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
;
1143 * If this is discard request, set bi_vcnt 0. We don't
1144 * want to confuse SCSI because SCSI will replace payload
1146 if (op
== REQ_OP_DISCARD
)
1149 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
1151 if (conf
->mddev
->gendisk
)
1152 trace_block_bio_remap(bi
->bi_disk
->queue
,
1153 bi
, disk_devt(conf
->mddev
->gendisk
),
1155 if (should_defer
&& op_is_write(op
))
1156 bio_list_add(&pending_bios
, bi
);
1158 generic_make_request(bi
);
1161 if (s
->syncing
|| s
->expanding
|| s
->expanded
1163 md_sync_acct(rrdev
->bdev
, STRIPE_SECTORS
);
1165 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
1167 bio_set_dev(rbi
, rrdev
->bdev
);
1168 bio_set_op_attrs(rbi
, op
, op_flags
);
1169 BUG_ON(!op_is_write(op
));
1170 rbi
->bi_end_io
= raid5_end_write_request
;
1171 rbi
->bi_private
= sh
;
1173 pr_debug("%s: for %llu schedule op %d on "
1174 "replacement disc %d\n",
1175 __func__
, (unsigned long long)sh
->sector
,
1177 atomic_inc(&sh
->count
);
1179 atomic_inc(&head_sh
->count
);
1180 if (use_new_offset(conf
, sh
))
1181 rbi
->bi_iter
.bi_sector
= (sh
->sector
1182 + rrdev
->new_data_offset
);
1184 rbi
->bi_iter
.bi_sector
= (sh
->sector
1185 + rrdev
->data_offset
);
1186 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1187 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1188 sh
->dev
[i
].rvec
.bv_page
= sh
->dev
[i
].page
;
1190 rbi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
1191 rbi
->bi_io_vec
[0].bv_offset
= 0;
1192 rbi
->bi_iter
.bi_size
= STRIPE_SIZE
;
1194 * If this is discard request, set bi_vcnt 0. We don't
1195 * want to confuse SCSI because SCSI will replace payload
1197 if (op
== REQ_OP_DISCARD
)
1199 if (conf
->mddev
->gendisk
)
1200 trace_block_bio_remap(rbi
->bi_disk
->queue
,
1201 rbi
, disk_devt(conf
->mddev
->gendisk
),
1203 if (should_defer
&& op_is_write(op
))
1204 bio_list_add(&pending_bios
, rbi
);
1206 generic_make_request(rbi
);
1208 if (!rdev
&& !rrdev
) {
1209 if (op_is_write(op
))
1210 set_bit(STRIPE_DEGRADED
, &sh
->state
);
1211 pr_debug("skip op %d on disc %d for sector %llu\n",
1212 bi
->bi_opf
, i
, (unsigned long long)sh
->sector
);
1213 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1214 set_bit(STRIPE_HANDLE
, &sh
->state
);
1217 if (!head_sh
->batch_head
)
1219 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1225 if (should_defer
&& !bio_list_empty(&pending_bios
))
1226 defer_issue_bios(conf
, head_sh
->sector
, &pending_bios
);
1229 static struct dma_async_tx_descriptor
*
1230 async_copy_data(int frombio
, struct bio
*bio
, struct page
**page
,
1231 sector_t sector
, struct dma_async_tx_descriptor
*tx
,
1232 struct stripe_head
*sh
, int no_skipcopy
)
1235 struct bvec_iter iter
;
1236 struct page
*bio_page
;
1238 struct async_submit_ctl submit
;
1239 enum async_tx_flags flags
= 0;
1241 if (bio
->bi_iter
.bi_sector
>= sector
)
1242 page_offset
= (signed)(bio
->bi_iter
.bi_sector
- sector
) * 512;
1244 page_offset
= (signed)(sector
- bio
->bi_iter
.bi_sector
) * -512;
1247 flags
|= ASYNC_TX_FENCE
;
1248 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
1250 bio_for_each_segment(bvl
, bio
, iter
) {
1251 int len
= bvl
.bv_len
;
1255 if (page_offset
< 0) {
1256 b_offset
= -page_offset
;
1257 page_offset
+= b_offset
;
1261 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
1262 clen
= STRIPE_SIZE
- page_offset
;
1267 b_offset
+= bvl
.bv_offset
;
1268 bio_page
= bvl
.bv_page
;
1270 if (sh
->raid_conf
->skip_copy
&&
1271 b_offset
== 0 && page_offset
== 0 &&
1272 clen
== STRIPE_SIZE
&&
1276 tx
= async_memcpy(*page
, bio_page
, page_offset
,
1277 b_offset
, clen
, &submit
);
1279 tx
= async_memcpy(bio_page
, *page
, b_offset
,
1280 page_offset
, clen
, &submit
);
1282 /* chain the operations */
1283 submit
.depend_tx
= tx
;
1285 if (clen
< len
) /* hit end of page */
1293 static void ops_complete_biofill(void *stripe_head_ref
)
1295 struct stripe_head
*sh
= stripe_head_ref
;
1298 pr_debug("%s: stripe %llu\n", __func__
,
1299 (unsigned long long)sh
->sector
);
1301 /* clear completed biofills */
1302 for (i
= sh
->disks
; i
--; ) {
1303 struct r5dev
*dev
= &sh
->dev
[i
];
1305 /* acknowledge completion of a biofill operation */
1306 /* and check if we need to reply to a read request,
1307 * new R5_Wantfill requests are held off until
1308 * !STRIPE_BIOFILL_RUN
1310 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
1311 struct bio
*rbi
, *rbi2
;
1316 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1317 dev
->sector
+ STRIPE_SECTORS
) {
1318 rbi2
= r5_next_bio(rbi
, dev
->sector
);
1324 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
1326 set_bit(STRIPE_HANDLE
, &sh
->state
);
1327 raid5_release_stripe(sh
);
1330 static void ops_run_biofill(struct stripe_head
*sh
)
1332 struct dma_async_tx_descriptor
*tx
= NULL
;
1333 struct async_submit_ctl submit
;
1336 BUG_ON(sh
->batch_head
);
1337 pr_debug("%s: stripe %llu\n", __func__
,
1338 (unsigned long long)sh
->sector
);
1340 for (i
= sh
->disks
; i
--; ) {
1341 struct r5dev
*dev
= &sh
->dev
[i
];
1342 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
1344 spin_lock_irq(&sh
->stripe_lock
);
1345 dev
->read
= rbi
= dev
->toread
;
1347 spin_unlock_irq(&sh
->stripe_lock
);
1348 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1349 dev
->sector
+ STRIPE_SECTORS
) {
1350 tx
= async_copy_data(0, rbi
, &dev
->page
,
1351 dev
->sector
, tx
, sh
, 0);
1352 rbi
= r5_next_bio(rbi
, dev
->sector
);
1357 atomic_inc(&sh
->count
);
1358 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
1359 async_trigger_callback(&submit
);
1362 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
1369 tgt
= &sh
->dev
[target
];
1370 set_bit(R5_UPTODATE
, &tgt
->flags
);
1371 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1372 clear_bit(R5_Wantcompute
, &tgt
->flags
);
1375 static void ops_complete_compute(void *stripe_head_ref
)
1377 struct stripe_head
*sh
= stripe_head_ref
;
1379 pr_debug("%s: stripe %llu\n", __func__
,
1380 (unsigned long long)sh
->sector
);
1382 /* mark the computed target(s) as uptodate */
1383 mark_target_uptodate(sh
, sh
->ops
.target
);
1384 mark_target_uptodate(sh
, sh
->ops
.target2
);
1386 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
1387 if (sh
->check_state
== check_state_compute_run
)
1388 sh
->check_state
= check_state_compute_result
;
1389 set_bit(STRIPE_HANDLE
, &sh
->state
);
1390 raid5_release_stripe(sh
);
1393 /* return a pointer to the address conversion region of the scribble buffer */
1394 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
1395 struct raid5_percpu
*percpu
, int i
)
1399 addr
= flex_array_get(percpu
->scribble
, i
);
1400 return addr
+ sizeof(struct page
*) * (sh
->disks
+ 2);
1403 /* return a pointer to the address conversion region of the scribble buffer */
1404 static struct page
**to_addr_page(struct raid5_percpu
*percpu
, int i
)
1408 addr
= flex_array_get(percpu
->scribble
, i
);
1412 static struct dma_async_tx_descriptor
*
1413 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1415 int disks
= sh
->disks
;
1416 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1417 int target
= sh
->ops
.target
;
1418 struct r5dev
*tgt
= &sh
->dev
[target
];
1419 struct page
*xor_dest
= tgt
->page
;
1421 struct dma_async_tx_descriptor
*tx
;
1422 struct async_submit_ctl submit
;
1425 BUG_ON(sh
->batch_head
);
1427 pr_debug("%s: stripe %llu block: %d\n",
1428 __func__
, (unsigned long long)sh
->sector
, target
);
1429 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1431 for (i
= disks
; i
--; )
1433 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1435 atomic_inc(&sh
->count
);
1437 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
1438 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
, 0));
1439 if (unlikely(count
== 1))
1440 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1442 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1447 /* set_syndrome_sources - populate source buffers for gen_syndrome
1448 * @srcs - (struct page *) array of size sh->disks
1449 * @sh - stripe_head to parse
1451 * Populates srcs in proper layout order for the stripe and returns the
1452 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1453 * destination buffer is recorded in srcs[count] and the Q destination
1454 * is recorded in srcs[count+1]].
1456 static int set_syndrome_sources(struct page
**srcs
,
1457 struct stripe_head
*sh
,
1460 int disks
= sh
->disks
;
1461 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
1462 int d0_idx
= raid6_d0(sh
);
1466 for (i
= 0; i
< disks
; i
++)
1472 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1473 struct r5dev
*dev
= &sh
->dev
[i
];
1475 if (i
== sh
->qd_idx
|| i
== sh
->pd_idx
||
1476 (srctype
== SYNDROME_SRC_ALL
) ||
1477 (srctype
== SYNDROME_SRC_WANT_DRAIN
&&
1478 (test_bit(R5_Wantdrain
, &dev
->flags
) ||
1479 test_bit(R5_InJournal
, &dev
->flags
))) ||
1480 (srctype
== SYNDROME_SRC_WRITTEN
&&
1482 test_bit(R5_InJournal
, &dev
->flags
)))) {
1483 if (test_bit(R5_InJournal
, &dev
->flags
))
1484 srcs
[slot
] = sh
->dev
[i
].orig_page
;
1486 srcs
[slot
] = sh
->dev
[i
].page
;
1488 i
= raid6_next_disk(i
, disks
);
1489 } while (i
!= d0_idx
);
1491 return syndrome_disks
;
1494 static struct dma_async_tx_descriptor
*
1495 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1497 int disks
= sh
->disks
;
1498 struct page
**blocks
= to_addr_page(percpu
, 0);
1500 int qd_idx
= sh
->qd_idx
;
1501 struct dma_async_tx_descriptor
*tx
;
1502 struct async_submit_ctl submit
;
1508 BUG_ON(sh
->batch_head
);
1509 if (sh
->ops
.target
< 0)
1510 target
= sh
->ops
.target2
;
1511 else if (sh
->ops
.target2
< 0)
1512 target
= sh
->ops
.target
;
1514 /* we should only have one valid target */
1517 pr_debug("%s: stripe %llu block: %d\n",
1518 __func__
, (unsigned long long)sh
->sector
, target
);
1520 tgt
= &sh
->dev
[target
];
1521 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1524 atomic_inc(&sh
->count
);
1526 if (target
== qd_idx
) {
1527 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1528 blocks
[count
] = NULL
; /* regenerating p is not necessary */
1529 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
1530 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1531 ops_complete_compute
, sh
,
1532 to_addr_conv(sh
, percpu
, 0));
1533 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1535 /* Compute any data- or p-drive using XOR */
1537 for (i
= disks
; i
-- ; ) {
1538 if (i
== target
|| i
== qd_idx
)
1540 blocks
[count
++] = sh
->dev
[i
].page
;
1543 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1544 NULL
, ops_complete_compute
, sh
,
1545 to_addr_conv(sh
, percpu
, 0));
1546 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
1552 static struct dma_async_tx_descriptor
*
1553 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1555 int i
, count
, disks
= sh
->disks
;
1556 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1557 int d0_idx
= raid6_d0(sh
);
1558 int faila
= -1, failb
= -1;
1559 int target
= sh
->ops
.target
;
1560 int target2
= sh
->ops
.target2
;
1561 struct r5dev
*tgt
= &sh
->dev
[target
];
1562 struct r5dev
*tgt2
= &sh
->dev
[target2
];
1563 struct dma_async_tx_descriptor
*tx
;
1564 struct page
**blocks
= to_addr_page(percpu
, 0);
1565 struct async_submit_ctl submit
;
1567 BUG_ON(sh
->batch_head
);
1568 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1569 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
1570 BUG_ON(target
< 0 || target2
< 0);
1571 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1572 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
1574 /* we need to open-code set_syndrome_sources to handle the
1575 * slot number conversion for 'faila' and 'failb'
1577 for (i
= 0; i
< disks
; i
++)
1582 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1584 blocks
[slot
] = sh
->dev
[i
].page
;
1590 i
= raid6_next_disk(i
, disks
);
1591 } while (i
!= d0_idx
);
1593 BUG_ON(faila
== failb
);
1596 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1597 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1599 atomic_inc(&sh
->count
);
1601 if (failb
== syndrome_disks
+1) {
1602 /* Q disk is one of the missing disks */
1603 if (faila
== syndrome_disks
) {
1604 /* Missing P+Q, just recompute */
1605 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1606 ops_complete_compute
, sh
,
1607 to_addr_conv(sh
, percpu
, 0));
1608 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
1609 STRIPE_SIZE
, &submit
);
1613 int qd_idx
= sh
->qd_idx
;
1615 /* Missing D+Q: recompute D from P, then recompute Q */
1616 if (target
== qd_idx
)
1617 data_target
= target2
;
1619 data_target
= target
;
1622 for (i
= disks
; i
-- ; ) {
1623 if (i
== data_target
|| i
== qd_idx
)
1625 blocks
[count
++] = sh
->dev
[i
].page
;
1627 dest
= sh
->dev
[data_target
].page
;
1628 init_async_submit(&submit
,
1629 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1631 to_addr_conv(sh
, percpu
, 0));
1632 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
1635 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1636 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1637 ops_complete_compute
, sh
,
1638 to_addr_conv(sh
, percpu
, 0));
1639 return async_gen_syndrome(blocks
, 0, count
+2,
1640 STRIPE_SIZE
, &submit
);
1643 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1644 ops_complete_compute
, sh
,
1645 to_addr_conv(sh
, percpu
, 0));
1646 if (failb
== syndrome_disks
) {
1647 /* We're missing D+P. */
1648 return async_raid6_datap_recov(syndrome_disks
+2,
1652 /* We're missing D+D. */
1653 return async_raid6_2data_recov(syndrome_disks
+2,
1654 STRIPE_SIZE
, faila
, failb
,
1660 static void ops_complete_prexor(void *stripe_head_ref
)
1662 struct stripe_head
*sh
= stripe_head_ref
;
1664 pr_debug("%s: stripe %llu\n", __func__
,
1665 (unsigned long long)sh
->sector
);
1667 if (r5c_is_writeback(sh
->raid_conf
->log
))
1669 * raid5-cache write back uses orig_page during prexor.
1670 * After prexor, it is time to free orig_page
1672 r5c_release_extra_page(sh
);
1675 static struct dma_async_tx_descriptor
*
1676 ops_run_prexor5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1677 struct dma_async_tx_descriptor
*tx
)
1679 int disks
= sh
->disks
;
1680 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1681 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1682 struct async_submit_ctl submit
;
1684 /* existing parity data subtracted */
1685 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1687 BUG_ON(sh
->batch_head
);
1688 pr_debug("%s: stripe %llu\n", __func__
,
1689 (unsigned long long)sh
->sector
);
1691 for (i
= disks
; i
--; ) {
1692 struct r5dev
*dev
= &sh
->dev
[i
];
1693 /* Only process blocks that are known to be uptodate */
1694 if (test_bit(R5_InJournal
, &dev
->flags
))
1695 xor_srcs
[count
++] = dev
->orig_page
;
1696 else if (test_bit(R5_Wantdrain
, &dev
->flags
))
1697 xor_srcs
[count
++] = dev
->page
;
1700 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1701 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1702 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1707 static struct dma_async_tx_descriptor
*
1708 ops_run_prexor6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1709 struct dma_async_tx_descriptor
*tx
)
1711 struct page
**blocks
= to_addr_page(percpu
, 0);
1713 struct async_submit_ctl submit
;
1715 pr_debug("%s: stripe %llu\n", __func__
,
1716 (unsigned long long)sh
->sector
);
1718 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_WANT_DRAIN
);
1720 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_PQ_XOR_DST
, tx
,
1721 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1722 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1727 static struct dma_async_tx_descriptor
*
1728 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1730 struct r5conf
*conf
= sh
->raid_conf
;
1731 int disks
= sh
->disks
;
1733 struct stripe_head
*head_sh
= sh
;
1735 pr_debug("%s: stripe %llu\n", __func__
,
1736 (unsigned long long)sh
->sector
);
1738 for (i
= disks
; i
--; ) {
1743 if (test_and_clear_bit(R5_Wantdrain
, &head_sh
->dev
[i
].flags
)) {
1749 * clear R5_InJournal, so when rewriting a page in
1750 * journal, it is not skipped by r5l_log_stripe()
1752 clear_bit(R5_InJournal
, &dev
->flags
);
1753 spin_lock_irq(&sh
->stripe_lock
);
1754 chosen
= dev
->towrite
;
1755 dev
->towrite
= NULL
;
1756 sh
->overwrite_disks
= 0;
1757 BUG_ON(dev
->written
);
1758 wbi
= dev
->written
= chosen
;
1759 spin_unlock_irq(&sh
->stripe_lock
);
1760 WARN_ON(dev
->page
!= dev
->orig_page
);
1762 while (wbi
&& wbi
->bi_iter
.bi_sector
<
1763 dev
->sector
+ STRIPE_SECTORS
) {
1764 if (wbi
->bi_opf
& REQ_FUA
)
1765 set_bit(R5_WantFUA
, &dev
->flags
);
1766 if (wbi
->bi_opf
& REQ_SYNC
)
1767 set_bit(R5_SyncIO
, &dev
->flags
);
1768 if (bio_op(wbi
) == REQ_OP_DISCARD
)
1769 set_bit(R5_Discard
, &dev
->flags
);
1771 tx
= async_copy_data(1, wbi
, &dev
->page
,
1772 dev
->sector
, tx
, sh
,
1773 r5c_is_writeback(conf
->log
));
1774 if (dev
->page
!= dev
->orig_page
&&
1775 !r5c_is_writeback(conf
->log
)) {
1776 set_bit(R5_SkipCopy
, &dev
->flags
);
1777 clear_bit(R5_UPTODATE
, &dev
->flags
);
1778 clear_bit(R5_OVERWRITE
, &dev
->flags
);
1781 wbi
= r5_next_bio(wbi
, dev
->sector
);
1784 if (head_sh
->batch_head
) {
1785 sh
= list_first_entry(&sh
->batch_list
,
1798 static void ops_complete_reconstruct(void *stripe_head_ref
)
1800 struct stripe_head
*sh
= stripe_head_ref
;
1801 int disks
= sh
->disks
;
1802 int pd_idx
= sh
->pd_idx
;
1803 int qd_idx
= sh
->qd_idx
;
1805 bool fua
= false, sync
= false, discard
= false;
1807 pr_debug("%s: stripe %llu\n", __func__
,
1808 (unsigned long long)sh
->sector
);
1810 for (i
= disks
; i
--; ) {
1811 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1812 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
1813 discard
|= test_bit(R5_Discard
, &sh
->dev
[i
].flags
);
1816 for (i
= disks
; i
--; ) {
1817 struct r5dev
*dev
= &sh
->dev
[i
];
1819 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1820 if (!discard
&& !test_bit(R5_SkipCopy
, &dev
->flags
)) {
1821 set_bit(R5_UPTODATE
, &dev
->flags
);
1822 if (test_bit(STRIPE_EXPAND_READY
, &sh
->state
))
1823 set_bit(R5_Expanded
, &dev
->flags
);
1826 set_bit(R5_WantFUA
, &dev
->flags
);
1828 set_bit(R5_SyncIO
, &dev
->flags
);
1832 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1833 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1834 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1835 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1837 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1838 sh
->reconstruct_state
= reconstruct_state_result
;
1841 set_bit(STRIPE_HANDLE
, &sh
->state
);
1842 raid5_release_stripe(sh
);
1846 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1847 struct dma_async_tx_descriptor
*tx
)
1849 int disks
= sh
->disks
;
1850 struct page
**xor_srcs
;
1851 struct async_submit_ctl submit
;
1852 int count
, pd_idx
= sh
->pd_idx
, i
;
1853 struct page
*xor_dest
;
1855 unsigned long flags
;
1857 struct stripe_head
*head_sh
= sh
;
1860 pr_debug("%s: stripe %llu\n", __func__
,
1861 (unsigned long long)sh
->sector
);
1863 for (i
= 0; i
< sh
->disks
; i
++) {
1866 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1869 if (i
>= sh
->disks
) {
1870 atomic_inc(&sh
->count
);
1871 set_bit(R5_Discard
, &sh
->dev
[pd_idx
].flags
);
1872 ops_complete_reconstruct(sh
);
1877 xor_srcs
= to_addr_page(percpu
, j
);
1878 /* check if prexor is active which means only process blocks
1879 * that are part of a read-modify-write (written)
1881 if (head_sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1883 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1884 for (i
= disks
; i
--; ) {
1885 struct r5dev
*dev
= &sh
->dev
[i
];
1886 if (head_sh
->dev
[i
].written
||
1887 test_bit(R5_InJournal
, &head_sh
->dev
[i
].flags
))
1888 xor_srcs
[count
++] = dev
->page
;
1891 xor_dest
= sh
->dev
[pd_idx
].page
;
1892 for (i
= disks
; i
--; ) {
1893 struct r5dev
*dev
= &sh
->dev
[i
];
1895 xor_srcs
[count
++] = dev
->page
;
1899 /* 1/ if we prexor'd then the dest is reused as a source
1900 * 2/ if we did not prexor then we are redoing the parity
1901 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1902 * for the synchronous xor case
1904 last_stripe
= !head_sh
->batch_head
||
1905 list_first_entry(&sh
->batch_list
,
1906 struct stripe_head
, batch_list
) == head_sh
;
1908 flags
= ASYNC_TX_ACK
|
1909 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1911 atomic_inc(&head_sh
->count
);
1912 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, head_sh
,
1913 to_addr_conv(sh
, percpu
, j
));
1915 flags
= prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
;
1916 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
,
1917 to_addr_conv(sh
, percpu
, j
));
1920 if (unlikely(count
== 1))
1921 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1923 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1926 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1933 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1934 struct dma_async_tx_descriptor
*tx
)
1936 struct async_submit_ctl submit
;
1937 struct page
**blocks
;
1938 int count
, i
, j
= 0;
1939 struct stripe_head
*head_sh
= sh
;
1942 unsigned long txflags
;
1944 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1946 for (i
= 0; i
< sh
->disks
; i
++) {
1947 if (sh
->pd_idx
== i
|| sh
->qd_idx
== i
)
1949 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1952 if (i
>= sh
->disks
) {
1953 atomic_inc(&sh
->count
);
1954 set_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
1955 set_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
1956 ops_complete_reconstruct(sh
);
1961 blocks
= to_addr_page(percpu
, j
);
1963 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1964 synflags
= SYNDROME_SRC_WRITTEN
;
1965 txflags
= ASYNC_TX_ACK
| ASYNC_TX_PQ_XOR_DST
;
1967 synflags
= SYNDROME_SRC_ALL
;
1968 txflags
= ASYNC_TX_ACK
;
1971 count
= set_syndrome_sources(blocks
, sh
, synflags
);
1972 last_stripe
= !head_sh
->batch_head
||
1973 list_first_entry(&sh
->batch_list
,
1974 struct stripe_head
, batch_list
) == head_sh
;
1977 atomic_inc(&head_sh
->count
);
1978 init_async_submit(&submit
, txflags
, tx
, ops_complete_reconstruct
,
1979 head_sh
, to_addr_conv(sh
, percpu
, j
));
1981 init_async_submit(&submit
, 0, tx
, NULL
, NULL
,
1982 to_addr_conv(sh
, percpu
, j
));
1983 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1986 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1992 static void ops_complete_check(void *stripe_head_ref
)
1994 struct stripe_head
*sh
= stripe_head_ref
;
1996 pr_debug("%s: stripe %llu\n", __func__
,
1997 (unsigned long long)sh
->sector
);
1999 sh
->check_state
= check_state_check_result
;
2000 set_bit(STRIPE_HANDLE
, &sh
->state
);
2001 raid5_release_stripe(sh
);
2004 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
2006 int disks
= sh
->disks
;
2007 int pd_idx
= sh
->pd_idx
;
2008 int qd_idx
= sh
->qd_idx
;
2009 struct page
*xor_dest
;
2010 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
2011 struct dma_async_tx_descriptor
*tx
;
2012 struct async_submit_ctl submit
;
2016 pr_debug("%s: stripe %llu\n", __func__
,
2017 (unsigned long long)sh
->sector
);
2019 BUG_ON(sh
->batch_head
);
2021 xor_dest
= sh
->dev
[pd_idx
].page
;
2022 xor_srcs
[count
++] = xor_dest
;
2023 for (i
= disks
; i
--; ) {
2024 if (i
== pd_idx
|| i
== qd_idx
)
2026 xor_srcs
[count
++] = sh
->dev
[i
].page
;
2029 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
2030 to_addr_conv(sh
, percpu
, 0));
2031 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
2032 &sh
->ops
.zero_sum_result
, &submit
);
2034 atomic_inc(&sh
->count
);
2035 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
2036 tx
= async_trigger_callback(&submit
);
2039 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
2041 struct page
**srcs
= to_addr_page(percpu
, 0);
2042 struct async_submit_ctl submit
;
2045 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
2046 (unsigned long long)sh
->sector
, checkp
);
2048 BUG_ON(sh
->batch_head
);
2049 count
= set_syndrome_sources(srcs
, sh
, SYNDROME_SRC_ALL
);
2053 atomic_inc(&sh
->count
);
2054 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
2055 sh
, to_addr_conv(sh
, percpu
, 0));
2056 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
2057 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
2060 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
2062 int overlap_clear
= 0, i
, disks
= sh
->disks
;
2063 struct dma_async_tx_descriptor
*tx
= NULL
;
2064 struct r5conf
*conf
= sh
->raid_conf
;
2065 int level
= conf
->level
;
2066 struct raid5_percpu
*percpu
;
2070 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
2071 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
2072 ops_run_biofill(sh
);
2076 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
2078 tx
= ops_run_compute5(sh
, percpu
);
2080 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
2081 tx
= ops_run_compute6_1(sh
, percpu
);
2083 tx
= ops_run_compute6_2(sh
, percpu
);
2085 /* terminate the chain if reconstruct is not set to be run */
2086 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
2090 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
)) {
2092 tx
= ops_run_prexor5(sh
, percpu
, tx
);
2094 tx
= ops_run_prexor6(sh
, percpu
, tx
);
2097 if (test_bit(STRIPE_OP_PARTIAL_PARITY
, &ops_request
))
2098 tx
= ops_run_partial_parity(sh
, percpu
, tx
);
2100 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
2101 tx
= ops_run_biodrain(sh
, tx
);
2105 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
2107 ops_run_reconstruct5(sh
, percpu
, tx
);
2109 ops_run_reconstruct6(sh
, percpu
, tx
);
2112 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
2113 if (sh
->check_state
== check_state_run
)
2114 ops_run_check_p(sh
, percpu
);
2115 else if (sh
->check_state
== check_state_run_q
)
2116 ops_run_check_pq(sh
, percpu
, 0);
2117 else if (sh
->check_state
== check_state_run_pq
)
2118 ops_run_check_pq(sh
, percpu
, 1);
2123 if (overlap_clear
&& !sh
->batch_head
)
2124 for (i
= disks
; i
--; ) {
2125 struct r5dev
*dev
= &sh
->dev
[i
];
2126 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
2127 wake_up(&sh
->raid_conf
->wait_for_overlap
);
2132 static void free_stripe(struct kmem_cache
*sc
, struct stripe_head
*sh
)
2135 __free_page(sh
->ppl_page
);
2136 kmem_cache_free(sc
, sh
);
2139 static struct stripe_head
*alloc_stripe(struct kmem_cache
*sc
, gfp_t gfp
,
2140 int disks
, struct r5conf
*conf
)
2142 struct stripe_head
*sh
;
2145 sh
= kmem_cache_zalloc(sc
, gfp
);
2147 spin_lock_init(&sh
->stripe_lock
);
2148 spin_lock_init(&sh
->batch_lock
);
2149 INIT_LIST_HEAD(&sh
->batch_list
);
2150 INIT_LIST_HEAD(&sh
->lru
);
2151 INIT_LIST_HEAD(&sh
->r5c
);
2152 INIT_LIST_HEAD(&sh
->log_list
);
2153 atomic_set(&sh
->count
, 1);
2154 sh
->raid_conf
= conf
;
2155 sh
->log_start
= MaxSector
;
2156 for (i
= 0; i
< disks
; i
++) {
2157 struct r5dev
*dev
= &sh
->dev
[i
];
2159 bio_init(&dev
->req
, &dev
->vec
, 1);
2160 bio_init(&dev
->rreq
, &dev
->rvec
, 1);
2163 if (raid5_has_ppl(conf
)) {
2164 sh
->ppl_page
= alloc_page(gfp
);
2165 if (!sh
->ppl_page
) {
2166 free_stripe(sc
, sh
);
2173 static int grow_one_stripe(struct r5conf
*conf
, gfp_t gfp
)
2175 struct stripe_head
*sh
;
2177 sh
= alloc_stripe(conf
->slab_cache
, gfp
, conf
->pool_size
, conf
);
2181 if (grow_buffers(sh
, gfp
)) {
2183 free_stripe(conf
->slab_cache
, sh
);
2186 sh
->hash_lock_index
=
2187 conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
;
2188 /* we just created an active stripe so... */
2189 atomic_inc(&conf
->active_stripes
);
2191 raid5_release_stripe(sh
);
2192 conf
->max_nr_stripes
++;
2196 static int grow_stripes(struct r5conf
*conf
, int num
)
2198 struct kmem_cache
*sc
;
2199 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
2201 if (conf
->mddev
->gendisk
)
2202 sprintf(conf
->cache_name
[0],
2203 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
2205 sprintf(conf
->cache_name
[0],
2206 "raid%d-%p", conf
->level
, conf
->mddev
);
2207 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
2209 conf
->active_name
= 0;
2210 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
2211 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
2215 conf
->slab_cache
= sc
;
2216 conf
->pool_size
= devs
;
2218 if (!grow_one_stripe(conf
, GFP_KERNEL
))
2225 * scribble_len - return the required size of the scribble region
2226 * @num - total number of disks in the array
2228 * The size must be enough to contain:
2229 * 1/ a struct page pointer for each device in the array +2
2230 * 2/ room to convert each entry in (1) to its corresponding dma
2231 * (dma_map_page()) or page (page_address()) address.
2233 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2234 * calculate over all devices (not just the data blocks), using zeros in place
2235 * of the P and Q blocks.
2237 static struct flex_array
*scribble_alloc(int num
, int cnt
, gfp_t flags
)
2239 struct flex_array
*ret
;
2242 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
2243 ret
= flex_array_alloc(len
, cnt
, flags
);
2246 /* always prealloc all elements, so no locking is required */
2247 if (flex_array_prealloc(ret
, 0, cnt
, flags
)) {
2248 flex_array_free(ret
);
2254 static int resize_chunks(struct r5conf
*conf
, int new_disks
, int new_sectors
)
2260 * Never shrink. And mddev_suspend() could deadlock if this is called
2261 * from raid5d. In that case, scribble_disks and scribble_sectors
2262 * should equal to new_disks and new_sectors
2264 if (conf
->scribble_disks
>= new_disks
&&
2265 conf
->scribble_sectors
>= new_sectors
)
2267 mddev_suspend(conf
->mddev
);
2269 for_each_present_cpu(cpu
) {
2270 struct raid5_percpu
*percpu
;
2271 struct flex_array
*scribble
;
2273 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
2274 scribble
= scribble_alloc(new_disks
,
2275 new_sectors
/ STRIPE_SECTORS
,
2279 flex_array_free(percpu
->scribble
);
2280 percpu
->scribble
= scribble
;
2287 mddev_resume(conf
->mddev
);
2289 conf
->scribble_disks
= new_disks
;
2290 conf
->scribble_sectors
= new_sectors
;
2295 static int resize_stripes(struct r5conf
*conf
, int newsize
)
2297 /* Make all the stripes able to hold 'newsize' devices.
2298 * New slots in each stripe get 'page' set to a new page.
2300 * This happens in stages:
2301 * 1/ create a new kmem_cache and allocate the required number of
2303 * 2/ gather all the old stripe_heads and transfer the pages across
2304 * to the new stripe_heads. This will have the side effect of
2305 * freezing the array as once all stripe_heads have been collected,
2306 * no IO will be possible. Old stripe heads are freed once their
2307 * pages have been transferred over, and the old kmem_cache is
2308 * freed when all stripes are done.
2309 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2310 * we simple return a failure status - no need to clean anything up.
2311 * 4/ allocate new pages for the new slots in the new stripe_heads.
2312 * If this fails, we don't bother trying the shrink the
2313 * stripe_heads down again, we just leave them as they are.
2314 * As each stripe_head is processed the new one is released into
2317 * Once step2 is started, we cannot afford to wait for a write,
2318 * so we use GFP_NOIO allocations.
2320 struct stripe_head
*osh
, *nsh
;
2321 LIST_HEAD(newstripes
);
2322 struct disk_info
*ndisks
;
2324 struct kmem_cache
*sc
;
2328 md_allow_write(conf
->mddev
);
2331 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
2332 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
2337 /* Need to ensure auto-resizing doesn't interfere */
2338 mutex_lock(&conf
->cache_size_mutex
);
2340 for (i
= conf
->max_nr_stripes
; i
; i
--) {
2341 nsh
= alloc_stripe(sc
, GFP_KERNEL
, newsize
, conf
);
2345 list_add(&nsh
->lru
, &newstripes
);
2348 /* didn't get enough, give up */
2349 while (!list_empty(&newstripes
)) {
2350 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2351 list_del(&nsh
->lru
);
2352 free_stripe(sc
, nsh
);
2354 kmem_cache_destroy(sc
);
2355 mutex_unlock(&conf
->cache_size_mutex
);
2358 /* Step 2 - Must use GFP_NOIO now.
2359 * OK, we have enough stripes, start collecting inactive
2360 * stripes and copying them over
2364 list_for_each_entry(nsh
, &newstripes
, lru
) {
2365 lock_device_hash_lock(conf
, hash
);
2366 wait_event_cmd(conf
->wait_for_stripe
,
2367 !list_empty(conf
->inactive_list
+ hash
),
2368 unlock_device_hash_lock(conf
, hash
),
2369 lock_device_hash_lock(conf
, hash
));
2370 osh
= get_free_stripe(conf
, hash
);
2371 unlock_device_hash_lock(conf
, hash
);
2373 for(i
=0; i
<conf
->pool_size
; i
++) {
2374 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
2375 nsh
->dev
[i
].orig_page
= osh
->dev
[i
].page
;
2377 nsh
->hash_lock_index
= hash
;
2378 free_stripe(conf
->slab_cache
, osh
);
2380 if (cnt
>= conf
->max_nr_stripes
/ NR_STRIPE_HASH_LOCKS
+
2381 !!((conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
) > hash
)) {
2386 kmem_cache_destroy(conf
->slab_cache
);
2389 * At this point, we are holding all the stripes so the array
2390 * is completely stalled, so now is a good time to resize
2391 * conf->disks and the scribble region
2393 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
2395 for (i
= 0; i
< conf
->pool_size
; i
++)
2396 ndisks
[i
] = conf
->disks
[i
];
2398 for (i
= conf
->pool_size
; i
< newsize
; i
++) {
2399 ndisks
[i
].extra_page
= alloc_page(GFP_NOIO
);
2400 if (!ndisks
[i
].extra_page
)
2405 for (i
= conf
->pool_size
; i
< newsize
; i
++)
2406 if (ndisks
[i
].extra_page
)
2407 put_page(ndisks
[i
].extra_page
);
2411 conf
->disks
= ndisks
;
2416 mutex_unlock(&conf
->cache_size_mutex
);
2418 conf
->slab_cache
= sc
;
2419 conf
->active_name
= 1-conf
->active_name
;
2421 /* Step 4, return new stripes to service */
2422 while(!list_empty(&newstripes
)) {
2423 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2424 list_del_init(&nsh
->lru
);
2426 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
2427 if (nsh
->dev
[i
].page
== NULL
) {
2428 struct page
*p
= alloc_page(GFP_NOIO
);
2429 nsh
->dev
[i
].page
= p
;
2430 nsh
->dev
[i
].orig_page
= p
;
2434 raid5_release_stripe(nsh
);
2436 /* critical section pass, GFP_NOIO no longer needed */
2439 conf
->pool_size
= newsize
;
2443 static int drop_one_stripe(struct r5conf
*conf
)
2445 struct stripe_head
*sh
;
2446 int hash
= (conf
->max_nr_stripes
- 1) & STRIPE_HASH_LOCKS_MASK
;
2448 spin_lock_irq(conf
->hash_locks
+ hash
);
2449 sh
= get_free_stripe(conf
, hash
);
2450 spin_unlock_irq(conf
->hash_locks
+ hash
);
2453 BUG_ON(atomic_read(&sh
->count
));
2455 free_stripe(conf
->slab_cache
, sh
);
2456 atomic_dec(&conf
->active_stripes
);
2457 conf
->max_nr_stripes
--;
2461 static void shrink_stripes(struct r5conf
*conf
)
2463 while (conf
->max_nr_stripes
&&
2464 drop_one_stripe(conf
))
2467 kmem_cache_destroy(conf
->slab_cache
);
2468 conf
->slab_cache
= NULL
;
2471 static void raid5_end_read_request(struct bio
* bi
)
2473 struct stripe_head
*sh
= bi
->bi_private
;
2474 struct r5conf
*conf
= sh
->raid_conf
;
2475 int disks
= sh
->disks
, i
;
2476 char b
[BDEVNAME_SIZE
];
2477 struct md_rdev
*rdev
= NULL
;
2480 for (i
=0 ; i
<disks
; i
++)
2481 if (bi
== &sh
->dev
[i
].req
)
2484 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2485 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2492 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2493 /* If replacement finished while this request was outstanding,
2494 * 'replacement' might be NULL already.
2495 * In that case it moved down to 'rdev'.
2496 * rdev is not removed until all requests are finished.
2498 rdev
= conf
->disks
[i
].replacement
;
2500 rdev
= conf
->disks
[i
].rdev
;
2502 if (use_new_offset(conf
, sh
))
2503 s
= sh
->sector
+ rdev
->new_data_offset
;
2505 s
= sh
->sector
+ rdev
->data_offset
;
2506 if (!bi
->bi_status
) {
2507 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2508 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2509 /* Note that this cannot happen on a
2510 * replacement device. We just fail those on
2513 pr_info_ratelimited(
2514 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2515 mdname(conf
->mddev
), STRIPE_SECTORS
,
2516 (unsigned long long)s
,
2517 bdevname(rdev
->bdev
, b
));
2518 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
2519 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2520 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2521 } else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2522 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2524 if (test_bit(R5_InJournal
, &sh
->dev
[i
].flags
))
2526 * end read for a page in journal, this
2527 * must be preparing for prexor in rmw
2529 set_bit(R5_OrigPageUPTDODATE
, &sh
->dev
[i
].flags
);
2531 if (atomic_read(&rdev
->read_errors
))
2532 atomic_set(&rdev
->read_errors
, 0);
2534 const char *bdn
= bdevname(rdev
->bdev
, b
);
2538 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2539 atomic_inc(&rdev
->read_errors
);
2540 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2541 pr_warn_ratelimited(
2542 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2543 mdname(conf
->mddev
),
2544 (unsigned long long)s
,
2546 else if (conf
->mddev
->degraded
>= conf
->max_degraded
) {
2548 pr_warn_ratelimited(
2549 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2550 mdname(conf
->mddev
),
2551 (unsigned long long)s
,
2553 } else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
)) {
2556 pr_warn_ratelimited(
2557 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2558 mdname(conf
->mddev
),
2559 (unsigned long long)s
,
2561 } else if (atomic_read(&rdev
->read_errors
)
2562 > conf
->max_nr_stripes
)
2563 pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2564 mdname(conf
->mddev
), bdn
);
2567 if (set_bad
&& test_bit(In_sync
, &rdev
->flags
)
2568 && !test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2571 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
)) {
2572 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2573 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2575 set_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2577 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2578 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2580 && test_bit(In_sync
, &rdev
->flags
)
2581 && rdev_set_badblocks(
2582 rdev
, sh
->sector
, STRIPE_SECTORS
, 0)))
2583 md_error(conf
->mddev
, rdev
);
2586 rdev_dec_pending(rdev
, conf
->mddev
);
2588 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2589 set_bit(STRIPE_HANDLE
, &sh
->state
);
2590 raid5_release_stripe(sh
);
2593 static void raid5_end_write_request(struct bio
*bi
)
2595 struct stripe_head
*sh
= bi
->bi_private
;
2596 struct r5conf
*conf
= sh
->raid_conf
;
2597 int disks
= sh
->disks
, i
;
2598 struct md_rdev
*uninitialized_var(rdev
);
2601 int replacement
= 0;
2603 for (i
= 0 ; i
< disks
; i
++) {
2604 if (bi
== &sh
->dev
[i
].req
) {
2605 rdev
= conf
->disks
[i
].rdev
;
2608 if (bi
== &sh
->dev
[i
].rreq
) {
2609 rdev
= conf
->disks
[i
].replacement
;
2613 /* rdev was removed and 'replacement'
2614 * replaced it. rdev is not removed
2615 * until all requests are finished.
2617 rdev
= conf
->disks
[i
].rdev
;
2621 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2622 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2632 md_error(conf
->mddev
, rdev
);
2633 else if (is_badblock(rdev
, sh
->sector
,
2635 &first_bad
, &bad_sectors
))
2636 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
2638 if (bi
->bi_status
) {
2639 set_bit(STRIPE_DEGRADED
, &sh
->state
);
2640 set_bit(WriteErrorSeen
, &rdev
->flags
);
2641 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
2642 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2643 set_bit(MD_RECOVERY_NEEDED
,
2644 &rdev
->mddev
->recovery
);
2645 } else if (is_badblock(rdev
, sh
->sector
,
2647 &first_bad
, &bad_sectors
)) {
2648 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
2649 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))
2650 /* That was a successful write so make
2651 * sure it looks like we already did
2654 set_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2657 rdev_dec_pending(rdev
, conf
->mddev
);
2659 if (sh
->batch_head
&& bi
->bi_status
&& !replacement
)
2660 set_bit(STRIPE_BATCH_ERR
, &sh
->batch_head
->state
);
2663 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
2664 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2665 set_bit(STRIPE_HANDLE
, &sh
->state
);
2666 raid5_release_stripe(sh
);
2668 if (sh
->batch_head
&& sh
!= sh
->batch_head
)
2669 raid5_release_stripe(sh
->batch_head
);
2672 static void raid5_error(struct mddev
*mddev
, struct md_rdev
*rdev
)
2674 char b
[BDEVNAME_SIZE
];
2675 struct r5conf
*conf
= mddev
->private;
2676 unsigned long flags
;
2677 pr_debug("raid456: error called\n");
2679 spin_lock_irqsave(&conf
->device_lock
, flags
);
2680 set_bit(Faulty
, &rdev
->flags
);
2681 clear_bit(In_sync
, &rdev
->flags
);
2682 mddev
->degraded
= raid5_calc_degraded(conf
);
2683 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2684 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2686 set_bit(Blocked
, &rdev
->flags
);
2687 set_mask_bits(&mddev
->sb_flags
, 0,
2688 BIT(MD_SB_CHANGE_DEVS
) | BIT(MD_SB_CHANGE_PENDING
));
2689 pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2690 "md/raid:%s: Operation continuing on %d devices.\n",
2692 bdevname(rdev
->bdev
, b
),
2694 conf
->raid_disks
- mddev
->degraded
);
2695 r5c_update_on_rdev_error(mddev
, rdev
);
2699 * Input: a 'big' sector number,
2700 * Output: index of the data and parity disk, and the sector # in them.
2702 sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
2703 int previous
, int *dd_idx
,
2704 struct stripe_head
*sh
)
2706 sector_t stripe
, stripe2
;
2707 sector_t chunk_number
;
2708 unsigned int chunk_offset
;
2711 sector_t new_sector
;
2712 int algorithm
= previous
? conf
->prev_algo
2714 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2715 : conf
->chunk_sectors
;
2716 int raid_disks
= previous
? conf
->previous_raid_disks
2718 int data_disks
= raid_disks
- conf
->max_degraded
;
2720 /* First compute the information on this sector */
2723 * Compute the chunk number and the sector offset inside the chunk
2725 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
2726 chunk_number
= r_sector
;
2729 * Compute the stripe number
2731 stripe
= chunk_number
;
2732 *dd_idx
= sector_div(stripe
, data_disks
);
2735 * Select the parity disk based on the user selected algorithm.
2737 pd_idx
= qd_idx
= -1;
2738 switch(conf
->level
) {
2740 pd_idx
= data_disks
;
2743 switch (algorithm
) {
2744 case ALGORITHM_LEFT_ASYMMETRIC
:
2745 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2746 if (*dd_idx
>= pd_idx
)
2749 case ALGORITHM_RIGHT_ASYMMETRIC
:
2750 pd_idx
= sector_div(stripe2
, raid_disks
);
2751 if (*dd_idx
>= pd_idx
)
2754 case ALGORITHM_LEFT_SYMMETRIC
:
2755 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2756 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2758 case ALGORITHM_RIGHT_SYMMETRIC
:
2759 pd_idx
= sector_div(stripe2
, raid_disks
);
2760 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2762 case ALGORITHM_PARITY_0
:
2766 case ALGORITHM_PARITY_N
:
2767 pd_idx
= data_disks
;
2775 switch (algorithm
) {
2776 case ALGORITHM_LEFT_ASYMMETRIC
:
2777 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2778 qd_idx
= pd_idx
+ 1;
2779 if (pd_idx
== raid_disks
-1) {
2780 (*dd_idx
)++; /* Q D D D P */
2782 } else if (*dd_idx
>= pd_idx
)
2783 (*dd_idx
) += 2; /* D D P Q D */
2785 case ALGORITHM_RIGHT_ASYMMETRIC
:
2786 pd_idx
= sector_div(stripe2
, raid_disks
);
2787 qd_idx
= pd_idx
+ 1;
2788 if (pd_idx
== raid_disks
-1) {
2789 (*dd_idx
)++; /* Q D D D P */
2791 } else if (*dd_idx
>= pd_idx
)
2792 (*dd_idx
) += 2; /* D D P Q D */
2794 case ALGORITHM_LEFT_SYMMETRIC
:
2795 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2796 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2797 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2799 case ALGORITHM_RIGHT_SYMMETRIC
:
2800 pd_idx
= sector_div(stripe2
, raid_disks
);
2801 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2802 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2805 case ALGORITHM_PARITY_0
:
2810 case ALGORITHM_PARITY_N
:
2811 pd_idx
= data_disks
;
2812 qd_idx
= data_disks
+ 1;
2815 case ALGORITHM_ROTATING_ZERO_RESTART
:
2816 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2817 * of blocks for computing Q is different.
2819 pd_idx
= sector_div(stripe2
, raid_disks
);
2820 qd_idx
= pd_idx
+ 1;
2821 if (pd_idx
== raid_disks
-1) {
2822 (*dd_idx
)++; /* Q D D D P */
2824 } else if (*dd_idx
>= pd_idx
)
2825 (*dd_idx
) += 2; /* D D P Q D */
2829 case ALGORITHM_ROTATING_N_RESTART
:
2830 /* Same a left_asymmetric, by first stripe is
2831 * D D D P Q rather than
2835 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2836 qd_idx
= pd_idx
+ 1;
2837 if (pd_idx
== raid_disks
-1) {
2838 (*dd_idx
)++; /* Q D D D P */
2840 } else if (*dd_idx
>= pd_idx
)
2841 (*dd_idx
) += 2; /* D D P Q D */
2845 case ALGORITHM_ROTATING_N_CONTINUE
:
2846 /* Same as left_symmetric but Q is before P */
2847 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2848 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
2849 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2853 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2854 /* RAID5 left_asymmetric, with Q on last device */
2855 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2856 if (*dd_idx
>= pd_idx
)
2858 qd_idx
= raid_disks
- 1;
2861 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2862 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2863 if (*dd_idx
>= pd_idx
)
2865 qd_idx
= raid_disks
- 1;
2868 case ALGORITHM_LEFT_SYMMETRIC_6
:
2869 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2870 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2871 qd_idx
= raid_disks
- 1;
2874 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2875 pd_idx
= 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_PARITY_0_6
:
2883 qd_idx
= raid_disks
- 1;
2893 sh
->pd_idx
= pd_idx
;
2894 sh
->qd_idx
= qd_idx
;
2895 sh
->ddf_layout
= ddf_layout
;
2898 * Finally, compute the new sector number
2900 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
2904 sector_t
raid5_compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
2906 struct r5conf
*conf
= sh
->raid_conf
;
2907 int raid_disks
= sh
->disks
;
2908 int data_disks
= raid_disks
- conf
->max_degraded
;
2909 sector_t new_sector
= sh
->sector
, check
;
2910 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2911 : conf
->chunk_sectors
;
2912 int algorithm
= previous
? conf
->prev_algo
2916 sector_t chunk_number
;
2917 int dummy1
, dd_idx
= i
;
2919 struct stripe_head sh2
;
2921 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
2922 stripe
= new_sector
;
2924 if (i
== sh
->pd_idx
)
2926 switch(conf
->level
) {
2929 switch (algorithm
) {
2930 case ALGORITHM_LEFT_ASYMMETRIC
:
2931 case ALGORITHM_RIGHT_ASYMMETRIC
:
2935 case ALGORITHM_LEFT_SYMMETRIC
:
2936 case ALGORITHM_RIGHT_SYMMETRIC
:
2939 i
-= (sh
->pd_idx
+ 1);
2941 case ALGORITHM_PARITY_0
:
2944 case ALGORITHM_PARITY_N
:
2951 if (i
== sh
->qd_idx
)
2952 return 0; /* It is the Q disk */
2953 switch (algorithm
) {
2954 case ALGORITHM_LEFT_ASYMMETRIC
:
2955 case ALGORITHM_RIGHT_ASYMMETRIC
:
2956 case ALGORITHM_ROTATING_ZERO_RESTART
:
2957 case ALGORITHM_ROTATING_N_RESTART
:
2958 if (sh
->pd_idx
== raid_disks
-1)
2959 i
--; /* Q D D D P */
2960 else if (i
> sh
->pd_idx
)
2961 i
-= 2; /* D D P Q D */
2963 case ALGORITHM_LEFT_SYMMETRIC
:
2964 case ALGORITHM_RIGHT_SYMMETRIC
:
2965 if (sh
->pd_idx
== raid_disks
-1)
2966 i
--; /* Q D D D P */
2971 i
-= (sh
->pd_idx
+ 2);
2974 case ALGORITHM_PARITY_0
:
2977 case ALGORITHM_PARITY_N
:
2979 case ALGORITHM_ROTATING_N_CONTINUE
:
2980 /* Like left_symmetric, but P is before Q */
2981 if (sh
->pd_idx
== 0)
2982 i
--; /* P D D D Q */
2987 i
-= (sh
->pd_idx
+ 1);
2990 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2991 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2995 case ALGORITHM_LEFT_SYMMETRIC_6
:
2996 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2998 i
+= data_disks
+ 1;
2999 i
-= (sh
->pd_idx
+ 1);
3001 case ALGORITHM_PARITY_0_6
:
3010 chunk_number
= stripe
* data_disks
+ i
;
3011 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
3013 check
= raid5_compute_sector(conf
, r_sector
,
3014 previous
, &dummy1
, &sh2
);
3015 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
3016 || sh2
.qd_idx
!= sh
->qd_idx
) {
3017 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3018 mdname(conf
->mddev
));
3025 * There are cases where we want handle_stripe_dirtying() and
3026 * schedule_reconstruction() to delay towrite to some dev of a stripe.
3028 * This function checks whether we want to delay the towrite. Specifically,
3029 * we delay the towrite when:
3031 * 1. degraded stripe has a non-overwrite to the missing dev, AND this
3032 * stripe has data in journal (for other devices).
3034 * In this case, when reading data for the non-overwrite dev, it is
3035 * necessary to handle complex rmw of write back cache (prexor with
3036 * orig_page, and xor with page). To keep read path simple, we would
3037 * like to flush data in journal to RAID disks first, so complex rmw
3038 * is handled in the write patch (handle_stripe_dirtying).
3040 * 2. when journal space is critical (R5C_LOG_CRITICAL=1)
3042 * It is important to be able to flush all stripes in raid5-cache.
3043 * Therefore, we need reserve some space on the journal device for
3044 * these flushes. If flush operation includes pending writes to the
3045 * stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3046 * for the flush out. If we exclude these pending writes from flush
3047 * operation, we only need (conf->max_degraded + 1) pages per stripe.
3048 * Therefore, excluding pending writes in these cases enables more
3049 * efficient use of the journal device.
3051 * Note: To make sure the stripe makes progress, we only delay
3052 * towrite for stripes with data already in journal (injournal > 0).
3053 * When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3054 * no_space_stripes list.
3056 * 3. during journal failure
3057 * In journal failure, we try to flush all cached data to raid disks
3058 * based on data in stripe cache. The array is read-only to upper
3059 * layers, so we would skip all pending writes.
3062 static inline bool delay_towrite(struct r5conf
*conf
,
3064 struct stripe_head_state
*s
)
3067 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3068 !test_bit(R5_Insync
, &dev
->flags
) && s
->injournal
)
3071 if (test_bit(R5C_LOG_CRITICAL
, &conf
->cache_state
) &&
3075 if (s
->log_failed
&& s
->injournal
)
3081 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3082 int rcw
, int expand
)
3084 int i
, pd_idx
= sh
->pd_idx
, qd_idx
= sh
->qd_idx
, disks
= sh
->disks
;
3085 struct r5conf
*conf
= sh
->raid_conf
;
3086 int level
= conf
->level
;
3090 * In some cases, handle_stripe_dirtying initially decided to
3091 * run rmw and allocates extra page for prexor. However, rcw is
3092 * cheaper later on. We need to free the extra page now,
3093 * because we won't be able to do that in ops_complete_prexor().
3095 r5c_release_extra_page(sh
);
3097 for (i
= disks
; i
--; ) {
3098 struct r5dev
*dev
= &sh
->dev
[i
];
3100 if (dev
->towrite
&& !delay_towrite(conf
, dev
, s
)) {
3101 set_bit(R5_LOCKED
, &dev
->flags
);
3102 set_bit(R5_Wantdrain
, &dev
->flags
);
3104 clear_bit(R5_UPTODATE
, &dev
->flags
);
3106 } else if (test_bit(R5_InJournal
, &dev
->flags
)) {
3107 set_bit(R5_LOCKED
, &dev
->flags
);
3111 /* if we are not expanding this is a proper write request, and
3112 * there will be bios with new data to be drained into the
3117 /* False alarm, nothing to do */
3119 sh
->reconstruct_state
= reconstruct_state_drain_run
;
3120 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
3122 sh
->reconstruct_state
= reconstruct_state_run
;
3124 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
3126 if (s
->locked
+ conf
->max_degraded
== disks
)
3127 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3128 atomic_inc(&conf
->pending_full_writes
);
3130 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
3131 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
3132 BUG_ON(level
== 6 &&
3133 (!(test_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
) ||
3134 test_bit(R5_Wantcompute
, &sh
->dev
[qd_idx
].flags
))));
3136 for (i
= disks
; i
--; ) {
3137 struct r5dev
*dev
= &sh
->dev
[i
];
3138 if (i
== pd_idx
|| i
== qd_idx
)
3142 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
3143 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3144 set_bit(R5_Wantdrain
, &dev
->flags
);
3145 set_bit(R5_LOCKED
, &dev
->flags
);
3146 clear_bit(R5_UPTODATE
, &dev
->flags
);
3148 } else if (test_bit(R5_InJournal
, &dev
->flags
)) {
3149 set_bit(R5_LOCKED
, &dev
->flags
);
3154 /* False alarm - nothing to do */
3156 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
3157 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
3158 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
3159 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
3162 /* keep the parity disk(s) locked while asynchronous operations
3165 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
3166 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
3170 int qd_idx
= sh
->qd_idx
;
3171 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
3173 set_bit(R5_LOCKED
, &dev
->flags
);
3174 clear_bit(R5_UPTODATE
, &dev
->flags
);
3178 if (raid5_has_ppl(sh
->raid_conf
) && sh
->ppl_page
&&
3179 test_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
) &&
3180 !test_bit(STRIPE_FULL_WRITE
, &sh
->state
) &&
3181 test_bit(R5_Insync
, &sh
->dev
[pd_idx
].flags
))
3182 set_bit(STRIPE_OP_PARTIAL_PARITY
, &s
->ops_request
);
3184 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3185 __func__
, (unsigned long long)sh
->sector
,
3186 s
->locked
, s
->ops_request
);
3190 * Each stripe/dev can have one or more bion attached.
3191 * toread/towrite point to the first in a chain.
3192 * The bi_next chain must be in order.
3194 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
,
3195 int forwrite
, int previous
)
3198 struct r5conf
*conf
= sh
->raid_conf
;
3201 pr_debug("adding bi b#%llu to stripe s#%llu\n",
3202 (unsigned long long)bi
->bi_iter
.bi_sector
,
3203 (unsigned long long)sh
->sector
);
3205 spin_lock_irq(&sh
->stripe_lock
);
3206 /* Don't allow new IO added to stripes in batch list */
3210 bip
= &sh
->dev
[dd_idx
].towrite
;
3214 bip
= &sh
->dev
[dd_idx
].toread
;
3215 while (*bip
&& (*bip
)->bi_iter
.bi_sector
< bi
->bi_iter
.bi_sector
) {
3216 if (bio_end_sector(*bip
) > bi
->bi_iter
.bi_sector
)
3218 bip
= & (*bip
)->bi_next
;
3220 if (*bip
&& (*bip
)->bi_iter
.bi_sector
< bio_end_sector(bi
))
3223 if (forwrite
&& raid5_has_ppl(conf
)) {
3225 * With PPL only writes to consecutive data chunks within a
3226 * stripe are allowed because for a single stripe_head we can
3227 * only have one PPL entry at a time, which describes one data
3228 * range. Not really an overlap, but wait_for_overlap can be
3229 * used to handle this.
3237 for (i
= 0; i
< sh
->disks
; i
++) {
3238 if (i
!= sh
->pd_idx
&&
3239 (i
== dd_idx
|| sh
->dev
[i
].towrite
)) {
3240 sector
= sh
->dev
[i
].sector
;
3241 if (count
== 0 || sector
< first
)
3249 if (first
+ conf
->chunk_sectors
* (count
- 1) != last
)
3253 if (!forwrite
|| previous
)
3254 clear_bit(STRIPE_BATCH_READY
, &sh
->state
);
3256 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
3260 bio_inc_remaining(bi
);
3261 md_write_inc(conf
->mddev
, bi
);
3264 /* check if page is covered */
3265 sector_t sector
= sh
->dev
[dd_idx
].sector
;
3266 for (bi
=sh
->dev
[dd_idx
].towrite
;
3267 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
3268 bi
&& bi
->bi_iter
.bi_sector
<= sector
;
3269 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
3270 if (bio_end_sector(bi
) >= sector
)
3271 sector
= bio_end_sector(bi
);
3273 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
3274 if (!test_and_set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
))
3275 sh
->overwrite_disks
++;
3278 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3279 (unsigned long long)(*bip
)->bi_iter
.bi_sector
,
3280 (unsigned long long)sh
->sector
, dd_idx
);
3282 if (conf
->mddev
->bitmap
&& firstwrite
) {
3283 /* Cannot hold spinlock over bitmap_startwrite,
3284 * but must ensure this isn't added to a batch until
3285 * we have added to the bitmap and set bm_seq.
3286 * So set STRIPE_BITMAP_PENDING to prevent
3288 * If multiple add_stripe_bio() calls race here they
3289 * much all set STRIPE_BITMAP_PENDING. So only the first one
3290 * to complete "bitmap_startwrite" gets to set
3291 * STRIPE_BIT_DELAY. This is important as once a stripe
3292 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3295 set_bit(STRIPE_BITMAP_PENDING
, &sh
->state
);
3296 spin_unlock_irq(&sh
->stripe_lock
);
3297 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
3299 spin_lock_irq(&sh
->stripe_lock
);
3300 clear_bit(STRIPE_BITMAP_PENDING
, &sh
->state
);
3301 if (!sh
->batch_head
) {
3302 sh
->bm_seq
= conf
->seq_flush
+1;
3303 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
3306 spin_unlock_irq(&sh
->stripe_lock
);
3308 if (stripe_can_batch(sh
))
3309 stripe_add_to_batch_list(conf
, sh
);
3313 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
3314 spin_unlock_irq(&sh
->stripe_lock
);
3318 static void end_reshape(struct r5conf
*conf
);
3320 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
3321 struct stripe_head
*sh
)
3323 int sectors_per_chunk
=
3324 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
3326 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
3327 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
3329 raid5_compute_sector(conf
,
3330 stripe
* (disks
- conf
->max_degraded
)
3331 *sectors_per_chunk
+ chunk_offset
,
3337 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
3338 struct stripe_head_state
*s
, int disks
)
3341 BUG_ON(sh
->batch_head
);
3342 for (i
= disks
; i
--; ) {
3346 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
3347 struct md_rdev
*rdev
;
3349 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3350 if (rdev
&& test_bit(In_sync
, &rdev
->flags
) &&
3351 !test_bit(Faulty
, &rdev
->flags
))
3352 atomic_inc(&rdev
->nr_pending
);
3357 if (!rdev_set_badblocks(
3361 md_error(conf
->mddev
, rdev
);
3362 rdev_dec_pending(rdev
, conf
->mddev
);
3365 spin_lock_irq(&sh
->stripe_lock
);
3366 /* fail all writes first */
3367 bi
= sh
->dev
[i
].towrite
;
3368 sh
->dev
[i
].towrite
= NULL
;
3369 sh
->overwrite_disks
= 0;
3370 spin_unlock_irq(&sh
->stripe_lock
);
3374 log_stripe_write_finished(sh
);
3376 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3377 wake_up(&conf
->wait_for_overlap
);
3379 while (bi
&& bi
->bi_iter
.bi_sector
<
3380 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3381 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3383 md_write_end(conf
->mddev
);
3388 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3389 STRIPE_SECTORS
, 0, 0);
3391 /* and fail all 'written' */
3392 bi
= sh
->dev
[i
].written
;
3393 sh
->dev
[i
].written
= NULL
;
3394 if (test_and_clear_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
)) {
3395 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
3396 sh
->dev
[i
].page
= sh
->dev
[i
].orig_page
;
3399 if (bi
) bitmap_end
= 1;
3400 while (bi
&& bi
->bi_iter
.bi_sector
<
3401 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3402 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3404 md_write_end(conf
->mddev
);
3409 /* fail any reads if this device is non-operational and
3410 * the data has not reached the cache yet.
3412 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
3413 s
->failed
> conf
->max_degraded
&&
3414 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
3415 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
3416 spin_lock_irq(&sh
->stripe_lock
);
3417 bi
= sh
->dev
[i
].toread
;
3418 sh
->dev
[i
].toread
= NULL
;
3419 spin_unlock_irq(&sh
->stripe_lock
);
3420 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3421 wake_up(&conf
->wait_for_overlap
);
3424 while (bi
&& bi
->bi_iter
.bi_sector
<
3425 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3426 struct bio
*nextbi
=
3427 r5_next_bio(bi
, sh
->dev
[i
].sector
);
3434 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3435 STRIPE_SECTORS
, 0, 0);
3436 /* If we were in the middle of a write the parity block might
3437 * still be locked - so just clear all R5_LOCKED flags
3439 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3444 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3445 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3446 md_wakeup_thread(conf
->mddev
->thread
);
3450 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
3451 struct stripe_head_state
*s
)
3456 BUG_ON(sh
->batch_head
);
3457 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3458 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
3459 wake_up(&conf
->wait_for_overlap
);
3462 /* There is nothing more to do for sync/check/repair.
3463 * Don't even need to abort as that is handled elsewhere
3464 * if needed, and not always wanted e.g. if there is a known
3466 * For recover/replace we need to record a bad block on all
3467 * non-sync devices, or abort the recovery
3469 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
3470 /* During recovery devices cannot be removed, so
3471 * locking and refcounting of rdevs is not needed
3474 for (i
= 0; i
< conf
->raid_disks
; i
++) {
3475 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3477 && !test_bit(Faulty
, &rdev
->flags
)
3478 && !test_bit(In_sync
, &rdev
->flags
)
3479 && !rdev_set_badblocks(rdev
, sh
->sector
,
3482 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
3484 && !test_bit(Faulty
, &rdev
->flags
)
3485 && !test_bit(In_sync
, &rdev
->flags
)
3486 && !rdev_set_badblocks(rdev
, sh
->sector
,
3492 conf
->recovery_disabled
=
3493 conf
->mddev
->recovery_disabled
;
3495 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, !abort
);
3498 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
3500 struct md_rdev
*rdev
;
3504 rdev
= rcu_dereference(sh
->raid_conf
->disks
[disk_idx
].replacement
);
3506 && !test_bit(Faulty
, &rdev
->flags
)
3507 && !test_bit(In_sync
, &rdev
->flags
)
3508 && (rdev
->recovery_offset
<= sh
->sector
3509 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
3515 static int need_this_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3516 int disk_idx
, int disks
)
3518 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3519 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
3520 &sh
->dev
[s
->failed_num
[1]] };
3524 if (test_bit(R5_LOCKED
, &dev
->flags
) ||
3525 test_bit(R5_UPTODATE
, &dev
->flags
))
3526 /* No point reading this as we already have it or have
3527 * decided to get it.
3532 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)))
3533 /* We need this block to directly satisfy a request */
3536 if (s
->syncing
|| s
->expanding
||
3537 (s
->replacing
&& want_replace(sh
, disk_idx
)))
3538 /* When syncing, or expanding we read everything.
3539 * When replacing, we need the replaced block.
3543 if ((s
->failed
>= 1 && fdev
[0]->toread
) ||
3544 (s
->failed
>= 2 && fdev
[1]->toread
))
3545 /* If we want to read from a failed device, then
3546 * we need to actually read every other device.
3550 /* Sometimes neither read-modify-write nor reconstruct-write
3551 * cycles can work. In those cases we read every block we
3552 * can. Then the parity-update is certain to have enough to
3554 * This can only be a problem when we need to write something,
3555 * and some device has failed. If either of those tests
3556 * fail we need look no further.
3558 if (!s
->failed
|| !s
->to_write
)
3561 if (test_bit(R5_Insync
, &dev
->flags
) &&
3562 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3563 /* Pre-reads at not permitted until after short delay
3564 * to gather multiple requests. However if this
3565 * device is no Insync, the block could only be computed
3566 * and there is no need to delay that.
3570 for (i
= 0; i
< s
->failed
&& i
< 2; i
++) {
3571 if (fdev
[i
]->towrite
&&
3572 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3573 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3574 /* If we have a partial write to a failed
3575 * device, then we will need to reconstruct
3576 * the content of that device, so all other
3577 * devices must be read.
3582 /* If we are forced to do a reconstruct-write, either because
3583 * the current RAID6 implementation only supports that, or
3584 * because parity cannot be trusted and we are currently
3585 * recovering it, there is extra need to be careful.
3586 * If one of the devices that we would need to read, because
3587 * it is not being overwritten (and maybe not written at all)
3588 * is missing/faulty, then we need to read everything we can.
3590 if (sh
->raid_conf
->level
!= 6 &&
3591 sh
->sector
< sh
->raid_conf
->mddev
->recovery_cp
)
3592 /* reconstruct-write isn't being forced */
3594 for (i
= 0; i
< s
->failed
&& i
< 2; i
++) {
3595 if (s
->failed_num
[i
] != sh
->pd_idx
&&
3596 s
->failed_num
[i
] != sh
->qd_idx
&&
3597 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3598 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3605 /* fetch_block - checks the given member device to see if its data needs
3606 * to be read or computed to satisfy a request.
3608 * Returns 1 when no more member devices need to be checked, otherwise returns
3609 * 0 to tell the loop in handle_stripe_fill to continue
3611 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3612 int disk_idx
, int disks
)
3614 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3616 /* is the data in this block needed, and can we get it? */
3617 if (need_this_block(sh
, s
, disk_idx
, disks
)) {
3618 /* we would like to get this block, possibly by computing it,
3619 * otherwise read it if the backing disk is insync
3621 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
3622 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
3623 BUG_ON(sh
->batch_head
);
3626 * In the raid6 case if the only non-uptodate disk is P
3627 * then we already trusted P to compute the other failed
3628 * drives. It is safe to compute rather than re-read P.
3629 * In other cases we only compute blocks from failed
3630 * devices, otherwise check/repair might fail to detect
3631 * a real inconsistency.
3634 if ((s
->uptodate
== disks
- 1) &&
3635 ((sh
->qd_idx
>= 0 && sh
->pd_idx
== disk_idx
) ||
3636 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
3637 disk_idx
== s
->failed_num
[1])))) {
3638 /* have disk failed, and we're requested to fetch it;
3641 pr_debug("Computing stripe %llu block %d\n",
3642 (unsigned long long)sh
->sector
, disk_idx
);
3643 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3644 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3645 set_bit(R5_Wantcompute
, &dev
->flags
);
3646 sh
->ops
.target
= disk_idx
;
3647 sh
->ops
.target2
= -1; /* no 2nd target */
3649 /* Careful: from this point on 'uptodate' is in the eye
3650 * of raid_run_ops which services 'compute' operations
3651 * before writes. R5_Wantcompute flags a block that will
3652 * be R5_UPTODATE by the time it is needed for a
3653 * subsequent operation.
3657 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
3658 /* Computing 2-failure is *very* expensive; only
3659 * do it if failed >= 2
3662 for (other
= disks
; other
--; ) {
3663 if (other
== disk_idx
)
3665 if (!test_bit(R5_UPTODATE
,
3666 &sh
->dev
[other
].flags
))
3670 pr_debug("Computing stripe %llu blocks %d,%d\n",
3671 (unsigned long long)sh
->sector
,
3673 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3674 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3675 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
3676 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
3677 sh
->ops
.target
= disk_idx
;
3678 sh
->ops
.target2
= other
;
3682 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
3683 set_bit(R5_LOCKED
, &dev
->flags
);
3684 set_bit(R5_Wantread
, &dev
->flags
);
3686 pr_debug("Reading block %d (sync=%d)\n",
3687 disk_idx
, s
->syncing
);
3695 * handle_stripe_fill - read or compute data to satisfy pending requests.
3697 static void handle_stripe_fill(struct stripe_head
*sh
,
3698 struct stripe_head_state
*s
,
3703 /* look for blocks to read/compute, skip this if a compute
3704 * is already in flight, or if the stripe contents are in the
3705 * midst of changing due to a write
3707 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
3708 !sh
->reconstruct_state
) {
3711 * For degraded stripe with data in journal, do not handle
3712 * read requests yet, instead, flush the stripe to raid
3713 * disks first, this avoids handling complex rmw of write
3714 * back cache (prexor with orig_page, and then xor with
3715 * page) in the read path
3717 if (s
->injournal
&& s
->failed
) {
3718 if (test_bit(STRIPE_R5C_CACHING
, &sh
->state
))
3719 r5c_make_stripe_write_out(sh
);
3723 for (i
= disks
; i
--; )
3724 if (fetch_block(sh
, s
, i
, disks
))
3728 set_bit(STRIPE_HANDLE
, &sh
->state
);
3731 static void break_stripe_batch_list(struct stripe_head
*head_sh
,
3732 unsigned long handle_flags
);
3733 /* handle_stripe_clean_event
3734 * any written block on an uptodate or failed drive can be returned.
3735 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3736 * never LOCKED, so we don't need to test 'failed' directly.
3738 static void handle_stripe_clean_event(struct r5conf
*conf
,
3739 struct stripe_head
*sh
, int disks
)
3743 int discard_pending
= 0;
3744 struct stripe_head
*head_sh
= sh
;
3745 bool do_endio
= false;
3747 for (i
= disks
; i
--; )
3748 if (sh
->dev
[i
].written
) {
3750 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
3751 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
3752 test_bit(R5_Discard
, &dev
->flags
) ||
3753 test_bit(R5_SkipCopy
, &dev
->flags
))) {
3754 /* We can return any write requests */
3755 struct bio
*wbi
, *wbi2
;
3756 pr_debug("Return write for disc %d\n", i
);
3757 if (test_and_clear_bit(R5_Discard
, &dev
->flags
))
3758 clear_bit(R5_UPTODATE
, &dev
->flags
);
3759 if (test_and_clear_bit(R5_SkipCopy
, &dev
->flags
)) {
3760 WARN_ON(test_bit(R5_UPTODATE
, &dev
->flags
));
3765 dev
->page
= dev
->orig_page
;
3767 dev
->written
= NULL
;
3768 while (wbi
&& wbi
->bi_iter
.bi_sector
<
3769 dev
->sector
+ STRIPE_SECTORS
) {
3770 wbi2
= r5_next_bio(wbi
, dev
->sector
);
3771 md_write_end(conf
->mddev
);
3775 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3777 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
3779 if (head_sh
->batch_head
) {
3780 sh
= list_first_entry(&sh
->batch_list
,
3783 if (sh
!= head_sh
) {
3790 } else if (test_bit(R5_Discard
, &dev
->flags
))
3791 discard_pending
= 1;
3794 log_stripe_write_finished(sh
);
3796 if (!discard_pending
&&
3797 test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
)) {
3799 clear_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
3800 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3801 if (sh
->qd_idx
>= 0) {
3802 clear_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
3803 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
);
3805 /* now that discard is done we can proceed with any sync */
3806 clear_bit(STRIPE_DISCARD
, &sh
->state
);
3808 * SCSI discard will change some bio fields and the stripe has
3809 * no updated data, so remove it from hash list and the stripe
3810 * will be reinitialized
3813 hash
= sh
->hash_lock_index
;
3814 spin_lock_irq(conf
->hash_locks
+ hash
);
3816 spin_unlock_irq(conf
->hash_locks
+ hash
);
3817 if (head_sh
->batch_head
) {
3818 sh
= list_first_entry(&sh
->batch_list
,
3819 struct stripe_head
, batch_list
);
3825 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
))
3826 set_bit(STRIPE_HANDLE
, &sh
->state
);
3830 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3831 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3832 md_wakeup_thread(conf
->mddev
->thread
);
3834 if (head_sh
->batch_head
&& do_endio
)
3835 break_stripe_batch_list(head_sh
, STRIPE_EXPAND_SYNC_FLAGS
);
3839 * For RMW in write back cache, we need extra page in prexor to store the
3840 * old data. This page is stored in dev->orig_page.
3842 * This function checks whether we have data for prexor. The exact logic
3844 * R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3846 static inline bool uptodate_for_rmw(struct r5dev
*dev
)
3848 return (test_bit(R5_UPTODATE
, &dev
->flags
)) &&
3849 (!test_bit(R5_InJournal
, &dev
->flags
) ||
3850 test_bit(R5_OrigPageUPTDODATE
, &dev
->flags
));
3853 static int handle_stripe_dirtying(struct r5conf
*conf
,
3854 struct stripe_head
*sh
,
3855 struct stripe_head_state
*s
,
3858 int rmw
= 0, rcw
= 0, i
;
3859 sector_t recovery_cp
= conf
->mddev
->recovery_cp
;
3861 /* Check whether resync is now happening or should start.
3862 * If yes, then the array is dirty (after unclean shutdown or
3863 * initial creation), so parity in some stripes might be inconsistent.
3864 * In this case, we need to always do reconstruct-write, to ensure
3865 * that in case of drive failure or read-error correction, we
3866 * generate correct data from the parity.
3868 if (conf
->rmw_level
== PARITY_DISABLE_RMW
||
3869 (recovery_cp
< MaxSector
&& sh
->sector
>= recovery_cp
&&
3871 /* Calculate the real rcw later - for now make it
3872 * look like rcw is cheaper
3875 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3876 conf
->rmw_level
, (unsigned long long)recovery_cp
,
3877 (unsigned long long)sh
->sector
);
3878 } else for (i
= disks
; i
--; ) {
3879 /* would I have to read this buffer for read_modify_write */
3880 struct r5dev
*dev
= &sh
->dev
[i
];
3881 if (((dev
->towrite
&& !delay_towrite(conf
, dev
, s
)) ||
3882 i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3883 test_bit(R5_InJournal
, &dev
->flags
)) &&
3884 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3885 !(uptodate_for_rmw(dev
) ||
3886 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3887 if (test_bit(R5_Insync
, &dev
->flags
))
3890 rmw
+= 2*disks
; /* cannot read it */
3892 /* Would I have to read this buffer for reconstruct_write */
3893 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3894 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3895 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3896 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3897 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3898 if (test_bit(R5_Insync
, &dev
->flags
))
3905 pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3906 (unsigned long long)sh
->sector
, sh
->state
, rmw
, rcw
);
3907 set_bit(STRIPE_HANDLE
, &sh
->state
);
3908 if ((rmw
< rcw
|| (rmw
== rcw
&& conf
->rmw_level
== PARITY_PREFER_RMW
)) && rmw
> 0) {
3909 /* prefer read-modify-write, but need to get some data */
3910 if (conf
->mddev
->queue
)
3911 blk_add_trace_msg(conf
->mddev
->queue
,
3912 "raid5 rmw %llu %d",
3913 (unsigned long long)sh
->sector
, rmw
);
3914 for (i
= disks
; i
--; ) {
3915 struct r5dev
*dev
= &sh
->dev
[i
];
3916 if (test_bit(R5_InJournal
, &dev
->flags
) &&
3917 dev
->page
== dev
->orig_page
&&
3918 !test_bit(R5_LOCKED
, &sh
->dev
[sh
->pd_idx
].flags
)) {
3919 /* alloc page for prexor */
3920 struct page
*p
= alloc_page(GFP_NOIO
);
3928 * alloc_page() failed, try use
3929 * disk_info->extra_page
3931 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE
,
3932 &conf
->cache_state
)) {
3933 r5c_use_extra_page(sh
);
3937 /* extra_page in use, add to delayed_list */
3938 set_bit(STRIPE_DELAYED
, &sh
->state
);
3939 s
->waiting_extra_page
= 1;
3944 for (i
= disks
; i
--; ) {
3945 struct r5dev
*dev
= &sh
->dev
[i
];
3946 if (((dev
->towrite
&& !delay_towrite(conf
, dev
, s
)) ||
3947 i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3948 test_bit(R5_InJournal
, &dev
->flags
)) &&
3949 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3950 !(uptodate_for_rmw(dev
) ||
3951 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
3952 test_bit(R5_Insync
, &dev
->flags
)) {
3953 if (test_bit(STRIPE_PREREAD_ACTIVE
,
3955 pr_debug("Read_old block %d for r-m-w\n",
3957 set_bit(R5_LOCKED
, &dev
->flags
);
3958 set_bit(R5_Wantread
, &dev
->flags
);
3961 set_bit(STRIPE_DELAYED
, &sh
->state
);
3962 set_bit(STRIPE_HANDLE
, &sh
->state
);
3967 if ((rcw
< rmw
|| (rcw
== rmw
&& conf
->rmw_level
!= PARITY_PREFER_RMW
)) && rcw
> 0) {
3968 /* want reconstruct write, but need to get some data */
3971 for (i
= disks
; i
--; ) {
3972 struct r5dev
*dev
= &sh
->dev
[i
];
3973 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3974 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3975 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3976 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3977 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3979 if (test_bit(R5_Insync
, &dev
->flags
) &&
3980 test_bit(STRIPE_PREREAD_ACTIVE
,
3982 pr_debug("Read_old block "
3983 "%d for Reconstruct\n", i
);
3984 set_bit(R5_LOCKED
, &dev
->flags
);
3985 set_bit(R5_Wantread
, &dev
->flags
);
3989 set_bit(STRIPE_DELAYED
, &sh
->state
);
3990 set_bit(STRIPE_HANDLE
, &sh
->state
);
3994 if (rcw
&& conf
->mddev
->queue
)
3995 blk_add_trace_msg(conf
->mddev
->queue
, "raid5 rcw %llu %d %d %d",
3996 (unsigned long long)sh
->sector
,
3997 rcw
, qread
, test_bit(STRIPE_DELAYED
, &sh
->state
));
4000 if (rcw
> disks
&& rmw
> disks
&&
4001 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4002 set_bit(STRIPE_DELAYED
, &sh
->state
);
4004 /* now if nothing is locked, and if we have enough data,
4005 * we can start a write request
4007 /* since handle_stripe can be called at any time we need to handle the
4008 * case where a compute block operation has been submitted and then a
4009 * subsequent call wants to start a write request. raid_run_ops only
4010 * handles the case where compute block and reconstruct are requested
4011 * simultaneously. If this is not the case then new writes need to be
4012 * held off until the compute completes.
4014 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
4015 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
4016 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
4017 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
4021 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
4022 struct stripe_head_state
*s
, int disks
)
4024 struct r5dev
*dev
= NULL
;
4026 BUG_ON(sh
->batch_head
);
4027 set_bit(STRIPE_HANDLE
, &sh
->state
);
4029 switch (sh
->check_state
) {
4030 case check_state_idle
:
4031 /* start a new check operation if there are no failures */
4032 if (s
->failed
== 0) {
4033 BUG_ON(s
->uptodate
!= disks
);
4034 sh
->check_state
= check_state_run
;
4035 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
4036 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
4040 dev
= &sh
->dev
[s
->failed_num
[0]];
4042 case check_state_compute_result
:
4043 sh
->check_state
= check_state_idle
;
4045 dev
= &sh
->dev
[sh
->pd_idx
];
4047 /* check that a write has not made the stripe insync */
4048 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
4051 /* either failed parity check, or recovery is happening */
4052 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
4053 BUG_ON(s
->uptodate
!= disks
);
4055 set_bit(R5_LOCKED
, &dev
->flags
);
4057 set_bit(R5_Wantwrite
, &dev
->flags
);
4059 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
4060 set_bit(STRIPE_INSYNC
, &sh
->state
);
4062 case check_state_run
:
4063 break; /* we will be called again upon completion */
4064 case check_state_check_result
:
4065 sh
->check_state
= check_state_idle
;
4067 /* if a failure occurred during the check operation, leave
4068 * STRIPE_INSYNC not set and let the stripe be handled again
4073 /* handle a successful check operation, if parity is correct
4074 * we are done. Otherwise update the mismatch count and repair
4075 * parity if !MD_RECOVERY_CHECK
4077 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
4078 /* parity is correct (on disc,
4079 * not in buffer any more)
4081 set_bit(STRIPE_INSYNC
, &sh
->state
);
4083 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
4084 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
)) {
4085 /* don't try to repair!! */
4086 set_bit(STRIPE_INSYNC
, &sh
->state
);
4087 pr_warn_ratelimited("%s: mismatch sector in range "
4088 "%llu-%llu\n", mdname(conf
->mddev
),
4089 (unsigned long long) sh
->sector
,
4090 (unsigned long long) sh
->sector
+
4093 sh
->check_state
= check_state_compute_run
;
4094 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
4095 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
4096 set_bit(R5_Wantcompute
,
4097 &sh
->dev
[sh
->pd_idx
].flags
);
4098 sh
->ops
.target
= sh
->pd_idx
;
4099 sh
->ops
.target2
= -1;
4104 case check_state_compute_run
:
4107 pr_err("%s: unknown check_state: %d sector: %llu\n",
4108 __func__
, sh
->check_state
,
4109 (unsigned long long) sh
->sector
);
4114 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
4115 struct stripe_head_state
*s
,
4118 int pd_idx
= sh
->pd_idx
;
4119 int qd_idx
= sh
->qd_idx
;
4122 BUG_ON(sh
->batch_head
);
4123 set_bit(STRIPE_HANDLE
, &sh
->state
);
4125 BUG_ON(s
->failed
> 2);
4127 /* Want to check and possibly repair P and Q.
4128 * However there could be one 'failed' device, in which
4129 * case we can only check one of them, possibly using the
4130 * other to generate missing data
4133 switch (sh
->check_state
) {
4134 case check_state_idle
:
4135 /* start a new check operation if there are < 2 failures */
4136 if (s
->failed
== s
->q_failed
) {
4137 /* The only possible failed device holds Q, so it
4138 * makes sense to check P (If anything else were failed,
4139 * we would have used P to recreate it).
4141 sh
->check_state
= check_state_run
;
4143 if (!s
->q_failed
&& s
->failed
< 2) {
4144 /* Q is not failed, and we didn't use it to generate
4145 * anything, so it makes sense to check it
4147 if (sh
->check_state
== check_state_run
)
4148 sh
->check_state
= check_state_run_pq
;
4150 sh
->check_state
= check_state_run_q
;
4153 /* discard potentially stale zero_sum_result */
4154 sh
->ops
.zero_sum_result
= 0;
4156 if (sh
->check_state
== check_state_run
) {
4157 /* async_xor_zero_sum destroys the contents of P */
4158 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
4161 if (sh
->check_state
>= check_state_run
&&
4162 sh
->check_state
<= check_state_run_pq
) {
4163 /* async_syndrome_zero_sum preserves P and Q, so
4164 * no need to mark them !uptodate here
4166 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
4170 /* we have 2-disk failure */
4171 BUG_ON(s
->failed
!= 2);
4173 case check_state_compute_result
:
4174 sh
->check_state
= check_state_idle
;
4176 /* check that a write has not made the stripe insync */
4177 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
4180 /* now write out any block on a failed drive,
4181 * or P or Q if they were recomputed
4183 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
4184 if (s
->failed
== 2) {
4185 dev
= &sh
->dev
[s
->failed_num
[1]];
4187 set_bit(R5_LOCKED
, &dev
->flags
);
4188 set_bit(R5_Wantwrite
, &dev
->flags
);
4190 if (s
->failed
>= 1) {
4191 dev
= &sh
->dev
[s
->failed_num
[0]];
4193 set_bit(R5_LOCKED
, &dev
->flags
);
4194 set_bit(R5_Wantwrite
, &dev
->flags
);
4196 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
4197 dev
= &sh
->dev
[pd_idx
];
4199 set_bit(R5_LOCKED
, &dev
->flags
);
4200 set_bit(R5_Wantwrite
, &dev
->flags
);
4202 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
4203 dev
= &sh
->dev
[qd_idx
];
4205 set_bit(R5_LOCKED
, &dev
->flags
);
4206 set_bit(R5_Wantwrite
, &dev
->flags
);
4208 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
4210 set_bit(STRIPE_INSYNC
, &sh
->state
);
4212 case check_state_run
:
4213 case check_state_run_q
:
4214 case check_state_run_pq
:
4215 break; /* we will be called again upon completion */
4216 case check_state_check_result
:
4217 sh
->check_state
= check_state_idle
;
4219 /* handle a successful check operation, if parity is correct
4220 * we are done. Otherwise update the mismatch count and repair
4221 * parity if !MD_RECOVERY_CHECK
4223 if (sh
->ops
.zero_sum_result
== 0) {
4224 /* both parities are correct */
4226 set_bit(STRIPE_INSYNC
, &sh
->state
);
4228 /* in contrast to the raid5 case we can validate
4229 * parity, but still have a failure to write
4232 sh
->check_state
= check_state_compute_result
;
4233 /* Returning at this point means that we may go
4234 * off and bring p and/or q uptodate again so
4235 * we make sure to check zero_sum_result again
4236 * to verify if p or q need writeback
4240 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
4241 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
)) {
4242 /* don't try to repair!! */
4243 set_bit(STRIPE_INSYNC
, &sh
->state
);
4244 pr_warn_ratelimited("%s: mismatch sector in range "
4245 "%llu-%llu\n", mdname(conf
->mddev
),
4246 (unsigned long long) sh
->sector
,
4247 (unsigned long long) sh
->sector
+
4250 int *target
= &sh
->ops
.target
;
4252 sh
->ops
.target
= -1;
4253 sh
->ops
.target2
= -1;
4254 sh
->check_state
= check_state_compute_run
;
4255 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
4256 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
4257 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
4258 set_bit(R5_Wantcompute
,
4259 &sh
->dev
[pd_idx
].flags
);
4261 target
= &sh
->ops
.target2
;
4264 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
4265 set_bit(R5_Wantcompute
,
4266 &sh
->dev
[qd_idx
].flags
);
4273 case check_state_compute_run
:
4276 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4277 __func__
, sh
->check_state
,
4278 (unsigned long long) sh
->sector
);
4283 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
4287 /* We have read all the blocks in this stripe and now we need to
4288 * copy some of them into a target stripe for expand.
4290 struct dma_async_tx_descriptor
*tx
= NULL
;
4291 BUG_ON(sh
->batch_head
);
4292 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
4293 for (i
= 0; i
< sh
->disks
; i
++)
4294 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
4296 struct stripe_head
*sh2
;
4297 struct async_submit_ctl submit
;
4299 sector_t bn
= raid5_compute_blocknr(sh
, i
, 1);
4300 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
4302 sh2
= raid5_get_active_stripe(conf
, s
, 0, 1, 1);
4304 /* so far only the early blocks of this stripe
4305 * have been requested. When later blocks
4306 * get requested, we will try again
4309 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
4310 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
4311 /* must have already done this block */
4312 raid5_release_stripe(sh2
);
4316 /* place all the copies on one channel */
4317 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
4318 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
4319 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
4322 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
4323 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
4324 for (j
= 0; j
< conf
->raid_disks
; j
++)
4325 if (j
!= sh2
->pd_idx
&&
4327 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
4329 if (j
== conf
->raid_disks
) {
4330 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
4331 set_bit(STRIPE_HANDLE
, &sh2
->state
);
4333 raid5_release_stripe(sh2
);
4336 /* done submitting copies, wait for them to complete */
4337 async_tx_quiesce(&tx
);
4341 * handle_stripe - do things to a stripe.
4343 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4344 * state of various bits to see what needs to be done.
4346 * return some read requests which now have data
4347 * return some write requests which are safely on storage
4348 * schedule a read on some buffers
4349 * schedule a write of some buffers
4350 * return confirmation of parity correctness
4354 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
4356 struct r5conf
*conf
= sh
->raid_conf
;
4357 int disks
= sh
->disks
;
4360 int do_recovery
= 0;
4362 memset(s
, 0, sizeof(*s
));
4364 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
) && !sh
->batch_head
;
4365 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
) && !sh
->batch_head
;
4366 s
->failed_num
[0] = -1;
4367 s
->failed_num
[1] = -1;
4368 s
->log_failed
= r5l_log_disk_error(conf
);
4370 /* Now to look around and see what can be done */
4372 for (i
=disks
; i
--; ) {
4373 struct md_rdev
*rdev
;
4380 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4382 dev
->toread
, dev
->towrite
, dev
->written
);
4383 /* maybe we can reply to a read
4385 * new wantfill requests are only permitted while
4386 * ops_complete_biofill is guaranteed to be inactive
4388 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
4389 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
4390 set_bit(R5_Wantfill
, &dev
->flags
);
4392 /* now count some things */
4393 if (test_bit(R5_LOCKED
, &dev
->flags
))
4395 if (test_bit(R5_UPTODATE
, &dev
->flags
))
4397 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
4399 BUG_ON(s
->compute
> 2);
4402 if (test_bit(R5_Wantfill
, &dev
->flags
))
4404 else if (dev
->toread
)
4408 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
4413 /* Prefer to use the replacement for reads, but only
4414 * if it is recovered enough and has no bad blocks.
4416 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
4417 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
4418 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
4419 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4420 &first_bad
, &bad_sectors
))
4421 set_bit(R5_ReadRepl
, &dev
->flags
);
4423 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4424 set_bit(R5_NeedReplace
, &dev
->flags
);
4426 clear_bit(R5_NeedReplace
, &dev
->flags
);
4427 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
4428 clear_bit(R5_ReadRepl
, &dev
->flags
);
4430 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
4433 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4434 &first_bad
, &bad_sectors
);
4435 if (s
->blocked_rdev
== NULL
4436 && (test_bit(Blocked
, &rdev
->flags
)
4439 set_bit(BlockedBadBlocks
,
4441 s
->blocked_rdev
= rdev
;
4442 atomic_inc(&rdev
->nr_pending
);
4445 clear_bit(R5_Insync
, &dev
->flags
);
4449 /* also not in-sync */
4450 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
4451 test_bit(R5_UPTODATE
, &dev
->flags
)) {
4452 /* treat as in-sync, but with a read error
4453 * which we can now try to correct
4455 set_bit(R5_Insync
, &dev
->flags
);
4456 set_bit(R5_ReadError
, &dev
->flags
);
4458 } else if (test_bit(In_sync
, &rdev
->flags
))
4459 set_bit(R5_Insync
, &dev
->flags
);
4460 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
4461 /* in sync if before recovery_offset */
4462 set_bit(R5_Insync
, &dev
->flags
);
4463 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
4464 test_bit(R5_Expanded
, &dev
->flags
))
4465 /* If we've reshaped into here, we assume it is Insync.
4466 * We will shortly update recovery_offset to make
4469 set_bit(R5_Insync
, &dev
->flags
);
4471 if (test_bit(R5_WriteError
, &dev
->flags
)) {
4472 /* This flag does not apply to '.replacement'
4473 * only to .rdev, so make sure to check that*/
4474 struct md_rdev
*rdev2
= rcu_dereference(
4475 conf
->disks
[i
].rdev
);
4477 clear_bit(R5_Insync
, &dev
->flags
);
4478 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4479 s
->handle_bad_blocks
= 1;
4480 atomic_inc(&rdev2
->nr_pending
);
4482 clear_bit(R5_WriteError
, &dev
->flags
);
4484 if (test_bit(R5_MadeGood
, &dev
->flags
)) {
4485 /* This flag does not apply to '.replacement'
4486 * only to .rdev, so make sure to check that*/
4487 struct md_rdev
*rdev2
= rcu_dereference(
4488 conf
->disks
[i
].rdev
);
4489 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4490 s
->handle_bad_blocks
= 1;
4491 atomic_inc(&rdev2
->nr_pending
);
4493 clear_bit(R5_MadeGood
, &dev
->flags
);
4495 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
4496 struct md_rdev
*rdev2
= rcu_dereference(
4497 conf
->disks
[i
].replacement
);
4498 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4499 s
->handle_bad_blocks
= 1;
4500 atomic_inc(&rdev2
->nr_pending
);
4502 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
4504 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4505 /* The ReadError flag will just be confusing now */
4506 clear_bit(R5_ReadError
, &dev
->flags
);
4507 clear_bit(R5_ReWrite
, &dev
->flags
);
4509 if (test_bit(R5_ReadError
, &dev
->flags
))
4510 clear_bit(R5_Insync
, &dev
->flags
);
4511 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4513 s
->failed_num
[s
->failed
] = i
;
4515 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4519 if (test_bit(R5_InJournal
, &dev
->flags
))
4521 if (test_bit(R5_InJournal
, &dev
->flags
) && dev
->written
)
4524 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
4525 /* If there is a failed device being replaced,
4526 * we must be recovering.
4527 * else if we are after recovery_cp, we must be syncing
4528 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4529 * else we can only be replacing
4530 * sync and recovery both need to read all devices, and so
4531 * use the same flag.
4534 sh
->sector
>= conf
->mddev
->recovery_cp
||
4535 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
4543 static int clear_batch_ready(struct stripe_head
*sh
)
4545 /* Return '1' if this is a member of batch, or
4546 * '0' if it is a lone stripe or a head which can now be
4549 struct stripe_head
*tmp
;
4550 if (!test_and_clear_bit(STRIPE_BATCH_READY
, &sh
->state
))
4551 return (sh
->batch_head
&& sh
->batch_head
!= sh
);
4552 spin_lock(&sh
->stripe_lock
);
4553 if (!sh
->batch_head
) {
4554 spin_unlock(&sh
->stripe_lock
);
4559 * this stripe could be added to a batch list before we check
4560 * BATCH_READY, skips it
4562 if (sh
->batch_head
!= sh
) {
4563 spin_unlock(&sh
->stripe_lock
);
4566 spin_lock(&sh
->batch_lock
);
4567 list_for_each_entry(tmp
, &sh
->batch_list
, batch_list
)
4568 clear_bit(STRIPE_BATCH_READY
, &tmp
->state
);
4569 spin_unlock(&sh
->batch_lock
);
4570 spin_unlock(&sh
->stripe_lock
);
4573 * BATCH_READY is cleared, no new stripes can be added.
4574 * batch_list can be accessed without lock
4579 static void break_stripe_batch_list(struct stripe_head
*head_sh
,
4580 unsigned long handle_flags
)
4582 struct stripe_head
*sh
, *next
;
4586 list_for_each_entry_safe(sh
, next
, &head_sh
->batch_list
, batch_list
) {
4588 list_del_init(&sh
->batch_list
);
4590 WARN_ONCE(sh
->state
& ((1 << STRIPE_ACTIVE
) |
4591 (1 << STRIPE_SYNCING
) |
4592 (1 << STRIPE_REPLACED
) |
4593 (1 << STRIPE_DELAYED
) |
4594 (1 << STRIPE_BIT_DELAY
) |
4595 (1 << STRIPE_FULL_WRITE
) |
4596 (1 << STRIPE_BIOFILL_RUN
) |
4597 (1 << STRIPE_COMPUTE_RUN
) |
4598 (1 << STRIPE_OPS_REQ_PENDING
) |
4599 (1 << STRIPE_DISCARD
) |
4600 (1 << STRIPE_BATCH_READY
) |
4601 (1 << STRIPE_BATCH_ERR
) |
4602 (1 << STRIPE_BITMAP_PENDING
)),
4603 "stripe state: %lx\n", sh
->state
);
4604 WARN_ONCE(head_sh
->state
& ((1 << STRIPE_DISCARD
) |
4605 (1 << STRIPE_REPLACED
)),
4606 "head stripe state: %lx\n", head_sh
->state
);
4608 set_mask_bits(&sh
->state
, ~(STRIPE_EXPAND_SYNC_FLAGS
|
4609 (1 << STRIPE_PREREAD_ACTIVE
) |
4610 (1 << STRIPE_DEGRADED
) |
4611 (1 << STRIPE_ON_UNPLUG_LIST
)),
4612 head_sh
->state
& (1 << STRIPE_INSYNC
));
4614 sh
->check_state
= head_sh
->check_state
;
4615 sh
->reconstruct_state
= head_sh
->reconstruct_state
;
4616 for (i
= 0; i
< sh
->disks
; i
++) {
4617 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
4619 sh
->dev
[i
].flags
= head_sh
->dev
[i
].flags
&
4620 (~((1 << R5_WriteError
) | (1 << R5_Overlap
)));
4622 spin_lock_irq(&sh
->stripe_lock
);
4623 sh
->batch_head
= NULL
;
4624 spin_unlock_irq(&sh
->stripe_lock
);
4625 if (handle_flags
== 0 ||
4626 sh
->state
& handle_flags
)
4627 set_bit(STRIPE_HANDLE
, &sh
->state
);
4628 raid5_release_stripe(sh
);
4630 spin_lock_irq(&head_sh
->stripe_lock
);
4631 head_sh
->batch_head
= NULL
;
4632 spin_unlock_irq(&head_sh
->stripe_lock
);
4633 for (i
= 0; i
< head_sh
->disks
; i
++)
4634 if (test_and_clear_bit(R5_Overlap
, &head_sh
->dev
[i
].flags
))
4636 if (head_sh
->state
& handle_flags
)
4637 set_bit(STRIPE_HANDLE
, &head_sh
->state
);
4640 wake_up(&head_sh
->raid_conf
->wait_for_overlap
);
4643 static void handle_stripe(struct stripe_head
*sh
)
4645 struct stripe_head_state s
;
4646 struct r5conf
*conf
= sh
->raid_conf
;
4649 int disks
= sh
->disks
;
4650 struct r5dev
*pdev
, *qdev
;
4652 clear_bit(STRIPE_HANDLE
, &sh
->state
);
4653 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
4654 /* already being handled, ensure it gets handled
4655 * again when current action finishes */
4656 set_bit(STRIPE_HANDLE
, &sh
->state
);
4660 if (clear_batch_ready(sh
) ) {
4661 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
4665 if (test_and_clear_bit(STRIPE_BATCH_ERR
, &sh
->state
))
4666 break_stripe_batch_list(sh
, 0);
4668 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
) && !sh
->batch_head
) {
4669 spin_lock(&sh
->stripe_lock
);
4671 * Cannot process 'sync' concurrently with 'discard'.
4672 * Flush data in r5cache before 'sync'.
4674 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE
, &sh
->state
) &&
4675 !test_bit(STRIPE_R5C_FULL_STRIPE
, &sh
->state
) &&
4676 !test_bit(STRIPE_DISCARD
, &sh
->state
) &&
4677 test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
4678 set_bit(STRIPE_SYNCING
, &sh
->state
);
4679 clear_bit(STRIPE_INSYNC
, &sh
->state
);
4680 clear_bit(STRIPE_REPLACED
, &sh
->state
);
4682 spin_unlock(&sh
->stripe_lock
);
4684 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4686 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4687 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4688 (unsigned long long)sh
->sector
, sh
->state
,
4689 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
4690 sh
->check_state
, sh
->reconstruct_state
);
4692 analyse_stripe(sh
, &s
);
4694 if (test_bit(STRIPE_LOG_TRAPPED
, &sh
->state
))
4697 if (s
.handle_bad_blocks
||
4698 test_bit(MD_SB_CHANGE_PENDING
, &conf
->mddev
->sb_flags
)) {
4699 set_bit(STRIPE_HANDLE
, &sh
->state
);
4703 if (unlikely(s
.blocked_rdev
)) {
4704 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
4705 s
.replacing
|| s
.to_write
|| s
.written
) {
4706 set_bit(STRIPE_HANDLE
, &sh
->state
);
4709 /* There is nothing for the blocked_rdev to block */
4710 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
4711 s
.blocked_rdev
= NULL
;
4714 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
4715 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
4716 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
4719 pr_debug("locked=%d uptodate=%d to_read=%d"
4720 " to_write=%d failed=%d failed_num=%d,%d\n",
4721 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
4722 s
.failed_num
[0], s
.failed_num
[1]);
4724 * check if the array has lost more than max_degraded devices and,
4725 * if so, some requests might need to be failed.
4727 * When journal device failed (log_failed), we will only process
4728 * the stripe if there is data need write to raid disks
4730 if (s
.failed
> conf
->max_degraded
||
4731 (s
.log_failed
&& s
.injournal
== 0)) {
4732 sh
->check_state
= 0;
4733 sh
->reconstruct_state
= 0;
4734 break_stripe_batch_list(sh
, 0);
4735 if (s
.to_read
+s
.to_write
+s
.written
)
4736 handle_failed_stripe(conf
, sh
, &s
, disks
);
4737 if (s
.syncing
+ s
.replacing
)
4738 handle_failed_sync(conf
, sh
, &s
);
4741 /* Now we check to see if any write operations have recently
4745 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
4747 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
4748 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
4749 sh
->reconstruct_state
= reconstruct_state_idle
;
4751 /* All the 'written' buffers and the parity block are ready to
4752 * be written back to disk
4754 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
) &&
4755 !test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
));
4756 BUG_ON(sh
->qd_idx
>= 0 &&
4757 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
) &&
4758 !test_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
));
4759 for (i
= disks
; i
--; ) {
4760 struct r5dev
*dev
= &sh
->dev
[i
];
4761 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
4762 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
4763 dev
->written
|| test_bit(R5_InJournal
,
4765 pr_debug("Writing block %d\n", i
);
4766 set_bit(R5_Wantwrite
, &dev
->flags
);
4771 if (!test_bit(R5_Insync
, &dev
->flags
) ||
4772 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
4774 set_bit(STRIPE_INSYNC
, &sh
->state
);
4777 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4778 s
.dec_preread_active
= 1;
4782 * might be able to return some write requests if the parity blocks
4783 * are safe, or on a failed drive
4785 pdev
= &sh
->dev
[sh
->pd_idx
];
4786 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
4787 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
4788 qdev
= &sh
->dev
[sh
->qd_idx
];
4789 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
4790 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
4794 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
4795 && !test_bit(R5_LOCKED
, &pdev
->flags
)
4796 && (test_bit(R5_UPTODATE
, &pdev
->flags
) ||
4797 test_bit(R5_Discard
, &pdev
->flags
))))) &&
4798 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
4799 && !test_bit(R5_LOCKED
, &qdev
->flags
)
4800 && (test_bit(R5_UPTODATE
, &qdev
->flags
) ||
4801 test_bit(R5_Discard
, &qdev
->flags
))))))
4802 handle_stripe_clean_event(conf
, sh
, disks
);
4805 r5c_handle_cached_data_endio(conf
, sh
, disks
);
4806 log_stripe_write_finished(sh
);
4808 /* Now we might consider reading some blocks, either to check/generate
4809 * parity, or to satisfy requests
4810 * or to load a block that is being partially written.
4812 if (s
.to_read
|| s
.non_overwrite
4813 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
4814 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
4817 handle_stripe_fill(sh
, &s
, disks
);
4820 * When the stripe finishes full journal write cycle (write to journal
4821 * and raid disk), this is the clean up procedure so it is ready for
4824 r5c_finish_stripe_write_out(conf
, sh
, &s
);
4827 * Now to consider new write requests, cache write back and what else,
4828 * if anything should be read. We do not handle new writes when:
4829 * 1/ A 'write' operation (copy+xor) is already in flight.
4830 * 2/ A 'check' operation is in flight, as it may clobber the parity
4832 * 3/ A r5c cache log write is in flight.
4835 if (!sh
->reconstruct_state
&& !sh
->check_state
&& !sh
->log_io
) {
4836 if (!r5c_is_writeback(conf
->log
)) {
4838 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
4839 } else { /* write back cache */
4842 /* First, try handle writes in caching phase */
4844 ret
= r5c_try_caching_write(conf
, sh
, &s
,
4847 * If caching phase failed: ret == -EAGAIN
4849 * stripe under reclaim: !caching && injournal
4851 * fall back to handle_stripe_dirtying()
4853 if (ret
== -EAGAIN
||
4854 /* stripe under reclaim: !caching && injournal */
4855 (!test_bit(STRIPE_R5C_CACHING
, &sh
->state
) &&
4857 ret
= handle_stripe_dirtying(conf
, sh
, &s
,
4865 /* maybe we need to check and possibly fix the parity for this stripe
4866 * Any reads will already have been scheduled, so we just see if enough
4867 * data is available. The parity check is held off while parity
4868 * dependent operations are in flight.
4870 if (sh
->check_state
||
4871 (s
.syncing
&& s
.locked
== 0 &&
4872 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4873 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
4874 if (conf
->level
== 6)
4875 handle_parity_checks6(conf
, sh
, &s
, disks
);
4877 handle_parity_checks5(conf
, sh
, &s
, disks
);
4880 if ((s
.replacing
|| s
.syncing
) && s
.locked
== 0
4881 && !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)
4882 && !test_bit(STRIPE_REPLACED
, &sh
->state
)) {
4883 /* Write out to replacement devices where possible */
4884 for (i
= 0; i
< conf
->raid_disks
; i
++)
4885 if (test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
4886 WARN_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
4887 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
4888 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4892 set_bit(STRIPE_INSYNC
, &sh
->state
);
4893 set_bit(STRIPE_REPLACED
, &sh
->state
);
4895 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
4896 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4897 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
4898 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4899 clear_bit(STRIPE_SYNCING
, &sh
->state
);
4900 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
4901 wake_up(&conf
->wait_for_overlap
);
4904 /* If the failed drives are just a ReadError, then we might need
4905 * to progress the repair/check process
4907 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
4908 for (i
= 0; i
< s
.failed
; i
++) {
4909 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
4910 if (test_bit(R5_ReadError
, &dev
->flags
)
4911 && !test_bit(R5_LOCKED
, &dev
->flags
)
4912 && test_bit(R5_UPTODATE
, &dev
->flags
)
4914 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
4915 set_bit(R5_Wantwrite
, &dev
->flags
);
4916 set_bit(R5_ReWrite
, &dev
->flags
);
4917 set_bit(R5_LOCKED
, &dev
->flags
);
4920 /* let's read it back */
4921 set_bit(R5_Wantread
, &dev
->flags
);
4922 set_bit(R5_LOCKED
, &dev
->flags
);
4928 /* Finish reconstruct operations initiated by the expansion process */
4929 if (sh
->reconstruct_state
== reconstruct_state_result
) {
4930 struct stripe_head
*sh_src
4931 = raid5_get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
4932 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
4933 /* sh cannot be written until sh_src has been read.
4934 * so arrange for sh to be delayed a little
4936 set_bit(STRIPE_DELAYED
, &sh
->state
);
4937 set_bit(STRIPE_HANDLE
, &sh
->state
);
4938 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
4940 atomic_inc(&conf
->preread_active_stripes
);
4941 raid5_release_stripe(sh_src
);
4945 raid5_release_stripe(sh_src
);
4947 sh
->reconstruct_state
= reconstruct_state_idle
;
4948 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
4949 for (i
= conf
->raid_disks
; i
--; ) {
4950 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
4951 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4956 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
4957 !sh
->reconstruct_state
) {
4958 /* Need to write out all blocks after computing parity */
4959 sh
->disks
= conf
->raid_disks
;
4960 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
4961 schedule_reconstruction(sh
, &s
, 1, 1);
4962 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
4963 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4964 atomic_dec(&conf
->reshape_stripes
);
4965 wake_up(&conf
->wait_for_overlap
);
4966 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4969 if (s
.expanding
&& s
.locked
== 0 &&
4970 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
4971 handle_stripe_expansion(conf
, sh
);
4974 /* wait for this device to become unblocked */
4975 if (unlikely(s
.blocked_rdev
)) {
4976 if (conf
->mddev
->external
)
4977 md_wait_for_blocked_rdev(s
.blocked_rdev
,
4980 /* Internal metadata will immediately
4981 * be written by raid5d, so we don't
4982 * need to wait here.
4984 rdev_dec_pending(s
.blocked_rdev
,
4988 if (s
.handle_bad_blocks
)
4989 for (i
= disks
; i
--; ) {
4990 struct md_rdev
*rdev
;
4991 struct r5dev
*dev
= &sh
->dev
[i
];
4992 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
4993 /* We own a safe reference to the rdev */
4994 rdev
= conf
->disks
[i
].rdev
;
4995 if (!rdev_set_badblocks(rdev
, sh
->sector
,
4997 md_error(conf
->mddev
, rdev
);
4998 rdev_dec_pending(rdev
, conf
->mddev
);
5000 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
5001 rdev
= conf
->disks
[i
].rdev
;
5002 rdev_clear_badblocks(rdev
, sh
->sector
,
5004 rdev_dec_pending(rdev
, conf
->mddev
);
5006 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
5007 rdev
= conf
->disks
[i
].replacement
;
5009 /* rdev have been moved down */
5010 rdev
= conf
->disks
[i
].rdev
;
5011 rdev_clear_badblocks(rdev
, sh
->sector
,
5013 rdev_dec_pending(rdev
, conf
->mddev
);
5018 raid_run_ops(sh
, s
.ops_request
);
5022 if (s
.dec_preread_active
) {
5023 /* We delay this until after ops_run_io so that if make_request
5024 * is waiting on a flush, it won't continue until the writes
5025 * have actually been submitted.
5027 atomic_dec(&conf
->preread_active_stripes
);
5028 if (atomic_read(&conf
->preread_active_stripes
) <
5030 md_wakeup_thread(conf
->mddev
->thread
);
5033 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
5036 static void raid5_activate_delayed(struct r5conf
*conf
)
5038 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
5039 while (!list_empty(&conf
->delayed_list
)) {
5040 struct list_head
*l
= conf
->delayed_list
.next
;
5041 struct stripe_head
*sh
;
5042 sh
= list_entry(l
, struct stripe_head
, lru
);
5044 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5045 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5046 atomic_inc(&conf
->preread_active_stripes
);
5047 list_add_tail(&sh
->lru
, &conf
->hold_list
);
5048 raid5_wakeup_stripe_thread(sh
);
5053 static void activate_bit_delay(struct r5conf
*conf
,
5054 struct list_head
*temp_inactive_list
)
5056 /* device_lock is held */
5057 struct list_head head
;
5058 list_add(&head
, &conf
->bitmap_list
);
5059 list_del_init(&conf
->bitmap_list
);
5060 while (!list_empty(&head
)) {
5061 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
5063 list_del_init(&sh
->lru
);
5064 atomic_inc(&sh
->count
);
5065 hash
= sh
->hash_lock_index
;
5066 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
5070 static int raid5_congested(struct mddev
*mddev
, int bits
)
5072 struct r5conf
*conf
= mddev
->private;
5074 /* No difference between reads and writes. Just check
5075 * how busy the stripe_cache is
5078 if (test_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
))
5081 /* Also checks whether there is pressure on r5cache log space */
5082 if (test_bit(R5C_LOG_TIGHT
, &conf
->cache_state
))
5086 if (atomic_read(&conf
->empty_inactive_list_nr
))
5092 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
5094 struct r5conf
*conf
= mddev
->private;
5095 sector_t sector
= bio
->bi_iter
.bi_sector
;
5096 unsigned int chunk_sectors
;
5097 unsigned int bio_sectors
= bio_sectors(bio
);
5099 WARN_ON_ONCE(bio
->bi_partno
);
5101 chunk_sectors
= min(conf
->chunk_sectors
, conf
->prev_chunk_sectors
);
5102 return chunk_sectors
>=
5103 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
5107 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
5108 * later sampled by raid5d.
5110 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
5112 unsigned long flags
;
5114 spin_lock_irqsave(&conf
->device_lock
, flags
);
5116 bi
->bi_next
= conf
->retry_read_aligned_list
;
5117 conf
->retry_read_aligned_list
= bi
;
5119 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5120 md_wakeup_thread(conf
->mddev
->thread
);
5123 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
,
5124 unsigned int *offset
)
5128 bi
= conf
->retry_read_aligned
;
5130 *offset
= conf
->retry_read_offset
;
5131 conf
->retry_read_aligned
= NULL
;
5134 bi
= conf
->retry_read_aligned_list
;
5136 conf
->retry_read_aligned_list
= bi
->bi_next
;
5145 * The "raid5_align_endio" should check if the read succeeded and if it
5146 * did, call bio_endio on the original bio (having bio_put the new bio
5148 * If the read failed..
5150 static void raid5_align_endio(struct bio
*bi
)
5152 struct bio
* raid_bi
= bi
->bi_private
;
5153 struct mddev
*mddev
;
5154 struct r5conf
*conf
;
5155 struct md_rdev
*rdev
;
5156 blk_status_t error
= bi
->bi_status
;
5160 rdev
= (void*)raid_bi
->bi_next
;
5161 raid_bi
->bi_next
= NULL
;
5162 mddev
= rdev
->mddev
;
5163 conf
= mddev
->private;
5165 rdev_dec_pending(rdev
, conf
->mddev
);
5169 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
5170 wake_up(&conf
->wait_for_quiescent
);
5174 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5176 add_bio_to_retry(raid_bi
, conf
);
5179 static int raid5_read_one_chunk(struct mddev
*mddev
, struct bio
*raid_bio
)
5181 struct r5conf
*conf
= mddev
->private;
5183 struct bio
* align_bi
;
5184 struct md_rdev
*rdev
;
5185 sector_t end_sector
;
5187 if (!in_chunk_boundary(mddev
, raid_bio
)) {
5188 pr_debug("%s: non aligned\n", __func__
);
5192 * use bio_clone_fast to make a copy of the bio
5194 align_bi
= bio_clone_fast(raid_bio
, GFP_NOIO
, mddev
->bio_set
);
5198 * set bi_end_io to a new function, and set bi_private to the
5201 align_bi
->bi_end_io
= raid5_align_endio
;
5202 align_bi
->bi_private
= raid_bio
;
5206 align_bi
->bi_iter
.bi_sector
=
5207 raid5_compute_sector(conf
, raid_bio
->bi_iter
.bi_sector
,
5210 end_sector
= bio_end_sector(align_bi
);
5212 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
5213 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
5214 rdev
->recovery_offset
< end_sector
) {
5215 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
5217 (test_bit(Faulty
, &rdev
->flags
) ||
5218 !(test_bit(In_sync
, &rdev
->flags
) ||
5219 rdev
->recovery_offset
>= end_sector
)))
5223 if (r5c_big_stripe_cached(conf
, align_bi
->bi_iter
.bi_sector
)) {
5233 atomic_inc(&rdev
->nr_pending
);
5235 raid_bio
->bi_next
= (void*)rdev
;
5236 bio_set_dev(align_bi
, rdev
->bdev
);
5237 bio_clear_flag(align_bi
, BIO_SEG_VALID
);
5239 if (is_badblock(rdev
, align_bi
->bi_iter
.bi_sector
,
5240 bio_sectors(align_bi
),
5241 &first_bad
, &bad_sectors
)) {
5243 rdev_dec_pending(rdev
, mddev
);
5247 /* No reshape active, so we can trust rdev->data_offset */
5248 align_bi
->bi_iter
.bi_sector
+= rdev
->data_offset
;
5250 spin_lock_irq(&conf
->device_lock
);
5251 wait_event_lock_irq(conf
->wait_for_quiescent
,
5254 atomic_inc(&conf
->active_aligned_reads
);
5255 spin_unlock_irq(&conf
->device_lock
);
5258 trace_block_bio_remap(align_bi
->bi_disk
->queue
,
5259 align_bi
, disk_devt(mddev
->gendisk
),
5260 raid_bio
->bi_iter
.bi_sector
);
5261 generic_make_request(align_bi
);
5270 static struct bio
*chunk_aligned_read(struct mddev
*mddev
, struct bio
*raid_bio
)
5273 sector_t sector
= raid_bio
->bi_iter
.bi_sector
;
5274 unsigned chunk_sects
= mddev
->chunk_sectors
;
5275 unsigned sectors
= chunk_sects
- (sector
& (chunk_sects
-1));
5277 if (sectors
< bio_sectors(raid_bio
)) {
5278 struct r5conf
*conf
= mddev
->private;
5279 split
= bio_split(raid_bio
, sectors
, GFP_NOIO
, conf
->bio_split
);
5280 bio_chain(split
, raid_bio
);
5281 generic_make_request(raid_bio
);
5285 if (!raid5_read_one_chunk(mddev
, raid_bio
))
5291 /* __get_priority_stripe - get the next stripe to process
5293 * Full stripe writes are allowed to pass preread active stripes up until
5294 * the bypass_threshold is exceeded. In general the bypass_count
5295 * increments when the handle_list is handled before the hold_list; however, it
5296 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5297 * stripe with in flight i/o. The bypass_count will be reset when the
5298 * head of the hold_list has changed, i.e. the head was promoted to the
5301 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
, int group
)
5303 struct stripe_head
*sh
, *tmp
;
5304 struct list_head
*handle_list
= NULL
;
5305 struct r5worker_group
*wg
;
5306 bool second_try
= !r5c_is_writeback(conf
->log
) &&
5307 !r5l_log_disk_error(conf
);
5308 bool try_loprio
= test_bit(R5C_LOG_TIGHT
, &conf
->cache_state
) ||
5309 r5l_log_disk_error(conf
);
5314 if (conf
->worker_cnt_per_group
== 0) {
5315 handle_list
= try_loprio
? &conf
->loprio_list
:
5317 } else if (group
!= ANY_GROUP
) {
5318 handle_list
= try_loprio
? &conf
->worker_groups
[group
].loprio_list
:
5319 &conf
->worker_groups
[group
].handle_list
;
5320 wg
= &conf
->worker_groups
[group
];
5323 for (i
= 0; i
< conf
->group_cnt
; i
++) {
5324 handle_list
= try_loprio
? &conf
->worker_groups
[i
].loprio_list
:
5325 &conf
->worker_groups
[i
].handle_list
;
5326 wg
= &conf
->worker_groups
[i
];
5327 if (!list_empty(handle_list
))
5332 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5334 list_empty(handle_list
) ? "empty" : "busy",
5335 list_empty(&conf
->hold_list
) ? "empty" : "busy",
5336 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
5338 if (!list_empty(handle_list
)) {
5339 sh
= list_entry(handle_list
->next
, typeof(*sh
), lru
);
5341 if (list_empty(&conf
->hold_list
))
5342 conf
->bypass_count
= 0;
5343 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
5344 if (conf
->hold_list
.next
== conf
->last_hold
)
5345 conf
->bypass_count
++;
5347 conf
->last_hold
= conf
->hold_list
.next
;
5348 conf
->bypass_count
-= conf
->bypass_threshold
;
5349 if (conf
->bypass_count
< 0)
5350 conf
->bypass_count
= 0;
5353 } else if (!list_empty(&conf
->hold_list
) &&
5354 ((conf
->bypass_threshold
&&
5355 conf
->bypass_count
> conf
->bypass_threshold
) ||
5356 atomic_read(&conf
->pending_full_writes
) == 0)) {
5358 list_for_each_entry(tmp
, &conf
->hold_list
, lru
) {
5359 if (conf
->worker_cnt_per_group
== 0 ||
5360 group
== ANY_GROUP
||
5361 !cpu_online(tmp
->cpu
) ||
5362 cpu_to_group(tmp
->cpu
) == group
) {
5369 conf
->bypass_count
-= conf
->bypass_threshold
;
5370 if (conf
->bypass_count
< 0)
5371 conf
->bypass_count
= 0;
5380 try_loprio
= !try_loprio
;
5388 list_del_init(&sh
->lru
);
5389 BUG_ON(atomic_inc_return(&sh
->count
) != 1);
5393 struct raid5_plug_cb
{
5394 struct blk_plug_cb cb
;
5395 struct list_head list
;
5396 struct list_head temp_inactive_list
[NR_STRIPE_HASH_LOCKS
];
5399 static void raid5_unplug(struct blk_plug_cb
*blk_cb
, bool from_schedule
)
5401 struct raid5_plug_cb
*cb
= container_of(
5402 blk_cb
, struct raid5_plug_cb
, cb
);
5403 struct stripe_head
*sh
;
5404 struct mddev
*mddev
= cb
->cb
.data
;
5405 struct r5conf
*conf
= mddev
->private;
5409 if (cb
->list
.next
&& !list_empty(&cb
->list
)) {
5410 spin_lock_irq(&conf
->device_lock
);
5411 while (!list_empty(&cb
->list
)) {
5412 sh
= list_first_entry(&cb
->list
, struct stripe_head
, lru
);
5413 list_del_init(&sh
->lru
);
5415 * avoid race release_stripe_plug() sees
5416 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5417 * is still in our list
5419 smp_mb__before_atomic();
5420 clear_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
);
5422 * STRIPE_ON_RELEASE_LIST could be set here. In that
5423 * case, the count is always > 1 here
5425 hash
= sh
->hash_lock_index
;
5426 __release_stripe(conf
, sh
, &cb
->temp_inactive_list
[hash
]);
5429 spin_unlock_irq(&conf
->device_lock
);
5431 release_inactive_stripe_list(conf
, cb
->temp_inactive_list
,
5432 NR_STRIPE_HASH_LOCKS
);
5434 trace_block_unplug(mddev
->queue
, cnt
, !from_schedule
);
5438 static void release_stripe_plug(struct mddev
*mddev
,
5439 struct stripe_head
*sh
)
5441 struct blk_plug_cb
*blk_cb
= blk_check_plugged(
5442 raid5_unplug
, mddev
,
5443 sizeof(struct raid5_plug_cb
));
5444 struct raid5_plug_cb
*cb
;
5447 raid5_release_stripe(sh
);
5451 cb
= container_of(blk_cb
, struct raid5_plug_cb
, cb
);
5453 if (cb
->list
.next
== NULL
) {
5455 INIT_LIST_HEAD(&cb
->list
);
5456 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5457 INIT_LIST_HEAD(cb
->temp_inactive_list
+ i
);
5460 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
))
5461 list_add_tail(&sh
->lru
, &cb
->list
);
5463 raid5_release_stripe(sh
);
5466 static void make_discard_request(struct mddev
*mddev
, struct bio
*bi
)
5468 struct r5conf
*conf
= mddev
->private;
5469 sector_t logical_sector
, last_sector
;
5470 struct stripe_head
*sh
;
5473 if (mddev
->reshape_position
!= MaxSector
)
5474 /* Skip discard while reshape is happening */
5477 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5478 last_sector
= bi
->bi_iter
.bi_sector
+ (bi
->bi_iter
.bi_size
>>9);
5482 stripe_sectors
= conf
->chunk_sectors
*
5483 (conf
->raid_disks
- conf
->max_degraded
);
5484 logical_sector
= DIV_ROUND_UP_SECTOR_T(logical_sector
,
5486 sector_div(last_sector
, stripe_sectors
);
5488 logical_sector
*= conf
->chunk_sectors
;
5489 last_sector
*= conf
->chunk_sectors
;
5491 for (; logical_sector
< last_sector
;
5492 logical_sector
+= STRIPE_SECTORS
) {
5496 sh
= raid5_get_active_stripe(conf
, logical_sector
, 0, 0, 0);
5497 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5498 TASK_UNINTERRUPTIBLE
);
5499 set_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5500 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
5501 raid5_release_stripe(sh
);
5505 clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5506 spin_lock_irq(&sh
->stripe_lock
);
5507 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5508 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5510 if (sh
->dev
[d
].towrite
|| sh
->dev
[d
].toread
) {
5511 set_bit(R5_Overlap
, &sh
->dev
[d
].flags
);
5512 spin_unlock_irq(&sh
->stripe_lock
);
5513 raid5_release_stripe(sh
);
5518 set_bit(STRIPE_DISCARD
, &sh
->state
);
5519 finish_wait(&conf
->wait_for_overlap
, &w
);
5520 sh
->overwrite_disks
= 0;
5521 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5522 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5524 sh
->dev
[d
].towrite
= bi
;
5525 set_bit(R5_OVERWRITE
, &sh
->dev
[d
].flags
);
5526 bio_inc_remaining(bi
);
5527 md_write_inc(mddev
, bi
);
5528 sh
->overwrite_disks
++;
5530 spin_unlock_irq(&sh
->stripe_lock
);
5531 if (conf
->mddev
->bitmap
) {
5533 d
< conf
->raid_disks
- conf
->max_degraded
;
5535 bitmap_startwrite(mddev
->bitmap
,
5539 sh
->bm_seq
= conf
->seq_flush
+ 1;
5540 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
5543 set_bit(STRIPE_HANDLE
, &sh
->state
);
5544 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5545 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5546 atomic_inc(&conf
->preread_active_stripes
);
5547 release_stripe_plug(mddev
, sh
);
5553 static bool raid5_make_request(struct mddev
*mddev
, struct bio
* bi
)
5555 struct r5conf
*conf
= mddev
->private;
5557 sector_t new_sector
;
5558 sector_t logical_sector
, last_sector
;
5559 struct stripe_head
*sh
;
5560 const int rw
= bio_data_dir(bi
);
5563 bool do_flush
= false;
5565 if (unlikely(bi
->bi_opf
& REQ_PREFLUSH
)) {
5566 int ret
= r5l_handle_flush_request(conf
->log
, bi
);
5570 if (ret
== -ENODEV
) {
5571 md_flush_request(mddev
, bi
);
5574 /* ret == -EAGAIN, fallback */
5576 * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5577 * we need to flush journal device
5579 do_flush
= bi
->bi_opf
& REQ_PREFLUSH
;
5582 if (!md_write_start(mddev
, bi
))
5585 * If array is degraded, better not do chunk aligned read because
5586 * later we might have to read it again in order to reconstruct
5587 * data on failed drives.
5589 if (rw
== READ
&& mddev
->degraded
== 0 &&
5590 mddev
->reshape_position
== MaxSector
) {
5591 bi
= chunk_aligned_read(mddev
, bi
);
5596 if (unlikely(bio_op(bi
) == REQ_OP_DISCARD
)) {
5597 make_discard_request(mddev
, bi
);
5598 md_write_end(mddev
);
5602 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5603 last_sector
= bio_end_sector(bi
);
5606 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
5607 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
5613 seq
= read_seqcount_begin(&conf
->gen_lock
);
5616 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5617 TASK_UNINTERRUPTIBLE
);
5618 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
5619 /* spinlock is needed as reshape_progress may be
5620 * 64bit on a 32bit platform, and so it might be
5621 * possible to see a half-updated value
5622 * Of course reshape_progress could change after
5623 * the lock is dropped, so once we get a reference
5624 * to the stripe that we think it is, we will have
5627 spin_lock_irq(&conf
->device_lock
);
5628 if (mddev
->reshape_backwards
5629 ? logical_sector
< conf
->reshape_progress
5630 : logical_sector
>= conf
->reshape_progress
) {
5633 if (mddev
->reshape_backwards
5634 ? logical_sector
< conf
->reshape_safe
5635 : logical_sector
>= conf
->reshape_safe
) {
5636 spin_unlock_irq(&conf
->device_lock
);
5642 spin_unlock_irq(&conf
->device_lock
);
5645 new_sector
= raid5_compute_sector(conf
, logical_sector
,
5648 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5649 (unsigned long long)new_sector
,
5650 (unsigned long long)logical_sector
);
5652 sh
= raid5_get_active_stripe(conf
, new_sector
, previous
,
5653 (bi
->bi_opf
& REQ_RAHEAD
), 0);
5655 if (unlikely(previous
)) {
5656 /* expansion might have moved on while waiting for a
5657 * stripe, so we must do the range check again.
5658 * Expansion could still move past after this
5659 * test, but as we are holding a reference to
5660 * 'sh', we know that if that happens,
5661 * STRIPE_EXPANDING will get set and the expansion
5662 * won't proceed until we finish with the stripe.
5665 spin_lock_irq(&conf
->device_lock
);
5666 if (mddev
->reshape_backwards
5667 ? logical_sector
>= conf
->reshape_progress
5668 : logical_sector
< conf
->reshape_progress
)
5669 /* mismatch, need to try again */
5671 spin_unlock_irq(&conf
->device_lock
);
5673 raid5_release_stripe(sh
);
5679 if (read_seqcount_retry(&conf
->gen_lock
, seq
)) {
5680 /* Might have got the wrong stripe_head
5683 raid5_release_stripe(sh
);
5687 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
5688 !add_stripe_bio(sh
, bi
, dd_idx
, rw
, previous
)) {
5689 /* Stripe is busy expanding or
5690 * add failed due to overlap. Flush everything
5693 md_wakeup_thread(mddev
->thread
);
5694 raid5_release_stripe(sh
);
5700 set_bit(STRIPE_R5C_PREFLUSH
, &sh
->state
);
5701 /* we only need flush for one stripe */
5705 set_bit(STRIPE_HANDLE
, &sh
->state
);
5706 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5707 if ((!sh
->batch_head
|| sh
== sh
->batch_head
) &&
5708 (bi
->bi_opf
& REQ_SYNC
) &&
5709 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5710 atomic_inc(&conf
->preread_active_stripes
);
5711 release_stripe_plug(mddev
, sh
);
5713 /* cannot get stripe for read-ahead, just give-up */
5714 bi
->bi_status
= BLK_STS_IOERR
;
5718 finish_wait(&conf
->wait_for_overlap
, &w
);
5721 md_write_end(mddev
);
5726 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
5728 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
5730 /* reshaping is quite different to recovery/resync so it is
5731 * handled quite separately ... here.
5733 * On each call to sync_request, we gather one chunk worth of
5734 * destination stripes and flag them as expanding.
5735 * Then we find all the source stripes and request reads.
5736 * As the reads complete, handle_stripe will copy the data
5737 * into the destination stripe and release that stripe.
5739 struct r5conf
*conf
= mddev
->private;
5740 struct stripe_head
*sh
;
5741 struct md_rdev
*rdev
;
5742 sector_t first_sector
, last_sector
;
5743 int raid_disks
= conf
->previous_raid_disks
;
5744 int data_disks
= raid_disks
- conf
->max_degraded
;
5745 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5748 sector_t writepos
, readpos
, safepos
;
5749 sector_t stripe_addr
;
5750 int reshape_sectors
;
5751 struct list_head stripes
;
5754 if (sector_nr
== 0) {
5755 /* If restarting in the middle, skip the initial sectors */
5756 if (mddev
->reshape_backwards
&&
5757 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
5758 sector_nr
= raid5_size(mddev
, 0, 0)
5759 - conf
->reshape_progress
;
5760 } else if (mddev
->reshape_backwards
&&
5761 conf
->reshape_progress
== MaxSector
) {
5762 /* shouldn't happen, but just in case, finish up.*/
5763 sector_nr
= MaxSector
;
5764 } else if (!mddev
->reshape_backwards
&&
5765 conf
->reshape_progress
> 0)
5766 sector_nr
= conf
->reshape_progress
;
5767 sector_div(sector_nr
, new_data_disks
);
5769 mddev
->curr_resync_completed
= sector_nr
;
5770 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5777 /* We need to process a full chunk at a time.
5778 * If old and new chunk sizes differ, we need to process the
5782 reshape_sectors
= max(conf
->chunk_sectors
, conf
->prev_chunk_sectors
);
5784 /* We update the metadata at least every 10 seconds, or when
5785 * the data about to be copied would over-write the source of
5786 * the data at the front of the range. i.e. one new_stripe
5787 * along from reshape_progress new_maps to after where
5788 * reshape_safe old_maps to
5790 writepos
= conf
->reshape_progress
;
5791 sector_div(writepos
, new_data_disks
);
5792 readpos
= conf
->reshape_progress
;
5793 sector_div(readpos
, data_disks
);
5794 safepos
= conf
->reshape_safe
;
5795 sector_div(safepos
, data_disks
);
5796 if (mddev
->reshape_backwards
) {
5797 BUG_ON(writepos
< reshape_sectors
);
5798 writepos
-= reshape_sectors
;
5799 readpos
+= reshape_sectors
;
5800 safepos
+= reshape_sectors
;
5802 writepos
+= reshape_sectors
;
5803 /* readpos and safepos are worst-case calculations.
5804 * A negative number is overly pessimistic, and causes
5805 * obvious problems for unsigned storage. So clip to 0.
5807 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
5808 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
5811 /* Having calculated the 'writepos' possibly use it
5812 * to set 'stripe_addr' which is where we will write to.
5814 if (mddev
->reshape_backwards
) {
5815 BUG_ON(conf
->reshape_progress
== 0);
5816 stripe_addr
= writepos
;
5817 BUG_ON((mddev
->dev_sectors
&
5818 ~((sector_t
)reshape_sectors
- 1))
5819 - reshape_sectors
- stripe_addr
5822 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
5823 stripe_addr
= sector_nr
;
5826 /* 'writepos' is the most advanced device address we might write.
5827 * 'readpos' is the least advanced device address we might read.
5828 * 'safepos' is the least address recorded in the metadata as having
5830 * If there is a min_offset_diff, these are adjusted either by
5831 * increasing the safepos/readpos if diff is negative, or
5832 * increasing writepos if diff is positive.
5833 * If 'readpos' is then behind 'writepos', there is no way that we can
5834 * ensure safety in the face of a crash - that must be done by userspace
5835 * making a backup of the data. So in that case there is no particular
5836 * rush to update metadata.
5837 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5838 * update the metadata to advance 'safepos' to match 'readpos' so that
5839 * we can be safe in the event of a crash.
5840 * So we insist on updating metadata if safepos is behind writepos and
5841 * readpos is beyond writepos.
5842 * In any case, update the metadata every 10 seconds.
5843 * Maybe that number should be configurable, but I'm not sure it is
5844 * worth it.... maybe it could be a multiple of safemode_delay???
5846 if (conf
->min_offset_diff
< 0) {
5847 safepos
+= -conf
->min_offset_diff
;
5848 readpos
+= -conf
->min_offset_diff
;
5850 writepos
+= conf
->min_offset_diff
;
5852 if ((mddev
->reshape_backwards
5853 ? (safepos
> writepos
&& readpos
< writepos
)
5854 : (safepos
< writepos
&& readpos
> writepos
)) ||
5855 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
5856 /* Cannot proceed until we've updated the superblock... */
5857 wait_event(conf
->wait_for_overlap
,
5858 atomic_read(&conf
->reshape_stripes
)==0
5859 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5860 if (atomic_read(&conf
->reshape_stripes
) != 0)
5862 mddev
->reshape_position
= conf
->reshape_progress
;
5863 mddev
->curr_resync_completed
= sector_nr
;
5864 if (!mddev
->reshape_backwards
)
5865 /* Can update recovery_offset */
5866 rdev_for_each(rdev
, mddev
)
5867 if (rdev
->raid_disk
>= 0 &&
5868 !test_bit(Journal
, &rdev
->flags
) &&
5869 !test_bit(In_sync
, &rdev
->flags
) &&
5870 rdev
->recovery_offset
< sector_nr
)
5871 rdev
->recovery_offset
= sector_nr
;
5873 conf
->reshape_checkpoint
= jiffies
;
5874 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
5875 md_wakeup_thread(mddev
->thread
);
5876 wait_event(mddev
->sb_wait
, mddev
->sb_flags
== 0 ||
5877 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5878 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5880 spin_lock_irq(&conf
->device_lock
);
5881 conf
->reshape_safe
= mddev
->reshape_position
;
5882 spin_unlock_irq(&conf
->device_lock
);
5883 wake_up(&conf
->wait_for_overlap
);
5884 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5887 INIT_LIST_HEAD(&stripes
);
5888 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
5890 int skipped_disk
= 0;
5891 sh
= raid5_get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
5892 set_bit(STRIPE_EXPANDING
, &sh
->state
);
5893 atomic_inc(&conf
->reshape_stripes
);
5894 /* If any of this stripe is beyond the end of the old
5895 * array, then we need to zero those blocks
5897 for (j
=sh
->disks
; j
--;) {
5899 if (j
== sh
->pd_idx
)
5901 if (conf
->level
== 6 &&
5904 s
= raid5_compute_blocknr(sh
, j
, 0);
5905 if (s
< raid5_size(mddev
, 0, 0)) {
5909 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
5910 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
5911 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
5913 if (!skipped_disk
) {
5914 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
5915 set_bit(STRIPE_HANDLE
, &sh
->state
);
5917 list_add(&sh
->lru
, &stripes
);
5919 spin_lock_irq(&conf
->device_lock
);
5920 if (mddev
->reshape_backwards
)
5921 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
5923 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
5924 spin_unlock_irq(&conf
->device_lock
);
5925 /* Ok, those stripe are ready. We can start scheduling
5926 * reads on the source stripes.
5927 * The source stripes are determined by mapping the first and last
5928 * block on the destination stripes.
5931 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
5934 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
5935 * new_data_disks
- 1),
5937 if (last_sector
>= mddev
->dev_sectors
)
5938 last_sector
= mddev
->dev_sectors
- 1;
5939 while (first_sector
<= last_sector
) {
5940 sh
= raid5_get_active_stripe(conf
, first_sector
, 1, 0, 1);
5941 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
5942 set_bit(STRIPE_HANDLE
, &sh
->state
);
5943 raid5_release_stripe(sh
);
5944 first_sector
+= STRIPE_SECTORS
;
5946 /* Now that the sources are clearly marked, we can release
5947 * the destination stripes
5949 while (!list_empty(&stripes
)) {
5950 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
5951 list_del_init(&sh
->lru
);
5952 raid5_release_stripe(sh
);
5954 /* If this takes us to the resync_max point where we have to pause,
5955 * then we need to write out the superblock.
5957 sector_nr
+= reshape_sectors
;
5958 retn
= reshape_sectors
;
5960 if (mddev
->curr_resync_completed
> mddev
->resync_max
||
5961 (sector_nr
- mddev
->curr_resync_completed
) * 2
5962 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
5963 /* Cannot proceed until we've updated the superblock... */
5964 wait_event(conf
->wait_for_overlap
,
5965 atomic_read(&conf
->reshape_stripes
) == 0
5966 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5967 if (atomic_read(&conf
->reshape_stripes
) != 0)
5969 mddev
->reshape_position
= conf
->reshape_progress
;
5970 mddev
->curr_resync_completed
= sector_nr
;
5971 if (!mddev
->reshape_backwards
)
5972 /* Can update recovery_offset */
5973 rdev_for_each(rdev
, mddev
)
5974 if (rdev
->raid_disk
>= 0 &&
5975 !test_bit(Journal
, &rdev
->flags
) &&
5976 !test_bit(In_sync
, &rdev
->flags
) &&
5977 rdev
->recovery_offset
< sector_nr
)
5978 rdev
->recovery_offset
= sector_nr
;
5979 conf
->reshape_checkpoint
= jiffies
;
5980 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
5981 md_wakeup_thread(mddev
->thread
);
5982 wait_event(mddev
->sb_wait
,
5983 !test_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
)
5984 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5985 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5987 spin_lock_irq(&conf
->device_lock
);
5988 conf
->reshape_safe
= mddev
->reshape_position
;
5989 spin_unlock_irq(&conf
->device_lock
);
5990 wake_up(&conf
->wait_for_overlap
);
5991 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5997 static inline sector_t
raid5_sync_request(struct mddev
*mddev
, sector_t sector_nr
,
6000 struct r5conf
*conf
= mddev
->private;
6001 struct stripe_head
*sh
;
6002 sector_t max_sector
= mddev
->dev_sectors
;
6003 sector_t sync_blocks
;
6004 int still_degraded
= 0;
6007 if (sector_nr
>= max_sector
) {
6008 /* just being told to finish up .. nothing much to do */
6010 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
6015 if (mddev
->curr_resync
< max_sector
) /* aborted */
6016 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
6018 else /* completed sync */
6020 bitmap_close_sync(mddev
->bitmap
);
6025 /* Allow raid5_quiesce to complete */
6026 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
6028 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
6029 return reshape_request(mddev
, sector_nr
, skipped
);
6031 /* No need to check resync_max as we never do more than one
6032 * stripe, and as resync_max will always be on a chunk boundary,
6033 * if the check in md_do_sync didn't fire, there is no chance
6034 * of overstepping resync_max here
6037 /* if there is too many failed drives and we are trying
6038 * to resync, then assert that we are finished, because there is
6039 * nothing we can do.
6041 if (mddev
->degraded
>= conf
->max_degraded
&&
6042 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
6043 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
6047 if (!test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
6049 !bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
6050 sync_blocks
>= STRIPE_SECTORS
) {
6051 /* we can skip this block, and probably more */
6052 sync_blocks
/= STRIPE_SECTORS
;
6054 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
6057 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
, false);
6059 sh
= raid5_get_active_stripe(conf
, sector_nr
, 0, 1, 0);
6061 sh
= raid5_get_active_stripe(conf
, sector_nr
, 0, 0, 0);
6062 /* make sure we don't swamp the stripe cache if someone else
6063 * is trying to get access
6065 schedule_timeout_uninterruptible(1);
6067 /* Need to check if array will still be degraded after recovery/resync
6068 * Note in case of > 1 drive failures it's possible we're rebuilding
6069 * one drive while leaving another faulty drive in array.
6072 for (i
= 0; i
< conf
->raid_disks
; i
++) {
6073 struct md_rdev
*rdev
= READ_ONCE(conf
->disks
[i
].rdev
);
6075 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
))
6080 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
6082 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
6083 set_bit(STRIPE_HANDLE
, &sh
->state
);
6085 raid5_release_stripe(sh
);
6087 return STRIPE_SECTORS
;
6090 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
,
6091 unsigned int offset
)
6093 /* We may not be able to submit a whole bio at once as there
6094 * may not be enough stripe_heads available.
6095 * We cannot pre-allocate enough stripe_heads as we may need
6096 * more than exist in the cache (if we allow ever large chunks).
6097 * So we do one stripe head at a time and record in
6098 * ->bi_hw_segments how many have been done.
6100 * We *know* that this entire raid_bio is in one chunk, so
6101 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6103 struct stripe_head
*sh
;
6105 sector_t sector
, logical_sector
, last_sector
;
6109 logical_sector
= raid_bio
->bi_iter
.bi_sector
&
6110 ~((sector_t
)STRIPE_SECTORS
-1);
6111 sector
= raid5_compute_sector(conf
, logical_sector
,
6113 last_sector
= bio_end_sector(raid_bio
);
6115 for (; logical_sector
< last_sector
;
6116 logical_sector
+= STRIPE_SECTORS
,
6117 sector
+= STRIPE_SECTORS
,
6121 /* already done this stripe */
6124 sh
= raid5_get_active_stripe(conf
, sector
, 0, 1, 1);
6127 /* failed to get a stripe - must wait */
6128 conf
->retry_read_aligned
= raid_bio
;
6129 conf
->retry_read_offset
= scnt
;
6133 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0, 0)) {
6134 raid5_release_stripe(sh
);
6135 conf
->retry_read_aligned
= raid_bio
;
6136 conf
->retry_read_offset
= scnt
;
6140 set_bit(R5_ReadNoMerge
, &sh
->dev
[dd_idx
].flags
);
6142 raid5_release_stripe(sh
);
6146 bio_endio(raid_bio
);
6148 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
6149 wake_up(&conf
->wait_for_quiescent
);
6153 static int handle_active_stripes(struct r5conf
*conf
, int group
,
6154 struct r5worker
*worker
,
6155 struct list_head
*temp_inactive_list
)
6157 struct stripe_head
*batch
[MAX_STRIPE_BATCH
], *sh
;
6158 int i
, batch_size
= 0, hash
;
6159 bool release_inactive
= false;
6161 while (batch_size
< MAX_STRIPE_BATCH
&&
6162 (sh
= __get_priority_stripe(conf
, group
)) != NULL
)
6163 batch
[batch_size
++] = sh
;
6165 if (batch_size
== 0) {
6166 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6167 if (!list_empty(temp_inactive_list
+ i
))
6169 if (i
== NR_STRIPE_HASH_LOCKS
) {
6170 spin_unlock_irq(&conf
->device_lock
);
6171 r5l_flush_stripe_to_raid(conf
->log
);
6172 spin_lock_irq(&conf
->device_lock
);
6175 release_inactive
= true;
6177 spin_unlock_irq(&conf
->device_lock
);
6179 release_inactive_stripe_list(conf
, temp_inactive_list
,
6180 NR_STRIPE_HASH_LOCKS
);
6182 r5l_flush_stripe_to_raid(conf
->log
);
6183 if (release_inactive
) {
6184 spin_lock_irq(&conf
->device_lock
);
6188 for (i
= 0; i
< batch_size
; i
++)
6189 handle_stripe(batch
[i
]);
6190 log_write_stripe_run(conf
);
6194 spin_lock_irq(&conf
->device_lock
);
6195 for (i
= 0; i
< batch_size
; i
++) {
6196 hash
= batch
[i
]->hash_lock_index
;
6197 __release_stripe(conf
, batch
[i
], &temp_inactive_list
[hash
]);
6202 static void raid5_do_work(struct work_struct
*work
)
6204 struct r5worker
*worker
= container_of(work
, struct r5worker
, work
);
6205 struct r5worker_group
*group
= worker
->group
;
6206 struct r5conf
*conf
= group
->conf
;
6207 struct mddev
*mddev
= conf
->mddev
;
6208 int group_id
= group
- conf
->worker_groups
;
6210 struct blk_plug plug
;
6212 pr_debug("+++ raid5worker active\n");
6214 blk_start_plug(&plug
);
6216 spin_lock_irq(&conf
->device_lock
);
6218 int batch_size
, released
;
6220 released
= release_stripe_list(conf
, worker
->temp_inactive_list
);
6222 batch_size
= handle_active_stripes(conf
, group_id
, worker
,
6223 worker
->temp_inactive_list
);
6224 worker
->working
= false;
6225 if (!batch_size
&& !released
)
6227 handled
+= batch_size
;
6228 wait_event_lock_irq(mddev
->sb_wait
,
6229 !test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
),
6232 pr_debug("%d stripes handled\n", handled
);
6234 spin_unlock_irq(&conf
->device_lock
);
6236 flush_deferred_bios(conf
);
6238 r5l_flush_stripe_to_raid(conf
->log
);
6240 async_tx_issue_pending_all();
6241 blk_finish_plug(&plug
);
6243 pr_debug("--- raid5worker inactive\n");
6247 * This is our raid5 kernel thread.
6249 * We scan the hash table for stripes which can be handled now.
6250 * During the scan, completed stripes are saved for us by the interrupt
6251 * handler, so that they will not have to wait for our next wakeup.
6253 static void raid5d(struct md_thread
*thread
)
6255 struct mddev
*mddev
= thread
->mddev
;
6256 struct r5conf
*conf
= mddev
->private;
6258 struct blk_plug plug
;
6260 pr_debug("+++ raid5d active\n");
6262 md_check_recovery(mddev
);
6264 blk_start_plug(&plug
);
6266 spin_lock_irq(&conf
->device_lock
);
6269 int batch_size
, released
;
6270 unsigned int offset
;
6272 released
= release_stripe_list(conf
, conf
->temp_inactive_list
);
6274 clear_bit(R5_DID_ALLOC
, &conf
->cache_state
);
6277 !list_empty(&conf
->bitmap_list
)) {
6278 /* Now is a good time to flush some bitmap updates */
6280 spin_unlock_irq(&conf
->device_lock
);
6281 bitmap_unplug(mddev
->bitmap
);
6282 spin_lock_irq(&conf
->device_lock
);
6283 conf
->seq_write
= conf
->seq_flush
;
6284 activate_bit_delay(conf
, conf
->temp_inactive_list
);
6286 raid5_activate_delayed(conf
);
6288 while ((bio
= remove_bio_from_retry(conf
, &offset
))) {
6290 spin_unlock_irq(&conf
->device_lock
);
6291 ok
= retry_aligned_read(conf
, bio
, offset
);
6292 spin_lock_irq(&conf
->device_lock
);
6298 batch_size
= handle_active_stripes(conf
, ANY_GROUP
, NULL
,
6299 conf
->temp_inactive_list
);
6300 if (!batch_size
&& !released
)
6302 handled
+= batch_size
;
6304 if (mddev
->sb_flags
& ~(1 << MD_SB_CHANGE_PENDING
)) {
6305 spin_unlock_irq(&conf
->device_lock
);
6306 md_check_recovery(mddev
);
6307 spin_lock_irq(&conf
->device_lock
);
6310 pr_debug("%d stripes handled\n", handled
);
6312 spin_unlock_irq(&conf
->device_lock
);
6313 if (test_and_clear_bit(R5_ALLOC_MORE
, &conf
->cache_state
) &&
6314 mutex_trylock(&conf
->cache_size_mutex
)) {
6315 grow_one_stripe(conf
, __GFP_NOWARN
);
6316 /* Set flag even if allocation failed. This helps
6317 * slow down allocation requests when mem is short
6319 set_bit(R5_DID_ALLOC
, &conf
->cache_state
);
6320 mutex_unlock(&conf
->cache_size_mutex
);
6323 flush_deferred_bios(conf
);
6325 r5l_flush_stripe_to_raid(conf
->log
);
6327 async_tx_issue_pending_all();
6328 blk_finish_plug(&plug
);
6330 pr_debug("--- raid5d inactive\n");
6334 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
6336 struct r5conf
*conf
;
6338 spin_lock(&mddev
->lock
);
6339 conf
= mddev
->private;
6341 ret
= sprintf(page
, "%d\n", conf
->min_nr_stripes
);
6342 spin_unlock(&mddev
->lock
);
6347 raid5_set_cache_size(struct mddev
*mddev
, int size
)
6349 struct r5conf
*conf
= mddev
->private;
6351 if (size
<= 16 || size
> 32768)
6354 conf
->min_nr_stripes
= size
;
6355 mutex_lock(&conf
->cache_size_mutex
);
6356 while (size
< conf
->max_nr_stripes
&&
6357 drop_one_stripe(conf
))
6359 mutex_unlock(&conf
->cache_size_mutex
);
6361 md_allow_write(mddev
);
6363 mutex_lock(&conf
->cache_size_mutex
);
6364 while (size
> conf
->max_nr_stripes
)
6365 if (!grow_one_stripe(conf
, GFP_KERNEL
))
6367 mutex_unlock(&conf
->cache_size_mutex
);
6371 EXPORT_SYMBOL(raid5_set_cache_size
);
6374 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
6376 struct r5conf
*conf
;
6380 if (len
>= PAGE_SIZE
)
6382 if (kstrtoul(page
, 10, &new))
6384 err
= mddev_lock(mddev
);
6387 conf
= mddev
->private;
6391 err
= raid5_set_cache_size(mddev
, new);
6392 mddev_unlock(mddev
);
6397 static struct md_sysfs_entry
6398 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
6399 raid5_show_stripe_cache_size
,
6400 raid5_store_stripe_cache_size
);
6403 raid5_show_rmw_level(struct mddev
*mddev
, char *page
)
6405 struct r5conf
*conf
= mddev
->private;
6407 return sprintf(page
, "%d\n", conf
->rmw_level
);
6413 raid5_store_rmw_level(struct mddev
*mddev
, const char *page
, size_t len
)
6415 struct r5conf
*conf
= mddev
->private;
6421 if (len
>= PAGE_SIZE
)
6424 if (kstrtoul(page
, 10, &new))
6427 if (new != PARITY_DISABLE_RMW
&& !raid6_call
.xor_syndrome
)
6430 if (new != PARITY_DISABLE_RMW
&&
6431 new != PARITY_ENABLE_RMW
&&
6432 new != PARITY_PREFER_RMW
)
6435 conf
->rmw_level
= new;
6439 static struct md_sysfs_entry
6440 raid5_rmw_level
= __ATTR(rmw_level
, S_IRUGO
| S_IWUSR
,
6441 raid5_show_rmw_level
,
6442 raid5_store_rmw_level
);
6446 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
6448 struct r5conf
*conf
;
6450 spin_lock(&mddev
->lock
);
6451 conf
= mddev
->private;
6453 ret
= sprintf(page
, "%d\n", conf
->bypass_threshold
);
6454 spin_unlock(&mddev
->lock
);
6459 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
6461 struct r5conf
*conf
;
6465 if (len
>= PAGE_SIZE
)
6467 if (kstrtoul(page
, 10, &new))
6470 err
= mddev_lock(mddev
);
6473 conf
= mddev
->private;
6476 else if (new > conf
->min_nr_stripes
)
6479 conf
->bypass_threshold
= new;
6480 mddev_unlock(mddev
);
6484 static struct md_sysfs_entry
6485 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
6487 raid5_show_preread_threshold
,
6488 raid5_store_preread_threshold
);
6491 raid5_show_skip_copy(struct mddev
*mddev
, char *page
)
6493 struct r5conf
*conf
;
6495 spin_lock(&mddev
->lock
);
6496 conf
= mddev
->private;
6498 ret
= sprintf(page
, "%d\n", conf
->skip_copy
);
6499 spin_unlock(&mddev
->lock
);
6504 raid5_store_skip_copy(struct mddev
*mddev
, const char *page
, size_t len
)
6506 struct r5conf
*conf
;
6510 if (len
>= PAGE_SIZE
)
6512 if (kstrtoul(page
, 10, &new))
6516 err
= mddev_lock(mddev
);
6519 conf
= mddev
->private;
6522 else if (new != conf
->skip_copy
) {
6523 mddev_suspend(mddev
);
6524 conf
->skip_copy
= new;
6526 mddev
->queue
->backing_dev_info
->capabilities
|=
6527 BDI_CAP_STABLE_WRITES
;
6529 mddev
->queue
->backing_dev_info
->capabilities
&=
6530 ~BDI_CAP_STABLE_WRITES
;
6531 mddev_resume(mddev
);
6533 mddev_unlock(mddev
);
6537 static struct md_sysfs_entry
6538 raid5_skip_copy
= __ATTR(skip_copy
, S_IRUGO
| S_IWUSR
,
6539 raid5_show_skip_copy
,
6540 raid5_store_skip_copy
);
6543 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
6545 struct r5conf
*conf
= mddev
->private;
6547 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
6552 static struct md_sysfs_entry
6553 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
6556 raid5_show_group_thread_cnt(struct mddev
*mddev
, char *page
)
6558 struct r5conf
*conf
;
6560 spin_lock(&mddev
->lock
);
6561 conf
= mddev
->private;
6563 ret
= sprintf(page
, "%d\n", conf
->worker_cnt_per_group
);
6564 spin_unlock(&mddev
->lock
);
6568 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6570 int *worker_cnt_per_group
,
6571 struct r5worker_group
**worker_groups
);
6573 raid5_store_group_thread_cnt(struct mddev
*mddev
, const char *page
, size_t len
)
6575 struct r5conf
*conf
;
6578 struct r5worker_group
*new_groups
, *old_groups
;
6579 int group_cnt
, worker_cnt_per_group
;
6581 if (len
>= PAGE_SIZE
)
6583 if (kstrtouint(page
, 10, &new))
6585 /* 8192 should be big enough */
6589 err
= mddev_lock(mddev
);
6592 conf
= mddev
->private;
6595 else if (new != conf
->worker_cnt_per_group
) {
6596 mddev_suspend(mddev
);
6598 old_groups
= conf
->worker_groups
;
6600 flush_workqueue(raid5_wq
);
6602 err
= alloc_thread_groups(conf
, new,
6603 &group_cnt
, &worker_cnt_per_group
,
6606 spin_lock_irq(&conf
->device_lock
);
6607 conf
->group_cnt
= group_cnt
;
6608 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6609 conf
->worker_groups
= new_groups
;
6610 spin_unlock_irq(&conf
->device_lock
);
6613 kfree(old_groups
[0].workers
);
6616 mddev_resume(mddev
);
6618 mddev_unlock(mddev
);
6623 static struct md_sysfs_entry
6624 raid5_group_thread_cnt
= __ATTR(group_thread_cnt
, S_IRUGO
| S_IWUSR
,
6625 raid5_show_group_thread_cnt
,
6626 raid5_store_group_thread_cnt
);
6628 static struct attribute
*raid5_attrs
[] = {
6629 &raid5_stripecache_size
.attr
,
6630 &raid5_stripecache_active
.attr
,
6631 &raid5_preread_bypass_threshold
.attr
,
6632 &raid5_group_thread_cnt
.attr
,
6633 &raid5_skip_copy
.attr
,
6634 &raid5_rmw_level
.attr
,
6635 &r5c_journal_mode
.attr
,
6638 static struct attribute_group raid5_attrs_group
= {
6640 .attrs
= raid5_attrs
,
6643 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6645 int *worker_cnt_per_group
,
6646 struct r5worker_group
**worker_groups
)
6650 struct r5worker
*workers
;
6652 *worker_cnt_per_group
= cnt
;
6655 *worker_groups
= NULL
;
6658 *group_cnt
= num_possible_nodes();
6659 size
= sizeof(struct r5worker
) * cnt
;
6660 workers
= kzalloc(size
* *group_cnt
, GFP_NOIO
);
6661 *worker_groups
= kzalloc(sizeof(struct r5worker_group
) *
6662 *group_cnt
, GFP_NOIO
);
6663 if (!*worker_groups
|| !workers
) {
6665 kfree(*worker_groups
);
6669 for (i
= 0; i
< *group_cnt
; i
++) {
6670 struct r5worker_group
*group
;
6672 group
= &(*worker_groups
)[i
];
6673 INIT_LIST_HEAD(&group
->handle_list
);
6674 INIT_LIST_HEAD(&group
->loprio_list
);
6676 group
->workers
= workers
+ i
* cnt
;
6678 for (j
= 0; j
< cnt
; j
++) {
6679 struct r5worker
*worker
= group
->workers
+ j
;
6680 worker
->group
= group
;
6681 INIT_WORK(&worker
->work
, raid5_do_work
);
6683 for (k
= 0; k
< NR_STRIPE_HASH_LOCKS
; k
++)
6684 INIT_LIST_HEAD(worker
->temp_inactive_list
+ k
);
6691 static void free_thread_groups(struct r5conf
*conf
)
6693 if (conf
->worker_groups
)
6694 kfree(conf
->worker_groups
[0].workers
);
6695 kfree(conf
->worker_groups
);
6696 conf
->worker_groups
= NULL
;
6700 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
6702 struct r5conf
*conf
= mddev
->private;
6705 sectors
= mddev
->dev_sectors
;
6707 /* size is defined by the smallest of previous and new size */
6708 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
6710 sectors
&= ~((sector_t
)conf
->chunk_sectors
- 1);
6711 sectors
&= ~((sector_t
)conf
->prev_chunk_sectors
- 1);
6712 return sectors
* (raid_disks
- conf
->max_degraded
);
6715 static void free_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6717 safe_put_page(percpu
->spare_page
);
6718 if (percpu
->scribble
)
6719 flex_array_free(percpu
->scribble
);
6720 percpu
->spare_page
= NULL
;
6721 percpu
->scribble
= NULL
;
6724 static int alloc_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6726 if (conf
->level
== 6 && !percpu
->spare_page
)
6727 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
6728 if (!percpu
->scribble
)
6729 percpu
->scribble
= scribble_alloc(max(conf
->raid_disks
,
6730 conf
->previous_raid_disks
),
6731 max(conf
->chunk_sectors
,
6732 conf
->prev_chunk_sectors
)
6736 if (!percpu
->scribble
|| (conf
->level
== 6 && !percpu
->spare_page
)) {
6737 free_scratch_buffer(conf
, percpu
);
6744 static int raid456_cpu_dead(unsigned int cpu
, struct hlist_node
*node
)
6746 struct r5conf
*conf
= hlist_entry_safe(node
, struct r5conf
, node
);
6748 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6752 static void raid5_free_percpu(struct r5conf
*conf
)
6757 cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE
, &conf
->node
);
6758 free_percpu(conf
->percpu
);
6761 static void free_conf(struct r5conf
*conf
)
6767 if (conf
->shrinker
.nr_deferred
)
6768 unregister_shrinker(&conf
->shrinker
);
6770 free_thread_groups(conf
);
6771 shrink_stripes(conf
);
6772 raid5_free_percpu(conf
);
6773 for (i
= 0; i
< conf
->pool_size
; i
++)
6774 if (conf
->disks
[i
].extra_page
)
6775 put_page(conf
->disks
[i
].extra_page
);
6777 if (conf
->bio_split
)
6778 bioset_free(conf
->bio_split
);
6779 kfree(conf
->stripe_hashtbl
);
6780 kfree(conf
->pending_data
);
6784 static int raid456_cpu_up_prepare(unsigned int cpu
, struct hlist_node
*node
)
6786 struct r5conf
*conf
= hlist_entry_safe(node
, struct r5conf
, node
);
6787 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
6789 if (alloc_scratch_buffer(conf
, percpu
)) {
6790 pr_warn("%s: failed memory allocation for cpu%u\n",
6797 static int raid5_alloc_percpu(struct r5conf
*conf
)
6801 conf
->percpu
= alloc_percpu(struct raid5_percpu
);
6805 err
= cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE
, &conf
->node
);
6807 conf
->scribble_disks
= max(conf
->raid_disks
,
6808 conf
->previous_raid_disks
);
6809 conf
->scribble_sectors
= max(conf
->chunk_sectors
,
6810 conf
->prev_chunk_sectors
);
6815 static unsigned long raid5_cache_scan(struct shrinker
*shrink
,
6816 struct shrink_control
*sc
)
6818 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
6819 unsigned long ret
= SHRINK_STOP
;
6821 if (mutex_trylock(&conf
->cache_size_mutex
)) {
6823 while (ret
< sc
->nr_to_scan
&&
6824 conf
->max_nr_stripes
> conf
->min_nr_stripes
) {
6825 if (drop_one_stripe(conf
) == 0) {
6831 mutex_unlock(&conf
->cache_size_mutex
);
6836 static unsigned long raid5_cache_count(struct shrinker
*shrink
,
6837 struct shrink_control
*sc
)
6839 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
6841 if (conf
->max_nr_stripes
< conf
->min_nr_stripes
)
6842 /* unlikely, but not impossible */
6844 return conf
->max_nr_stripes
- conf
->min_nr_stripes
;
6847 static struct r5conf
*setup_conf(struct mddev
*mddev
)
6849 struct r5conf
*conf
;
6850 int raid_disk
, memory
, max_disks
;
6851 struct md_rdev
*rdev
;
6852 struct disk_info
*disk
;
6855 int group_cnt
, worker_cnt_per_group
;
6856 struct r5worker_group
*new_group
;
6858 if (mddev
->new_level
!= 5
6859 && mddev
->new_level
!= 4
6860 && mddev
->new_level
!= 6) {
6861 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6862 mdname(mddev
), mddev
->new_level
);
6863 return ERR_PTR(-EIO
);
6865 if ((mddev
->new_level
== 5
6866 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
6867 (mddev
->new_level
== 6
6868 && !algorithm_valid_raid6(mddev
->new_layout
))) {
6869 pr_warn("md/raid:%s: layout %d not supported\n",
6870 mdname(mddev
), mddev
->new_layout
);
6871 return ERR_PTR(-EIO
);
6873 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
6874 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6875 mdname(mddev
), mddev
->raid_disks
);
6876 return ERR_PTR(-EINVAL
);
6879 if (!mddev
->new_chunk_sectors
||
6880 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
6881 !is_power_of_2(mddev
->new_chunk_sectors
)) {
6882 pr_warn("md/raid:%s: invalid chunk size %d\n",
6883 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
6884 return ERR_PTR(-EINVAL
);
6887 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
6890 INIT_LIST_HEAD(&conf
->free_list
);
6891 INIT_LIST_HEAD(&conf
->pending_list
);
6892 conf
->pending_data
= kzalloc(sizeof(struct r5pending_data
) *
6893 PENDING_IO_MAX
, GFP_KERNEL
);
6894 if (!conf
->pending_data
)
6896 for (i
= 0; i
< PENDING_IO_MAX
; i
++)
6897 list_add(&conf
->pending_data
[i
].sibling
, &conf
->free_list
);
6898 /* Don't enable multi-threading by default*/
6899 if (!alloc_thread_groups(conf
, 0, &group_cnt
, &worker_cnt_per_group
,
6901 conf
->group_cnt
= group_cnt
;
6902 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6903 conf
->worker_groups
= new_group
;
6906 spin_lock_init(&conf
->device_lock
);
6907 seqcount_init(&conf
->gen_lock
);
6908 mutex_init(&conf
->cache_size_mutex
);
6909 init_waitqueue_head(&conf
->wait_for_quiescent
);
6910 init_waitqueue_head(&conf
->wait_for_stripe
);
6911 init_waitqueue_head(&conf
->wait_for_overlap
);
6912 INIT_LIST_HEAD(&conf
->handle_list
);
6913 INIT_LIST_HEAD(&conf
->loprio_list
);
6914 INIT_LIST_HEAD(&conf
->hold_list
);
6915 INIT_LIST_HEAD(&conf
->delayed_list
);
6916 INIT_LIST_HEAD(&conf
->bitmap_list
);
6917 init_llist_head(&conf
->released_stripes
);
6918 atomic_set(&conf
->active_stripes
, 0);
6919 atomic_set(&conf
->preread_active_stripes
, 0);
6920 atomic_set(&conf
->active_aligned_reads
, 0);
6921 spin_lock_init(&conf
->pending_bios_lock
);
6922 conf
->batch_bio_dispatch
= true;
6923 rdev_for_each(rdev
, mddev
) {
6924 if (test_bit(Journal
, &rdev
->flags
))
6926 if (blk_queue_nonrot(bdev_get_queue(rdev
->bdev
))) {
6927 conf
->batch_bio_dispatch
= false;
6932 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
6933 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
6935 conf
->raid_disks
= mddev
->raid_disks
;
6936 if (mddev
->reshape_position
== MaxSector
)
6937 conf
->previous_raid_disks
= mddev
->raid_disks
;
6939 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
6940 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
6942 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
6948 for (i
= 0; i
< max_disks
; i
++) {
6949 conf
->disks
[i
].extra_page
= alloc_page(GFP_KERNEL
);
6950 if (!conf
->disks
[i
].extra_page
)
6954 conf
->bio_split
= bioset_create(BIO_POOL_SIZE
, 0, 0);
6955 if (!conf
->bio_split
)
6957 conf
->mddev
= mddev
;
6959 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
6962 /* We init hash_locks[0] separately to that it can be used
6963 * as the reference lock in the spin_lock_nest_lock() call
6964 * in lock_all_device_hash_locks_irq in order to convince
6965 * lockdep that we know what we are doing.
6967 spin_lock_init(conf
->hash_locks
);
6968 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6969 spin_lock_init(conf
->hash_locks
+ i
);
6971 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6972 INIT_LIST_HEAD(conf
->inactive_list
+ i
);
6974 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6975 INIT_LIST_HEAD(conf
->temp_inactive_list
+ i
);
6977 atomic_set(&conf
->r5c_cached_full_stripes
, 0);
6978 INIT_LIST_HEAD(&conf
->r5c_full_stripe_list
);
6979 atomic_set(&conf
->r5c_cached_partial_stripes
, 0);
6980 INIT_LIST_HEAD(&conf
->r5c_partial_stripe_list
);
6981 atomic_set(&conf
->r5c_flushing_full_stripes
, 0);
6982 atomic_set(&conf
->r5c_flushing_partial_stripes
, 0);
6984 conf
->level
= mddev
->new_level
;
6985 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
6986 if (raid5_alloc_percpu(conf
) != 0)
6989 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
6991 rdev_for_each(rdev
, mddev
) {
6992 raid_disk
= rdev
->raid_disk
;
6993 if (raid_disk
>= max_disks
6994 || raid_disk
< 0 || test_bit(Journal
, &rdev
->flags
))
6996 disk
= conf
->disks
+ raid_disk
;
6998 if (test_bit(Replacement
, &rdev
->flags
)) {
6999 if (disk
->replacement
)
7001 disk
->replacement
= rdev
;
7008 if (test_bit(In_sync
, &rdev
->flags
)) {
7009 char b
[BDEVNAME_SIZE
];
7010 pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7011 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
7012 } else if (rdev
->saved_raid_disk
!= raid_disk
)
7013 /* Cannot rely on bitmap to complete recovery */
7017 conf
->level
= mddev
->new_level
;
7018 if (conf
->level
== 6) {
7019 conf
->max_degraded
= 2;
7020 if (raid6_call
.xor_syndrome
)
7021 conf
->rmw_level
= PARITY_ENABLE_RMW
;
7023 conf
->rmw_level
= PARITY_DISABLE_RMW
;
7025 conf
->max_degraded
= 1;
7026 conf
->rmw_level
= PARITY_ENABLE_RMW
;
7028 conf
->algorithm
= mddev
->new_layout
;
7029 conf
->reshape_progress
= mddev
->reshape_position
;
7030 if (conf
->reshape_progress
!= MaxSector
) {
7031 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
7032 conf
->prev_algo
= mddev
->layout
;
7034 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
7035 conf
->prev_algo
= conf
->algorithm
;
7038 conf
->min_nr_stripes
= NR_STRIPES
;
7039 if (mddev
->reshape_position
!= MaxSector
) {
7040 int stripes
= max_t(int,
7041 ((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4,
7042 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4);
7043 conf
->min_nr_stripes
= max(NR_STRIPES
, stripes
);
7044 if (conf
->min_nr_stripes
!= NR_STRIPES
)
7045 pr_info("md/raid:%s: force stripe size %d for reshape\n",
7046 mdname(mddev
), conf
->min_nr_stripes
);
7048 memory
= conf
->min_nr_stripes
* (sizeof(struct stripe_head
) +
7049 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
7050 atomic_set(&conf
->empty_inactive_list_nr
, NR_STRIPE_HASH_LOCKS
);
7051 if (grow_stripes(conf
, conf
->min_nr_stripes
)) {
7052 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7053 mdname(mddev
), memory
);
7056 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev
), memory
);
7058 * Losing a stripe head costs more than the time to refill it,
7059 * it reduces the queue depth and so can hurt throughput.
7060 * So set it rather large, scaled by number of devices.
7062 conf
->shrinker
.seeks
= DEFAULT_SEEKS
* conf
->raid_disks
* 4;
7063 conf
->shrinker
.scan_objects
= raid5_cache_scan
;
7064 conf
->shrinker
.count_objects
= raid5_cache_count
;
7065 conf
->shrinker
.batch
= 128;
7066 conf
->shrinker
.flags
= 0;
7067 if (register_shrinker(&conf
->shrinker
)) {
7068 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7073 sprintf(pers_name
, "raid%d", mddev
->new_level
);
7074 conf
->thread
= md_register_thread(raid5d
, mddev
, pers_name
);
7075 if (!conf
->thread
) {
7076 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7086 return ERR_PTR(-EIO
);
7088 return ERR_PTR(-ENOMEM
);
7091 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
7094 case ALGORITHM_PARITY_0
:
7095 if (raid_disk
< max_degraded
)
7098 case ALGORITHM_PARITY_N
:
7099 if (raid_disk
>= raid_disks
- max_degraded
)
7102 case ALGORITHM_PARITY_0_6
:
7103 if (raid_disk
== 0 ||
7104 raid_disk
== raid_disks
- 1)
7107 case ALGORITHM_LEFT_ASYMMETRIC_6
:
7108 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
7109 case ALGORITHM_LEFT_SYMMETRIC_6
:
7110 case ALGORITHM_RIGHT_SYMMETRIC_6
:
7111 if (raid_disk
== raid_disks
- 1)
7117 static int raid5_run(struct mddev
*mddev
)
7119 struct r5conf
*conf
;
7120 int working_disks
= 0;
7121 int dirty_parity_disks
= 0;
7122 struct md_rdev
*rdev
;
7123 struct md_rdev
*journal_dev
= NULL
;
7124 sector_t reshape_offset
= 0;
7126 long long min_offset_diff
= 0;
7129 if (mddev_init_writes_pending(mddev
) < 0)
7132 if (mddev
->recovery_cp
!= MaxSector
)
7133 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7136 rdev_for_each(rdev
, mddev
) {
7139 if (test_bit(Journal
, &rdev
->flags
)) {
7143 if (rdev
->raid_disk
< 0)
7145 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
7147 min_offset_diff
= diff
;
7149 } else if (mddev
->reshape_backwards
&&
7150 diff
< min_offset_diff
)
7151 min_offset_diff
= diff
;
7152 else if (!mddev
->reshape_backwards
&&
7153 diff
> min_offset_diff
)
7154 min_offset_diff
= diff
;
7157 if ((test_bit(MD_HAS_JOURNAL
, &mddev
->flags
) || journal_dev
) &&
7158 (mddev
->bitmap_info
.offset
|| mddev
->bitmap_info
.file
)) {
7159 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7164 if (mddev
->reshape_position
!= MaxSector
) {
7165 /* Check that we can continue the reshape.
7166 * Difficulties arise if the stripe we would write to
7167 * next is at or after the stripe we would read from next.
7168 * For a reshape that changes the number of devices, this
7169 * is only possible for a very short time, and mdadm makes
7170 * sure that time appears to have past before assembling
7171 * the array. So we fail if that time hasn't passed.
7172 * For a reshape that keeps the number of devices the same
7173 * mdadm must be monitoring the reshape can keeping the
7174 * critical areas read-only and backed up. It will start
7175 * the array in read-only mode, so we check for that.
7177 sector_t here_new
, here_old
;
7179 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
7184 pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7189 if (mddev
->new_level
!= mddev
->level
) {
7190 pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7194 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
7195 /* reshape_position must be on a new-stripe boundary, and one
7196 * further up in new geometry must map after here in old
7198 * If the chunk sizes are different, then as we perform reshape
7199 * in units of the largest of the two, reshape_position needs
7200 * be a multiple of the largest chunk size times new data disks.
7202 here_new
= mddev
->reshape_position
;
7203 chunk_sectors
= max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
);
7204 new_data_disks
= mddev
->raid_disks
- max_degraded
;
7205 if (sector_div(here_new
, chunk_sectors
* new_data_disks
)) {
7206 pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7210 reshape_offset
= here_new
* chunk_sectors
;
7211 /* here_new is the stripe we will write to */
7212 here_old
= mddev
->reshape_position
;
7213 sector_div(here_old
, chunk_sectors
* (old_disks
-max_degraded
));
7214 /* here_old is the first stripe that we might need to read
7216 if (mddev
->delta_disks
== 0) {
7217 /* We cannot be sure it is safe to start an in-place
7218 * reshape. It is only safe if user-space is monitoring
7219 * and taking constant backups.
7220 * mdadm always starts a situation like this in
7221 * readonly mode so it can take control before
7222 * allowing any writes. So just check for that.
7224 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
7225 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
7226 /* not really in-place - so OK */;
7227 else if (mddev
->ro
== 0) {
7228 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7232 } else if (mddev
->reshape_backwards
7233 ? (here_new
* chunk_sectors
+ min_offset_diff
<=
7234 here_old
* chunk_sectors
)
7235 : (here_new
* chunk_sectors
>=
7236 here_old
* chunk_sectors
+ (-min_offset_diff
))) {
7237 /* Reading from the same stripe as writing to - bad */
7238 pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7242 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev
));
7243 /* OK, we should be able to continue; */
7245 BUG_ON(mddev
->level
!= mddev
->new_level
);
7246 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
7247 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
7248 BUG_ON(mddev
->delta_disks
!= 0);
7251 if (test_bit(MD_HAS_JOURNAL
, &mddev
->flags
) &&
7252 test_bit(MD_HAS_PPL
, &mddev
->flags
)) {
7253 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7255 clear_bit(MD_HAS_PPL
, &mddev
->flags
);
7256 clear_bit(MD_HAS_MULTIPLE_PPLS
, &mddev
->flags
);
7259 if (mddev
->private == NULL
)
7260 conf
= setup_conf(mddev
);
7262 conf
= mddev
->private;
7265 return PTR_ERR(conf
);
7267 if (test_bit(MD_HAS_JOURNAL
, &mddev
->flags
)) {
7269 pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7272 set_disk_ro(mddev
->gendisk
, 1);
7273 } else if (mddev
->recovery_cp
== MaxSector
)
7274 set_bit(MD_JOURNAL_CLEAN
, &mddev
->flags
);
7277 conf
->min_offset_diff
= min_offset_diff
;
7278 mddev
->thread
= conf
->thread
;
7279 conf
->thread
= NULL
;
7280 mddev
->private = conf
;
7282 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
7284 rdev
= conf
->disks
[i
].rdev
;
7285 if (!rdev
&& conf
->disks
[i
].replacement
) {
7286 /* The replacement is all we have yet */
7287 rdev
= conf
->disks
[i
].replacement
;
7288 conf
->disks
[i
].replacement
= NULL
;
7289 clear_bit(Replacement
, &rdev
->flags
);
7290 conf
->disks
[i
].rdev
= rdev
;
7294 if (conf
->disks
[i
].replacement
&&
7295 conf
->reshape_progress
!= MaxSector
) {
7296 /* replacements and reshape simply do not mix. */
7297 pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7300 if (test_bit(In_sync
, &rdev
->flags
)) {
7304 /* This disc is not fully in-sync. However if it
7305 * just stored parity (beyond the recovery_offset),
7306 * when we don't need to be concerned about the
7307 * array being dirty.
7308 * When reshape goes 'backwards', we never have
7309 * partially completed devices, so we only need
7310 * to worry about reshape going forwards.
7312 /* Hack because v0.91 doesn't store recovery_offset properly. */
7313 if (mddev
->major_version
== 0 &&
7314 mddev
->minor_version
> 90)
7315 rdev
->recovery_offset
= reshape_offset
;
7317 if (rdev
->recovery_offset
< reshape_offset
) {
7318 /* We need to check old and new layout */
7319 if (!only_parity(rdev
->raid_disk
,
7322 conf
->max_degraded
))
7325 if (!only_parity(rdev
->raid_disk
,
7327 conf
->previous_raid_disks
,
7328 conf
->max_degraded
))
7330 dirty_parity_disks
++;
7334 * 0 for a fully functional array, 1 or 2 for a degraded array.
7336 mddev
->degraded
= raid5_calc_degraded(conf
);
7338 if (has_failed(conf
)) {
7339 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7340 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
7344 /* device size must be a multiple of chunk size */
7345 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
7346 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
7348 if (mddev
->degraded
> dirty_parity_disks
&&
7349 mddev
->recovery_cp
!= MaxSector
) {
7350 if (test_bit(MD_HAS_PPL
, &mddev
->flags
))
7351 pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7353 else if (mddev
->ok_start_degraded
)
7354 pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7357 pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7363 pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7364 mdname(mddev
), conf
->level
,
7365 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
7368 print_raid5_conf(conf
);
7370 if (conf
->reshape_progress
!= MaxSector
) {
7371 conf
->reshape_safe
= conf
->reshape_progress
;
7372 atomic_set(&conf
->reshape_stripes
, 0);
7373 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
7374 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
7375 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
7376 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
7377 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
7381 /* Ok, everything is just fine now */
7382 if (mddev
->to_remove
== &raid5_attrs_group
)
7383 mddev
->to_remove
= NULL
;
7384 else if (mddev
->kobj
.sd
&&
7385 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
7386 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7388 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
7392 /* read-ahead size must cover two whole stripes, which
7393 * is 2 * (datadisks) * chunksize where 'n' is the
7394 * number of raid devices
7396 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
7397 int stripe
= data_disks
*
7398 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
7399 if (mddev
->queue
->backing_dev_info
->ra_pages
< 2 * stripe
)
7400 mddev
->queue
->backing_dev_info
->ra_pages
= 2 * stripe
;
7402 chunk_size
= mddev
->chunk_sectors
<< 9;
7403 blk_queue_io_min(mddev
->queue
, chunk_size
);
7404 blk_queue_io_opt(mddev
->queue
, chunk_size
*
7405 (conf
->raid_disks
- conf
->max_degraded
));
7406 mddev
->queue
->limits
.raid_partial_stripes_expensive
= 1;
7408 * We can only discard a whole stripe. It doesn't make sense to
7409 * discard data disk but write parity disk
7411 stripe
= stripe
* PAGE_SIZE
;
7412 /* Round up to power of 2, as discard handling
7413 * currently assumes that */
7414 while ((stripe
-1) & stripe
)
7415 stripe
= (stripe
| (stripe
-1)) + 1;
7416 mddev
->queue
->limits
.discard_alignment
= stripe
;
7417 mddev
->queue
->limits
.discard_granularity
= stripe
;
7419 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
7420 blk_queue_max_write_zeroes_sectors(mddev
->queue
, 0);
7422 rdev_for_each(rdev
, mddev
) {
7423 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
7424 rdev
->data_offset
<< 9);
7425 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
7426 rdev
->new_data_offset
<< 9);
7430 * zeroing is required, otherwise data
7431 * could be lost. Consider a scenario: discard a stripe
7432 * (the stripe could be inconsistent if
7433 * discard_zeroes_data is 0); write one disk of the
7434 * stripe (the stripe could be inconsistent again
7435 * depending on which disks are used to calculate
7436 * parity); the disk is broken; The stripe data of this
7439 * We only allow DISCARD if the sysadmin has confirmed that
7440 * only safe devices are in use by setting a module parameter.
7441 * A better idea might be to turn DISCARD into WRITE_ZEROES
7442 * requests, as that is required to be safe.
7444 if (devices_handle_discard_safely
&&
7445 mddev
->queue
->limits
.max_discard_sectors
>= (stripe
>> 9) &&
7446 mddev
->queue
->limits
.discard_granularity
>= stripe
)
7447 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
7450 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
7453 blk_queue_max_hw_sectors(mddev
->queue
, UINT_MAX
);
7456 if (log_init(conf
, journal_dev
, raid5_has_ppl(conf
)))
7461 md_unregister_thread(&mddev
->thread
);
7462 print_raid5_conf(conf
);
7464 mddev
->private = NULL
;
7465 pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev
));
7469 static void raid5_free(struct mddev
*mddev
, void *priv
)
7471 struct r5conf
*conf
= priv
;
7474 mddev
->to_remove
= &raid5_attrs_group
;
7477 static void raid5_status(struct seq_file
*seq
, struct mddev
*mddev
)
7479 struct r5conf
*conf
= mddev
->private;
7482 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
7483 conf
->chunk_sectors
/ 2, mddev
->layout
);
7484 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
7486 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7487 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
7488 seq_printf (seq
, "%s", rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
7491 seq_printf (seq
, "]");
7494 static void print_raid5_conf (struct r5conf
*conf
)
7497 struct disk_info
*tmp
;
7499 pr_debug("RAID conf printout:\n");
7501 pr_debug("(conf==NULL)\n");
7504 pr_debug(" --- level:%d rd:%d wd:%d\n", conf
->level
,
7506 conf
->raid_disks
- conf
->mddev
->degraded
);
7508 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7509 char b
[BDEVNAME_SIZE
];
7510 tmp
= conf
->disks
+ i
;
7512 pr_debug(" disk %d, o:%d, dev:%s\n",
7513 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
7514 bdevname(tmp
->rdev
->bdev
, b
));
7518 static int raid5_spare_active(struct mddev
*mddev
)
7521 struct r5conf
*conf
= mddev
->private;
7522 struct disk_info
*tmp
;
7524 unsigned long flags
;
7526 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7527 tmp
= conf
->disks
+ i
;
7528 if (tmp
->replacement
7529 && tmp
->replacement
->recovery_offset
== MaxSector
7530 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
7531 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
7532 /* Replacement has just become active. */
7534 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
7537 /* Replaced device not technically faulty,
7538 * but we need to be sure it gets removed
7539 * and never re-added.
7541 set_bit(Faulty
, &tmp
->rdev
->flags
);
7542 sysfs_notify_dirent_safe(
7543 tmp
->rdev
->sysfs_state
);
7545 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
7546 } else if (tmp
->rdev
7547 && tmp
->rdev
->recovery_offset
== MaxSector
7548 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
7549 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
7551 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
7554 spin_lock_irqsave(&conf
->device_lock
, flags
);
7555 mddev
->degraded
= raid5_calc_degraded(conf
);
7556 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7557 print_raid5_conf(conf
);
7561 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7563 struct r5conf
*conf
= mddev
->private;
7565 int number
= rdev
->raid_disk
;
7566 struct md_rdev
**rdevp
;
7567 struct disk_info
*p
= conf
->disks
+ number
;
7569 print_raid5_conf(conf
);
7570 if (test_bit(Journal
, &rdev
->flags
) && conf
->log
) {
7572 * we can't wait pending write here, as this is called in
7573 * raid5d, wait will deadlock.
7574 * neilb: there is no locking about new writes here,
7575 * so this cannot be safe.
7577 if (atomic_read(&conf
->active_stripes
) ||
7578 atomic_read(&conf
->r5c_cached_full_stripes
) ||
7579 atomic_read(&conf
->r5c_cached_partial_stripes
)) {
7585 if (rdev
== p
->rdev
)
7587 else if (rdev
== p
->replacement
)
7588 rdevp
= &p
->replacement
;
7592 if (number
>= conf
->raid_disks
&&
7593 conf
->reshape_progress
== MaxSector
)
7594 clear_bit(In_sync
, &rdev
->flags
);
7596 if (test_bit(In_sync
, &rdev
->flags
) ||
7597 atomic_read(&rdev
->nr_pending
)) {
7601 /* Only remove non-faulty devices if recovery
7604 if (!test_bit(Faulty
, &rdev
->flags
) &&
7605 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
7606 !has_failed(conf
) &&
7607 (!p
->replacement
|| p
->replacement
== rdev
) &&
7608 number
< conf
->raid_disks
) {
7613 if (!test_bit(RemoveSynchronized
, &rdev
->flags
)) {
7615 if (atomic_read(&rdev
->nr_pending
)) {
7616 /* lost the race, try later */
7622 err
= log_modify(conf
, rdev
, false);
7626 if (p
->replacement
) {
7627 /* We must have just cleared 'rdev' */
7628 p
->rdev
= p
->replacement
;
7629 clear_bit(Replacement
, &p
->replacement
->flags
);
7630 smp_mb(); /* Make sure other CPUs may see both as identical
7631 * but will never see neither - if they are careful
7633 p
->replacement
= NULL
;
7636 err
= log_modify(conf
, p
->rdev
, true);
7639 clear_bit(WantReplacement
, &rdev
->flags
);
7642 print_raid5_conf(conf
);
7646 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7648 struct r5conf
*conf
= mddev
->private;
7651 struct disk_info
*p
;
7653 int last
= conf
->raid_disks
- 1;
7655 if (test_bit(Journal
, &rdev
->flags
)) {
7659 rdev
->raid_disk
= 0;
7661 * The array is in readonly mode if journal is missing, so no
7662 * write requests running. We should be safe
7664 log_init(conf
, rdev
, false);
7667 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
7670 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
7671 /* no point adding a device */
7674 if (rdev
->raid_disk
>= 0)
7675 first
= last
= rdev
->raid_disk
;
7678 * find the disk ... but prefer rdev->saved_raid_disk
7681 if (rdev
->saved_raid_disk
>= 0 &&
7682 rdev
->saved_raid_disk
>= first
&&
7683 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
7684 first
= rdev
->saved_raid_disk
;
7686 for (disk
= first
; disk
<= last
; disk
++) {
7687 p
= conf
->disks
+ disk
;
7688 if (p
->rdev
== NULL
) {
7689 clear_bit(In_sync
, &rdev
->flags
);
7690 rdev
->raid_disk
= disk
;
7691 if (rdev
->saved_raid_disk
!= disk
)
7693 rcu_assign_pointer(p
->rdev
, rdev
);
7695 err
= log_modify(conf
, rdev
, true);
7700 for (disk
= first
; disk
<= last
; disk
++) {
7701 p
= conf
->disks
+ disk
;
7702 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
7703 p
->replacement
== NULL
) {
7704 clear_bit(In_sync
, &rdev
->flags
);
7705 set_bit(Replacement
, &rdev
->flags
);
7706 rdev
->raid_disk
= disk
;
7709 rcu_assign_pointer(p
->replacement
, rdev
);
7714 print_raid5_conf(conf
);
7718 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
7720 /* no resync is happening, and there is enough space
7721 * on all devices, so we can resize.
7722 * We need to make sure resync covers any new space.
7723 * If the array is shrinking we should possibly wait until
7724 * any io in the removed space completes, but it hardly seems
7728 struct r5conf
*conf
= mddev
->private;
7730 if (conf
->log
|| raid5_has_ppl(conf
))
7732 sectors
&= ~((sector_t
)conf
->chunk_sectors
- 1);
7733 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
7734 if (mddev
->external_size
&&
7735 mddev
->array_sectors
> newsize
)
7737 if (mddev
->bitmap
) {
7738 int ret
= bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
7742 md_set_array_sectors(mddev
, newsize
);
7743 if (sectors
> mddev
->dev_sectors
&&
7744 mddev
->recovery_cp
> mddev
->dev_sectors
) {
7745 mddev
->recovery_cp
= mddev
->dev_sectors
;
7746 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
7748 mddev
->dev_sectors
= sectors
;
7749 mddev
->resync_max_sectors
= sectors
;
7753 static int check_stripe_cache(struct mddev
*mddev
)
7755 /* Can only proceed if there are plenty of stripe_heads.
7756 * We need a minimum of one full stripe,, and for sensible progress
7757 * it is best to have about 4 times that.
7758 * If we require 4 times, then the default 256 4K stripe_heads will
7759 * allow for chunk sizes up to 256K, which is probably OK.
7760 * If the chunk size is greater, user-space should request more
7761 * stripe_heads first.
7763 struct r5conf
*conf
= mddev
->private;
7764 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7765 > conf
->min_nr_stripes
||
7766 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7767 > conf
->min_nr_stripes
) {
7768 pr_warn("md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7770 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
7777 static int check_reshape(struct mddev
*mddev
)
7779 struct r5conf
*conf
= mddev
->private;
7781 if (conf
->log
|| raid5_has_ppl(conf
))
7783 if (mddev
->delta_disks
== 0 &&
7784 mddev
->new_layout
== mddev
->layout
&&
7785 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
7786 return 0; /* nothing to do */
7787 if (has_failed(conf
))
7789 if (mddev
->delta_disks
< 0 && mddev
->reshape_position
== MaxSector
) {
7790 /* We might be able to shrink, but the devices must
7791 * be made bigger first.
7792 * For raid6, 4 is the minimum size.
7793 * Otherwise 2 is the minimum
7796 if (mddev
->level
== 6)
7798 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
7802 if (!check_stripe_cache(mddev
))
7805 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
||
7806 mddev
->delta_disks
> 0)
7807 if (resize_chunks(conf
,
7808 conf
->previous_raid_disks
7809 + max(0, mddev
->delta_disks
),
7810 max(mddev
->new_chunk_sectors
,
7811 mddev
->chunk_sectors
)
7815 if (conf
->previous_raid_disks
+ mddev
->delta_disks
<= conf
->pool_size
)
7816 return 0; /* never bother to shrink */
7817 return resize_stripes(conf
, (conf
->previous_raid_disks
7818 + mddev
->delta_disks
));
7821 static int raid5_start_reshape(struct mddev
*mddev
)
7823 struct r5conf
*conf
= mddev
->private;
7824 struct md_rdev
*rdev
;
7826 unsigned long flags
;
7828 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
7831 if (!check_stripe_cache(mddev
))
7834 if (has_failed(conf
))
7837 rdev_for_each(rdev
, mddev
) {
7838 if (!test_bit(In_sync
, &rdev
->flags
)
7839 && !test_bit(Faulty
, &rdev
->flags
))
7843 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
7844 /* Not enough devices even to make a degraded array
7849 /* Refuse to reduce size of the array. Any reductions in
7850 * array size must be through explicit setting of array_size
7853 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
7854 < mddev
->array_sectors
) {
7855 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7860 atomic_set(&conf
->reshape_stripes
, 0);
7861 spin_lock_irq(&conf
->device_lock
);
7862 write_seqcount_begin(&conf
->gen_lock
);
7863 conf
->previous_raid_disks
= conf
->raid_disks
;
7864 conf
->raid_disks
+= mddev
->delta_disks
;
7865 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
7866 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
7867 conf
->prev_algo
= conf
->algorithm
;
7868 conf
->algorithm
= mddev
->new_layout
;
7870 /* Code that selects data_offset needs to see the generation update
7871 * if reshape_progress has been set - so a memory barrier needed.
7874 if (mddev
->reshape_backwards
)
7875 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
7877 conf
->reshape_progress
= 0;
7878 conf
->reshape_safe
= conf
->reshape_progress
;
7879 write_seqcount_end(&conf
->gen_lock
);
7880 spin_unlock_irq(&conf
->device_lock
);
7882 /* Now make sure any requests that proceeded on the assumption
7883 * the reshape wasn't running - like Discard or Read - have
7886 mddev_suspend(mddev
);
7887 mddev_resume(mddev
);
7889 /* Add some new drives, as many as will fit.
7890 * We know there are enough to make the newly sized array work.
7891 * Don't add devices if we are reducing the number of
7892 * devices in the array. This is because it is not possible
7893 * to correctly record the "partially reconstructed" state of
7894 * such devices during the reshape and confusion could result.
7896 if (mddev
->delta_disks
>= 0) {
7897 rdev_for_each(rdev
, mddev
)
7898 if (rdev
->raid_disk
< 0 &&
7899 !test_bit(Faulty
, &rdev
->flags
)) {
7900 if (raid5_add_disk(mddev
, rdev
) == 0) {
7902 >= conf
->previous_raid_disks
)
7903 set_bit(In_sync
, &rdev
->flags
);
7905 rdev
->recovery_offset
= 0;
7907 if (sysfs_link_rdev(mddev
, rdev
))
7908 /* Failure here is OK */;
7910 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
7911 && !test_bit(Faulty
, &rdev
->flags
)) {
7912 /* This is a spare that was manually added */
7913 set_bit(In_sync
, &rdev
->flags
);
7916 /* When a reshape changes the number of devices,
7917 * ->degraded is measured against the larger of the
7918 * pre and post number of devices.
7920 spin_lock_irqsave(&conf
->device_lock
, flags
);
7921 mddev
->degraded
= raid5_calc_degraded(conf
);
7922 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7924 mddev
->raid_disks
= conf
->raid_disks
;
7925 mddev
->reshape_position
= conf
->reshape_progress
;
7926 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
7928 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
7929 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
7930 clear_bit(MD_RECOVERY_DONE
, &mddev
->recovery
);
7931 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
7932 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
7933 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
7935 if (!mddev
->sync_thread
) {
7936 mddev
->recovery
= 0;
7937 spin_lock_irq(&conf
->device_lock
);
7938 write_seqcount_begin(&conf
->gen_lock
);
7939 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
7940 mddev
->new_chunk_sectors
=
7941 conf
->chunk_sectors
= conf
->prev_chunk_sectors
;
7942 mddev
->new_layout
= conf
->algorithm
= conf
->prev_algo
;
7943 rdev_for_each(rdev
, mddev
)
7944 rdev
->new_data_offset
= rdev
->data_offset
;
7946 conf
->generation
--;
7947 conf
->reshape_progress
= MaxSector
;
7948 mddev
->reshape_position
= MaxSector
;
7949 write_seqcount_end(&conf
->gen_lock
);
7950 spin_unlock_irq(&conf
->device_lock
);
7953 conf
->reshape_checkpoint
= jiffies
;
7954 md_wakeup_thread(mddev
->sync_thread
);
7955 md_new_event(mddev
);
7959 /* This is called from the reshape thread and should make any
7960 * changes needed in 'conf'
7962 static void end_reshape(struct r5conf
*conf
)
7965 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
7966 struct md_rdev
*rdev
;
7968 spin_lock_irq(&conf
->device_lock
);
7969 conf
->previous_raid_disks
= conf
->raid_disks
;
7970 md_finish_reshape(conf
->mddev
);
7972 conf
->reshape_progress
= MaxSector
;
7973 conf
->mddev
->reshape_position
= MaxSector
;
7974 rdev_for_each(rdev
, conf
->mddev
)
7975 if (rdev
->raid_disk
>= 0 &&
7976 !test_bit(Journal
, &rdev
->flags
) &&
7977 !test_bit(In_sync
, &rdev
->flags
))
7978 rdev
->recovery_offset
= MaxSector
;
7979 spin_unlock_irq(&conf
->device_lock
);
7980 wake_up(&conf
->wait_for_overlap
);
7982 /* read-ahead size must cover two whole stripes, which is
7983 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7985 if (conf
->mddev
->queue
) {
7986 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
7987 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
7989 if (conf
->mddev
->queue
->backing_dev_info
->ra_pages
< 2 * stripe
)
7990 conf
->mddev
->queue
->backing_dev_info
->ra_pages
= 2 * stripe
;
7995 /* This is called from the raid5d thread with mddev_lock held.
7996 * It makes config changes to the device.
7998 static void raid5_finish_reshape(struct mddev
*mddev
)
8000 struct r5conf
*conf
= mddev
->private;
8002 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
8004 if (mddev
->delta_disks
> 0) {
8005 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
8007 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
8008 revalidate_disk(mddev
->gendisk
);
8012 spin_lock_irq(&conf
->device_lock
);
8013 mddev
->degraded
= raid5_calc_degraded(conf
);
8014 spin_unlock_irq(&conf
->device_lock
);
8015 for (d
= conf
->raid_disks
;
8016 d
< conf
->raid_disks
- mddev
->delta_disks
;
8018 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
8020 clear_bit(In_sync
, &rdev
->flags
);
8021 rdev
= conf
->disks
[d
].replacement
;
8023 clear_bit(In_sync
, &rdev
->flags
);
8026 mddev
->layout
= conf
->algorithm
;
8027 mddev
->chunk_sectors
= conf
->chunk_sectors
;
8028 mddev
->reshape_position
= MaxSector
;
8029 mddev
->delta_disks
= 0;
8030 mddev
->reshape_backwards
= 0;
8034 static void raid5_quiesce(struct mddev
*mddev
, int quiesce
)
8036 struct r5conf
*conf
= mddev
->private;
8039 /* stop all writes */
8040 lock_all_device_hash_locks_irq(conf
);
8041 /* '2' tells resync/reshape to pause so that all
8042 * active stripes can drain
8044 r5c_flush_cache(conf
, INT_MAX
);
8046 wait_event_cmd(conf
->wait_for_quiescent
,
8047 atomic_read(&conf
->active_stripes
) == 0 &&
8048 atomic_read(&conf
->active_aligned_reads
) == 0,
8049 unlock_all_device_hash_locks_irq(conf
),
8050 lock_all_device_hash_locks_irq(conf
));
8052 unlock_all_device_hash_locks_irq(conf
);
8053 /* allow reshape to continue */
8054 wake_up(&conf
->wait_for_overlap
);
8056 /* re-enable writes */
8057 lock_all_device_hash_locks_irq(conf
);
8059 wake_up(&conf
->wait_for_quiescent
);
8060 wake_up(&conf
->wait_for_overlap
);
8061 unlock_all_device_hash_locks_irq(conf
);
8063 r5l_quiesce(conf
->log
, quiesce
);
8066 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
8068 struct r0conf
*raid0_conf
= mddev
->private;
8071 /* for raid0 takeover only one zone is supported */
8072 if (raid0_conf
->nr_strip_zones
> 1) {
8073 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8075 return ERR_PTR(-EINVAL
);
8078 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
8079 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
8080 mddev
->dev_sectors
= sectors
;
8081 mddev
->new_level
= level
;
8082 mddev
->new_layout
= ALGORITHM_PARITY_N
;
8083 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
8084 mddev
->raid_disks
+= 1;
8085 mddev
->delta_disks
= 1;
8086 /* make sure it will be not marked as dirty */
8087 mddev
->recovery_cp
= MaxSector
;
8089 return setup_conf(mddev
);
8092 static void *raid5_takeover_raid1(struct mddev
*mddev
)
8097 if (mddev
->raid_disks
!= 2 ||
8098 mddev
->degraded
> 1)
8099 return ERR_PTR(-EINVAL
);
8101 /* Should check if there are write-behind devices? */
8103 chunksect
= 64*2; /* 64K by default */
8105 /* The array must be an exact multiple of chunksize */
8106 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
8109 if ((chunksect
<<9) < STRIPE_SIZE
)
8110 /* array size does not allow a suitable chunk size */
8111 return ERR_PTR(-EINVAL
);
8113 mddev
->new_level
= 5;
8114 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
8115 mddev
->new_chunk_sectors
= chunksect
;
8117 ret
= setup_conf(mddev
);
8119 mddev_clear_unsupported_flags(mddev
,
8120 UNSUPPORTED_MDDEV_FLAGS
);
8124 static void *raid5_takeover_raid6(struct mddev
*mddev
)
8128 switch (mddev
->layout
) {
8129 case ALGORITHM_LEFT_ASYMMETRIC_6
:
8130 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
8132 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
8133 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
8135 case ALGORITHM_LEFT_SYMMETRIC_6
:
8136 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
8138 case ALGORITHM_RIGHT_SYMMETRIC_6
:
8139 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
8141 case ALGORITHM_PARITY_0_6
:
8142 new_layout
= ALGORITHM_PARITY_0
;
8144 case ALGORITHM_PARITY_N
:
8145 new_layout
= ALGORITHM_PARITY_N
;
8148 return ERR_PTR(-EINVAL
);
8150 mddev
->new_level
= 5;
8151 mddev
->new_layout
= new_layout
;
8152 mddev
->delta_disks
= -1;
8153 mddev
->raid_disks
-= 1;
8154 return setup_conf(mddev
);
8157 static int raid5_check_reshape(struct mddev
*mddev
)
8159 /* For a 2-drive array, the layout and chunk size can be changed
8160 * immediately as not restriping is needed.
8161 * For larger arrays we record the new value - after validation
8162 * to be used by a reshape pass.
8164 struct r5conf
*conf
= mddev
->private;
8165 int new_chunk
= mddev
->new_chunk_sectors
;
8167 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
8169 if (new_chunk
> 0) {
8170 if (!is_power_of_2(new_chunk
))
8172 if (new_chunk
< (PAGE_SIZE
>>9))
8174 if (mddev
->array_sectors
& (new_chunk
-1))
8175 /* not factor of array size */
8179 /* They look valid */
8181 if (mddev
->raid_disks
== 2) {
8182 /* can make the change immediately */
8183 if (mddev
->new_layout
>= 0) {
8184 conf
->algorithm
= mddev
->new_layout
;
8185 mddev
->layout
= mddev
->new_layout
;
8187 if (new_chunk
> 0) {
8188 conf
->chunk_sectors
= new_chunk
;
8189 mddev
->chunk_sectors
= new_chunk
;
8191 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
8192 md_wakeup_thread(mddev
->thread
);
8194 return check_reshape(mddev
);
8197 static int raid6_check_reshape(struct mddev
*mddev
)
8199 int new_chunk
= mddev
->new_chunk_sectors
;
8201 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
8203 if (new_chunk
> 0) {
8204 if (!is_power_of_2(new_chunk
))
8206 if (new_chunk
< (PAGE_SIZE
>> 9))
8208 if (mddev
->array_sectors
& (new_chunk
-1))
8209 /* not factor of array size */
8213 /* They look valid */
8214 return check_reshape(mddev
);
8217 static void *raid5_takeover(struct mddev
*mddev
)
8219 /* raid5 can take over:
8220 * raid0 - if there is only one strip zone - make it a raid4 layout
8221 * raid1 - if there are two drives. We need to know the chunk size
8222 * raid4 - trivial - just use a raid4 layout.
8223 * raid6 - Providing it is a *_6 layout
8225 if (mddev
->level
== 0)
8226 return raid45_takeover_raid0(mddev
, 5);
8227 if (mddev
->level
== 1)
8228 return raid5_takeover_raid1(mddev
);
8229 if (mddev
->level
== 4) {
8230 mddev
->new_layout
= ALGORITHM_PARITY_N
;
8231 mddev
->new_level
= 5;
8232 return setup_conf(mddev
);
8234 if (mddev
->level
== 6)
8235 return raid5_takeover_raid6(mddev
);
8237 return ERR_PTR(-EINVAL
);
8240 static void *raid4_takeover(struct mddev
*mddev
)
8242 /* raid4 can take over:
8243 * raid0 - if there is only one strip zone
8244 * raid5 - if layout is right
8246 if (mddev
->level
== 0)
8247 return raid45_takeover_raid0(mddev
, 4);
8248 if (mddev
->level
== 5 &&
8249 mddev
->layout
== ALGORITHM_PARITY_N
) {
8250 mddev
->new_layout
= 0;
8251 mddev
->new_level
= 4;
8252 return setup_conf(mddev
);
8254 return ERR_PTR(-EINVAL
);
8257 static struct md_personality raid5_personality
;
8259 static void *raid6_takeover(struct mddev
*mddev
)
8261 /* Currently can only take over a raid5. We map the
8262 * personality to an equivalent raid6 personality
8263 * with the Q block at the end.
8267 if (mddev
->pers
!= &raid5_personality
)
8268 return ERR_PTR(-EINVAL
);
8269 if (mddev
->degraded
> 1)
8270 return ERR_PTR(-EINVAL
);
8271 if (mddev
->raid_disks
> 253)
8272 return ERR_PTR(-EINVAL
);
8273 if (mddev
->raid_disks
< 3)
8274 return ERR_PTR(-EINVAL
);
8276 switch (mddev
->layout
) {
8277 case ALGORITHM_LEFT_ASYMMETRIC
:
8278 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
8280 case ALGORITHM_RIGHT_ASYMMETRIC
:
8281 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
8283 case ALGORITHM_LEFT_SYMMETRIC
:
8284 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
8286 case ALGORITHM_RIGHT_SYMMETRIC
:
8287 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
8289 case ALGORITHM_PARITY_0
:
8290 new_layout
= ALGORITHM_PARITY_0_6
;
8292 case ALGORITHM_PARITY_N
:
8293 new_layout
= ALGORITHM_PARITY_N
;
8296 return ERR_PTR(-EINVAL
);
8298 mddev
->new_level
= 6;
8299 mddev
->new_layout
= new_layout
;
8300 mddev
->delta_disks
= 1;
8301 mddev
->raid_disks
+= 1;
8302 return setup_conf(mddev
);
8305 static int raid5_change_consistency_policy(struct mddev
*mddev
, const char *buf
)
8307 struct r5conf
*conf
;
8310 err
= mddev_lock(mddev
);
8313 conf
= mddev
->private;
8315 mddev_unlock(mddev
);
8319 if (strncmp(buf
, "ppl", 3) == 0) {
8320 /* ppl only works with RAID 5 */
8321 if (!raid5_has_ppl(conf
) && conf
->level
== 5) {
8322 err
= log_init(conf
, NULL
, true);
8324 err
= resize_stripes(conf
, conf
->pool_size
);
8330 } else if (strncmp(buf
, "resync", 6) == 0) {
8331 if (raid5_has_ppl(conf
)) {
8332 mddev_suspend(mddev
);
8334 mddev_resume(mddev
);
8335 err
= resize_stripes(conf
, conf
->pool_size
);
8336 } else if (test_bit(MD_HAS_JOURNAL
, &conf
->mddev
->flags
) &&
8337 r5l_log_disk_error(conf
)) {
8338 bool journal_dev_exists
= false;
8339 struct md_rdev
*rdev
;
8341 rdev_for_each(rdev
, mddev
)
8342 if (test_bit(Journal
, &rdev
->flags
)) {
8343 journal_dev_exists
= true;
8347 if (!journal_dev_exists
) {
8348 mddev_suspend(mddev
);
8349 clear_bit(MD_HAS_JOURNAL
, &mddev
->flags
);
8350 mddev_resume(mddev
);
8351 } else /* need remove journal device first */
8360 md_update_sb(mddev
, 1);
8362 mddev_unlock(mddev
);
8367 static struct md_personality raid6_personality
=
8371 .owner
= THIS_MODULE
,
8372 .make_request
= raid5_make_request
,
8375 .status
= raid5_status
,
8376 .error_handler
= raid5_error
,
8377 .hot_add_disk
= raid5_add_disk
,
8378 .hot_remove_disk
= raid5_remove_disk
,
8379 .spare_active
= raid5_spare_active
,
8380 .sync_request
= raid5_sync_request
,
8381 .resize
= raid5_resize
,
8383 .check_reshape
= raid6_check_reshape
,
8384 .start_reshape
= raid5_start_reshape
,
8385 .finish_reshape
= raid5_finish_reshape
,
8386 .quiesce
= raid5_quiesce
,
8387 .takeover
= raid6_takeover
,
8388 .congested
= raid5_congested
,
8389 .change_consistency_policy
= raid5_change_consistency_policy
,
8391 static struct md_personality raid5_personality
=
8395 .owner
= THIS_MODULE
,
8396 .make_request
= raid5_make_request
,
8399 .status
= raid5_status
,
8400 .error_handler
= raid5_error
,
8401 .hot_add_disk
= raid5_add_disk
,
8402 .hot_remove_disk
= raid5_remove_disk
,
8403 .spare_active
= raid5_spare_active
,
8404 .sync_request
= raid5_sync_request
,
8405 .resize
= raid5_resize
,
8407 .check_reshape
= raid5_check_reshape
,
8408 .start_reshape
= raid5_start_reshape
,
8409 .finish_reshape
= raid5_finish_reshape
,
8410 .quiesce
= raid5_quiesce
,
8411 .takeover
= raid5_takeover
,
8412 .congested
= raid5_congested
,
8413 .change_consistency_policy
= raid5_change_consistency_policy
,
8416 static struct md_personality raid4_personality
=
8420 .owner
= THIS_MODULE
,
8421 .make_request
= raid5_make_request
,
8424 .status
= raid5_status
,
8425 .error_handler
= raid5_error
,
8426 .hot_add_disk
= raid5_add_disk
,
8427 .hot_remove_disk
= raid5_remove_disk
,
8428 .spare_active
= raid5_spare_active
,
8429 .sync_request
= raid5_sync_request
,
8430 .resize
= raid5_resize
,
8432 .check_reshape
= raid5_check_reshape
,
8433 .start_reshape
= raid5_start_reshape
,
8434 .finish_reshape
= raid5_finish_reshape
,
8435 .quiesce
= raid5_quiesce
,
8436 .takeover
= raid4_takeover
,
8437 .congested
= raid5_congested
,
8438 .change_consistency_policy
= raid5_change_consistency_policy
,
8441 static int __init
raid5_init(void)
8445 raid5_wq
= alloc_workqueue("raid5wq",
8446 WQ_UNBOUND
|WQ_MEM_RECLAIM
|WQ_CPU_INTENSIVE
|WQ_SYSFS
, 0);
8450 ret
= cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE
,
8452 raid456_cpu_up_prepare
,
8455 destroy_workqueue(raid5_wq
);
8458 register_md_personality(&raid6_personality
);
8459 register_md_personality(&raid5_personality
);
8460 register_md_personality(&raid4_personality
);
8464 static void raid5_exit(void)
8466 unregister_md_personality(&raid6_personality
);
8467 unregister_md_personality(&raid5_personality
);
8468 unregister_md_personality(&raid4_personality
);
8469 cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE
);
8470 destroy_workqueue(raid5_wq
);
8473 module_init(raid5_init
);
8474 module_exit(raid5_exit
);
8475 MODULE_LICENSE("GPL");
8476 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8477 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8478 MODULE_ALIAS("md-raid5");
8479 MODULE_ALIAS("md-raid4");
8480 MODULE_ALIAS("md-level-5");
8481 MODULE_ALIAS("md-level-4");
8482 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8483 MODULE_ALIAS("md-raid6");
8484 MODULE_ALIAS("md-level-6");
8486 /* This used to be two separate modules, they were: */
8487 MODULE_ALIAS("raid5");
8488 MODULE_ALIAS("raid6");