pd_idx
= sector_div(stripe2
, raid_disks
);
2793 qd_idx
= pd_idx
+ 1;
2794 if (pd_idx
== raid_disks
-1) {
2795 (*dd_idx
)++; /* Q D D D P */
2797 } else if (*dd_idx
>= pd_idx
)
2798 (*dd_idx
) += 2; /* D D P Q D */
2800 case ALGORITHM_LEFT_SYMMETRIC
:
2801 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2802 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2803 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2805 case ALGORITHM_RIGHT_SYMMETRIC
:
2806 pd_idx
= sector_div(stripe2
, raid_disks
);
2807 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2808 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2811 case ALGORITHM_PARITY_0
:
2816 case ALGORITHM_PARITY_N
:
2817 pd_idx
= data_disks
;
2818 qd_idx
= data_disks
+ 1;
2821 case ALGORITHM_ROTATING_ZERO_RESTART
:
2822 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2823 * of blocks for computing Q is different.
2825 pd_idx
= sector_div(stripe2
, raid_disks
);
2826 qd_idx
= pd_idx
+ 1;
2827 if (pd_idx
== raid_disks
-1) {
2828 (*dd_idx
)++; /* Q D D D P */
2830 } else if (*dd_idx
>= pd_idx
)
2831 (*dd_idx
) += 2; /* D D P Q D */
2835 case ALGORITHM_ROTATING_N_RESTART
:
2836 /* Same a left_asymmetric, by first stripe is
2837 * D D D P Q rather than
2841 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2842 qd_idx
= pd_idx
+ 1;
2843 if (pd_idx
== raid_disks
-1) {
2844 (*dd_idx
)++; /* Q D D D P */
2846 } else if (*dd_idx
>= pd_idx
)
2847 (*dd_idx
) += 2; /* D D P Q D */
2851 case ALGORITHM_ROTATING_N_CONTINUE
:
2852 /* Same as left_symmetric but Q is before P */
2853 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2854 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
2855 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2859 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2860 /* RAID5 left_asymmetric, with Q on last device */
2861 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2862 if (*dd_idx
>= pd_idx
)
2864 qd_idx
= raid_disks
- 1;
2867 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2868 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2869 if (*dd_idx
>= pd_idx
)
2871 qd_idx
= raid_disks
- 1;
2874 case ALGORITHM_LEFT_SYMMETRIC_6
:
2875 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2876 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2877 qd_idx
= raid_disks
- 1;
2880 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2881 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2882 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2883 qd_idx
= raid_disks
- 1;
2886 case ALGORITHM_PARITY_0_6
:
2889 qd_idx
= raid_disks
- 1;
2899 sh
->pd_idx
= pd_idx
;
2900 sh
->qd_idx
= qd_idx
;
2901 sh
->ddf_layout
= ddf_layout
;
2904 * Finally, compute the new sector number
2906 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
2910 sector_t
raid5_compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
2912 struct r5conf
*conf
= sh
->raid_conf
;
2913 int raid_disks
= sh
->disks
;
2914 int data_disks
= raid_disks
- conf
->max_degraded
;
2915 sector_t new_sector
= sh
->sector
, check
;
2916 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2917 : conf
->chunk_sectors
;
2918 int algorithm
= previous
? conf
->prev_algo
2922 sector_t chunk_number
;
2923 int dummy1
, dd_idx
= i
;
2925 struct stripe_head sh2
;
2927 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
2928 stripe
= new_sector
;
2930 if (i
== sh
->pd_idx
)
2932 switch(conf
->level
) {
2935 switch (algorithm
) {
2936 case ALGORITHM_LEFT_ASYMMETRIC
:
2937 case ALGORITHM_RIGHT_ASYMMETRIC
:
2941 case ALGORITHM_LEFT_SYMMETRIC
:
2942 case ALGORITHM_RIGHT_SYMMETRIC
:
2945 i
-= (sh
->pd_idx
+ 1);
2947 case ALGORITHM_PARITY_0
:
2950 case ALGORITHM_PARITY_N
:
2957 if (i
== sh
->qd_idx
)
2958 return 0; /* It is the Q disk */
2959 switch (algorithm
) {
2960 case ALGORITHM_LEFT_ASYMMETRIC
:
2961 case ALGORITHM_RIGHT_ASYMMETRIC
:
2962 case ALGORITHM_ROTATING_ZERO_RESTART
:
2963 case ALGORITHM_ROTATING_N_RESTART
:
2964 if (sh
->pd_idx
== raid_disks
-1)
2965 i
--; /* Q D D D P */
2966 else if (i
> sh
->pd_idx
)
2967 i
-= 2; /* D D P Q D */
2969 case ALGORITHM_LEFT_SYMMETRIC
:
2970 case ALGORITHM_RIGHT_SYMMETRIC
:
2971 if (sh
->pd_idx
== raid_disks
-1)
2972 i
--; /* Q D D D P */
2977 i
-= (sh
->pd_idx
+ 2);
2980 case ALGORITHM_PARITY_0
:
2983 case ALGORITHM_PARITY_N
:
2985 case ALGORITHM_ROTATING_N_CONTINUE
:
2986 /* Like left_symmetric, but P is before Q */
2987 if (sh
->pd_idx
== 0)
2988 i
--; /* P D D D Q */
2993 i
-= (sh
->pd_idx
+ 1);
2996 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2997 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
3001 case ALGORITHM_LEFT_SYMMETRIC_6
:
3002 case ALGORITHM_RIGHT_SYMMETRIC_6
:
3004 i
+= data_disks
+ 1;
3005 i
-= (sh
->pd_idx
+ 1);
3007 case ALGORITHM_PARITY_0_6
:
3016 chunk_number
= stripe
* data_disks
+ i
;
3017 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
3019 check
= raid5_compute_sector(conf
, r_sector
,
3020 previous
, &dummy1
, &sh2
);
3021 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
3022 || sh2
.qd_idx
!= sh
->qd_idx
) {
3023 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3024 mdname(conf
->mddev
));
3031 * There are cases where we want handle_stripe_dirtying() and
3032 * schedule_reconstruction() to delay towrite to some dev of a stripe.
3034 * This function checks whether we want to delay the towrite. Specifically,
3035 * we delay the towrite when:
3037 * 1. degraded stripe has a non-overwrite to the missing dev, AND this
3038 * stripe has data in journal (for other devices).
3040 * In this case, when reading data for the non-overwrite dev, it is
3041 * necessary to handle complex rmw of write back cache (prexor with
3042 * orig_page, and xor with page). To keep read path simple, we would
3043 * like to flush data in journal to RAID disks first, so complex rmw
3044 * is handled in the write patch (handle_stripe_dirtying).
3046 * 2. when journal space is critical (R5C_LOG_CRITICAL=1)
3048 * It is important to be able to flush all stripes in raid5-cache.
3049 * Therefore, we need reserve some space on the journal device for
3050 * these flushes. If flush operation includes pending writes to the
3051 * stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3052 * for the flush out. If we exclude these pending writes from flush
3053 * operation, we only need (conf->max_degraded + 1) pages per stripe.
3054 * Therefore, excluding pending writes in these cases enables more
3055 * efficient use of the journal device.
3057 * Note: To make sure the stripe makes progress, we only delay
3058 * towrite for stripes with data already in journal (injournal > 0).
3059 * When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3060 * no_space_stripes list.
3062 * 3. during journal failure
3063 * In journal failure, we try to flush all cached data to raid disks
3064 * based on data in stripe cache. The array is read-only to upper
3065 * layers, so we would skip all pending writes.
3068 static inline bool delay_towrite(struct r5conf
*conf
,
3070 struct stripe_head_state
*s
)
3073 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3074 !test_bit(R5_Insync
, &dev
->flags
) && s
->injournal
)
3077 if (test_bit(R5C_LOG_CRITICAL
, &conf
->cache_state
) &&
3081 if (s
->log_failed
&& s
->injournal
)
3087 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3088 int rcw
, int expand
)
3090 int i
, pd_idx
= sh
->pd_idx
, qd_idx
= sh
->qd_idx
, disks
= sh
->disks
;
3091 struct r5conf
*conf
= sh
->raid_conf
;
3092 int level
= conf
->level
;
3096 * In some cases, handle_stripe_dirtying initially decided to
3097 * run rmw and allocates extra page for prexor. However, rcw is
3098 * cheaper later on. We need to free the extra page now,
3099 * because we won't be able to do that in ops_complete_prexor().
3101 r5c_release_extra_page(sh
);
3103 for (i
= disks
; i
--; ) {
3104 struct r5dev
*dev
= &sh
->dev
[i
];
3106 if (dev
->towrite
&& !delay_towrite(conf
, dev
, s
)) {
3107 set_bit(R5_LOCKED
, &dev
->flags
);
3108 set_bit(R5_Wantdrain
, &dev
->flags
);
3110 clear_bit(R5_UPTODATE
, &dev
->flags
);
3112 } else if (test_bit(R5_InJournal
, &dev
->flags
)) {
3113 set_bit(R5_LOCKED
, &dev
->flags
);
3117 /* if we are not expanding this is a proper write request, and
3118 * there will be bios with new data to be drained into the
3123 /* False alarm, nothing to do */
3125 sh
->reconstruct_state
= reconstruct_state_drain_run
;
3126 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
3128 sh
->reconstruct_state
= reconstruct_state_run
;
3130 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
3132 if (s
->locked
+ conf
->max_degraded
== disks
)
3133 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3134 atomic_inc(&conf
->pending_full_writes
);
3136 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
3137 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
3138 BUG_ON(level
== 6 &&
3139 (!(test_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
) ||
3140 test_bit(R5_Wantcompute
, &sh
->dev
[qd_idx
].flags
))));
3142 for (i
= disks
; i
--; ) {
3143 struct r5dev
*dev
= &sh
->dev
[i
];
3144 if (i
== pd_idx
|| i
== qd_idx
)
3148 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
3149 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3150 set_bit(R5_Wantdrain
, &dev
->flags
);
3151 set_bit(R5_LOCKED
, &dev
->flags
);
3152 clear_bit(R5_UPTODATE
, &dev
->flags
);
3154 } else if (test_bit(R5_InJournal
, &dev
->flags
)) {
3155 set_bit(R5_LOCKED
, &dev
->flags
);
3160 /* False alarm - nothing to do */
3162 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
3163 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
3164 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
3165 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
3168 /* keep the parity disk(s) locked while asynchronous operations
3171 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
3172 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
3176 int qd_idx
= sh
->qd_idx
;
3177 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
3179 set_bit(R5_LOCKED
, &dev
->flags
);
3180 clear_bit(R5_UPTODATE
, &dev
->flags
);
3184 if (raid5_has_ppl(sh
->raid_conf
) && sh
->ppl_page
&&
3185 test_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
) &&
3186 !test_bit(STRIPE_FULL_WRITE
, &sh
->state
) &&
3187 test_bit(R5_Insync
, &sh
->dev
[pd_idx
].flags
))
3188 set_bit(STRIPE_OP_PARTIAL_PARITY
, &s
->ops_request
);
3190 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3191 __func__
, (unsigned long long)sh
->sector
,
3192 s
->locked
, s
->ops_request
);
3196 * Each stripe/dev can have one or more bion attached.
3197 * toread/towrite point to the first in a chain.
3198 * The bi_next chain must be in order.
3200 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
,
3201 int forwrite
, int previous
)
3204 struct r5conf
*conf
= sh
->raid_conf
;
3207 pr_debug("adding bi b#%llu to stripe s#%llu\n",
3208 (unsigned long long)bi
->bi_iter
.bi_sector
,
3209 (unsigned long long)sh
->sector
);
3211 spin_lock_irq(&sh
->stripe_lock
);
3212 sh
->dev
[dd_idx
].write_hint
= bi
->bi_write_hint
;
3213 /* Don't allow new IO added to stripes in batch list */
3217 bip
= &sh
->dev
[dd_idx
].towrite
;
3221 bip
= &sh
->dev
[dd_idx
].toread
;
3222 while (*bip
&& (*bip
)->bi_iter
.bi_sector
< bi
->bi_iter
.bi_sector
) {
3223 if (bio_end_sector(*bip
) > bi
->bi_iter
.bi_sector
)
3225 bip
= & (*bip
)->bi_next
;
3227 if (*bip
&& (*bip
)->bi_iter
.bi_sector
< bio_end_sector(bi
))
3230 if (forwrite
&& raid5_has_ppl(conf
)) {
3232 * With PPL only writes to consecutive data chunks within a
3233 * stripe are allowed because for a single stripe_head we can
3234 * only have one PPL entry at a time, which describes one data
3235 * range. Not really an overlap, but wait_for_overlap can be
3236 * used to handle this.
3244 for (i
= 0; i
< sh
->disks
; i
++) {
3245 if (i
!= sh
->pd_idx
&&
3246 (i
== dd_idx
|| sh
->dev
[i
].towrite
)) {
3247 sector
= sh
->dev
[i
].sector
;
3248 if (count
== 0 || sector
< first
)
3256 if (first
+ conf
->chunk_sectors
* (count
- 1) != last
)
3260 if (!forwrite
|| previous
)
3261 clear_bit(STRIPE_BATCH_READY
, &sh
->state
);
3263 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
3267 bio_inc_remaining(bi
);
3268 md_write_inc(conf
->mddev
, bi
);
3271 /* check if page is covered */
3272 sector_t sector
= sh
->dev
[dd_idx
].sector
;
3273 for (bi
=sh
->dev
[dd_idx
].towrite
;
3274 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
3275 bi
&& bi
->bi_iter
.bi_sector
<= sector
;
3276 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
3277 if (bio_end_sector(bi
) >= sector
)
3278 sector
= bio_end_sector(bi
);
3280 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
3281 if (!test_and_set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
))
3282 sh
->overwrite_disks
++;
3285 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3286 (unsigned long long)(*bip
)->bi_iter
.bi_sector
,
3287 (unsigned long long)sh
->sector
, dd_idx
);
3289 if (conf
->mddev
->bitmap
&& firstwrite
) {
3290 /* Cannot hold spinlock over bitmap_startwrite,
3291 * but must ensure this isn't added to a batch until
3292 * we have added to the bitmap and set bm_seq.
3293 * So set STRIPE_BITMAP_PENDING to prevent
3295 * If multiple add_stripe_bio() calls race here they
3296 * much all set STRIPE_BITMAP_PENDING. So only the first one
3297 * to complete "bitmap_startwrite" gets to set
3298 * STRIPE_BIT_DELAY. This is important as once a stripe
3299 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3302 set_bit(STRIPE_BITMAP_PENDING
, &sh
->state
);
3303 spin_unlock_irq(&sh
->stripe_lock
);
3304 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
3306 spin_lock_irq(&sh
->stripe_lock
);
3307 clear_bit(STRIPE_BITMAP_PENDING
, &sh
->state
);
3308 if (!sh
->batch_head
) {
3309 sh
->bm_seq
= conf
->seq_flush
+1;
3310 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
3313 spin_unlock_irq(&sh
->stripe_lock
);
3315 if (stripe_can_batch(sh
))
3316 stripe_add_to_batch_list(conf
, sh
);
3320 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
3321 spin_unlock_irq(&sh
->stripe_lock
);
3325 static void end_reshape(struct r5conf
*conf
);
3327 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
3328 struct stripe_head
*sh
)
3330 int sectors_per_chunk
=
3331 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
3333 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
3334 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
3336 raid5_compute_sector(conf
,
3337 stripe
* (disks
- conf
->max_degraded
)
3338 *sectors_per_chunk
+ chunk_offset
,
3344 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
3345 struct stripe_head_state
*s
, int disks
)
3348 BUG_ON(sh
->batch_head
);
3349 for (i
= disks
; i
--; ) {
3353 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
3354 struct md_rdev
*rdev
;
3356 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3357 if (rdev
&& test_bit(In_sync
, &rdev
->flags
) &&
3358 !test_bit(Faulty
, &rdev
->flags
))
3359 atomic_inc(&rdev
->nr_pending
);
3364 if (!rdev_set_badblocks(
3368 md_error(conf
->mddev
, rdev
);
3369 rdev_dec_pending(rdev
, conf
->mddev
);
3372 spin_lock_irq(&sh
->stripe_lock
);
3373 /* fail all writes first */
3374 bi
= sh
->dev
[i
].towrite
;
3375 sh
->dev
[i
].towrite
= NULL
;
3376 sh
->overwrite_disks
= 0;
3377 spin_unlock_irq(&sh
->stripe_lock
);
3381 log_stripe_write_finished(sh
);
3383 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3384 wake_up(&conf
->wait_for_overlap
);
3386 while (bi
&& bi
->bi_iter
.bi_sector
<
3387 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3388 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3390 md_write_end(conf
->mddev
);
3395 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3396 STRIPE_SECTORS
, 0, 0);
3398 /* and fail all 'written' */
3399 bi
= sh
->dev
[i
].written
;
3400 sh
->dev
[i
].written
= NULL
;
3401 if (test_and_clear_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
)) {
3402 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
3403 sh
->dev
[i
].page
= sh
->dev
[i
].orig_page
;
3406 if (bi
) bitmap_end
= 1;
3407 while (bi
&& bi
->bi_iter
.bi_sector
<
3408 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3409 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3411 md_write_end(conf
->mddev
);
3416 /* fail any reads if this device is non-operational and
3417 * the data has not reached the cache yet.
3419 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
3420 s
->failed
> conf
->max_degraded
&&
3421 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
3422 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
3423 spin_lock_irq(&sh
->stripe_lock
);
3424 bi
= sh
->dev
[i
].toread
;
3425 sh
->dev
[i
].toread
= NULL
;
3426 spin_unlock_irq(&sh
->stripe_lock
);
3427 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3428 wake_up(&conf
->wait_for_overlap
);
3431 while (bi
&& bi
->bi_iter
.bi_sector
<
3432 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3433 struct bio
*nextbi
=
3434 r5_next_bio(bi
, sh
->dev
[i
].sector
);
3441 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3442 STRIPE_SECTORS
, 0, 0);
3443 /* If we were in the middle of a write the parity block might
3444 * still be locked - so just clear all R5_LOCKED flags
3446 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3451 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3452 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3453 md_wakeup_thread(conf
->mddev
->thread
);
3457 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
3458 struct stripe_head_state
*s
)
3463 BUG_ON(sh
->batch_head
);
3464 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3465 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
3466 wake_up(&conf
->wait_for_overlap
);
3469 /* There is nothing more to do for sync/check/repair.
3470 * Don't even need to abort as that is handled elsewhere
3471 * if needed, and not always wanted e.g. if there is a known
3473 * For recover/replace we need to record a bad block on all
3474 * non-sync devices, or abort the recovery
3476 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
3477 /* During recovery devices cannot be removed, so
3478 * locking and refcounting of rdevs is not needed
3481 for (i
= 0; i
< conf
->raid_disks
; i
++) {
3482 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3484 && !test_bit(Faulty
, &rdev
->flags
)
3485 && !test_bit(In_sync
, &rdev
->flags
)
3486 && !rdev_set_badblocks(rdev
, sh
->sector
,
3489 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
3491 && !test_bit(Faulty
, &rdev
->flags
)
3492 && !test_bit(In_sync
, &rdev
->flags
)
3493 && !rdev_set_badblocks(rdev
, sh
->sector
,
3499 conf
->recovery_disabled
=
3500 conf
->mddev
->recovery_disabled
;
3502 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, !abort
);
3505 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
3507 struct md_rdev
*rdev
;
3511 rdev
= rcu_dereference(sh
->raid_conf
->disks
[disk_idx
].replacement
);
3513 && !test_bit(Faulty
, &rdev
->flags
)
3514 && !test_bit(In_sync
, &rdev
->flags
)
3515 && (rdev
->recovery_offset
<= sh
->sector
3516 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
3522 static int need_this_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3523 int disk_idx
, int disks
)
3525 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3526 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
3527 &sh
->dev
[s
->failed_num
[1]] };
3531 if (test_bit(R5_LOCKED
, &dev
->flags
) ||
3532 test_bit(R5_UPTODATE
, &dev
->flags
))
3533 /* No point reading this as we already have it or have
3534 * decided to get it.
3539 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)))
3540 /* We need this block to directly satisfy a request */
3543 if (s
->syncing
|| s
->expanding
||
3544 (s
->replacing
&& want_replace(sh
, disk_idx
)))
3545 /* When syncing, or expanding we read everything.
3546 * When replacing, we need the replaced block.
3550 if ((s
->failed
>= 1 && fdev
[0]->toread
) ||
3551 (s
->failed
>= 2 && fdev
[1]->toread
))
3552 /* If we want to read from a failed device, then
3553 * we need to actually read every other device.
3557 /* Sometimes neither read-modify-write nor reconstruct-write
3558 * cycles can work. In those cases we read every block we
3559 * can. Then the parity-update is certain to have enough to
3561 * This can only be a problem when we need to write something,
3562 * and some device has failed. If either of those tests
3563 * fail we need look no further.
3565 if (!s
->failed
|| !s
->to_write
)
3568 if (test_bit(R5_Insync
, &dev
->flags
) &&
3569 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3570 /* Pre-reads at not permitted until after short delay
3571 * to gather multiple requests. However if this
3572 * device is no Insync, the block could only be computed
3573 * and there is no need to delay that.
3577 for (i
= 0; i
< s
->failed
&& i
< 2; i
++) {
3578 if (fdev
[i
]->towrite
&&
3579 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3580 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3581 /* If we have a partial write to a failed
3582 * device, then we will need to reconstruct
3583 * the content of that device, so all other
3584 * devices must be read.
3589 /* If we are forced to do a reconstruct-write, either because
3590 * the current RAID6 implementation only supports that, or
3591 * because parity cannot be trusted and we are currently
3592 * recovering it, there is extra need to be careful.
3593 * If one of the devices that we would need to read, because
3594 * it is not being overwritten (and maybe not written at all)
3595 * is missing/faulty, then we need to read everything we can.
3597 if (sh
->raid_conf
->level
!= 6 &&
3598 sh
->sector
< sh
->raid_conf
->mddev
->recovery_cp
)
3599 /* reconstruct-write isn't being forced */
3601 for (i
= 0; i
< s
->failed
&& i
< 2; i
++) {
3602 if (s
->failed_num
[i
] != sh
->pd_idx
&&
3603 s
->failed_num
[i
] != sh
->qd_idx
&&
3604 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3605 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3612 /* fetch_block - checks the given member device to see if its data needs
3613 * to be read or computed to satisfy a request.
3615 * Returns 1 when no more member devices need to be checked, otherwise returns
3616 * 0 to tell the loop in handle_stripe_fill to continue
3618 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3619 int disk_idx
, int disks
)
3621 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3623 /* is the data in this block needed, and can we get it? */
3624 if (need_this_block(sh
, s
, disk_idx
, disks
)) {
3625 /* we would like to get this block, possibly by computing it,
3626 * otherwise read it if the backing disk is insync
3628 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
3629 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
3630 BUG_ON(sh
->batch_head
);
3633 * In the raid6 case if the only non-uptodate disk is P
3634 * then we already trusted P to compute the other failed
3635 * drives. It is safe to compute rather than re-read P.
3636 * In other cases we only compute blocks from failed
3637 * devices, otherwise check/repair might fail to detect
3638 * a real inconsistency.
3641 if ((s
->uptodate
== disks
- 1) &&
3642 ((sh
->qd_idx
>= 0 && sh
->pd_idx
== disk_idx
) ||
3643 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
3644 disk_idx
== s
->failed_num
[1])))) {
3645 /* have disk failed, and we're requested to fetch it;
3648 pr_debug("Computing stripe %llu block %d\n",
3649 (unsigned long long)sh
->sector
, disk_idx
);
3650 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3651 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3652 set_bit(R5_Wantcompute
, &dev
->flags
);
3653 sh
->ops
.target
= disk_idx
;
3654 sh
->ops
.target2
= -1; /* no 2nd target */
3656 /* Careful: from this point on 'uptodate' is in the eye
3657 * of raid_run_ops which services 'compute' operations
3658 * before writes. R5_Wantcompute flags a block that will
3659 * be R5_UPTODATE by the time it is needed for a
3660 * subsequent operation.
3664 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
3665 /* Computing 2-failure is *very* expensive; only
3666 * do it if failed >= 2
3669 for (other
= disks
; other
--; ) {
3670 if (other
== disk_idx
)
3672 if (!test_bit(R5_UPTODATE
,
3673 &sh
->dev
[other
].flags
))
3677 pr_debug("Computing stripe %llu blocks %d,%d\n",
3678 (unsigned long long)sh
->sector
,
3680 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3681 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3682 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
3683 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
3684 sh
->ops
.target
= disk_idx
;
3685 sh
->ops
.target2
= other
;
3689 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
3690 set_bit(R5_LOCKED
, &dev
->flags
);
3691 set_bit(R5_Wantread
, &dev
->flags
);
3693 pr_debug("Reading block %d (sync=%d)\n",
3694 disk_idx
, s
->syncing
);
3702 * handle_stripe_fill - read or compute data to satisfy pending requests.
3704 static void handle_stripe_fill(struct stripe_head
*sh
,
3705 struct stripe_head_state
*s
,
3710 /* look for blocks to read/compute, skip this if a compute
3711 * is already in flight, or if the stripe contents are in the
3712 * midst of changing due to a write
3714 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
3715 !sh
->reconstruct_state
) {
3718 * For degraded stripe with data in journal, do not handle
3719 * read requests yet, instead, flush the stripe to raid
3720 * disks first, this avoids handling complex rmw of write
3721 * back cache (prexor with orig_page, and then xor with
3722 * page) in the read path
3724 if (s
->injournal
&& s
->failed
) {
3725 if (test_bit(STRIPE_R5C_CACHING
, &sh
->state
))
3726 r5c_make_stripe_write_out(sh
);
3730 for (i
= disks
; i
--; )
3731 if (fetch_block(sh
, s
, i
, disks
))
3735 set_bit(STRIPE_HANDLE
, &sh
->state
);
3738 static void break_stripe_batch_list(struct stripe_head
*head_sh
,
3739 unsigned long handle_flags
);
3740 /* handle_stripe_clean_event
3741 * any written block on an uptodate or failed drive can be returned.
3742 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3743 * never LOCKED, so we don't need to test 'failed' directly.
3745 static void handle_stripe_clean_event(struct r5conf
*conf
,
3746 struct stripe_head
*sh
, int disks
)
3750 int discard_pending
= 0;
3751 struct stripe_head
*head_sh
= sh
;
3752 bool do_endio
= false;
3754 for (i
= disks
; i
--; )
3755 if (sh
->dev
[i
].written
) {
3757 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
3758 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
3759 test_bit(R5_Discard
, &dev
->flags
) ||
3760 test_bit(R5_SkipCopy
, &dev
->flags
))) {
3761 /* We can return any write requests */
3762 struct bio
*wbi
, *wbi2
;
3763 pr_debug("Return write for disc %d\n", i
);
3764 if (test_and_clear_bit(R5_Discard
, &dev
->flags
))
3765 clear_bit(R5_UPTODATE
, &dev
->flags
);
3766 if (test_and_clear_bit(R5_SkipCopy
, &dev
->flags
)) {
3767 WARN_ON(test_bit(R5_UPTODATE
, &dev
->flags
));
3772 dev
->page
= dev
->orig_page
;
3774 dev
->written
= NULL
;
3775 while (wbi
&& wbi
->bi_iter
.bi_sector
<
3776 dev
->sector
+ STRIPE_SECTORS
) {
3777 wbi2
= r5_next_bio(wbi
, dev
->sector
);
3778 md_write_end(conf
->mddev
);
3782 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3784 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
3786 if (head_sh
->batch_head
) {
3787 sh
= list_first_entry(&sh
->batch_list
,
3790 if (sh
!= head_sh
) {
3797 } else if (test_bit(R5_Discard
, &dev
->flags
))
3798 discard_pending
= 1;
3801 log_stripe_write_finished(sh
);
3803 if (!discard_pending
&&
3804 test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
)) {
3806 clear_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
3807 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3808 if (sh
->qd_idx
>= 0) {
3809 clear_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
3810 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
);
3812 /* now that discard is done we can proceed with any sync */
3813 clear_bit(STRIPE_DISCARD
, &sh
->state
);
3815 * SCSI discard will change some bio fields and the stripe has
3816 * no updated data, so remove it from hash list and the stripe
3817 * will be reinitialized
3820 hash
= sh
->hash_lock_index
;
3821 spin_lock_irq(conf
->hash_locks
+ hash
);
3823 spin_unlock_irq(conf
->hash_locks
+ hash
);
3824 if (head_sh
->batch_head
) {
3825 sh
= list_first_entry(&sh
->batch_list
,
3826 struct stripe_head
, batch_list
);
3832 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
))
3833 set_bit(STRIPE_HANDLE
, &sh
->state
);
3837 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3838 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3839 md_wakeup_thread(conf
->mddev
->thread
);
3841 if (head_sh
->batch_head
&& do_endio
)
3842 break_stripe_batch_list(head_sh
, STRIPE_EXPAND_SYNC_FLAGS
);
3846 * For RMW in write back cache, we need extra page in prexor to store the
3847 * old data. This page is stored in dev->orig_page.
3849 * This function checks whether we have data for prexor. The exact logic
3851 * R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3853 static inline bool uptodate_for_rmw(struct r5dev
*dev
)
3855 return (test_bit(R5_UPTODATE
, &dev
->flags
)) &&
3856 (!test_bit(R5_InJournal
, &dev
->flags
) ||
3857 test_bit(R5_OrigPageUPTDODATE
, &dev
->flags
));
3860 static int handle_stripe_dirtying(struct r5conf
*conf
,
3861 struct stripe_head
*sh
,
3862 struct stripe_head_state
*s
,
3865 int rmw
= 0, rcw
= 0, i
;
3866 sector_t recovery_cp
= conf
->mddev
->recovery_cp
;
3868 /* Check whether resync is now happening or should start.
3869 * If yes, then the array is dirty (after unclean shutdown or
3870 * initial creation), so parity in some stripes might be inconsistent.
3871 * In this case, we need to always do reconstruct-write, to ensure
3872 * that in case of drive failure or read-error correction, we
3873 * generate correct data from the parity.
3875 if (conf
->rmw_level
== PARITY_DISABLE_RMW
||
3876 (recovery_cp
< MaxSector
&& sh
->sector
>= recovery_cp
&&
3878 /* Calculate the real rcw later - for now make it
3879 * look like rcw is cheaper
3882 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3883 conf
->rmw_level
, (unsigned long long)recovery_cp
,
3884 (unsigned long long)sh
->sector
);
3885 } else for (i
= disks
; i
--; ) {
3886 /* would I have to read this buffer for read_modify_write */
3887 struct r5dev
*dev
= &sh
->dev
[i
];
3888 if (((dev
->towrite
&& !delay_towrite(conf
, dev
, s
)) ||
3889 i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3890 test_bit(R5_InJournal
, &dev
->flags
)) &&
3891 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3892 !(uptodate_for_rmw(dev
) ||
3893 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3894 if (test_bit(R5_Insync
, &dev
->flags
))
3897 rmw
+= 2*disks
; /* cannot read it */
3899 /* Would I have to read this buffer for reconstruct_write */
3900 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3901 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3902 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3903 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3904 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3905 if (test_bit(R5_Insync
, &dev
->flags
))
3912 pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3913 (unsigned long long)sh
->sector
, sh
->state
, rmw
, rcw
);
3914 set_bit(STRIPE_HANDLE
, &sh
->state
);
3915 if ((rmw
< rcw
|| (rmw
== rcw
&& conf
->rmw_level
== PARITY_PREFER_RMW
)) && rmw
> 0) {
3916 /* prefer read-modify-write, but need to get some data */
3917 if (conf
->mddev
->queue
)
3918 blk_add_trace_msg(conf
->mddev
->queue
,
3919 "raid5 rmw %llu %d",
3920 (unsigned long long)sh
->sector
, rmw
);
3921 for (i
= disks
; i
--; ) {
3922 struct r5dev
*dev
= &sh
->dev
[i
];
3923 if (test_bit(R5_InJournal
, &dev
->flags
) &&
3924 dev
->page
== dev
->orig_page
&&
3925 !test_bit(R5_LOCKED
, &sh
->dev
[sh
->pd_idx
].flags
)) {
3926 /* alloc page for prexor */
3927 struct page
*p
= alloc_page(GFP_NOIO
);
3935 * alloc_page() failed, try use
3936 * disk_info->extra_page
3938 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE
,
3939 &conf
->cache_state
)) {
3940 r5c_use_extra_page(sh
);
3944 /* extra_page in use, add to delayed_list */
3945 set_bit(STRIPE_DELAYED
, &sh
->state
);
3946 s
->waiting_extra_page
= 1;
3951 for (i
= disks
; i
--; ) {
3952 struct r5dev
*dev
= &sh
->dev
[i
];
3953 if (((dev
->towrite
&& !delay_towrite(conf
, dev
, s
)) ||
3954 i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3955 test_bit(R5_InJournal
, &dev
->flags
)) &&
3956 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3957 !(uptodate_for_rmw(dev
) ||
3958 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
3959 test_bit(R5_Insync
, &dev
->flags
)) {
3960 if (test_bit(STRIPE_PREREAD_ACTIVE
,
3962 pr_debug("Read_old block %d for r-m-w\n",
3964 set_bit(R5_LOCKED
, &dev
->flags
);
3965 set_bit(R5_Wantread
, &dev
->flags
);
3968 set_bit(STRIPE_DELAYED
, &sh
->state
);
3969 set_bit(STRIPE_HANDLE
, &sh
->state
);
3974 if ((rcw
< rmw
|| (rcw
== rmw
&& conf
->rmw_level
!= PARITY_PREFER_RMW
)) && rcw
> 0) {
3975 /* want reconstruct write, but need to get some data */
3978 for (i
= disks
; i
--; ) {
3979 struct r5dev
*dev
= &sh
->dev
[i
];
3980 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3981 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3982 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3983 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3984 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3986 if (test_bit(R5_Insync
, &dev
->flags
) &&
3987 test_bit(STRIPE_PREREAD_ACTIVE
,
3989 pr_debug("Read_old block "
3990 "%d for Reconstruct\n", i
);
3991 set_bit(R5_LOCKED
, &dev
->flags
);
3992 set_bit(R5_Wantread
, &dev
->flags
);
3996 set_bit(STRIPE_DELAYED
, &sh
->state
);
3997 set_bit(STRIPE_HANDLE
, &sh
->state
);
4001 if (rcw
&& conf
->mddev
->queue
)
4002 blk_add_trace_msg(conf
->mddev
->queue
, "raid5 rcw %llu %d %d %d",
4003 (unsigned long long)sh
->sector
,
4004 rcw
, qread
, test_bit(STRIPE_DELAYED
, &sh
->state
));
4007 if (rcw
> disks
&& rmw
> disks
&&
4008 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4009 set_bit(STRIPE_DELAYED
, &sh
->state
);
4011 /* now if nothing is locked, and if we have enough data,
4012 * we can start a write request
4014 /* since handle_stripe can be called at any time we need to handle the
4015 * case where a compute block operation has been submitted and then a
4016 * subsequent call wants to start a write request. raid_run_ops only
4017 * handles the case where compute block and reconstruct are requested
4018 * simultaneously. If this is not the case then new writes need to be
4019 * held off until the compute completes.
4021 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
4022 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
4023 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
4024 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
4028 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
4029 struct stripe_head_state
*s
, int disks
)
4031 struct r5dev
*dev
= NULL
;
4033 BUG_ON(sh
->batch_head
);
4034 set_bit(STRIPE_HANDLE
, &sh
->state
);
4036 switch (sh
->check_state
) {
4037 case check_state_idle
:
4038 /* start a new check operation if there are no failures */
4039 if (s
->failed
== 0) {
4040 BUG_ON(s
->uptodate
!= disks
);
4041 sh
->check_state
= check_state_run
;
4042 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
4043 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
4047 dev
= &sh
->dev
[s
->failed_num
[0]];
4049 case check_state_compute_result
:
4050 sh
->check_state
= check_state_idle
;
4052 dev
= &sh
->dev
[sh
->pd_idx
];
4054 /* check that a write has not made the stripe insync */
4055 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
4058 /* either failed parity check, or recovery is happening */
4059 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
4060 BUG_ON(s
->uptodate
!= disks
);
4062 set_bit(R5_LOCKED
, &dev
->flags
);
4064 set_bit(R5_Wantwrite
, &dev
->flags
);
4066 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
4067 set_bit(STRIPE_INSYNC
, &sh
->state
);
4069 case check_state_run
:
4070 break; /* we will be called again upon completion */
4071 case check_state_check_result
:
4072 sh
->check_state
= check_state_idle
;
4074 /* if a failure occurred during the check operation, leave
4075 * STRIPE_INSYNC not set and let the stripe be handled again
4080 /* handle a successful check operation, if parity is correct
4081 * we are done. Otherwise update the mismatch count and repair
4082 * parity if !MD_RECOVERY_CHECK
4084 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
4085 /* parity is correct (on disc,
4086 * not in buffer any more)
4088 set_bit(STRIPE_INSYNC
, &sh
->state
);
4090 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
4091 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
)) {
4092 /* don't try to repair!! */
4093 set_bit(STRIPE_INSYNC
, &sh
->state
);
4094 pr_warn_ratelimited("%s: mismatch sector in range "
4095 "%llu-%llu\n", mdname(conf
->mddev
),
4096 (unsigned long long) sh
->sector
,
4097 (unsigned long long) sh
->sector
+
4100 sh
->check_state
= check_state_compute_run
;
4101 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
4102 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
4103 set_bit(R5_Wantcompute
,
4104 &sh
->dev
[sh
->pd_idx
].flags
);
4105 sh
->ops
.target
= sh
->pd_idx
;
4106 sh
->ops
.target2
= -1;
4111 case check_state_compute_run
:
4114 pr_err("%s: unknown check_state: %d sector: %llu\n",
4115 __func__
, sh
->check_state
,
4116 (unsigned long long) sh
->sector
);
4121 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
4122 struct stripe_head_state
*s
,
4125 int pd_idx
= sh
->pd_idx
;
4126 int qd_idx
= sh
->qd_idx
;
4129 BUG_ON(sh
->batch_head
);
4130 set_bit(STRIPE_HANDLE
, &sh
->state
);
4132 BUG_ON(s
->failed
> 2);
4134 /* Want to check and possibly repair P and Q.
4135 * However there could be one 'failed' device, in which
4136 * case we can only check one of them, possibly using the
4137 * other to generate missing data
4140 switch (sh
->check_state
) {
4141 case check_state_idle
:
4142 /* start a new check operation if there are < 2 failures */
4143 if (s
->failed
== s
->q_failed
) {
4144 /* The only possible failed device holds Q, so it
4145 * makes sense to check P (If anything else were failed,
4146 * we would have used P to recreate it).
4148 sh
->check_state
= check_state_run
;
4150 if (!s
->q_failed
&& s
->failed
< 2) {
4151 /* Q is not failed, and we didn't use it to generate
4152 * anything, so it makes sense to check it
4154 if (sh
->check_state
== check_state_run
)
4155 sh
->check_state
= check_state_run_pq
;
4157 sh
->check_state
= check_state_run_q
;
4160 /* discard potentially stale zero_sum_result */
4161 sh
->ops
.zero_sum_result
= 0;
4163 if (sh
->check_state
== check_state_run
) {
4164 /* async_xor_zero_sum destroys the contents of P */
4165 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
4168 if (sh
->check_state
>= check_state_run
&&
4169 sh
->check_state
<= check_state_run_pq
) {
4170 /* async_syndrome_zero_sum preserves P and Q, so
4171 * no need to mark them !uptodate here
4173 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
4177 /* we have 2-disk failure */
4178 BUG_ON(s
->failed
!= 2);
4180 case check_state_compute_result
:
4181 sh
->check_state
= check_state_idle
;
4183 /* check that a write has not made the stripe insync */
4184 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
4187 /* now write out any block on a failed drive,
4188 * or P or Q if they were recomputed
4190 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
4191 if (s
->failed
== 2) {
4192 dev
= &sh
->dev
[s
->failed_num
[1]];
4194 set_bit(R5_LOCKED
, &dev
->flags
);
4195 set_bit(R5_Wantwrite
, &dev
->flags
);
4197 if (s
->failed
>= 1) {
4198 dev
= &sh
->dev
[s
->failed_num
[0]];
4200 set_bit(R5_LOCKED
, &dev
->flags
);
4201 set_bit(R5_Wantwrite
, &dev
->flags
);
4203 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
4204 dev
= &sh
->dev
[pd_idx
];
4206 set_bit(R5_LOCKED
, &dev
->flags
);
4207 set_bit(R5_Wantwrite
, &dev
->flags
);
4209 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
4210 dev
= &sh
->dev
[qd_idx
];
4212 set_bit(R5_LOCKED
, &dev
->flags
);
4213 set_bit(R5_Wantwrite
, &dev
->flags
);
4215 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
4217 set_bit(STRIPE_INSYNC
, &sh
->state
);
4219 case check_state_run
:
4220 case check_state_run_q
:
4221 case check_state_run_pq
:
4222 break; /* we will be called again upon completion */
4223 case check_state_check_result
:
4224 sh
->check_state
= check_state_idle
;
4226 /* handle a successful check operation, if parity is correct
4227 * we are done. Otherwise update the mismatch count and repair
4228 * parity if !MD_RECOVERY_CHECK
4230 if (sh
->ops
.zero_sum_result
== 0) {
4231 /* both parities are correct */
4233 set_bit(STRIPE_INSYNC
, &sh
->state
);
4235 /* in contrast to the raid5 case we can validate
4236 * parity, but still have a failure to write
4239 sh
->check_state
= check_state_compute_result
;
4240 /* Returning at this point means that we may go
4241 * off and bring p and/or q uptodate again so
4242 * we make sure to check zero_sum_result again
4243 * to verify if p or q need writeback
4247 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
4248 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
)) {
4249 /* don't try to repair!! */
4250 set_bit(STRIPE_INSYNC
, &sh
->state
);
4251 pr_warn_ratelimited("%s: mismatch sector in range "
4252 "%llu-%llu\n", mdname(conf
->mddev
),
4253 (unsigned long long) sh
->sector
,
4254 (unsigned long long) sh
->sector
+
4257 int *target
= &sh
->ops
.target
;
4259 sh
->ops
.target
= -1;
4260 sh
->ops
.target2
= -1;
4261 sh
->check_state
= check_state_compute_run
;
4262 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
4263 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
4264 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
4265 set_bit(R5_Wantcompute
,
4266 &sh
->dev
[pd_idx
].flags
);
4268 target
= &sh
->ops
.target2
;
4271 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
4272 set_bit(R5_Wantcompute
,
4273 &sh
->dev
[qd_idx
].flags
);
4280 case check_state_compute_run
:
4283 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4284 __func__
, sh
->check_state
,
4285 (unsigned long long) sh
->sector
);
4290 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
4294 /* We have read all the blocks in this stripe and now we need to
4295 * copy some of them into a target stripe for expand.
4297 struct dma_async_tx_descriptor
*tx
= NULL
;
4298 BUG_ON(sh
->batch_head
);
4299 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
4300 for (i
= 0; i
< sh
->disks
; i
++)
4301 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
4303 struct stripe_head
*sh2
;
4304 struct async_submit_ctl submit
;
4306 sector_t bn
= raid5_compute_blocknr(sh
, i
, 1);
4307 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
4309 sh2
= raid5_get_active_stripe(conf
, s
, 0, 1, 1);
4311 /* so far only the early blocks of this stripe
4312 * have been requested. When later blocks
4313 * get requested, we will try again
4316 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
4317 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
4318 /* must have already done this block */
4319 raid5_release_stripe(sh2
);
4323 /* place all the copies on one channel */
4324 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
4325 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
4326 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
4329 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
4330 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
4331 for (j
= 0; j
< conf
->raid_disks
; j
++)
4332 if (j
!= sh2
->pd_idx
&&
4334 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
4336 if (j
== conf
->raid_disks
) {
4337 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
4338 set_bit(STRIPE_HANDLE
, &sh2
->state
);
4340 raid5_release_stripe(sh2
);
4343 /* done submitting copies, wait for them to complete */
4344 async_tx_quiesce(&tx
);
4348 * handle_stripe - do things to a stripe.
4350 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4351 * state of various bits to see what needs to be done.
4353 * return some read requests which now have data
4354 * return some write requests which are safely on storage
4355 * schedule a read on some buffers
4356 * schedule a write of some buffers
4357 * return confirmation of parity correctness
4361 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
4363 struct r5conf
*conf
= sh
->raid_conf
;
4364 int disks
= sh
->disks
;
4367 int do_recovery
= 0;
4369 memset(s
, 0, sizeof(*s
));
4371 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
) && !sh
->batch_head
;
4372 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
) && !sh
->batch_head
;
4373 s
->failed_num
[0] = -1;
4374 s
->failed_num
[1] = -1;
4375 s
->log_failed
= r5l_log_disk_error(conf
);
4377 /* Now to look around and see what can be done */
4379 for (i
=disks
; i
--; ) {
4380 struct md_rdev
*rdev
;
4387 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4389 dev
->toread
, dev
->towrite
, dev
->written
);
4390 /* maybe we can reply to a read
4392 * new wantfill requests are only permitted while
4393 * ops_complete_biofill is guaranteed to be inactive
4395 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
4396 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
4397 set_bit(R5_Wantfill
, &dev
->flags
);
4399 /* now count some things */
4400 if (test_bit(R5_LOCKED
, &dev
->flags
))
4402 if (test_bit(R5_UPTODATE
, &dev
->flags
))
4404 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
4406 BUG_ON(s
->compute
> 2);
4409 if (test_bit(R5_Wantfill
, &dev
->flags
))
4411 else if (dev
->toread
)
4415 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
4420 /* Prefer to use the replacement for reads, but only
4421 * if it is recovered enough and has no bad blocks.
4423 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
4424 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
4425 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
4426 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4427 &first_bad
, &bad_sectors
))
4428 set_bit(R5_ReadRepl
, &dev
->flags
);
4430 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4431 set_bit(R5_NeedReplace
, &dev
->flags
);
4433 clear_bit(R5_NeedReplace
, &dev
->flags
);
4434 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
4435 clear_bit(R5_ReadRepl
, &dev
->flags
);
4437 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
4440 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4441 &first_bad
, &bad_sectors
);
4442 if (s
->blocked_rdev
== NULL
4443 && (test_bit(Blocked
, &rdev
->flags
)
4446 set_bit(BlockedBadBlocks
,
4448 s
->blocked_rdev
= rdev
;
4449 atomic_inc(&rdev
->nr_pending
);
4452 clear_bit(R5_Insync
, &dev
->flags
);
4456 /* also not in-sync */
4457 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
4458 test_bit(R5_UPTODATE
, &dev
->flags
)) {
4459 /* treat as in-sync, but with a read error
4460 * which we can now try to correct
4462 set_bit(R5_Insync
, &dev
->flags
);
4463 set_bit(R5_ReadError
, &dev
->flags
);
4465 } else if (test_bit(In_sync
, &rdev
->flags
))
4466 set_bit(R5_Insync
, &dev
->flags
);
4467 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
4468 /* in sync if before recovery_offset */
4469 set_bit(R5_Insync
, &dev
->flags
);
4470 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
4471 test_bit(R5_Expanded
, &dev
->flags
))
4472 /* If we've reshaped into here, we assume it is Insync.
4473 * We will shortly update recovery_offset to make
4476 set_bit(R5_Insync
, &dev
->flags
);
4478 if (test_bit(R5_WriteError
, &dev
->flags
)) {
4479 /* This flag does not apply to '.replacement'
4480 * only to .rdev, so make sure to check that*/
4481 struct md_rdev
*rdev2
= rcu_dereference(
4482 conf
->disks
[i
].rdev
);
4484 clear_bit(R5_Insync
, &dev
->flags
);
4485 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4486 s
->handle_bad_blocks
= 1;
4487 atomic_inc(&rdev2
->nr_pending
);
4489 clear_bit(R5_WriteError
, &dev
->flags
);
4491 if (test_bit(R5_MadeGood
, &dev
->flags
)) {
4492 /* This flag does not apply to '.replacement'
4493 * only to .rdev, so make sure to check that*/
4494 struct md_rdev
*rdev2
= rcu_dereference(
4495 conf
->disks
[i
].rdev
);
4496 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4497 s
->handle_bad_blocks
= 1;
4498 atomic_inc(&rdev2
->nr_pending
);
4500 clear_bit(R5_MadeGood
, &dev
->flags
);
4502 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
4503 struct md_rdev
*rdev2
= rcu_dereference(
4504 conf
->disks
[i
].replacement
);
4505 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4506 s
->handle_bad_blocks
= 1;
4507 atomic_inc(&rdev2
->nr_pending
);
4509 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
4511 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4512 /* The ReadError flag will just be confusing now */
4513 clear_bit(R5_ReadError
, &dev
->flags
);
4514 clear_bit(R5_ReWrite
, &dev
->flags
);
4516 if (test_bit(R5_ReadError
, &dev
->flags
))
4517 clear_bit(R5_Insync
, &dev
->flags
);
4518 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4520 s
->failed_num
[s
->failed
] = i
;
4522 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4526 if (test_bit(R5_InJournal
, &dev
->flags
))
4528 if (test_bit(R5_InJournal
, &dev
->flags
) && dev
->written
)
4531 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
4532 /* If there is a failed device being replaced,
4533 * we must be recovering.
4534 * else if we are after recovery_cp, we must be syncing
4535 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4536 * else we can only be replacing
4537 * sync and recovery both need to read all devices, and so
4538 * use the same flag.
4541 sh
->sector
>= conf
->mddev
->recovery_cp
||
4542 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
4550 static int clear_batch_ready(struct stripe_head
*sh
)
4552 /* Return '1' if this is a member of batch, or
4553 * '0' if it is a lone stripe or a head which can now be
4556 struct stripe_head
*tmp
;
4557 if (!test_and_clear_bit(STRIPE_BATCH_READY
, &sh
->state
))
4558 return (sh
->batch_head
&& sh
->batch_head
!= sh
);
4559 spin_lock(&sh
->stripe_lock
);
4560 if (!sh
->batch_head
) {
4561 spin_unlock(&sh
->stripe_lock
);
4566 * this stripe could be added to a batch list before we check
4567 * BATCH_READY, skips it
4569 if (sh
->batch_head
!= sh
) {
4570 spin_unlock(&sh
->stripe_lock
);
4573 spin_lock(&sh
->batch_lock
);
4574 list_for_each_entry(tmp
, &sh
->batch_list
, batch_list
)
4575 clear_bit(STRIPE_BATCH_READY
, &tmp
->state
);
4576 spin_unlock(&sh
->batch_lock
);
4577 spin_unlock(&sh
->stripe_lock
);
4580 * BATCH_READY is cleared, no new stripes can be added.
4581 * batch_list can be accessed without lock
4586 static void break_stripe_batch_list(struct stripe_head
*head_sh
,
4587 unsigned long handle_flags
)
4589 struct stripe_head
*sh
, *next
;
4593 list_for_each_entry_safe(sh
, next
, &head_sh
->batch_list
, batch_list
) {
4595 list_del_init(&sh
->batch_list
);
4597 WARN_ONCE(sh
->state
& ((1 << STRIPE_ACTIVE
) |
4598 (1 << STRIPE_SYNCING
) |
4599 (1 << STRIPE_REPLACED
) |
4600 (1 << STRIPE_DELAYED
) |
4601 (1 << STRIPE_BIT_DELAY
) |
4602 (1 << STRIPE_FULL_WRITE
) |
4603 (1 << STRIPE_BIOFILL_RUN
) |
4604 (1 << STRIPE_COMPUTE_RUN
) |
4605 (1 << STRIPE_OPS_REQ_PENDING
) |
4606 (1 << STRIPE_DISCARD
) |
4607 (1 << STRIPE_BATCH_READY
) |
4608 (1 << STRIPE_BATCH_ERR
) |
4609 (1 << STRIPE_BITMAP_PENDING
)),
4610 "stripe state: %lx\n", sh
->state
);
4611 WARN_ONCE(head_sh
->state
& ((1 << STRIPE_DISCARD
) |
4612 (1 << STRIPE_REPLACED
)),
4613 "head stripe state: %lx\n", head_sh
->state
);
4615 set_mask_bits(&sh
->state
, ~(STRIPE_EXPAND_SYNC_FLAGS
|
4616 (1 << STRIPE_PREREAD_ACTIVE
) |
4617 (1 << STRIPE_DEGRADED
) |
4618 (1 << STRIPE_ON_UNPLUG_LIST
)),
4619 head_sh
->state
& (1 << STRIPE_INSYNC
));
4621 sh
->check_state
= head_sh
->check_state
;
4622 sh
->reconstruct_state
= head_sh
->reconstruct_state
;
4623 spin_lock_irq(&sh
->stripe_lock
);
4624 sh
->batch_head
= NULL
;
4625 spin_unlock_irq(&sh
->stripe_lock
);
4626 for (i
= 0; i
< sh
->disks
; i
++) {
4627 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
4629 sh
->dev
[i
].flags
= head_sh
->dev
[i
].flags
&
4630 (~((1 << R5_WriteError
) | (1 << R5_Overlap
)));
4632 if (handle_flags
== 0 ||
4633 sh
->state
& handle_flags
)
4634 set_bit(STRIPE_HANDLE
, &sh
->state
);
4635 raid5_release_stripe(sh
);
4637 spin_lock_irq(&head_sh
->stripe_lock
);
4638 head_sh
->batch_head
= NULL
;
4639 spin_unlock_irq(&head_sh
->stripe_lock
);
4640 for (i
= 0; i
< head_sh
->disks
; i
++)
4641 if (test_and_clear_bit(R5_Overlap
, &head_sh
->dev
[i
].flags
))
4643 if (head_sh
->state
& handle_flags
)
4644 set_bit(STRIPE_HANDLE
, &head_sh
->state
);
4647 wake_up(&head_sh
->raid_conf
->wait_for_overlap
);
4650 static void handle_stripe(struct stripe_head
*sh
)
4652 struct stripe_head_state s
;
4653 struct r5conf
*conf
= sh
->raid_conf
;
4656 int disks
= sh
->disks
;
4657 struct r5dev
*pdev
, *qdev
;
4659 clear_bit(STRIPE_HANDLE
, &sh
->state
);
4660 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
4661 /* already being handled, ensure it gets handled
4662 * again when current action finishes */
4663 set_bit(STRIPE_HANDLE
, &sh
->state
);
4667 if (clear_batch_ready(sh
) ) {
4668 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
4672 if (test_and_clear_bit(STRIPE_BATCH_ERR
, &sh
->state
))
4673 break_stripe_batch_list(sh
, 0);
4675 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
) && !sh
->batch_head
) {
4676 spin_lock(&sh
->stripe_lock
);
4678 * Cannot process 'sync' concurrently with 'discard'.
4679 * Flush data in r5cache before 'sync'.
4681 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE
, &sh
->state
) &&
4682 !test_bit(STRIPE_R5C_FULL_STRIPE
, &sh
->state
) &&
4683 !test_bit(STRIPE_DISCARD
, &sh
->state
) &&
4684 test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
4685 set_bit(STRIPE_SYNCING
, &sh
->state
);
4686 clear_bit(STRIPE_INSYNC
, &sh
->state
);
4687 clear_bit(STRIPE_REPLACED
, &sh
->state
);
4689 spin_unlock(&sh
->stripe_lock
);
4691 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4693 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4694 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4695 (unsigned long long)sh
->sector
, sh
->state
,
4696 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
4697 sh
->check_state
, sh
->reconstruct_state
);
4699 analyse_stripe(sh
, &s
);
4701 if (test_bit(STRIPE_LOG_TRAPPED
, &sh
->state
))
4704 if (s
.handle_bad_blocks
||
4705 test_bit(MD_SB_CHANGE_PENDING
, &conf
->mddev
->sb_flags
)) {
4706 set_bit(STRIPE_HANDLE
, &sh
->state
);
4710 if (unlikely(s
.blocked_rdev
)) {
4711 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
4712 s
.replacing
|| s
.to_write
|| s
.written
) {
4713 set_bit(STRIPE_HANDLE
, &sh
->state
);
4716 /* There is nothing for the blocked_rdev to block */
4717 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
4718 s
.blocked_rdev
= NULL
;
4721 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
4722 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
4723 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
4726 pr_debug("locked=%d uptodate=%d to_read=%d"
4727 " to_write=%d failed=%d failed_num=%d,%d\n",
4728 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
4729 s
.failed_num
[0], s
.failed_num
[1]);
4731 * check if the array has lost more than max_degraded devices and,
4732 * if so, some requests might need to be failed.
4734 * When journal device failed (log_failed), we will only process
4735 * the stripe if there is data need write to raid disks
4737 if (s
.failed
> conf
->max_degraded
||
4738 (s
.log_failed
&& s
.injournal
== 0)) {
4739 sh
->check_state
= 0;
4740 sh
->reconstruct_state
= 0;
4741 break_stripe_batch_list(sh
, 0);
4742 if (s
.to_read
+s
.to_write
+s
.written
)
4743 handle_failed_stripe(conf
, sh
, &s
, disks
);
4744 if (s
.syncing
+ s
.replacing
)
4745 handle_failed_sync(conf
, sh
, &s
);
4748 /* Now we check to see if any write operations have recently
4752 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
4754 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
4755 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
4756 sh
->reconstruct_state
= reconstruct_state_idle
;
4758 /* All the 'written' buffers and the parity block are ready to
4759 * be written back to disk
4761 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
) &&
4762 !test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
));
4763 BUG_ON(sh
->qd_idx
>= 0 &&
4764 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
) &&
4765 !test_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
));
4766 for (i
= disks
; i
--; ) {
4767 struct r5dev
*dev
= &sh
->dev
[i
];
4768 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
4769 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
4770 dev
->written
|| test_bit(R5_InJournal
,
4772 pr_debug("Writing block %d\n", i
);
4773 set_bit(R5_Wantwrite
, &dev
->flags
);
4778 if (!test_bit(R5_Insync
, &dev
->flags
) ||
4779 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
4781 set_bit(STRIPE_INSYNC
, &sh
->state
);
4784 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4785 s
.dec_preread_active
= 1;
4789 * might be able to return some write requests if the parity blocks
4790 * are safe, or on a failed drive
4792 pdev
= &sh
->dev
[sh
->pd_idx
];
4793 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
4794 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
4795 qdev
= &sh
->dev
[sh
->qd_idx
];
4796 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
4797 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
4801 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
4802 && !test_bit(R5_LOCKED
, &pdev
->flags
)
4803 && (test_bit(R5_UPTODATE
, &pdev
->flags
) ||
4804 test_bit(R5_Discard
, &pdev
->flags
))))) &&
4805 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
4806 && !test_bit(R5_LOCKED
, &qdev
->flags
)
4807 && (test_bit(R5_UPTODATE
, &qdev
->flags
) ||
4808 test_bit(R5_Discard
, &qdev
->flags
))))))
4809 handle_stripe_clean_event(conf
, sh
, disks
);
4812 r5c_handle_cached_data_endio(conf
, sh
, disks
);
4813 log_stripe_write_finished(sh
);
4815 /* Now we might consider reading some blocks, either to check/generate
4816 * parity, or to satisfy requests
4817 * or to load a block that is being partially written.
4819 if (s
.to_read
|| s
.non_overwrite
4820 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
4821 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
4824 handle_stripe_fill(sh
, &s
, disks
);
4827 * When the stripe finishes full journal write cycle (write to journal
4828 * and raid disk), this is the clean up procedure so it is ready for
4831 r5c_finish_stripe_write_out(conf
, sh
, &s
);
4834 * Now to consider new write requests, cache write back and what else,
4835 * if anything should be read. We do not handle new writes when:
4836 * 1/ A 'write' operation (copy+xor) is already in flight.
4837 * 2/ A 'check' operation is in flight, as it may clobber the parity
4839 * 3/ A r5c cache log write is in flight.
4842 if (!sh
->reconstruct_state
&& !sh
->check_state
&& !sh
->log_io
) {
4843 if (!r5c_is_writeback(conf
->log
)) {
4845 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
4846 } else { /* write back cache */
4849 /* First, try handle writes in caching phase */
4851 ret
= r5c_try_caching_write(conf
, sh
, &s
,
4854 * If caching phase failed: ret == -EAGAIN
4856 * stripe under reclaim: !caching && injournal
4858 * fall back to handle_stripe_dirtying()
4860 if (ret
== -EAGAIN
||
4861 /* stripe under reclaim: !caching && injournal */
4862 (!test_bit(STRIPE_R5C_CACHING
, &sh
->state
) &&
4864 ret
= handle_stripe_dirtying(conf
, sh
, &s
,
4872 /* maybe we need to check and possibly fix the parity for this stripe
4873 * Any reads will already have been scheduled, so we just see if enough
4874 * data is available. The parity check is held off while parity
4875 * dependent operations are in flight.
4877 if (sh
->check_state
||
4878 (s
.syncing
&& s
.locked
== 0 &&
4879 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4880 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
4881 if (conf
->level
== 6)
4882 handle_parity_checks6(conf
, sh
, &s
, disks
);
4884 handle_parity_checks5(conf
, sh
, &s
, disks
);
4887 if ((s
.replacing
|| s
.syncing
) && s
.locked
== 0
4888 && !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)
4889 && !test_bit(STRIPE_REPLACED
, &sh
->state
)) {
4890 /* Write out to replacement devices where possible */
4891 for (i
= 0; i
< conf
->raid_disks
; i
++)
4892 if (test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
4893 WARN_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
4894 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
4895 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4899 set_bit(STRIPE_INSYNC
, &sh
->state
);
4900 set_bit(STRIPE_REPLACED
, &sh
->state
);
4902 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
4903 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4904 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
4905 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4906 clear_bit(STRIPE_SYNCING
, &sh
->state
);
4907 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
4908 wake_up(&conf
->wait_for_overlap
);
4911 /* If the failed drives are just a ReadError, then we might need
4912 * to progress the repair/check process
4914 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
4915 for (i
= 0; i
< s
.failed
; i
++) {
4916 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
4917 if (test_bit(R5_ReadError
, &dev
->flags
)
4918 && !test_bit(R5_LOCKED
, &dev
->flags
)
4919 && test_bit(R5_UPTODATE
, &dev
->flags
)
4921 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
4922 set_bit(R5_Wantwrite
, &dev
->flags
);
4923 set_bit(R5_ReWrite
, &dev
->flags
);
4924 set_bit(R5_LOCKED
, &dev
->flags
);
4927 /* let's read it back */
4928 set_bit(R5_Wantread
, &dev
->flags
);
4929 set_bit(R5_LOCKED
, &dev
->flags
);
4935 /* Finish reconstruct operations initiated by the expansion process */
4936 if (sh
->reconstruct_state
== reconstruct_state_result
) {
4937 struct stripe_head
*sh_src
4938 = raid5_get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
4939 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
4940 /* sh cannot be written until sh_src has been read.
4941 * so arrange for sh to be delayed a little
4943 set_bit(STRIPE_DELAYED
, &sh
->state
);
4944 set_bit(STRIPE_HANDLE
, &sh
->state
);
4945 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
4947 atomic_inc(&conf
->preread_active_stripes
);
4948 raid5_release_stripe(sh_src
);
4952 raid5_release_stripe(sh_src
);
4954 sh
->reconstruct_state
= reconstruct_state_idle
;
4955 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
4956 for (i
= conf
->raid_disks
; i
--; ) {
4957 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
4958 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4963 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
4964 !sh
->reconstruct_state
) {
4965 /* Need to write out all blocks after computing parity */
4966 sh
->disks
= conf
->raid_disks
;
4967 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
4968 schedule_reconstruction(sh
, &s
, 1, 1);
4969 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
4970 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4971 atomic_dec(&conf
->reshape_stripes
);
4972 wake_up(&conf
->wait_for_overlap
);
4973 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4976 if (s
.expanding
&& s
.locked
== 0 &&
4977 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
4978 handle_stripe_expansion(conf
, sh
);
4981 /* wait for this device to become unblocked */
4982 if (unlikely(s
.blocked_rdev
)) {
4983 if (conf
->mddev
->external
)
4984 md_wait_for_blocked_rdev(s
.blocked_rdev
,
4987 /* Internal metadata will immediately
4988 * be written by raid5d, so we don't
4989 * need to wait here.
4991 rdev_dec_pending(s
.blocked_rdev
,
4995 if (s
.handle_bad_blocks
)
4996 for (i
= disks
; i
--; ) {
4997 struct md_rdev
*rdev
;
4998 struct r5dev
*dev
= &sh
->dev
[i
];
4999 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
5000 /* We own a safe reference to the rdev */
5001 rdev
= conf
->disks
[i
].rdev
;
5002 if (!rdev_set_badblocks(rdev
, sh
->sector
,
5004 md_error(conf
->mddev
, rdev
);
5005 rdev_dec_pending(rdev
, conf
->mddev
);
5007 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
5008 rdev
= conf
->disks
[i
].rdev
;
5009 rdev_clear_badblocks(rdev
, sh
->sector
,
5011 rdev_dec_pending(rdev
, conf
->mddev
);
5013 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
5014 rdev
= conf
->disks
[i
].replacement
;
5016 /* rdev have been moved down */
5017 rdev
= conf
->disks
[i
].rdev
;
5018 rdev_clear_badblocks(rdev
, sh
->sector
,
5020 rdev_dec_pending(rdev
, conf
->mddev
);
5025 raid_run_ops(sh
, s
.ops_request
);
5029 if (s
.dec_preread_active
) {
5030 /* We delay this until after ops_run_io so that if make_request
5031 * is waiting on a flush, it won't continue until the writes
5032 * have actually been submitted.
5034 atomic_dec(&conf
->preread_active_stripes
);
5035 if (atomic_read(&conf
->preread_active_stripes
) <
5037 md_wakeup_thread(conf
->mddev
->thread
);
5040 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
5043 static void raid5_activate_delayed(struct r5conf
*conf
)
5045 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
5046 while (!list_empty(&conf
->delayed_list
)) {
5047 struct list_head
*l
= conf
->delayed_list
.next
;
5048 struct stripe_head
*sh
;
5049 sh
= list_entry(l
, struct stripe_head
, lru
);
5051 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5052 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5053 atomic_inc(&conf
->preread_active_stripes
);
5054 list_add_tail(&sh
->lru
, &conf
->hold_list
);
5055 raid5_wakeup_stripe_thread(sh
);
5060 static void activate_bit_delay(struct r5conf
*conf
,
5061 struct list_head
*temp_inactive_list
)
5063 /* device_lock is held */
5064 struct list_head head
;
5065 list_add(&head
, &conf
->bitmap_list
);
5066 list_del_init(&conf
->bitmap_list
);
5067 while (!list_empty(&head
)) {
5068 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
5070 list_del_init(&sh
->lru
);
5071 atomic_inc(&sh
->count
);
5072 hash
= sh
->hash_lock_index
;
5073 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
5077 static int raid5_congested(struct mddev
*mddev
, int bits
)
5079 struct r5conf
*conf
= mddev
->private;
5081 /* No difference between reads and writes. Just check
5082 * how busy the stripe_cache is
5085 if (test_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
))
5088 /* Also checks whether there is pressure on r5cache log space */
5089 if (test_bit(R5C_LOG_TIGHT
, &conf
->cache_state
))
5093 if (atomic_read(&conf
->empty_inactive_list_nr
))
5099 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
5101 struct r5conf
*conf
= mddev
->private;
5102 sector_t sector
= bio
->bi_iter
.bi_sector
;
5103 unsigned int chunk_sectors
;
5104 unsigned int bio_sectors
= bio_sectors(bio
);
5106 WARN_ON_ONCE(bio
->bi_partno
);
5108 chunk_sectors
= min(conf
->chunk_sectors
, conf
->prev_chunk_sectors
);
5109 return chunk_sectors
>=
5110 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
5114 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
5115 * later sampled by raid5d.
5117 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
5119 unsigned long flags
;
5121 spin_lock_irqsave(&conf
->device_lock
, flags
);
5123 bi
->bi_next
= conf
->retry_read_aligned_list
;
5124 conf
->retry_read_aligned_list
= bi
;
5126 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5127 md_wakeup_thread(conf
->mddev
->thread
);
5130 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
,
5131 unsigned int *offset
)
5135 bi
= conf
->retry_read_aligned
;
5137 *offset
= conf
->retry_read_offset
;
5138 conf
->retry_read_aligned
= NULL
;
5141 bi
= conf
->retry_read_aligned_list
;
5143 conf
->retry_read_aligned_list
= bi
->bi_next
;
5152 * The "raid5_align_endio" should check if the read succeeded and if it
5153 * did, call bio_endio on the original bio (having bio_put the new bio
5155 * If the read failed..
5157 static void raid5_align_endio(struct bio
*bi
)
5159 struct bio
* raid_bi
= bi
->bi_private
;
5160 struct mddev
*mddev
;
5161 struct r5conf
*conf
;
5162 struct md_rdev
*rdev
;
5163 blk_status_t error
= bi
->bi_status
;
5167 rdev
= (void*)raid_bi
->bi_next
;
5168 raid_bi
->bi_next
= NULL
;
5169 mddev
= rdev
->mddev
;
5170 conf
= mddev
->private;
5172 rdev_dec_pending(rdev
, conf
->mddev
);
5176 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
5177 wake_up(&conf
->wait_for_quiescent
);
5181 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5183 add_bio_to_retry(raid_bi
, conf
);
5186 static int raid5_read_one_chunk(struct mddev
*mddev
, struct bio
*raid_bio
)
5188 struct r5conf
*conf
= mddev
->private;
5190 struct bio
* align_bi
;
5191 struct md_rdev
*rdev
;
5192 sector_t end_sector
;
5194 if (!in_chunk_boundary(mddev
, raid_bio
)) {
5195 pr_debug("%s: non aligned\n", __func__
);
5199 * use bio_clone_fast to make a copy of the bio
5201 align_bi
= bio_clone_fast(raid_bio
, GFP_NOIO
, &mddev
->bio_set
);
5205 * set bi_end_io to a new function, and set bi_private to the
5208 align_bi
->bi_end_io
= raid5_align_endio
;
5209 align_bi
->bi_private
= raid_bio
;
5213 align_bi
->bi_iter
.bi_sector
=
5214 raid5_compute_sector(conf
, raid_bio
->bi_iter
.bi_sector
,
5217 end_sector
= bio_end_sector(align_bi
);
5219 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
5220 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
5221 rdev
->recovery_offset
< end_sector
) {
5222 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
5224 (test_bit(Faulty
, &rdev
->flags
) ||
5225 !(test_bit(In_sync
, &rdev
->flags
) ||
5226 rdev
->recovery_offset
>= end_sector
)))
5230 if (r5c_big_stripe_cached(conf
, align_bi
->bi_iter
.bi_sector
)) {
5240 atomic_inc(&rdev
->nr_pending
);
5242 raid_bio
->bi_next
= (void*)rdev
;
5243 bio_set_dev(align_bi
, rdev
->bdev
);
5244 bio_clear_flag(align_bi
, BIO_SEG_VALID
);
5246 if (is_badblock(rdev
, align_bi
->bi_iter
.bi_sector
,
5247 bio_sectors(align_bi
),
5248 &first_bad
, &bad_sectors
)) {
5250 rdev_dec_pending(rdev
, mddev
);
5254 /* No reshape active, so we can trust rdev->data_offset */
5255 align_bi
->bi_iter
.bi_sector
+= rdev
->data_offset
;
5257 spin_lock_irq(&conf
->device_lock
);
5258 wait_event_lock_irq(conf
->wait_for_quiescent
,
5261 atomic_inc(&conf
->active_aligned_reads
);
5262 spin_unlock_irq(&conf
->device_lock
);
5265 trace_block_bio_remap(align_bi
->bi_disk
->queue
,
5266 align_bi
, disk_devt(mddev
->gendisk
),
5267 raid_bio
->bi_iter
.bi_sector
);
5268 generic_make_request(align_bi
);
5277 static struct bio
*chunk_aligned_read(struct mddev
*mddev
, struct bio
*raid_bio
)
5280 sector_t sector
= raid_bio
->bi_iter
.bi_sector
;
5281 unsigned chunk_sects
= mddev
->chunk_sectors
;
5282 unsigned sectors
= chunk_sects
- (sector
& (chunk_sects
-1));
5284 if (sectors
< bio_sectors(raid_bio
)) {
5285 struct r5conf
*conf
= mddev
->private;
5286 split
= bio_split(raid_bio
, sectors
, GFP_NOIO
, &conf
->bio_split
);
5287 bio_chain(split
, raid_bio
);
5288 generic_make_request(raid_bio
);
5292 if (!raid5_read_one_chunk(mddev
, raid_bio
))
5298 /* __get_priority_stripe - get the next stripe to process
5300 * Full stripe writes are allowed to pass preread active stripes up until
5301 * the bypass_threshold is exceeded. In general the bypass_count
5302 * increments when the handle_list is handled before the hold_list; however, it
5303 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5304 * stripe with in flight i/o. The bypass_count will be reset when the
5305 * head of the hold_list has changed, i.e. the head was promoted to the
5308 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
, int group
)
5310 struct stripe_head
*sh
, *tmp
;
5311 struct list_head
*handle_list
= NULL
;
5312 struct r5worker_group
*wg
;
5313 bool second_try
= !r5c_is_writeback(conf
->log
) &&
5314 !r5l_log_disk_error(conf
);
5315 bool try_loprio
= test_bit(R5C_LOG_TIGHT
, &conf
->cache_state
) ||
5316 r5l_log_disk_error(conf
);
5321 if (conf
->worker_cnt_per_group
== 0) {
5322 handle_list
= try_loprio
? &conf
->loprio_list
:
5324 } else if (group
!= ANY_GROUP
) {
5325 handle_list
= try_loprio
? &conf
->worker_groups
[group
].loprio_list
:
5326 &conf
->worker_groups
[group
].handle_list
;
5327 wg
= &conf
->worker_groups
[group
];
5330 for (i
= 0; i
< conf
->group_cnt
; i
++) {
5331 handle_list
= try_loprio
? &conf
->worker_groups
[i
].loprio_list
:
5332 &conf
->worker_groups
[i
].handle_list
;
5333 wg
= &conf
->worker_groups
[i
];
5334 if (!list_empty(handle_list
))
5339 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5341 list_empty(handle_list
) ? "empty" : "busy",
5342 list_empty(&conf
->hold_list
) ? "empty" : "busy",
5343 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
5345 if (!list_empty(handle_list
)) {
5346 sh
= list_entry(handle_list
->next
, typeof(*sh
), lru
);
5348 if (list_empty(&conf
->hold_list
))
5349 conf
->bypass_count
= 0;
5350 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
5351 if (conf
->hold_list
.next
== conf
->last_hold
)
5352 conf
->bypass_count
++;
5354 conf
->last_hold
= conf
->hold_list
.next
;
5355 conf
->bypass_count
-= conf
->bypass_threshold
;
5356 if (conf
->bypass_count
< 0)
5357 conf
->bypass_count
= 0;
5360 } else if (!list_empty(&conf
->hold_list
) &&
5361 ((conf
->bypass_threshold
&&
5362 conf
->bypass_count
> conf
->bypass_threshold
) ||
5363 atomic_read(&conf
->pending_full_writes
) == 0)) {
5365 list_for_each_entry(tmp
, &conf
->hold_list
, lru
) {
5366 if (conf
->worker_cnt_per_group
== 0 ||
5367 group
== ANY_GROUP
||
5368 !cpu_online(tmp
->cpu
) ||
5369 cpu_to_group(tmp
->cpu
) == group
) {
5376 conf
->bypass_count
-= conf
->bypass_threshold
;
5377 if (conf
->bypass_count
< 0)
5378 conf
->bypass_count
= 0;
5387 try_loprio
= !try_loprio
;
5395 list_del_init(&sh
->lru
);
5396 BUG_ON(atomic_inc_return(&sh
->count
) != 1);
5400 struct raid5_plug_cb
{
5401 struct blk_plug_cb cb
;
5402 struct list_head list
;
5403 struct list_head temp_inactive_list
[NR_STRIPE_HASH_LOCKS
];
5406 static void raid5_unplug(struct blk_plug_cb
*blk_cb
, bool from_schedule
)
5408 struct raid5_plug_cb
*cb
= container_of(
5409 blk_cb
, struct raid5_plug_cb
, cb
);
5410 struct stripe_head
*sh
;
5411 struct mddev
*mddev
= cb
->cb
.data
;
5412 struct r5conf
*conf
= mddev
->private;
5416 if (cb
->list
.next
&& !list_empty(&cb
->list
)) {
5417 spin_lock_irq(&conf
->device_lock
);
5418 while (!list_empty(&cb
->list
)) {
5419 sh
= list_first_entry(&cb
->list
, struct stripe_head
, lru
);
5420 list_del_init(&sh
->lru
);
5422 * avoid race release_stripe_plug() sees
5423 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5424 * is still in our list
5426 smp_mb__before_atomic();
5427 clear_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
);
5429 * STRIPE_ON_RELEASE_LIST could be set here. In that
5430 * case, the count is always > 1 here
5432 hash
= sh
->hash_lock_index
;
5433 __release_stripe(conf
, sh
, &cb
->temp_inactive_list
[hash
]);
5436 spin_unlock_irq(&conf
->device_lock
);
5438 release_inactive_stripe_list(conf
, cb
->temp_inactive_list
,
5439 NR_STRIPE_HASH_LOCKS
);
5441 trace_block_unplug(mddev
->queue
, cnt
, !from_schedule
);
5445 static void release_stripe_plug(struct mddev
*mddev
,
5446 struct stripe_head
*sh
)
5448 struct blk_plug_cb
*blk_cb
= blk_check_plugged(
5449 raid5_unplug
, mddev
,
5450 sizeof(struct raid5_plug_cb
));
5451 struct raid5_plug_cb
*cb
;
5454 raid5_release_stripe(sh
);
5458 cb
= container_of(blk_cb
, struct raid5_plug_cb
, cb
);
5460 if (cb
->list
.next
== NULL
) {
5462 INIT_LIST_HEAD(&cb
->list
);
5463 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5464 INIT_LIST_HEAD(cb
->temp_inactive_list
+ i
);
5467 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
))
5468 list_add_tail(&sh
->lru
, &cb
->list
);
5470 raid5_release_stripe(sh
);
5473 static void make_discard_request(struct mddev
*mddev
, struct bio
*bi
)
5475 struct r5conf
*conf
= mddev
->private;
5476 sector_t logical_sector
, last_sector
;
5477 struct stripe_head
*sh
;
5480 if (mddev
->reshape_position
!= MaxSector
)
5481 /* Skip discard while reshape is happening */
5484 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5485 last_sector
= bi
->bi_iter
.bi_sector
+ (bi
->bi_iter
.bi_size
>>9);
5489 stripe_sectors
= conf
->chunk_sectors
*
5490 (conf
->raid_disks
- conf
->max_degraded
);
5491 logical_sector
= DIV_ROUND_UP_SECTOR_T(logical_sector
,
5493 sector_div(last_sector
, stripe_sectors
);
5495 logical_sector
*= conf
->chunk_sectors
;
5496 last_sector
*= conf
->chunk_sectors
;
5498 for (; logical_sector
< last_sector
;
5499 logical_sector
+= STRIPE_SECTORS
) {
5503 sh
= raid5_get_active_stripe(conf
, logical_sector
, 0, 0, 0);
5504 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5505 TASK_UNINTERRUPTIBLE
);
5506 set_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5507 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
5508 raid5_release_stripe(sh
);
5512 clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5513 spin_lock_irq(&sh
->stripe_lock
);
5514 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5515 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5517 if (sh
->dev
[d
].towrite
|| sh
->dev
[d
].toread
) {
5518 set_bit(R5_Overlap
, &sh
->dev
[d
].flags
);
5519 spin_unlock_irq(&sh
->stripe_lock
);
5520 raid5_release_stripe(sh
);
5525 set_bit(STRIPE_DISCARD
, &sh
->state
);
5526 finish_wait(&conf
->wait_for_overlap
, &w
);
5527 sh
->overwrite_disks
= 0;
5528 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5529 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5531 sh
->dev
[d
].towrite
= bi
;
5532 set_bit(R5_OVERWRITE
, &sh
->dev
[d
].flags
);
5533 bio_inc_remaining(bi
);
5534 md_write_inc(mddev
, bi
);
5535 sh
->overwrite_disks
++;
5537 spin_unlock_irq(&sh
->stripe_lock
);
5538 if (conf
->mddev
->bitmap
) {
5540 d
< conf
->raid_disks
- conf
->max_degraded
;
5542 bitmap_startwrite(mddev
->bitmap
,
5546 sh
->bm_seq
= conf
->seq_flush
+ 1;
5547 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
5550 set_bit(STRIPE_HANDLE
, &sh
->state
);
5551 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5552 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5553 atomic_inc(&conf
->preread_active_stripes
);
5554 release_stripe_plug(mddev
, sh
);
5560 static bool raid5_make_request(struct mddev
*mddev
, struct bio
* bi
)
5562 struct r5conf
*conf
= mddev
->private;
5564 sector_t new_sector
;
5565 sector_t logical_sector
, last_sector
;
5566 struct stripe_head
*sh
;
5567 const int rw
= bio_data_dir(bi
);
5570 bool do_flush
= false;
5572 if (unlikely(bi
->bi_opf
& REQ_PREFLUSH
)) {
5573 int ret
= log_handle_flush_request(conf
, bi
);
5577 if (ret
== -ENODEV
) {
5578 md_flush_request(mddev
, bi
);
5581 /* ret == -EAGAIN, fallback */
5583 * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5584 * we need to flush journal device
5586 do_flush
= bi
->bi_opf
& REQ_PREFLUSH
;
5589 if (!md_write_start(mddev
, bi
))
5592 * If array is degraded, better not do chunk aligned read because
5593 * later we might have to read it again in order to reconstruct
5594 * data on failed drives.
5596 if (rw
== READ
&& mddev
->degraded
== 0 &&
5597 mddev
->reshape_position
== MaxSector
) {
5598 bi
= chunk_aligned_read(mddev
, bi
);
5603 if (unlikely(bio_op(bi
) == REQ_OP_DISCARD
)) {
5604 make_discard_request(mddev
, bi
);
5605 md_write_end(mddev
);
5609 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5610 last_sector
= bio_end_sector(bi
);
5613 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
5614 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
5620 seq
= read_seqcount_begin(&conf
->gen_lock
);
5623 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5624 TASK_UNINTERRUPTIBLE
);
5625 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
5626 /* spinlock is needed as reshape_progress may be
5627 * 64bit on a 32bit platform, and so it might be
5628 * possible to see a half-updated value
5629 * Of course reshape_progress could change after
5630 * the lock is dropped, so once we get a reference
5631 * to the stripe that we think it is, we will have
5634 spin_lock_irq(&conf
->device_lock
);
5635 if (mddev
->reshape_backwards
5636 ? logical_sector
< conf
->reshape_progress
5637 : logical_sector
>= conf
->reshape_progress
) {
5640 if (mddev
->reshape_backwards
5641 ? logical_sector
< conf
->reshape_safe
5642 : logical_sector
>= conf
->reshape_safe
) {
5643 spin_unlock_irq(&conf
->device_lock
);
5649 spin_unlock_irq(&conf
->device_lock
);
5652 new_sector
= raid5_compute_sector(conf
, logical_sector
,
5655 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5656 (unsigned long long)new_sector
,
5657 (unsigned long long)logical_sector
);
5659 sh
= raid5_get_active_stripe(conf
, new_sector
, previous
,
5660 (bi
->bi_opf
& REQ_RAHEAD
), 0);
5662 if (unlikely(previous
)) {
5663 /* expansion might have moved on while waiting for a
5664 * stripe, so we must do the range check again.
5665 * Expansion could still move past after this
5666 * test, but as we are holding a reference to
5667 * 'sh', we know that if that happens,
5668 * STRIPE_EXPANDING will get set and the expansion
5669 * won't proceed until we finish with the stripe.
5672 spin_lock_irq(&conf
->device_lock
);
5673 if (mddev
->reshape_backwards
5674 ? logical_sector
>= conf
->reshape_progress
5675 : logical_sector
< conf
->reshape_progress
)
5676 /* mismatch, need to try again */
5678 spin_unlock_irq(&conf
->device_lock
);
5680 raid5_release_stripe(sh
);
5686 if (read_seqcount_retry(&conf
->gen_lock
, seq
)) {
5687 /* Might have got the wrong stripe_head
5690 raid5_release_stripe(sh
);
5694 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
5695 !add_stripe_bio(sh
, bi
, dd_idx
, rw
, previous
)) {
5696 /* Stripe is busy expanding or
5697 * add failed due to overlap. Flush everything
5700 md_wakeup_thread(mddev
->thread
);
5701 raid5_release_stripe(sh
);
5707 set_bit(STRIPE_R5C_PREFLUSH
, &sh
->state
);
5708 /* we only need flush for one stripe */
5712 set_bit(STRIPE_HANDLE
, &sh
->state
);
5713 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5714 if ((!sh
->batch_head
|| sh
== sh
->batch_head
) &&
5715 (bi
->bi_opf
& REQ_SYNC
) &&
5716 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5717 atomic_inc(&conf
->preread_active_stripes
);
5718 release_stripe_plug(mddev
, sh
);
5720 /* cannot get stripe for read-ahead, just give-up */
5721 bi
->bi_status
= BLK_STS_IOERR
;
5725 finish_wait(&conf
->wait_for_overlap
, &w
);
5728 md_write_end(mddev
);
5733 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
5735 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
5737 /* reshaping is quite different to recovery/resync so it is
5738 * handled quite separately ... here.
5740 * On each call to sync_request, we gather one chunk worth of
5741 * destination stripes and flag them as expanding.
5742 * Then we find all the source stripes and request reads.
5743 * As the reads complete, handle_stripe will copy the data
5744 * into the destination stripe and release that stripe.
5746 struct r5conf
*conf
= mddev
->private;
5747 struct stripe_head
*sh
;
5748 struct md_rdev
*rdev
;
5749 sector_t first_sector
, last_sector
;
5750 int raid_disks
= conf
->previous_raid_disks
;
5751 int data_disks
= raid_disks
- conf
->max_degraded
;
5752 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5755 sector_t writepos
, readpos
, safepos
;
5756 sector_t stripe_addr
;
5757 int reshape_sectors
;
5758 struct list_head stripes
;
5761 if (sector_nr
== 0) {
5762 /* If restarting in the middle, skip the initial sectors */
5763 if (mddev
->reshape_backwards
&&
5764 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
5765 sector_nr
= raid5_size(mddev
, 0, 0)
5766 - conf
->reshape_progress
;
5767 } else if (mddev
->reshape_backwards
&&
5768 conf
->reshape_progress
== MaxSector
) {
5769 /* shouldn't happen, but just in case, finish up.*/
5770 sector_nr
= MaxSector
;
5771 } else if (!mddev
->reshape_backwards
&&
5772 conf
->reshape_progress
> 0)
5773 sector_nr
= conf
->reshape_progress
;
5774 sector_div(sector_nr
, new_data_disks
);
5776 mddev
->curr_resync_completed
= sector_nr
;
5777 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5784 /* We need to process a full chunk at a time.
5785 * If old and new chunk sizes differ, we need to process the
5789 reshape_sectors
= max(conf
->chunk_sectors
, conf
->prev_chunk_sectors
);
5791 /* We update the metadata at least every 10 seconds, or when
5792 * the data about to be copied would over-write the source of
5793 * the data at the front of the range. i.e. one new_stripe
5794 * along from reshape_progress new_maps to after where
5795 * reshape_safe old_maps to
5797 writepos
= conf
->reshape_progress
;
5798 sector_div(writepos
, new_data_disks
);
5799 readpos
= conf
->reshape_progress
;
5800 sector_div(readpos
, data_disks
);
5801 safepos
= conf
->reshape_safe
;
5802 sector_div(safepos
, data_disks
);
5803 if (mddev
->reshape_backwards
) {
5804 BUG_ON(writepos
< reshape_sectors
);
5805 writepos
-= reshape_sectors
;
5806 readpos
+= reshape_sectors
;
5807 safepos
+= reshape_sectors
;
5809 writepos
+= reshape_sectors
;
5810 /* readpos and safepos are worst-case calculations.
5811 * A negative number is overly pessimistic, and causes
5812 * obvious problems for unsigned storage. So clip to 0.
5814 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
5815 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
5818 /* Having calculated the 'writepos' possibly use it
5819 * to set 'stripe_addr' which is where we will write to.
5821 if (mddev
->reshape_backwards
) {
5822 BUG_ON(conf
->reshape_progress
== 0);
5823 stripe_addr
= writepos
;
5824 BUG_ON((mddev
->dev_sectors
&
5825 ~((sector_t
)reshape_sectors
- 1))
5826 - reshape_sectors
- stripe_addr
5829 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
5830 stripe_addr
= sector_nr
;
5833 /* 'writepos' is the most advanced device address we might write.
5834 * 'readpos' is the least advanced device address we might read.
5835 * 'safepos' is the least address recorded in the metadata as having
5837 * If there is a min_offset_diff, these are adjusted either by
5838 * increasing the safepos/readpos if diff is negative, or
5839 * increasing writepos if diff is positive.
5840 * If 'readpos' is then behind 'writepos', there is no way that we can
5841 * ensure safety in the face of a crash - that must be done by userspace
5842 * making a backup of the data. So in that case there is no particular
5843 * rush to update metadata.
5844 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5845 * update the metadata to advance 'safepos' to match 'readpos' so that
5846 * we can be safe in the event of a crash.
5847 * So we insist on updating metadata if safepos is behind writepos and
5848 * readpos is beyond writepos.
5849 * In any case, update the metadata every 10 seconds.
5850 * Maybe that number should be configurable, but I'm not sure it is
5851 * worth it.... maybe it could be a multiple of safemode_delay???
5853 if (conf
->min_offset_diff
< 0) {
5854 safepos
+= -conf
->min_offset_diff
;
5855 readpos
+= -conf
->min_offset_diff
;
5857 writepos
+= conf
->min_offset_diff
;
5859 if ((mddev
->reshape_backwards
5860 ? (safepos
> writepos
&& readpos
< writepos
)
5861 : (safepos
< writepos
&& readpos
> writepos
)) ||
5862 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
5863 /* Cannot proceed until we've updated the superblock... */
5864 wait_event(conf
->wait_for_overlap
,
5865 atomic_read(&conf
->reshape_stripes
)==0
5866 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5867 if (atomic_read(&conf
->reshape_stripes
) != 0)
5869 mddev
->reshape_position
= conf
->reshape_progress
;
5870 mddev
->curr_resync_completed
= sector_nr
;
5871 if (!mddev
->reshape_backwards
)
5872 /* Can update recovery_offset */
5873 rdev_for_each(rdev
, mddev
)
5874 if (rdev
->raid_disk
>= 0 &&
5875 !test_bit(Journal
, &rdev
->flags
) &&
5876 !test_bit(In_sync
, &rdev
->flags
) &&
5877 rdev
->recovery_offset
< sector_nr
)
5878 rdev
->recovery_offset
= sector_nr
;
5880 conf
->reshape_checkpoint
= jiffies
;
5881 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
5882 md_wakeup_thread(mddev
->thread
);
5883 wait_event(mddev
->sb_wait
, mddev
->sb_flags
== 0 ||
5884 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5885 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5887 spin_lock_irq(&conf
->device_lock
);
5888 conf
->reshape_safe
= mddev
->reshape_position
;
5889 spin_unlock_irq(&conf
->device_lock
);
5890 wake_up(&conf
->wait_for_overlap
);
5891 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5894 INIT_LIST_HEAD(&stripes
);
5895 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
5897 int skipped_disk
= 0;
5898 sh
= raid5_get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
5899 set_bit(STRIPE_EXPANDING
, &sh
->state
);
5900 atomic_inc(&conf
->reshape_stripes
);
5901 /* If any of this stripe is beyond the end of the old
5902 * array, then we need to zero those blocks
5904 for (j
=sh
->disks
; j
--;) {
5906 if (j
== sh
->pd_idx
)
5908 if (conf
->level
== 6 &&
5911 s
= raid5_compute_blocknr(sh
, j
, 0);
5912 if (s
< raid5_size(mddev
, 0, 0)) {
5916 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
5917 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
5918 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
5920 if (!skipped_disk
) {
5921 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
5922 set_bit(STRIPE_HANDLE
, &sh
->state
);
5924 list_add(&sh
->lru
, &stripes
);
5926 spin_lock_irq(&conf
->device_lock
);
5927 if (mddev
->reshape_backwards
)
5928 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
5930 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
5931 spin_unlock_irq(&conf
->device_lock
);
5932 /* Ok, those stripe are ready. We can start scheduling
5933 * reads on the source stripes.
5934 * The source stripes are determined by mapping the first and last
5935 * block on the destination stripes.
5938 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
5941 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
5942 * new_data_disks
- 1),
5944 if (last_sector
>= mddev
->dev_sectors
)
5945 last_sector
= mddev
->dev_sectors
- 1;
5946 while (first_sector
<= last_sector
) {
5947 sh
= raid5_get_active_stripe(conf
, first_sector
, 1, 0, 1);
5948 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
5949 set_bit(STRIPE_HANDLE
, &sh
->state
);
5950 raid5_release_stripe(sh
);
5951 first_sector
+= STRIPE_SECTORS
;
5953 /* Now that the sources are clearly marked, we can release
5954 * the destination stripes
5956 while (!list_empty(&stripes
)) {
5957 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
5958 list_del_init(&sh
->lru
);
5959 raid5_release_stripe(sh
);
5961 /* If this takes us to the resync_max point where we have to pause,
5962 * then we need to write out the superblock.
5964 sector_nr
+= reshape_sectors
;
5965 retn
= reshape_sectors
;
5967 if (mddev
->curr_resync_completed
> mddev
->resync_max
||
5968 (sector_nr
- mddev
->curr_resync_completed
) * 2
5969 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
5970 /* Cannot proceed until we've updated the superblock... */
5971 wait_event(conf
->wait_for_overlap
,
5972 atomic_read(&conf
->reshape_stripes
) == 0
5973 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5974 if (atomic_read(&conf
->reshape_stripes
) != 0)
5976 mddev
->reshape_position
= conf
->reshape_progress
;
5977 mddev
->curr_resync_completed
= sector_nr
;
5978 if (!mddev
->reshape_backwards
)
5979 /* Can update recovery_offset */
5980 rdev_for_each(rdev
, mddev
)
5981 if (rdev
->raid_disk
>= 0 &&
5982 !test_bit(Journal
, &rdev
->flags
) &&
5983 !test_bit(In_sync
, &rdev
->flags
) &&
5984 rdev
->recovery_offset
< sector_nr
)
5985 rdev
->recovery_offset
= sector_nr
;
5986 conf
->reshape_checkpoint
= jiffies
;
5987 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
5988 md_wakeup_thread(mddev
->thread
);
5989 wait_event(mddev
->sb_wait
,
5990 !test_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
)
5991 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5992 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5994 spin_lock_irq(&conf
->device_lock
);
5995 conf
->reshape_safe
= mddev
->reshape_position
;
5996 spin_unlock_irq(&conf
->device_lock
);
5997 wake_up(&conf
->wait_for_overlap
);
5998 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
6004 static inline sector_t
raid5_sync_request(struct mddev
*mddev
, sector_t sector_nr
,
6007 struct r5conf
*conf
= mddev
->private;
6008 struct stripe_head
*sh
;
6009 sector_t max_sector
= mddev
->dev_sectors
;
6010 sector_t sync_blocks
;
6011 int still_degraded
= 0;
6014 if (sector_nr
>= max_sector
) {
6015 /* just being told to finish up .. nothing much to do */
6017 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
6022 if (mddev
->curr_resync
< max_sector
) /* aborted */
6023 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
6025 else /* completed sync */
6027 bitmap_close_sync(mddev
->bitmap
);
6032 /* Allow raid5_quiesce to complete */
6033 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
6035 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
6036 return reshape_request(mddev
, sector_nr
, skipped
);
6038 /* No need to check resync_max as we never do more than one
6039 * stripe, and as resync_max will always be on a chunk boundary,
6040 * if the check in md_do_sync didn't fire, there is no chance
6041 * of overstepping resync_max here
6044 /* if there is too many failed drives and we are trying
6045 * to resync, then assert that we are finished, because there is
6046 * nothing we can do.
6048 if (mddev
->degraded
>= conf
->max_degraded
&&
6049 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
6050 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
6054 if (!test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
6056 !bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
6057 sync_blocks
>= STRIPE_SECTORS
) {
6058 /* we can skip this block, and probably more */
6059 sync_blocks
/= STRIPE_SECTORS
;
6061 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
6064 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
, false);
6066 sh
= raid5_get_active_stripe(conf
, sector_nr
, 0, 1, 0);
6068 sh
= raid5_get_active_stripe(conf
, sector_nr
, 0, 0, 0);
6069 /* make sure we don't swamp the stripe cache if someone else
6070 * is trying to get access
6072 schedule_timeout_uninterruptible(1);
6074 /* Need to check if array will still be degraded after recovery/resync
6075 * Note in case of > 1 drive failures it's possible we're rebuilding
6076 * one drive while leaving another faulty drive in array.
6079 for (i
= 0; i
< conf
->raid_disks
; i
++) {
6080 struct md_rdev
*rdev
= READ_ONCE(conf
->disks
[i
].rdev
);
6082 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
))
6087 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
6089 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
6090 set_bit(STRIPE_HANDLE
, &sh
->state
);
6092 raid5_release_stripe(sh
);
6094 return STRIPE_SECTORS
;
6097 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
,
6098 unsigned int offset
)
6100 /* We may not be able to submit a whole bio at once as there
6101 * may not be enough stripe_heads available.
6102 * We cannot pre-allocate enough stripe_heads as we may need
6103 * more than exist in the cache (if we allow ever large chunks).
6104 * So we do one stripe head at a time and record in
6105 * ->bi_hw_segments how many have been done.
6107 * We *know* that this entire raid_bio is in one chunk, so
6108 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6110 struct stripe_head
*sh
;
6112 sector_t sector
, logical_sector
, last_sector
;
6116 logical_sector
= raid_bio
->bi_iter
.bi_sector
&
6117 ~((sector_t
)STRIPE_SECTORS
-1);
6118 sector
= raid5_compute_sector(conf
, logical_sector
,
6120 last_sector
= bio_end_sector(raid_bio
);
6122 for (; logical_sector
< last_sector
;
6123 logical_sector
+= STRIPE_SECTORS
,
6124 sector
+= STRIPE_SECTORS
,
6128 /* already done this stripe */
6131 sh
= raid5_get_active_stripe(conf
, sector
, 0, 1, 1);
6134 /* failed to get a stripe - must wait */
6135 conf
->retry_read_aligned
= raid_bio
;
6136 conf
->retry_read_offset
= scnt
;
6140 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0, 0)) {
6141 raid5_release_stripe(sh
);
6142 conf
->retry_read_aligned
= raid_bio
;
6143 conf
->retry_read_offset
= scnt
;
6147 set_bit(R5_ReadNoMerge
, &sh
->dev
[dd_idx
].flags
);
6149 raid5_release_stripe(sh
);
6153 bio_endio(raid_bio
);
6155 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
6156 wake_up(&conf
->wait_for_quiescent
);
6160 static int handle_active_stripes(struct r5conf
*conf
, int group
,
6161 struct r5worker
*worker
,
6162 struct list_head
*temp_inactive_list
)
6164 struct stripe_head
*batch
[MAX_STRIPE_BATCH
], *sh
;
6165 int i
, batch_size
= 0, hash
;
6166 bool release_inactive
= false;
6168 while (batch_size
< MAX_STRIPE_BATCH
&&
6169 (sh
= __get_priority_stripe(conf
, group
)) != NULL
)
6170 batch
[batch_size
++] = sh
;
6172 if (batch_size
== 0) {
6173 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6174 if (!list_empty(temp_inactive_list
+ i
))
6176 if (i
== NR_STRIPE_HASH_LOCKS
) {
6177 spin_unlock_irq(&conf
->device_lock
);
6178 log_flush_stripe_to_raid(conf
);
6179 spin_lock_irq(&conf
->device_lock
);
6182 release_inactive
= true;
6184 spin_unlock_irq(&conf
->device_lock
);
6186 release_inactive_stripe_list(conf
, temp_inactive_list
,
6187 NR_STRIPE_HASH_LOCKS
);
6189 r5l_flush_stripe_to_raid(conf
->log
);
6190 if (release_inactive
) {
6191 spin_lock_irq(&conf
->device_lock
);
6195 for (i
= 0; i
< batch_size
; i
++)
6196 handle_stripe(batch
[i
]);
6197 log_write_stripe_run(conf
);
6201 spin_lock_irq(&conf
->device_lock
);
6202 for (i
= 0; i
< batch_size
; i
++) {
6203 hash
= batch
[i
]->hash_lock_index
;
6204 __release_stripe(conf
, batch
[i
], &temp_inactive_list
[hash
]);
6209 static void raid5_do_work(struct work_struct
*work
)
6211 struct r5worker
*worker
= container_of(work
, struct r5worker
, work
);
6212 struct r5worker_group
*group
= worker
->group
;
6213 struct r5conf
*conf
= group
->conf
;
6214 struct mddev
*mddev
= conf
->mddev
;
6215 int group_id
= group
- conf
->worker_groups
;
6217 struct blk_plug plug
;
6219 pr_debug("+++ raid5worker active\n");
6221 blk_start_plug(&plug
);
6223 spin_lock_irq(&conf
->device_lock
);
6225 int batch_size
, released
;
6227 released
= release_stripe_list(conf
, worker
->temp_inactive_list
);
6229 batch_size
= handle_active_stripes(conf
, group_id
, worker
,
6230 worker
->temp_inactive_list
);
6231 worker
->working
= false;
6232 if (!batch_size
&& !released
)
6234 handled
+= batch_size
;
6235 wait_event_lock_irq(mddev
->sb_wait
,
6236 !test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
),
6239 pr_debug("%d stripes handled\n", handled
);
6241 spin_unlock_irq(&conf
->device_lock
);
6243 flush_deferred_bios(conf
);
6245 r5l_flush_stripe_to_raid(conf
->log
);
6247 async_tx_issue_pending_all();
6248 blk_finish_plug(&plug
);
6250 pr_debug("--- raid5worker inactive\n");
6254 * This is our raid5 kernel thread.
6256 * We scan the hash table for stripes which can be handled now.
6257 * During the scan, completed stripes are saved for us by the interrupt
6258 * handler, so that they will not have to wait for our next wakeup.
6260 static void raid5d(struct md_thread
*thread
)
6262 struct mddev
*mddev
= thread
->mddev
;
6263 struct r5conf
*conf
= mddev
->private;
6265 struct blk_plug plug
;
6267 pr_debug("+++ raid5d active\n");
6269 md_check_recovery(mddev
);
6271 blk_start_plug(&plug
);
6273 spin_lock_irq(&conf
->device_lock
);
6276 int batch_size
, released
;
6277 unsigned int offset
;
6279 released
= release_stripe_list(conf
, conf
->temp_inactive_list
);
6281 clear_bit(R5_DID_ALLOC
, &conf
->cache_state
);
6284 !list_empty(&conf
->bitmap_list
)) {
6285 /* Now is a good time to flush some bitmap updates */
6287 spin_unlock_irq(&conf
->device_lock
);
6288 bitmap_unplug(mddev
->bitmap
);
6289 spin_lock_irq(&conf
->device_lock
);
6290 conf
->seq_write
= conf
->seq_flush
;
6291 activate_bit_delay(conf
, conf
->temp_inactive_list
);
6293 raid5_activate_delayed(conf
);
6295 while ((bio
= remove_bio_from_retry(conf
, &offset
))) {
6297 spin_unlock_irq(&conf
->device_lock
);
6298 ok
= retry_aligned_read(conf
, bio
, offset
);
6299 spin_lock_irq(&conf
->device_lock
);
6305 batch_size
= handle_active_stripes(conf
, ANY_GROUP
, NULL
,
6306 conf
->temp_inactive_list
);
6307 if (!batch_size
&& !released
)
6309 handled
+= batch_size
;
6311 if (mddev
->sb_flags
& ~(1 << MD_SB_CHANGE_PENDING
)) {
6312 spin_unlock_irq(&conf
->device_lock
);
6313 md_check_recovery(mddev
);
6314 spin_lock_irq(&conf
->device_lock
);
6317 pr_debug("%d stripes handled\n", handled
);
6319 spin_unlock_irq(&conf
->device_lock
);
6320 if (test_and_clear_bit(R5_ALLOC_MORE
, &conf
->cache_state
) &&
6321 mutex_trylock(&conf
->cache_size_mutex
)) {
6322 grow_one_stripe(conf
, __GFP_NOWARN
);
6323 /* Set flag even if allocation failed. This helps
6324 * slow down allocation requests when mem is short
6326 set_bit(R5_DID_ALLOC
, &conf
->cache_state
);
6327 mutex_unlock(&conf
->cache_size_mutex
);
6330 flush_deferred_bios(conf
);
6332 r5l_flush_stripe_to_raid(conf
->log
);
6334 async_tx_issue_pending_all();
6335 blk_finish_plug(&plug
);
6337 pr_debug("--- raid5d inactive\n");
6341 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
6343 struct r5conf
*conf
;
6345 spin_lock(&mddev
->lock
);
6346 conf
= mddev
->private;
6348 ret
= sprintf(page
, "%d\n", conf
->min_nr_stripes
);
6349 spin_unlock(&mddev
->lock
);
6354 raid5_set_cache_size(struct mddev
*mddev
, int size
)
6356 struct r5conf
*conf
= mddev
->private;
6358 if (size
<= 16 || size
> 32768)
6361 conf
->min_nr_stripes
= size
;
6362 mutex_lock(&conf
->cache_size_mutex
);
6363 while (size
< conf
->max_nr_stripes
&&
6364 drop_one_stripe(conf
))
6366 mutex_unlock(&conf
->cache_size_mutex
);
6368 md_allow_write(mddev
);
6370 mutex_lock(&conf
->cache_size_mutex
);
6371 while (size
> conf
->max_nr_stripes
)
6372 if (!grow_one_stripe(conf
, GFP_KERNEL
))
6374 mutex_unlock(&conf
->cache_size_mutex
);
6378 EXPORT_SYMBOL(raid5_set_cache_size
);
6381 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
6383 struct r5conf
*conf
;
6387 if (len
>= PAGE_SIZE
)
6389 if (kstrtoul(page
, 10, &new))
6391 err
= mddev_lock(mddev
);
6394 conf
= mddev
->private;
6398 err
= raid5_set_cache_size(mddev
, new);
6399 mddev_unlock(mddev
);
6404 static struct md_sysfs_entry
6405 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
6406 raid5_show_stripe_cache_size
,
6407 raid5_store_stripe_cache_size
);
6410 raid5_show_rmw_level(struct mddev
*mddev
, char *page
)
6412 struct r5conf
*conf
= mddev
->private;
6414 return sprintf(page
, "%d\n", conf
->rmw_level
);
6420 raid5_store_rmw_level(struct mddev
*mddev
, const char *page
, size_t len
)
6422 struct r5conf
*conf
= mddev
->private;
6428 if (len
>= PAGE_SIZE
)
6431 if (kstrtoul(page
, 10, &new))
6434 if (new != PARITY_DISABLE_RMW
&& !raid6_call
.xor_syndrome
)
6437 if (new != PARITY_DISABLE_RMW
&&
6438 new != PARITY_ENABLE_RMW
&&
6439 new != PARITY_PREFER_RMW
)
6442 conf
->rmw_level
= new;
6446 static struct md_sysfs_entry
6447 raid5_rmw_level
= __ATTR(rmw_level
, S_IRUGO
| S_IWUSR
,
6448 raid5_show_rmw_level
,
6449 raid5_store_rmw_level
);
6453 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
6455 struct r5conf
*conf
;
6457 spin_lock(&mddev
->lock
);
6458 conf
= mddev
->private;
6460 ret
= sprintf(page
, "%d\n", conf
->bypass_threshold
);
6461 spin_unlock(&mddev
->lock
);
6466 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
6468 struct r5conf
*conf
;
6472 if (len
>= PAGE_SIZE
)
6474 if (kstrtoul(page
, 10, &new))
6477 err
= mddev_lock(mddev
);
6480 conf
= mddev
->private;
6483 else if (new > conf
->min_nr_stripes
)
6486 conf
->bypass_threshold
= new;
6487 mddev_unlock(mddev
);
6491 static struct md_sysfs_entry
6492 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
6494 raid5_show_preread_threshold
,
6495 raid5_store_preread_threshold
);
6498 raid5_show_skip_copy(struct mddev
*mddev
, char *page
)
6500 struct r5conf
*conf
;
6502 spin_lock(&mddev
->lock
);
6503 conf
= mddev
->private;
6505 ret
= sprintf(page
, "%d\n", conf
->skip_copy
);
6506 spin_unlock(&mddev
->lock
);
6511 raid5_store_skip_copy(struct mddev
*mddev
, const char *page
, size_t len
)
6513 struct r5conf
*conf
;
6517 if (len
>= PAGE_SIZE
)
6519 if (kstrtoul(page
, 10, &new))
6523 err
= mddev_lock(mddev
);
6526 conf
= mddev
->private;
6529 else if (new != conf
->skip_copy
) {
6530 mddev_suspend(mddev
);
6531 conf
->skip_copy
= new;
6533 mddev
->queue
->backing_dev_info
->capabilities
|=
6534 BDI_CAP_STABLE_WRITES
;
6536 mddev
->queue
->backing_dev_info
->capabilities
&=
6537 ~BDI_CAP_STABLE_WRITES
;
6538 mddev_resume(mddev
);
6540 mddev_unlock(mddev
);
6544 static struct md_sysfs_entry
6545 raid5_skip_copy
= __ATTR(skip_copy
, S_IRUGO
| S_IWUSR
,
6546 raid5_show_skip_copy
,
6547 raid5_store_skip_copy
);
6550 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
6552 struct r5conf
*conf
= mddev
->private;
6554 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
6559 static struct md_sysfs_entry
6560 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
6563 raid5_show_group_thread_cnt(struct mddev
*mddev
, char *page
)
6565 struct r5conf
*conf
;
6567 spin_lock(&mddev
->lock
);
6568 conf
= mddev
->private;
6570 ret
= sprintf(page
, "%d\n", conf
->worker_cnt_per_group
);
6571 spin_unlock(&mddev
->lock
);
6575 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6577 int *worker_cnt_per_group
,
6578 struct r5worker_group
**worker_groups
);
6580 raid5_store_group_thread_cnt(struct mddev
*mddev
, const char *page
, size_t len
)
6582 struct r5conf
*conf
;
6585 struct r5worker_group
*new_groups
, *old_groups
;
6586 int group_cnt
, worker_cnt_per_group
;
6588 if (len
>= PAGE_SIZE
)
6590 if (kstrtouint(page
, 10, &new))
6592 /* 8192 should be big enough */
6596 err
= mddev_lock(mddev
);
6599 conf
= mddev
->private;
6602 else if (new != conf
->worker_cnt_per_group
) {
6603 mddev_suspend(mddev
);
6605 old_groups
= conf
->worker_groups
;
6607 flush_workqueue(raid5_wq
);
6609 err
= alloc_thread_groups(conf
, new,
6610 &group_cnt
, &worker_cnt_per_group
,
6613 spin_lock_irq(&conf
->device_lock
);
6614 conf
->group_cnt
= group_cnt
;
6615 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6616 conf
->worker_groups
= new_groups
;
6617 spin_unlock_irq(&conf
->device_lock
);
6620 kfree(old_groups
[0].workers
);
6623 mddev_resume(mddev
);
6625 mddev_unlock(mddev
);
6630 static struct md_sysfs_entry
6631 raid5_group_thread_cnt
= __ATTR(group_thread_cnt
, S_IRUGO
| S_IWUSR
,
6632 raid5_show_group_thread_cnt
,
6633 raid5_store_group_thread_cnt
);
6635 static struct attribute
*raid5_attrs
[] = {
6636 &raid5_stripecache_size
.attr
,
6637 &raid5_stripecache_active
.attr
,
6638 &raid5_preread_bypass_threshold
.attr
,
6639 &raid5_group_thread_cnt
.attr
,
6640 &raid5_skip_copy
.attr
,
6641 &raid5_rmw_level
.attr
,
6642 &r5c_journal_mode
.attr
,
6645 static struct attribute_group raid5_attrs_group
= {
6647 .attrs
= raid5_attrs
,
6650 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6652 int *worker_cnt_per_group
,
6653 struct r5worker_group
**worker_groups
)
6657 struct r5worker
*workers
;
6659 *worker_cnt_per_group
= cnt
;
6662 *worker_groups
= NULL
;
6665 *group_cnt
= num_possible_nodes();
6666 size
= sizeof(struct r5worker
) * cnt
;
6667 workers
= kzalloc(size
* *group_cnt
, GFP_NOIO
);
6668 *worker_groups
= kzalloc(sizeof(struct r5worker_group
) *
6669 *group_cnt
, GFP_NOIO
);
6670 if (!*worker_groups
|| !workers
) {
6672 kfree(*worker_groups
);
6676 for (i
= 0; i
< *group_cnt
; i
++) {
6677 struct r5worker_group
*group
;
6679 group
= &(*worker_groups
)[i
];
6680 INIT_LIST_HEAD(&group
->handle_list
);
6681 INIT_LIST_HEAD(&group
->loprio_list
);
6683 group
->workers
= workers
+ i
* cnt
;
6685 for (j
= 0; j
< cnt
; j
++) {
6686 struct r5worker
*worker
= group
->workers
+ j
;
6687 worker
->group
= group
;
6688 INIT_WORK(&worker
->work
, raid5_do_work
);
6690 for (k
= 0; k
< NR_STRIPE_HASH_LOCKS
; k
++)
6691 INIT_LIST_HEAD(worker
->temp_inactive_list
+ k
);
6698 static void free_thread_groups(struct r5conf
*conf
)
6700 if (conf
->worker_groups
)
6701 kfree(conf
->worker_groups
[0].workers
);
6702 kfree(conf
->worker_groups
);
6703 conf
->worker_groups
= NULL
;
6707 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
6709 struct r5conf
*conf
= mddev
->private;
6712 sectors
= mddev
->dev_sectors
;
6714 /* size is defined by the smallest of previous and new size */
6715 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
6717 sectors
&= ~((sector_t
)conf
->chunk_sectors
- 1);
6718 sectors
&= ~((sector_t
)conf
->prev_chunk_sectors
- 1);
6719 return sectors
* (raid_disks
- conf
->max_degraded
);
6722 static void free_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6724 safe_put_page(percpu
->spare_page
);
6725 if (percpu
->scribble
)
6726 flex_array_free(percpu
->scribble
);
6727 percpu
->spare_page
= NULL
;
6728 percpu
->scribble
= NULL
;
6731 static int alloc_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6733 if (conf
->level
== 6 && !percpu
->spare_page
)
6734 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
6735 if (!percpu
->scribble
)
6736 percpu
->scribble
= scribble_alloc(max(conf
->raid_disks
,
6737 conf
->previous_raid_disks
),
6738 max(conf
->chunk_sectors
,
6739 conf
->prev_chunk_sectors
)
6743 if (!percpu
->scribble
|| (conf
->level
== 6 && !percpu
->spare_page
)) {
6744 free_scratch_buffer(conf
, percpu
);
6751 static int raid456_cpu_dead(unsigned int cpu
, struct hlist_node
*node
)
6753 struct r5conf
*conf
= hlist_entry_safe(node
, struct r5conf
, node
);
6755 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6759 static void raid5_free_percpu(struct r5conf
*conf
)
6764 cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE
, &conf
->node
);
6765 free_percpu(conf
->percpu
);
6768 static void free_conf(struct r5conf
*conf
)
6774 unregister_shrinker(&conf
->shrinker
);
6775 free_thread_groups(conf
);
6776 shrink_stripes(conf
);
6777 raid5_free_percpu(conf
);
6778 for (i
= 0; i
< conf
->pool_size
; i
++)
6779 if (conf
->disks
[i
].extra_page
)
6780 put_page(conf
->disks
[i
].extra_page
);
6782 bioset_exit(&conf
->bio_split
);
6783 kfree(conf
->stripe_hashtbl
);
6784 kfree(conf
->pending_data
);
6788 static int raid456_cpu_up_prepare(unsigned int cpu
, struct hlist_node
*node
)
6790 struct r5conf
*conf
= hlist_entry_safe(node
, struct r5conf
, node
);
6791 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
6793 if (alloc_scratch_buffer(conf
, percpu
)) {
6794 pr_warn("%s: failed memory allocation for cpu%u\n",
6801 static int raid5_alloc_percpu(struct r5conf
*conf
)
6805 conf
->percpu
= alloc_percpu(struct raid5_percpu
);
6809 err
= cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE
, &conf
->node
);
6811 conf
->scribble_disks
= max(conf
->raid_disks
,
6812 conf
->previous_raid_disks
);
6813 conf
->scribble_sectors
= max(conf
->chunk_sectors
,
6814 conf
->prev_chunk_sectors
);
6819 static unsigned long raid5_cache_scan(struct shrinker
*shrink
,
6820 struct shrink_control
*sc
)
6822 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
6823 unsigned long ret
= SHRINK_STOP
;
6825 if (mutex_trylock(&conf
->cache_size_mutex
)) {
6827 while (ret
< sc
->nr_to_scan
&&
6828 conf
->max_nr_stripes
> conf
->min_nr_stripes
) {
6829 if (drop_one_stripe(conf
) == 0) {
6835 mutex_unlock(&conf
->cache_size_mutex
);
6840 static unsigned long raid5_cache_count(struct shrinker
*shrink
,
6841 struct shrink_control
*sc
)
6843 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
6845 if (conf
->max_nr_stripes
< conf
->min_nr_stripes
)
6846 /* unlikely, but not impossible */
6848 return conf
->max_nr_stripes
- conf
->min_nr_stripes
;
6851 static struct r5conf
*setup_conf(struct mddev
*mddev
)
6853 struct r5conf
*conf
;
6854 int raid_disk
, memory
, max_disks
;
6855 struct md_rdev
*rdev
;
6856 struct disk_info
*disk
;
6859 int group_cnt
, worker_cnt_per_group
;
6860 struct r5worker_group
*new_group
;
6863 if (mddev
->new_level
!= 5
6864 && mddev
->new_level
!= 4
6865 && mddev
->new_level
!= 6) {
6866 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6867 mdname(mddev
), mddev
->new_level
);
6868 return ERR_PTR(-EIO
);
6870 if ((mddev
->new_level
== 5
6871 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
6872 (mddev
->new_level
== 6
6873 && !algorithm_valid_raid6(mddev
->new_layout
))) {
6874 pr_warn("md/raid:%s: layout %d not supported\n",
6875 mdname(mddev
), mddev
->new_layout
);
6876 return ERR_PTR(-EIO
);
6878 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
6879 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6880 mdname(mddev
), mddev
->raid_disks
);
6881 return ERR_PTR(-EINVAL
);
6884 if (!mddev
->new_chunk_sectors
||
6885 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
6886 !is_power_of_2(mddev
->new_chunk_sectors
)) {
6887 pr_warn("md/raid:%s: invalid chunk size %d\n",
6888 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
6889 return ERR_PTR(-EINVAL
);
6892 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
6895 INIT_LIST_HEAD(&conf
->free_list
);
6896 INIT_LIST_HEAD(&conf
->pending_list
);
6897 conf
->pending_data
= kzalloc(sizeof(struct r5pending_data
) *
6898 PENDING_IO_MAX
, GFP_KERNEL
);
6899 if (!conf
->pending_data
)
6901 for (i
= 0; i
< PENDING_IO_MAX
; i
++)
6902 list_add(&conf
->pending_data
[i
].sibling
, &conf
->free_list
);
6903 /* Don't enable multi-threading by default*/
6904 if (!alloc_thread_groups(conf
, 0, &group_cnt
, &worker_cnt_per_group
,
6906 conf
->group_cnt
= group_cnt
;
6907 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6908 conf
->worker_groups
= new_group
;
6911 spin_lock_init(&conf
->device_lock
);
6912 seqcount_init(&conf
->gen_lock
);
6913 mutex_init(&conf
->cache_size_mutex
);
6914 init_waitqueue_head(&conf
->wait_for_quiescent
);
6915 init_waitqueue_head(&conf
->wait_for_stripe
);
6916 init_waitqueue_head(&conf
->wait_for_overlap
);
6917 INIT_LIST_HEAD(&conf
->handle_list
);
6918 INIT_LIST_HEAD(&conf
->loprio_list
);
6919 INIT_LIST_HEAD(&conf
->hold_list
);
6920 INIT_LIST_HEAD(&conf
->delayed_list
);
6921 INIT_LIST_HEAD(&conf
->bitmap_list
);
6922 init_llist_head(&conf
->released_stripes
);
6923 atomic_set(&conf
->active_stripes
, 0);
6924 atomic_set(&conf
->preread_active_stripes
, 0);
6925 atomic_set(&conf
->active_aligned_reads
, 0);
6926 spin_lock_init(&conf
->pending_bios_lock
);
6927 conf
->batch_bio_dispatch
= true;
6928 rdev_for_each(rdev
, mddev
) {
6929 if (test_bit(Journal
, &rdev
->flags
))
6931 if (blk_queue_nonrot(bdev_get_queue(rdev
->bdev
))) {
6932 conf
->batch_bio_dispatch
= false;
6937 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
6938 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
6940 conf
->raid_disks
= mddev
->raid_disks
;
6941 if (mddev
->reshape_position
== MaxSector
)
6942 conf
->previous_raid_disks
= mddev
->raid_disks
;
6944 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
6945 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
6947 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
6953 for (i
= 0; i
< max_disks
; i
++) {
6954 conf
->disks
[i
].extra_page
= alloc_page(GFP_KERNEL
);
6955 if (!conf
->disks
[i
].extra_page
)
6959 ret
= bioset_init(&conf
->bio_split
, BIO_POOL_SIZE
, 0, 0);
6962 conf
->mddev
= mddev
;
6964 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
6967 /* We init hash_locks[0] separately to that it can be used
6968 * as the reference lock in the spin_lock_nest_lock() call
6969 * in lock_all_device_hash_locks_irq in order to convince
6970 * lockdep that we know what we are doing.
6972 spin_lock_init(conf
->hash_locks
);
6973 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6974 spin_lock_init(conf
->hash_locks
+ i
);
6976 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6977 INIT_LIST_HEAD(conf
->inactive_list
+ i
);
6979 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6980 INIT_LIST_HEAD(conf
->temp_inactive_list
+ i
);
6982 atomic_set(&conf
->r5c_cached_full_stripes
, 0);
6983 INIT_LIST_HEAD(&conf
->r5c_full_stripe_list
);
6984 atomic_set(&conf
->r5c_cached_partial_stripes
, 0);
6985 INIT_LIST_HEAD(&conf
->r5c_partial_stripe_list
);
6986 atomic_set(&conf
->r5c_flushing_full_stripes
, 0);
6987 atomic_set(&conf
->r5c_flushing_partial_stripes
, 0);
6989 conf
->level
= mddev
->new_level
;
6990 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
6991 if (raid5_alloc_percpu(conf
) != 0)
6994 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
6996 rdev_for_each(rdev
, mddev
) {
6997 raid_disk
= rdev
->raid_disk
;
6998 if (raid_disk
>= max_disks
6999 || raid_disk
< 0 || test_bit(Journal
, &rdev
->flags
))
7001 disk
= conf
->disks
+ raid_disk
;
7003 if (test_bit(Replacement
, &rdev
->flags
)) {
7004 if (disk
->replacement
)
7006 disk
->replacement
= rdev
;
7013 if (test_bit(In_sync
, &rdev
->flags
)) {
7014 char b
[BDEVNAME_SIZE
];
7015 pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7016 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
7017 } else if (rdev
->saved_raid_disk
!= raid_disk
)
7018 /* Cannot rely on bitmap to complete recovery */
7022 conf
->level
= mddev
->new_level
;
7023 if (conf
->level
== 6) {
7024 conf
->max_degraded
= 2;
7025 if (raid6_call
.xor_syndrome
)
7026 conf
->rmw_level
= PARITY_ENABLE_RMW
;
7028 conf
->rmw_level
= PARITY_DISABLE_RMW
;
7030 conf
->max_degraded
= 1;
7031 conf
->rmw_level
= PARITY_ENABLE_RMW
;
7033 conf
->algorithm
= mddev
->new_layout
;
7034 conf
->reshape_progress
= mddev
->reshape_position
;
7035 if (conf
->reshape_progress
!= MaxSector
) {
7036 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
7037 conf
->prev_algo
= mddev
->layout
;
7039 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
7040 conf
->prev_algo
= conf
->algorithm
;
7043 conf
->min_nr_stripes
= NR_STRIPES
;
7044 if (mddev
->reshape_position
!= MaxSector
) {
7045 int stripes
= max_t(int,
7046 ((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4,
7047 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4);
7048 conf
->min_nr_stripes
= max(NR_STRIPES
, stripes
);
7049 if (conf
->min_nr_stripes
!= NR_STRIPES
)
7050 pr_info("md/raid:%s: force stripe size %d for reshape\n",
7051 mdname(mddev
), conf
->min_nr_stripes
);
7053 memory
= conf
->min_nr_stripes
* (sizeof(struct stripe_head
) +
7054 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
7055 atomic_set(&conf
->empty_inactive_list_nr
, NR_STRIPE_HASH_LOCKS
);
7056 if (grow_stripes(conf
, conf
->min_nr_stripes
)) {
7057 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7058 mdname(mddev
), memory
);
7061 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev
), memory
);
7063 * Losing a stripe head costs more than the time to refill it,
7064 * it reduces the queue depth and so can hurt throughput.
7065 * So set it rather large, scaled by number of devices.
7067 conf
->shrinker
.seeks
= DEFAULT_SEEKS
* conf
->raid_disks
* 4;
7068 conf
->shrinker
.scan_objects
= raid5_cache_scan
;
7069 conf
->shrinker
.count_objects
= raid5_cache_count
;
7070 conf
->shrinker
.batch
= 128;
7071 conf
->shrinker
.flags
= 0;
7072 if (register_shrinker(&conf
->shrinker
)) {
7073 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7078 sprintf(pers_name
, "raid%d", mddev
->new_level
);
7079 conf
->thread
= md_register_thread(raid5d
, mddev
, pers_name
);
7080 if (!conf
->thread
) {
7081 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7091 return ERR_PTR(-EIO
);
7093 return ERR_PTR(-ENOMEM
);
7096 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
7099 case ALGORITHM_PARITY_0
:
7100 if (raid_disk
< max_degraded
)
7103 case ALGORITHM_PARITY_N
:
7104 if (raid_disk
>= raid_disks
- max_degraded
)
7107 case ALGORITHM_PARITY_0_6
:
7108 if (raid_disk
== 0 ||
7109 raid_disk
== raid_disks
- 1)
7112 case ALGORITHM_LEFT_ASYMMETRIC_6
:
7113 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
7114 case ALGORITHM_LEFT_SYMMETRIC_6
:
7115 case ALGORITHM_RIGHT_SYMMETRIC_6
:
7116 if (raid_disk
== raid_disks
- 1)
7122 static int raid5_run(struct mddev
*mddev
)
7124 struct r5conf
*conf
;
7125 int working_disks
= 0;
7126 int dirty_parity_disks
= 0;
7127 struct md_rdev
*rdev
;
7128 struct md_rdev
*journal_dev
= NULL
;
7129 sector_t reshape_offset
= 0;
7131 long long min_offset_diff
= 0;
7134 if (mddev_init_writes_pending(mddev
) < 0)
7137 if (mddev
->recovery_cp
!= MaxSector
)
7138 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7141 rdev_for_each(rdev
, mddev
) {
7144 if (test_bit(Journal
, &rdev
->flags
)) {
7148 if (rdev
->raid_disk
< 0)
7150 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
7152 min_offset_diff
= diff
;
7154 } else if (mddev
->reshape_backwards
&&
7155 diff
< min_offset_diff
)
7156 min_offset_diff
= diff
;
7157 else if (!mddev
->reshape_backwards
&&
7158 diff
> min_offset_diff
)
7159 min_offset_diff
= diff
;
7162 if ((test_bit(MD_HAS_JOURNAL
, &mddev
->flags
) || journal_dev
) &&
7163 (mddev
->bitmap_info
.offset
|| mddev
->bitmap_info
.file
)) {
7164 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7169 if (mddev
->reshape_position
!= MaxSector
) {
7170 /* Check that we can continue the reshape.
7171 * Difficulties arise if the stripe we would write to
7172 * next is at or after the stripe we would read from next.
7173 * For a reshape that changes the number of devices, this
7174 * is only possible for a very short time, and mdadm makes
7175 * sure that time appears to have past before assembling
7176 * the array. So we fail if that time hasn't passed.
7177 * For a reshape that keeps the number of devices the same
7178 * mdadm must be monitoring the reshape can keeping the
7179 * critical areas read-only and backed up. It will start
7180 * the array in read-only mode, so we check for that.
7182 sector_t here_new
, here_old
;
7184 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
7189 pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7194 if (mddev
->new_level
!= mddev
->level
) {
7195 pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7199 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
7200 /* reshape_position must be on a new-stripe boundary, and one
7201 * further up in new geometry must map after here in old
7203 * If the chunk sizes are different, then as we perform reshape
7204 * in units of the largest of the two, reshape_position needs
7205 * be a multiple of the largest chunk size times new data disks.
7207 here_new
= mddev
->reshape_position
;
7208 chunk_sectors
= max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
);
7209 new_data_disks
= mddev
->raid_disks
- max_degraded
;
7210 if (sector_div(here_new
, chunk_sectors
* new_data_disks
)) {
7211 pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7215 reshape_offset
= here_new
* chunk_sectors
;
7216 /* here_new is the stripe we will write to */
7217 here_old
= mddev
->reshape_position
;
7218 sector_div(here_old
, chunk_sectors
* (old_disks
-max_degraded
));
7219 /* here_old is the first stripe that we might need to read
7221 if (mddev
->delta_disks
== 0) {
7222 /* We cannot be sure it is safe to start an in-place
7223 * reshape. It is only safe if user-space is monitoring
7224 * and taking constant backups.
7225 * mdadm always starts a situation like this in
7226 * readonly mode so it can take control before
7227 * allowing any writes. So just check for that.
7229 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
7230 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
7231 /* not really in-place - so OK */;
7232 else if (mddev
->ro
== 0) {
7233 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7237 } else if (mddev
->reshape_backwards
7238 ? (here_new
* chunk_sectors
+ min_offset_diff
<=
7239 here_old
* chunk_sectors
)
7240 : (here_new
* chunk_sectors
>=
7241 here_old
* chunk_sectors
+ (-min_offset_diff
))) {
7242 /* Reading from the same stripe as writing to - bad */
7243 pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7247 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev
));
7248 /* OK, we should be able to continue; */
7250 BUG_ON(mddev
->level
!= mddev
->new_level
);
7251 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
7252 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
7253 BUG_ON(mddev
->delta_disks
!= 0);
7256 if (test_bit(MD_HAS_JOURNAL
, &mddev
->flags
) &&
7257 test_bit(MD_HAS_PPL
, &mddev
->flags
)) {
7258 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7260 clear_bit(MD_HAS_PPL
, &mddev
->flags
);
7261 clear_bit(MD_HAS_MULTIPLE_PPLS
, &mddev
->flags
);
7264 if (mddev
->private == NULL
)
7265 conf
= setup_conf(mddev
);
7267 conf
= mddev
->private;
7270 return PTR_ERR(conf
);
7272 if (test_bit(MD_HAS_JOURNAL
, &mddev
->flags
)) {
7274 pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7277 set_disk_ro(mddev
->gendisk
, 1);
7278 } else if (mddev
->recovery_cp
== MaxSector
)
7279 set_bit(MD_JOURNAL_CLEAN
, &mddev
->flags
);
7282 conf
->min_offset_diff
= min_offset_diff
;
7283 mddev
->thread
= conf
->thread
;
7284 conf
->thread
= NULL
;
7285 mddev
->private = conf
;
7287 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
7289 rdev
= conf
->disks
[i
].rdev
;
7290 if (!rdev
&& conf
->disks
[i
].replacement
) {
7291 /* The replacement is all we have yet */
7292 rdev
= conf
->disks
[i
].replacement
;
7293 conf
->disks
[i
].replacement
= NULL
;
7294 clear_bit(Replacement
, &rdev
->flags
);
7295 conf
->disks
[i
].rdev
= rdev
;
7299 if (conf
->disks
[i
].replacement
&&
7300 conf
->reshape_progress
!= MaxSector
) {
7301 /* replacements and reshape simply do not mix. */
7302 pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7305 if (test_bit(In_sync
, &rdev
->flags
)) {
7309 /* This disc is not fully in-sync. However if it
7310 * just stored parity (beyond the recovery_offset),
7311 * when we don't need to be concerned about the
7312 * array being dirty.
7313 * When reshape goes 'backwards', we never have
7314 * partially completed devices, so we only need
7315 * to worry about reshape going forwards.
7317 /* Hack because v0.91 doesn't store recovery_offset properly. */
7318 if (mddev
->major_version
== 0 &&
7319 mddev
->minor_version
> 90)
7320 rdev
->recovery_offset
= reshape_offset
;
7322 if (rdev
->recovery_offset
< reshape_offset
) {
7323 /* We need to check old and new layout */
7324 if (!only_parity(rdev
->raid_disk
,
7327 conf
->max_degraded
))
7330 if (!only_parity(rdev
->raid_disk
,
7332 conf
->previous_raid_disks
,
7333 conf
->max_degraded
))
7335 dirty_parity_disks
++;
7339 * 0 for a fully functional array, 1 or 2 for a degraded array.
7341 mddev
->degraded
= raid5_calc_degraded(conf
);
7343 if (has_failed(conf
)) {
7344 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7345 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
7349 /* device size must be a multiple of chunk size */
7350 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
7351 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
7353 if (mddev
->degraded
> dirty_parity_disks
&&
7354 mddev
->recovery_cp
!= MaxSector
) {
7355 if (test_bit(MD_HAS_PPL
, &mddev
->flags
))
7356 pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7358 else if (mddev
->ok_start_degraded
)
7359 pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7362 pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7368 pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7369 mdname(mddev
), conf
->level
,
7370 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
7373 print_raid5_conf(conf
);
7375 if (conf
->reshape_progress
!= MaxSector
) {
7376 conf
->reshape_safe
= conf
->reshape_progress
;
7377 atomic_set(&conf
->reshape_stripes
, 0);
7378 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
7379 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
7380 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
7381 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
7382 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
7386 /* Ok, everything is just fine now */
7387 if (mddev
->to_remove
== &raid5_attrs_group
)
7388 mddev
->to_remove
= NULL
;
7389 else if (mddev
->kobj
.sd
&&
7390 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
7391 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7393 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
7397 /* read-ahead size must cover two whole stripes, which
7398 * is 2 * (datadisks) * chunksize where 'n' is the
7399 * number of raid devices
7401 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
7402 int stripe
= data_disks
*
7403 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
7404 if (mddev
->queue
->backing_dev_info
->ra_pages
< 2 * stripe
)
7405 mddev
->queue
->backing_dev_info
->ra_pages
= 2 * stripe
;
7407 chunk_size
= mddev
->chunk_sectors
<< 9;
7408 blk_queue_io_min(mddev
->queue
, chunk_size
);
7409 blk_queue_io_opt(mddev
->queue
, chunk_size
*
7410 (conf
->raid_disks
- conf
->max_degraded
));
7411 mddev
->queue
->limits
.raid_partial_stripes_expensive
= 1;
7413 * We can only discard a whole stripe. It doesn't make sense to
7414 * discard data disk but write parity disk
7416 stripe
= stripe
* PAGE_SIZE
;
7417 /* Round up to power of 2, as discard handling
7418 * currently assumes that */
7419 while ((stripe
-1) & stripe
)
7420 stripe
= (stripe
| (stripe
-1)) + 1;
7421 mddev
->queue
->limits
.discard_alignment
= stripe
;
7422 mddev
->queue
->limits
.discard_granularity
= stripe
;
7424 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
7425 blk_queue_max_write_zeroes_sectors(mddev
->queue
, 0);
7427 rdev_for_each(rdev
, mddev
) {
7428 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
7429 rdev
->data_offset
<< 9);
7430 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
7431 rdev
->new_data_offset
<< 9);
7435 * zeroing is required, otherwise data
7436 * could be lost. Consider a scenario: discard a stripe
7437 * (the stripe could be inconsistent if
7438 * discard_zeroes_data is 0); write one disk of the
7439 * stripe (the stripe could be inconsistent again
7440 * depending on which disks are used to calculate
7441 * parity); the disk is broken; The stripe data of this
7444 * We only allow DISCARD if the sysadmin has confirmed that
7445 * only safe devices are in use by setting a module parameter.
7446 * A better idea might be to turn DISCARD into WRITE_ZEROES
7447 * requests, as that is required to be safe.
7449 if (devices_handle_discard_safely
&&
7450 mddev
->queue
->limits
.max_discard_sectors
>= (stripe
>> 9) &&
7451 mddev
->queue
->limits
.discard_granularity
>= stripe
)
7452 blk_queue_flag_set(QUEUE_FLAG_DISCARD
,
7455 blk_queue_flag_clear(QUEUE_FLAG_DISCARD
,
7458 blk_queue_max_hw_sectors(mddev
->queue
, UINT_MAX
);
7461 if (log_init(conf
, journal_dev
, raid5_has_ppl(conf
)))
7466 md_unregister_thread(&mddev
->thread
);
7467 print_raid5_conf(conf
);
7469 mddev
->private = NULL
;
7470 pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev
));
7474 static void raid5_free(struct mddev
*mddev
, void *priv
)
7476 struct r5conf
*conf
= priv
;
7479 mddev
->to_remove
= &raid5_attrs_group
;
7482 static void raid5_status(struct seq_file
*seq
, struct mddev
*mddev
)
7484 struct r5conf
*conf
= mddev
->private;
7487 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
7488 conf
->chunk_sectors
/ 2, mddev
->layout
);
7489 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
7491 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7492 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
7493 seq_printf (seq
, "%s", rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
7496 seq_printf (seq
, "]");
7499 static void print_raid5_conf (struct r5conf
*conf
)
7502 struct disk_info
*tmp
;
7504 pr_debug("RAID conf printout:\n");
7506 pr_debug("(conf==NULL)\n");
7509 pr_debug(" --- level:%d rd:%d wd:%d\n", conf
->level
,
7511 conf
->raid_disks
- conf
->mddev
->degraded
);
7513 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7514 char b
[BDEVNAME_SIZE
];
7515 tmp
= conf
->disks
+ i
;
7517 pr_debug(" disk %d, o:%d, dev:%s\n",
7518 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
7519 bdevname(tmp
->rdev
->bdev
, b
));
7523 static int raid5_spare_active(struct mddev
*mddev
)
7526 struct r5conf
*conf
= mddev
->private;
7527 struct disk_info
*tmp
;
7529 unsigned long flags
;
7531 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7532 tmp
= conf
->disks
+ i
;
7533 if (tmp
->replacement
7534 && tmp
->replacement
->recovery_offset
== MaxSector
7535 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
7536 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
7537 /* Replacement has just become active. */
7539 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
7542 /* Replaced device not technically faulty,
7543 * but we need to be sure it gets removed
7544 * and never re-added.
7546 set_bit(Faulty
, &tmp
->rdev
->flags
);
7547 sysfs_notify_dirent_safe(
7548 tmp
->rdev
->sysfs_state
);
7550 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
7551 } else if (tmp
->rdev
7552 && tmp
->rdev
->recovery_offset
== MaxSector
7553 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
7554 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
7556 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
7559 spin_lock_irqsave(&conf
->device_lock
, flags
);
7560 mddev
->degraded
= raid5_calc_degraded(conf
);
7561 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7562 print_raid5_conf(conf
);
7566 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7568 struct r5conf
*conf
= mddev
->private;
7570 int number
= rdev
->raid_disk
;
7571 struct md_rdev
**rdevp
;
7572 struct disk_info
*p
= conf
->disks
+ number
;
7574 print_raid5_conf(conf
);
7575 if (test_bit(Journal
, &rdev
->flags
) && conf
->log
) {
7577 * we can't wait pending write here, as this is called in
7578 * raid5d, wait will deadlock.
7579 * neilb: there is no locking about new writes here,
7580 * so this cannot be safe.
7582 if (atomic_read(&conf
->active_stripes
) ||
7583 atomic_read(&conf
->r5c_cached_full_stripes
) ||
7584 atomic_read(&conf
->r5c_cached_partial_stripes
)) {
7590 if (rdev
== p
->rdev
)
7592 else if (rdev
== p
->replacement
)
7593 rdevp
= &p
->replacement
;
7597 if (number
>= conf
->raid_disks
&&
7598 conf
->reshape_progress
== MaxSector
)
7599 clear_bit(In_sync
, &rdev
->flags
);
7601 if (test_bit(In_sync
, &rdev
->flags
) ||
7602 atomic_read(&rdev
->nr_pending
)) {
7606 /* Only remove non-faulty devices if recovery
7609 if (!test_bit(Faulty
, &rdev
->flags
) &&
7610 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
7611 !has_failed(conf
) &&
7612 (!p
->replacement
|| p
->replacement
== rdev
) &&
7613 number
< conf
->raid_disks
) {
7618 if (!test_bit(RemoveSynchronized
, &rdev
->flags
)) {
7620 if (atomic_read(&rdev
->nr_pending
)) {
7621 /* lost the race, try later */
7627 err
= log_modify(conf
, rdev
, false);
7631 if (p
->replacement
) {
7632 /* We must have just cleared 'rdev' */
7633 p
->rdev
= p
->replacement
;
7634 clear_bit(Replacement
, &p
->replacement
->flags
);
7635 smp_mb(); /* Make sure other CPUs may see both as identical
7636 * but will never see neither - if they are careful
7638 p
->replacement
= NULL
;
7641 err
= log_modify(conf
, p
->rdev
, true);
7644 clear_bit(WantReplacement
, &rdev
->flags
);
7647 print_raid5_conf(conf
);
7651 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7653 struct r5conf
*conf
= mddev
->private;
7656 struct disk_info
*p
;
7658 int last
= conf
->raid_disks
- 1;
7660 if (test_bit(Journal
, &rdev
->flags
)) {
7664 rdev
->raid_disk
= 0;
7666 * The array is in readonly mode if journal is missing, so no
7667 * write requests running. We should be safe
7669 log_init(conf
, rdev
, false);
7672 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
7675 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
7676 /* no point adding a device */
7679 if (rdev
->raid_disk
>= 0)
7680 first
= last
= rdev
->raid_disk
;
7683 * find the disk ... but prefer rdev->saved_raid_disk
7686 if (rdev
->saved_raid_disk
>= 0 &&
7687 rdev
->saved_raid_disk
>= first
&&
7688 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
7689 first
= rdev
->saved_raid_disk
;
7691 for (disk
= first
; disk
<= last
; disk
++) {
7692 p
= conf
->disks
+ disk
;
7693 if (p
->rdev
== NULL
) {
7694 clear_bit(In_sync
, &rdev
->flags
);
7695 rdev
->raid_disk
= disk
;
7696 if (rdev
->saved_raid_disk
!= disk
)
7698 rcu_assign_pointer(p
->rdev
, rdev
);
7700 err
= log_modify(conf
, rdev
, true);
7705 for (disk
= first
; disk
<= last
; disk
++) {
7706 p
= conf
->disks
+ disk
;
7707 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
7708 p
->replacement
== NULL
) {
7709 clear_bit(In_sync
, &rdev
->flags
);
7710 set_bit(Replacement
, &rdev
->flags
);
7711 rdev
->raid_disk
= disk
;
7714 rcu_assign_pointer(p
->replacement
, rdev
);
7719 print_raid5_conf(conf
);
7723 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
7725 /* no resync is happening, and there is enough space
7726 * on all devices, so we can resize.
7727 * We need to make sure resync covers any new space.
7728 * If the array is shrinking we should possibly wait until
7729 * any io in the removed space completes, but it hardly seems
7733 struct r5conf
*conf
= mddev
->private;
7735 if (conf
->log
|| raid5_has_ppl(conf
))
7737 sectors
&= ~((sector_t
)conf
->chunk_sectors
- 1);
7738 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
7739 if (mddev
->external_size
&&
7740 mddev
->array_sectors
> newsize
)
7742 if (mddev
->bitmap
) {
7743 int ret
= bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
7747 md_set_array_sectors(mddev
, newsize
);
7748 if (sectors
> mddev
->dev_sectors
&&
7749 mddev
->recovery_cp
> mddev
->dev_sectors
) {
7750 mddev
->recovery_cp
= mddev
->dev_sectors
;
7751 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
7753 mddev
->dev_sectors
= sectors
;
7754 mddev
->resync_max_sectors
= sectors
;
7758 static int check_stripe_cache(struct mddev
*mddev
)
7760 /* Can only proceed if there are plenty of stripe_heads.
7761 * We need a minimum of one full stripe,, and for sensible progress
7762 * it is best to have about 4 times that.
7763 * If we require 4 times, then the default 256 4K stripe_heads will
7764 * allow for chunk sizes up to 256K, which is probably OK.
7765 * If the chunk size is greater, user-space should request more
7766 * stripe_heads first.
7768 struct r5conf
*conf
= mddev
->private;
7769 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7770 > conf
->min_nr_stripes
||
7771 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7772 > conf
->min_nr_stripes
) {
7773 pr_warn("md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7775 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
7782 static int check_reshape(struct mddev
*mddev
)
7784 struct r5conf
*conf
= mddev
->private;
7786 if (conf
->log
|| raid5_has_ppl(conf
))
7788 if (mddev
->delta_disks
== 0 &&
7789 mddev
->new_layout
== mddev
->layout
&&
7790 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
7791 return 0; /* nothing to do */
7792 if (has_failed(conf
))
7794 if (mddev
->delta_disks
< 0 && mddev
->reshape_position
== MaxSector
) {
7795 /* We might be able to shrink, but the devices must
7796 * be made bigger first.
7797 * For raid6, 4 is the minimum size.
7798 * Otherwise 2 is the minimum
7801 if (mddev
->level
== 6)
7803 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
7807 if (!check_stripe_cache(mddev
))
7810 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
||
7811 mddev
->delta_disks
> 0)
7812 if (resize_chunks(conf
,
7813 conf
->previous_raid_disks
7814 + max(0, mddev
->delta_disks
),
7815 max(mddev
->new_chunk_sectors
,
7816 mddev
->chunk_sectors
)
7820 if (conf
->previous_raid_disks
+ mddev
->delta_disks
<= conf
->pool_size
)
7821 return 0; /* never bother to shrink */
7822 return resize_stripes(conf
, (conf
->previous_raid_disks
7823 + mddev
->delta_disks
));
7826 static int raid5_start_reshape(struct mddev
*mddev
)
7828 struct r5conf
*conf
= mddev
->private;
7829 struct md_rdev
*rdev
;
7831 unsigned long flags
;
7833 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
7836 if (!check_stripe_cache(mddev
))
7839 if (has_failed(conf
))
7842 rdev_for_each(rdev
, mddev
) {
7843 if (!test_bit(In_sync
, &rdev
->flags
)
7844 && !test_bit(Faulty
, &rdev
->flags
))
7848 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
7849 /* Not enough devices even to make a degraded array
7854 /* Refuse to reduce size of the array. Any reductions in
7855 * array size must be through explicit setting of array_size
7858 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
7859 < mddev
->array_sectors
) {
7860 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7865 atomic_set(&conf
->reshape_stripes
, 0);
7866 spin_lock_irq(&conf
->device_lock
);
7867 write_seqcount_begin(&conf
->gen_lock
);
7868 conf
->previous_raid_disks
= conf
->raid_disks
;
7869 conf
->raid_disks
+= mddev
->delta_disks
;
7870 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
7871 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
7872 conf
->prev_algo
= conf
->algorithm
;
7873 conf
->algorithm
= mddev
->new_layout
;
7875 /* Code that selects data_offset needs to see the generation update
7876 * if reshape_progress has been set - so a memory barrier needed.
7879 if (mddev
->reshape_backwards
)
7880 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
7882 conf
->reshape_progress
= 0;
7883 conf
->reshape_safe
= conf
->reshape_progress
;
7884 write_seqcount_end(&conf
->gen_lock
);
7885 spin_unlock_irq(&conf
->device_lock
);
7887 /* Now make sure any requests that proceeded on the assumption
7888 * the reshape wasn't running - like Discard or Read - have
7891 mddev_suspend(mddev
);
7892 mddev_resume(mddev
);
7894 /* Add some new drives, as many as will fit.
7895 * We know there are enough to make the newly sized array work.
7896 * Don't add devices if we are reducing the number of
7897 * devices in the array. This is because it is not possible
7898 * to correctly record the "partially reconstructed" state of
7899 * such devices during the reshape and confusion could result.
7901 if (mddev
->delta_disks
>= 0) {
7902 rdev_for_each(rdev
, mddev
)
7903 if (rdev
->raid_disk
< 0 &&
7904 !test_bit(Faulty
, &rdev
->flags
)) {
7905 if (raid5_add_disk(mddev
, rdev
) == 0) {
7907 >= conf
->previous_raid_disks
)
7908 set_bit(In_sync
, &rdev
->flags
);
7910 rdev
->recovery_offset
= 0;
7912 if (sysfs_link_rdev(mddev
, rdev
))
7913 /* Failure here is OK */;
7915 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
7916 && !test_bit(Faulty
, &rdev
->flags
)) {
7917 /* This is a spare that was manually added */
7918 set_bit(In_sync
, &rdev
->flags
);
7921 /* When a reshape changes the number of devices,
7922 * ->degraded is measured against the larger of the
7923 * pre and post number of devices.
7925 spin_lock_irqsave(&conf
->device_lock
, flags
);
7926 mddev
->degraded
= raid5_calc_degraded(conf
);
7927 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7929 mddev
->raid_disks
= conf
->raid_disks
;
7930 mddev
->reshape_position
= conf
->reshape_progress
;
7931 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
7933 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
7934 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
7935 clear_bit(MD_RECOVERY_DONE
, &mddev
->recovery
);
7936 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
7937 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
7938 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
7940 if (!mddev
->sync_thread
) {
7941 mddev
->recovery
= 0;
7942 spin_lock_irq(&conf
->device_lock
);
7943 write_seqcount_begin(&conf
->gen_lock
);
7944 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
7945 mddev
->new_chunk_sectors
=
7946 conf
->chunk_sectors
= conf
->prev_chunk_sectors
;
7947 mddev
->new_layout
= conf
->algorithm
= conf
->prev_algo
;
7948 rdev_for_each(rdev
, mddev
)
7949 rdev
->new_data_offset
= rdev
->data_offset
;
7951 conf
->generation
--;
7952 conf
->reshape_progress
= MaxSector
;
7953 mddev
->reshape_position
= MaxSector
;
7954 write_seqcount_end(&conf
->gen_lock
);
7955 spin_unlock_irq(&conf
->device_lock
);
7958 conf
->reshape_checkpoint
= jiffies
;
7959 md_wakeup_thread(mddev
->sync_thread
);
7960 md_new_event(mddev
);
7964 /* This is called from the reshape thread and should make any
7965 * changes needed in 'conf'
7967 static void end_reshape(struct r5conf
*conf
)
7970 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
7971 struct md_rdev
*rdev
;
7973 spin_lock_irq(&conf
->device_lock
);
7974 conf
->previous_raid_disks
= conf
->raid_disks
;
7975 md_finish_reshape(conf
->mddev
);
7977 conf
->reshape_progress
= MaxSector
;
7978 conf
->mddev
->reshape_position
= MaxSector
;
7979 rdev_for_each(rdev
, conf
->mddev
)
7980 if (rdev
->raid_disk
>= 0 &&
7981 !test_bit(Journal
, &rdev
->flags
) &&
7982 !test_bit(In_sync
, &rdev
->flags
))
7983 rdev
->recovery_offset
= MaxSector
;
7984 spin_unlock_irq(&conf
->device_lock
);
7985 wake_up(&conf
->wait_for_overlap
);
7987 /* read-ahead size must cover two whole stripes, which is
7988 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7990 if (conf
->mddev
->queue
) {
7991 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
7992 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
7994 if (conf
->mddev
->queue
->backing_dev_info
->ra_pages
< 2 * stripe
)
7995 conf
->mddev
->queue
->backing_dev_info
->ra_pages
= 2 * stripe
;
8000 /* This is called from the raid5d thread with mddev_lock held.
8001 * It makes config changes to the device.
8003 static void raid5_finish_reshape(struct mddev
*mddev
)
8005 struct r5conf
*conf
= mddev
->private;
8007 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
8009 if (mddev
->delta_disks
<= 0) {
8011 spin_lock_irq(&conf
->device_lock
);
8012 mddev
->degraded
= raid5_calc_degraded(conf
);
8013 spin_unlock_irq(&conf
->device_lock
);
8014 for (d
= conf
->raid_disks
;
8015 d
< conf
->raid_disks
- mddev
->delta_disks
;
8017 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
8019 clear_bit(In_sync
, &rdev
->flags
);
8020 rdev
= conf
->disks
[d
].replacement
;
8022 clear_bit(In_sync
, &rdev
->flags
);
8025 mddev
->layout
= conf
->algorithm
;
8026 mddev
->chunk_sectors
= conf
->chunk_sectors
;
8027 mddev
->reshape_position
= MaxSector
;
8028 mddev
->delta_disks
= 0;
8029 mddev
->reshape_backwards
= 0;
8033 static void raid5_quiesce(struct mddev
*mddev
, int quiesce
)
8035 struct r5conf
*conf
= mddev
->private;
8038 /* stop all writes */
8039 lock_all_device_hash_locks_irq(conf
);
8040 /* '2' tells resync/reshape to pause so that all
8041 * active stripes can drain
8043 r5c_flush_cache(conf
, INT_MAX
);
8045 wait_event_cmd(conf
->wait_for_quiescent
,
8046 atomic_read(&conf
->active_stripes
) == 0 &&
8047 atomic_read(&conf
->active_aligned_reads
) == 0,
8048 unlock_all_device_hash_locks_irq(conf
),
8049 lock_all_device_hash_locks_irq(conf
));
8051 unlock_all_device_hash_locks_irq(conf
);
8052 /* allow reshape to continue */
8053 wake_up(&conf
->wait_for_overlap
);
8055 /* re-enable writes */
8056 lock_all_device_hash_locks_irq(conf
);
8058 wake_up(&conf
->wait_for_quiescent
);
8059 wake_up(&conf
->wait_for_overlap
);
8060 unlock_all_device_hash_locks_irq(conf
);
8062 log_quiesce(conf
, quiesce
);
8065 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
8067 struct r0conf
*raid0_conf
= mddev
->private;
8070 /* for raid0 takeover only one zone is supported */
8071 if (raid0_conf
->nr_strip_zones
> 1) {
8072 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8074 return ERR_PTR(-EINVAL
);
8077 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
8078 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
8079 mddev
->dev_sectors
= sectors
;
8080 mddev
->new_level
= level
;
8081 mddev
->new_layout
= ALGORITHM_PARITY_N
;
8082 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
8083 mddev
->raid_disks
+= 1;
8084 mddev
->delta_disks
= 1;
8085 /* make sure it will be not marked as dirty */
8086 mddev
->recovery_cp
= MaxSector
;
8088 return setup_conf(mddev
);
8091 static void *raid5_takeover_raid1(struct mddev
*mddev
)
8096 if (mddev
->raid_disks
!= 2 ||
8097 mddev
->degraded
> 1)
8098 return ERR_PTR(-EINVAL
);
8100 /* Should check if there are write-behind devices? */
8102 chunksect
= 64*2; /* 64K by default */
8104 /* The array must be an exact multiple of chunksize */
8105 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
8108 if ((chunksect
<<9) < STRIPE_SIZE
)
8109 /* array size does not allow a suitable chunk size */
8110 return ERR_PTR(-EINVAL
);
8112 mddev
->new_level
= 5;
8113 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
8114 mddev
->new_chunk_sectors
= chunksect
;
8116 ret
= setup_conf(mddev
);
8118 mddev_clear_unsupported_flags(mddev
,
8119 UNSUPPORTED_MDDEV_FLAGS
);
8123 static void *raid5_takeover_raid6(struct mddev
*mddev
)
8127 switch (mddev
->layout
) {
8128 case ALGORITHM_LEFT_ASYMMETRIC_6
:
8129 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
8131 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
8132 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
8134 case ALGORITHM_LEFT_SYMMETRIC_6
:
8135 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
8137 case ALGORITHM_RIGHT_SYMMETRIC_6
:
8138 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
8140 case ALGORITHM_PARITY_0_6
:
8141 new_layout
= ALGORITHM_PARITY_0
;
8143 case ALGORITHM_PARITY_N
:
8144 new_layout
= ALGORITHM_PARITY_N
;
8147 return ERR_PTR(-EINVAL
);
8149 mddev
->new_level
= 5;
8150 mddev
->new_layout
= new_layout
;
8151 mddev
->delta_disks
= -1;
8152 mddev
->raid_disks
-= 1;
8153 return setup_conf(mddev
);
8156 static int raid5_check_reshape(struct mddev
*mddev
)
8158 /* For a 2-drive array, the layout and chunk size can be changed
8159 * immediately as not restriping is needed.
8160 * For larger arrays we record the new value - after validation
8161 * to be used by a reshape pass.
8163 struct r5conf
*conf
= mddev
->private;
8164 int new_chunk
= mddev
->new_chunk_sectors
;
8166 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
8168 if (new_chunk
> 0) {
8169 if (!is_power_of_2(new_chunk
))
8171 if (new_chunk
< (PAGE_SIZE
>>9))
8173 if (mddev
->array_sectors
& (new_chunk
-1))
8174 /* not factor of array size */
8178 /* They look valid */
8180 if (mddev
->raid_disks
== 2) {
8181 /* can make the change immediately */
8182 if (mddev
->new_layout
>= 0) {
8183 conf
->algorithm
= mddev
->new_layout
;
8184 mddev
->layout
= mddev
->new_layout
;
8186 if (new_chunk
> 0) {
8187 conf
->chunk_sectors
= new_chunk
;
8188 mddev
->chunk_sectors
= new_chunk
;
8190 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
8191 md_wakeup_thread(mddev
->thread
);
8193 return check_reshape(mddev
);
8196 static int raid6_check_reshape(struct mddev
*mddev
)
8198 int new_chunk
= mddev
->new_chunk_sectors
;
8200 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
8202 if (new_chunk
> 0) {
8203 if (!is_power_of_2(new_chunk
))
8205 if (new_chunk
< (PAGE_SIZE
>> 9))
8207 if (mddev
->array_sectors
& (new_chunk
-1))
8208 /* not factor of array size */
8212 /* They look valid */
8213 return check_reshape(mddev
);
8216 static void *raid5_takeover(struct mddev
*mddev
)
8218 /* raid5 can take over:
8219 * raid0 - if there is only one strip zone - make it a raid4 layout
8220 * raid1 - if there are two drives. We need to know the chunk size
8221 * raid4 - trivial - just use a raid4 layout.
8222 * raid6 - Providing it is a *_6 layout
8224 if (mddev
->level
== 0)
8225 return raid45_takeover_raid0(mddev
, 5);
8226 if (mddev
->level
== 1)
8227 return raid5_takeover_raid1(mddev
);
8228 if (mddev
->level
== 4) {
8229 mddev
->new_layout
= ALGORITHM_PARITY_N
;
8230 mddev
->new_level
= 5;
8231 return setup_conf(mddev
);
8233 if (mddev
->level
== 6)
8234 return raid5_takeover_raid6(mddev
);
8236 return ERR_PTR(-EINVAL
);
8239 static void *raid4_takeover(struct mddev
*mddev
)
8241 /* raid4 can take over:
8242 * raid0 - if there is only one strip zone
8243 * raid5 - if layout is right
8245 if (mddev
->level
== 0)
8246 return raid45_takeover_raid0(mddev
, 4);
8247 if (mddev
->level
== 5 &&
8248 mddev
->layout
== ALGORITHM_PARITY_N
) {
8249 mddev
->new_layout
= 0;
8250 mddev
->new_level
= 4;
8251 return setup_conf(mddev
);
8253 return ERR_PTR(-EINVAL
);
8256 static struct md_personality raid5_personality
;
8258 static void *raid6_takeover(struct mddev
*mddev
)
8260 /* Currently can only take over a raid5. We map the
8261 * personality to an equivalent raid6 personality
8262 * with the Q block at the end.
8266 if (mddev
->pers
!= &raid5_personality
)
8267 return ERR_PTR(-EINVAL
);
8268 if (mddev
->degraded
> 1)
8269 return ERR_PTR(-EINVAL
);
8270 if (mddev
->raid_disks
> 253)
8271 return ERR_PTR(-EINVAL
);
8272 if (mddev
->raid_disks
< 3)
8273 return ERR_PTR(-EINVAL
);
8275 switch (mddev
->layout
) {
8276 case ALGORITHM_LEFT_ASYMMETRIC
:
8277 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
8279 case ALGORITHM_RIGHT_ASYMMETRIC
:
8280 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
8282 case ALGORITHM_LEFT_SYMMETRIC
:
8283 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
8285 case ALGORITHM_RIGHT_SYMMETRIC
:
8286 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
8288 case ALGORITHM_PARITY_0
:
8289 new_layout
= ALGORITHM_PARITY_0_6
;
8291 case ALGORITHM_PARITY_N
:
8292 new_layout
= ALGORITHM_PARITY_N
;
8295 return ERR_PTR(-EINVAL
);
8297 mddev
->new_level
= 6;
8298 mddev
->new_layout
= new_layout
;
8299 mddev
->delta_disks
= 1;
8300 mddev
->raid_disks
+= 1;
8301 return setup_conf(mddev
);
8304 static int raid5_change_consistency_policy(struct mddev
*mddev
, const char *buf
)
8306 struct r5conf
*conf
;
8309 err
= mddev_lock(mddev
);
8312 conf
= mddev
->private;
8314 mddev_unlock(mddev
);
8318 if (strncmp(buf
, "ppl", 3) == 0) {
8319 /* ppl only works with RAID 5 */
8320 if (!raid5_has_ppl(conf
) && conf
->level
== 5) {
8321 err
= log_init(conf
, NULL
, true);
8323 err
= resize_stripes(conf
, conf
->pool_size
);
8329 } else if (strncmp(buf
, "resync", 6) == 0) {
8330 if (raid5_has_ppl(conf
)) {
8331 mddev_suspend(mddev
);
8333 mddev_resume(mddev
);
8334 err
= resize_stripes(conf
, conf
->pool_size
);
8335 } else if (test_bit(MD_HAS_JOURNAL
, &conf
->mddev
->flags
) &&
8336 r5l_log_disk_error(conf
)) {
8337 bool journal_dev_exists
= false;
8338 struct md_rdev
*rdev
;
8340 rdev_for_each(rdev
, mddev
)
8341 if (test_bit(Journal
, &rdev
->flags
)) {
8342 journal_dev_exists
= true;
8346 if (!journal_dev_exists
) {
8347 mddev_suspend(mddev
);
8348 clear_bit(MD_HAS_JOURNAL
, &mddev
->flags
);
8349 mddev_resume(mddev
);
8350 } else /* need remove journal device first */
8359 md_update_sb(mddev
, 1);
8361 mddev_unlock(mddev
);
8366 static int raid5_start(struct mddev
*mddev
)
8368 struct r5conf
*conf
= mddev
->private;
8370 return r5l_start(conf
->log
);
8373 static struct md_personality raid6_personality
=
8377 .owner
= THIS_MODULE
,
8378 .make_request
= raid5_make_request
,
8380 .start
= raid5_start
,
8382 .status
= raid5_status
,
8383 .error_handler
= raid5_error
,
8384 .hot_add_disk
= raid5_add_disk
,
8385 .hot_remove_disk
= raid5_remove_disk
,
8386 .spare_active
= raid5_spare_active
,
8387 .sync_request
= raid5_sync_request
,
8388 .resize
= raid5_resize
,
8390 .check_reshape
= raid6_check_reshape
,
8391 .start_reshape
= raid5_start_reshape
,
8392 .finish_reshape
= raid5_finish_reshape
,
8393 .quiesce
= raid5_quiesce
,
8394 .takeover
= raid6_takeover
,
8395 .congested
= raid5_congested
,
8396 .change_consistency_policy
= raid5_change_consistency_policy
,
8398 static struct md_personality raid5_personality
=
8402 .owner
= THIS_MODULE
,
8403 .make_request
= raid5_make_request
,
8405 .start
= raid5_start
,
8407 .status
= raid5_status
,
8408 .error_handler
= raid5_error
,
8409 .hot_add_disk
= raid5_add_disk
,
8410 .hot_remove_disk
= raid5_remove_disk
,
8411 .spare_active
= raid5_spare_active
,
8412 .sync_request
= raid5_sync_request
,
8413 .resize
= raid5_resize
,
8415 .check_reshape
= raid5_check_reshape
,
8416 .start_reshape
= raid5_start_reshape
,
8417 .finish_reshape
= raid5_finish_reshape
,
8418 .quiesce
= raid5_quiesce
,
8419 .takeover
= raid5_takeover
,
8420 .congested
= raid5_congested
,
8421 .change_consistency_policy
= raid5_change_consistency_policy
,
8424 static struct md_personality raid4_personality
=
8428 .owner
= THIS_MODULE
,
8429 .make_request
= raid5_make_request
,
8431 .start
= raid5_start
,
8433 .status
= raid5_status
,
8434 .error_handler
= raid5_error
,
8435 .hot_add_disk
= raid5_add_disk
,
8436 .hot_remove_disk
= raid5_remove_disk
,
8437 .spare_active
= raid5_spare_active
,
8438 .sync_request
= raid5_sync_request
,
8439 .resize
= raid5_resize
,
8441 .check_reshape
= raid5_check_reshape
,
8442 .start_reshape
= raid5_start_reshape
,
8443 .finish_reshape
= raid5_finish_reshape
,
8444 .quiesce
= raid5_quiesce
,
8445 .takeover
= raid4_takeover
,
8446 .congested
= raid5_congested
,
8447 .change_consistency_policy
= raid5_change_consistency_policy
,
8450 static int __init
raid5_init(void)
8454 raid5_wq
= alloc_workqueue("raid5wq",
8455 WQ_UNBOUND
|WQ_MEM_RECLAIM
|WQ_CPU_INTENSIVE
|WQ_SYSFS
, 0);
8459 ret
= cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE
,
8461 raid456_cpu_up_prepare
,
8464 destroy_workqueue(raid5_wq
);
8467 register_md_personality(&raid6_personality
);
8468 register_md_personality(&raid5_personality
);
8469 register_md_personality(&raid4_personality
);
8473 static void raid5_exit(void)
8475 unregister_md_personality(&raid6_personality
);
8476 unregister_md_personality(&raid5_personality
);
8477 unregister_md_personality(&raid4_personality
);
8478 cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE
);
8479 destroy_workqueue(raid5_wq
);
8482 module_init(raid5_init
);
8483 module_exit(raid5_exit
);
8484 MODULE_LICENSE("GPL");
8485 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8486 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8487 MODULE_ALIAS("md-raid5");
8488 MODULE_ALIAS("md-raid4");
8489 MODULE_ALIAS("md-level-5");
8490 MODULE_ALIAS("md-level-4");
8491 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8492 MODULE_ALIAS("md-raid6");
8493 MODULE_ALIAS("md-level-6");
8495 /* This used to be two separate modules, they were: */
8496 MODULE_ALIAS("raid5");
8497 MODULE_ALIAS("raid6");