2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
56 #include <linux/nodemask.h>
57 #include <linux/flex_array.h>
58 #include <linux/sched/signal.h>
60 #include <trace/events/block.h>
61 #include <linux/list_sort.h>
67 #include "raid5-log.h"
69 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
71 #define cpu_to_group(cpu) cpu_to_node(cpu)
72 #define ANY_GROUP NUMA_NO_NODE
74 static bool devices_handle_discard_safely
= false;
75 module_param(devices_handle_discard_safely
, bool, 0644);
76 MODULE_PARM_DESC(devices_handle_discard_safely
,
77 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
78 static struct workqueue_struct
*raid5_wq
;
80 static inline struct hlist_head
*stripe_hash(struct r5conf
*conf
, sector_t sect
)
82 int hash
= (sect
>> STRIPE_SHIFT
) & HASH_MASK
;
83 return &conf
->stripe_hashtbl
[hash
];
86 static inline int stripe_hash_locks_hash(sector_t sect
)
88 return (sect
>> STRIPE_SHIFT
) & STRIPE_HASH_LOCKS_MASK
;
91 static inline void lock_device_hash_lock(struct r5conf
*conf
, int hash
)
93 spin_lock_irq(conf
->hash_locks
+ hash
);
94 spin_lock(&conf
->device_lock
);
97 static inline void unlock_device_hash_lock(struct r5conf
*conf
, int hash
)
99 spin_unlock(&conf
->device_lock
);
100 spin_unlock_irq(conf
->hash_locks
+ hash
);
103 static inline void lock_all_device_hash_locks_irq(struct r5conf
*conf
)
106 spin_lock_irq(conf
->hash_locks
);
107 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
108 spin_lock_nest_lock(conf
->hash_locks
+ i
, conf
->hash_locks
);
109 spin_lock(&conf
->device_lock
);
112 static inline void unlock_all_device_hash_locks_irq(struct r5conf
*conf
)
115 spin_unlock(&conf
->device_lock
);
116 for (i
= NR_STRIPE_HASH_LOCKS
- 1; i
; i
--)
117 spin_unlock(conf
->hash_locks
+ i
);
118 spin_unlock_irq(conf
->hash_locks
);
121 /* Find first data disk in a raid6 stripe */
122 static inline int raid6_d0(struct stripe_head
*sh
)
125 /* ddf always start from first device */
127 /* md starts just after Q block */
128 if (sh
->qd_idx
== sh
->disks
- 1)
131 return sh
->qd_idx
+ 1;
133 static inline int raid6_next_disk(int disk
, int raid_disks
)
136 return (disk
< raid_disks
) ? disk
: 0;
139 /* When walking through the disks in a raid5, starting at raid6_d0,
140 * We need to map each disk to a 'slot', where the data disks are slot
141 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
142 * is raid_disks-1. This help does that mapping.
144 static int raid6_idx_to_slot(int idx
, struct stripe_head
*sh
,
145 int *count
, int syndrome_disks
)
151 if (idx
== sh
->pd_idx
)
152 return syndrome_disks
;
153 if (idx
== sh
->qd_idx
)
154 return syndrome_disks
+ 1;
160 static void print_raid5_conf (struct r5conf
*conf
);
162 static int stripe_operations_active(struct stripe_head
*sh
)
164 return sh
->check_state
|| sh
->reconstruct_state
||
165 test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
) ||
166 test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
169 static bool stripe_is_lowprio(struct stripe_head
*sh
)
171 return (test_bit(STRIPE_R5C_FULL_STRIPE
, &sh
->state
) ||
172 test_bit(STRIPE_R5C_PARTIAL_STRIPE
, &sh
->state
)) &&
173 !test_bit(STRIPE_R5C_CACHING
, &sh
->state
);
176 static void raid5_wakeup_stripe_thread(struct stripe_head
*sh
)
178 struct r5conf
*conf
= sh
->raid_conf
;
179 struct r5worker_group
*group
;
181 int i
, cpu
= sh
->cpu
;
183 if (!cpu_online(cpu
)) {
184 cpu
= cpumask_any(cpu_online_mask
);
188 if (list_empty(&sh
->lru
)) {
189 struct r5worker_group
*group
;
190 group
= conf
->worker_groups
+ cpu_to_group(cpu
);
191 if (stripe_is_lowprio(sh
))
192 list_add_tail(&sh
->lru
, &group
->loprio_list
);
194 list_add_tail(&sh
->lru
, &group
->handle_list
);
195 group
->stripes_cnt
++;
199 if (conf
->worker_cnt_per_group
== 0) {
200 md_wakeup_thread(conf
->mddev
->thread
);
204 group
= conf
->worker_groups
+ cpu_to_group(sh
->cpu
);
206 group
->workers
[0].working
= true;
207 /* at least one worker should run to avoid race */
208 queue_work_on(sh
->cpu
, raid5_wq
, &group
->workers
[0].work
);
210 thread_cnt
= group
->stripes_cnt
/ MAX_STRIPE_BATCH
- 1;
211 /* wakeup more workers */
212 for (i
= 1; i
< conf
->worker_cnt_per_group
&& thread_cnt
> 0; i
++) {
213 if (group
->workers
[i
].working
== false) {
214 group
->workers
[i
].working
= true;
215 queue_work_on(sh
->cpu
, raid5_wq
,
216 &group
->workers
[i
].work
);
222 static void do_release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
223 struct list_head
*temp_inactive_list
)
226 int injournal
= 0; /* number of date pages with R5_InJournal */
228 BUG_ON(!list_empty(&sh
->lru
));
229 BUG_ON(atomic_read(&conf
->active_stripes
)==0);
231 if (r5c_is_writeback(conf
->log
))
232 for (i
= sh
->disks
; i
--; )
233 if (test_bit(R5_InJournal
, &sh
->dev
[i
].flags
))
236 * In the following cases, the stripe cannot be released to cached
237 * lists. Therefore, we make the stripe write out and set
239 * 1. when quiesce in r5c write back;
240 * 2. when resync is requested fot the stripe.
242 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
) ||
243 (conf
->quiesce
&& r5c_is_writeback(conf
->log
) &&
244 !test_bit(STRIPE_HANDLE
, &sh
->state
) && injournal
!= 0)) {
245 if (test_bit(STRIPE_R5C_CACHING
, &sh
->state
))
246 r5c_make_stripe_write_out(sh
);
247 set_bit(STRIPE_HANDLE
, &sh
->state
);
250 if (test_bit(STRIPE_HANDLE
, &sh
->state
)) {
251 if (test_bit(STRIPE_DELAYED
, &sh
->state
) &&
252 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
253 list_add_tail(&sh
->lru
, &conf
->delayed_list
);
254 else if (test_bit(STRIPE_BIT_DELAY
, &sh
->state
) &&
255 sh
->bm_seq
- conf
->seq_write
> 0)
256 list_add_tail(&sh
->lru
, &conf
->bitmap_list
);
258 clear_bit(STRIPE_DELAYED
, &sh
->state
);
259 clear_bit(STRIPE_BIT_DELAY
, &sh
->state
);
260 if (conf
->worker_cnt_per_group
== 0) {
261 if (stripe_is_lowprio(sh
))
262 list_add_tail(&sh
->lru
,
265 list_add_tail(&sh
->lru
,
268 raid5_wakeup_stripe_thread(sh
);
272 md_wakeup_thread(conf
->mddev
->thread
);
274 BUG_ON(stripe_operations_active(sh
));
275 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
276 if (atomic_dec_return(&conf
->preread_active_stripes
)
278 md_wakeup_thread(conf
->mddev
->thread
);
279 atomic_dec(&conf
->active_stripes
);
280 if (!test_bit(STRIPE_EXPANDING
, &sh
->state
)) {
281 if (!r5c_is_writeback(conf
->log
))
282 list_add_tail(&sh
->lru
, temp_inactive_list
);
284 WARN_ON(test_bit(R5_InJournal
, &sh
->dev
[sh
->pd_idx
].flags
));
286 list_add_tail(&sh
->lru
, temp_inactive_list
);
287 else if (injournal
== conf
->raid_disks
- conf
->max_degraded
) {
289 if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE
, &sh
->state
))
290 atomic_inc(&conf
->r5c_cached_full_stripes
);
291 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE
, &sh
->state
))
292 atomic_dec(&conf
->r5c_cached_partial_stripes
);
293 list_add_tail(&sh
->lru
, &conf
->r5c_full_stripe_list
);
294 r5c_check_cached_full_stripe(conf
);
297 * STRIPE_R5C_PARTIAL_STRIPE is set in
298 * r5c_try_caching_write(). No need to
301 list_add_tail(&sh
->lru
, &conf
->r5c_partial_stripe_list
);
307 static void __release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
308 struct list_head
*temp_inactive_list
)
310 if (atomic_dec_and_test(&sh
->count
))
311 do_release_stripe(conf
, sh
, temp_inactive_list
);
315 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
317 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
318 * given time. Adding stripes only takes device lock, while deleting stripes
319 * only takes hash lock.
321 static void release_inactive_stripe_list(struct r5conf
*conf
,
322 struct list_head
*temp_inactive_list
,
326 bool do_wakeup
= false;
329 if (hash
== NR_STRIPE_HASH_LOCKS
) {
330 size
= NR_STRIPE_HASH_LOCKS
;
331 hash
= NR_STRIPE_HASH_LOCKS
- 1;
335 struct list_head
*list
= &temp_inactive_list
[size
- 1];
338 * We don't hold any lock here yet, raid5_get_active_stripe() might
339 * remove stripes from the list
341 if (!list_empty_careful(list
)) {
342 spin_lock_irqsave(conf
->hash_locks
+ hash
, flags
);
343 if (list_empty(conf
->inactive_list
+ hash
) &&
345 atomic_dec(&conf
->empty_inactive_list_nr
);
346 list_splice_tail_init(list
, conf
->inactive_list
+ hash
);
348 spin_unlock_irqrestore(conf
->hash_locks
+ hash
, flags
);
355 wake_up(&conf
->wait_for_stripe
);
356 if (atomic_read(&conf
->active_stripes
) == 0)
357 wake_up(&conf
->wait_for_quiescent
);
358 if (conf
->retry_read_aligned
)
359 md_wakeup_thread(conf
->mddev
->thread
);
363 /* should hold conf->device_lock already */
364 static int release_stripe_list(struct r5conf
*conf
,
365 struct list_head
*temp_inactive_list
)
367 struct stripe_head
*sh
, *t
;
369 struct llist_node
*head
;
371 head
= llist_del_all(&conf
->released_stripes
);
372 head
= llist_reverse_order(head
);
373 llist_for_each_entry_safe(sh
, t
, head
, release_list
) {
376 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
378 clear_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
);
380 * Don't worry the bit is set here, because if the bit is set
381 * again, the count is always > 1. This is true for
382 * STRIPE_ON_UNPLUG_LIST bit too.
384 hash
= sh
->hash_lock_index
;
385 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
392 void raid5_release_stripe(struct stripe_head
*sh
)
394 struct r5conf
*conf
= sh
->raid_conf
;
396 struct list_head list
;
400 /* Avoid release_list until the last reference.
402 if (atomic_add_unless(&sh
->count
, -1, 1))
405 if (unlikely(!conf
->mddev
->thread
) ||
406 test_and_set_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
))
408 wakeup
= llist_add(&sh
->release_list
, &conf
->released_stripes
);
410 md_wakeup_thread(conf
->mddev
->thread
);
413 local_irq_save(flags
);
414 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
415 if (atomic_dec_and_lock(&sh
->count
, &conf
->device_lock
)) {
416 INIT_LIST_HEAD(&list
);
417 hash
= sh
->hash_lock_index
;
418 do_release_stripe(conf
, sh
, &list
);
419 spin_unlock(&conf
->device_lock
);
420 release_inactive_stripe_list(conf
, &list
, hash
);
422 local_irq_restore(flags
);
425 static inline void remove_hash(struct stripe_head
*sh
)
427 pr_debug("remove_hash(), stripe %llu\n",
428 (unsigned long long)sh
->sector
);
430 hlist_del_init(&sh
->hash
);
433 static inline void insert_hash(struct r5conf
*conf
, struct stripe_head
*sh
)
435 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
437 pr_debug("insert_hash(), stripe %llu\n",
438 (unsigned long long)sh
->sector
);
440 hlist_add_head(&sh
->hash
, hp
);
443 /* find an idle stripe, make sure it is unhashed, and return it. */
444 static struct stripe_head
*get_free_stripe(struct r5conf
*conf
, int hash
)
446 struct stripe_head
*sh
= NULL
;
447 struct list_head
*first
;
449 if (list_empty(conf
->inactive_list
+ hash
))
451 first
= (conf
->inactive_list
+ hash
)->next
;
452 sh
= list_entry(first
, struct stripe_head
, lru
);
453 list_del_init(first
);
455 atomic_inc(&conf
->active_stripes
);
456 BUG_ON(hash
!= sh
->hash_lock_index
);
457 if (list_empty(conf
->inactive_list
+ hash
))
458 atomic_inc(&conf
->empty_inactive_list_nr
);
463 static void shrink_buffers(struct stripe_head
*sh
)
467 int num
= sh
->raid_conf
->pool_size
;
469 for (i
= 0; i
< num
; i
++) {
470 WARN_ON(sh
->dev
[i
].page
!= sh
->dev
[i
].orig_page
);
474 sh
->dev
[i
].page
= NULL
;
479 static int grow_buffers(struct stripe_head
*sh
, gfp_t gfp
)
482 int num
= sh
->raid_conf
->pool_size
;
484 for (i
= 0; i
< num
; i
++) {
487 if (!(page
= alloc_page(gfp
))) {
490 sh
->dev
[i
].page
= page
;
491 sh
->dev
[i
].orig_page
= page
;
497 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
);
498 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
499 struct stripe_head
*sh
);
501 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int previous
)
503 struct r5conf
*conf
= sh
->raid_conf
;
506 BUG_ON(atomic_read(&sh
->count
) != 0);
507 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
508 BUG_ON(stripe_operations_active(sh
));
509 BUG_ON(sh
->batch_head
);
511 pr_debug("init_stripe called, stripe %llu\n",
512 (unsigned long long)sector
);
514 seq
= read_seqcount_begin(&conf
->gen_lock
);
515 sh
->generation
= conf
->generation
- previous
;
516 sh
->disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
518 stripe_set_idx(sector
, conf
, previous
, sh
);
521 for (i
= sh
->disks
; i
--; ) {
522 struct r5dev
*dev
= &sh
->dev
[i
];
524 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
525 test_bit(R5_LOCKED
, &dev
->flags
)) {
526 pr_err("sector=%llx i=%d %p %p %p %p %d\n",
527 (unsigned long long)sh
->sector
, i
, dev
->toread
,
528 dev
->read
, dev
->towrite
, dev
->written
,
529 test_bit(R5_LOCKED
, &dev
->flags
));
533 raid5_build_block(sh
, i
, previous
);
535 if (read_seqcount_retry(&conf
->gen_lock
, seq
))
537 sh
->overwrite_disks
= 0;
538 insert_hash(conf
, sh
);
539 sh
->cpu
= smp_processor_id();
540 set_bit(STRIPE_BATCH_READY
, &sh
->state
);
543 static struct stripe_head
*__find_stripe(struct r5conf
*conf
, sector_t sector
,
546 struct stripe_head
*sh
;
548 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
549 hlist_for_each_entry(sh
, stripe_hash(conf
, sector
), hash
)
550 if (sh
->sector
== sector
&& sh
->generation
== generation
)
552 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
557 * Need to check if array has failed when deciding whether to:
559 * - remove non-faulty devices
562 * This determination is simple when no reshape is happening.
563 * However if there is a reshape, we need to carefully check
564 * both the before and after sections.
565 * This is because some failed devices may only affect one
566 * of the two sections, and some non-in_sync devices may
567 * be insync in the section most affected by failed devices.
569 int raid5_calc_degraded(struct r5conf
*conf
)
571 int degraded
, degraded2
;
576 for (i
= 0; i
< conf
->previous_raid_disks
; i
++) {
577 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
578 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
579 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
580 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
582 else if (test_bit(In_sync
, &rdev
->flags
))
585 /* not in-sync or faulty.
586 * If the reshape increases the number of devices,
587 * this is being recovered by the reshape, so
588 * this 'previous' section is not in_sync.
589 * If the number of devices is being reduced however,
590 * the device can only be part of the array if
591 * we are reverting a reshape, so this section will
594 if (conf
->raid_disks
>= conf
->previous_raid_disks
)
598 if (conf
->raid_disks
== conf
->previous_raid_disks
)
602 for (i
= 0; i
< conf
->raid_disks
; i
++) {
603 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
604 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
605 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
606 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
608 else if (test_bit(In_sync
, &rdev
->flags
))
611 /* not in-sync or faulty.
612 * If reshape increases the number of devices, this
613 * section has already been recovered, else it
614 * almost certainly hasn't.
616 if (conf
->raid_disks
<= conf
->previous_raid_disks
)
620 if (degraded2
> degraded
)
625 static int has_failed(struct r5conf
*conf
)
629 if (conf
->mddev
->reshape_position
== MaxSector
)
630 return conf
->mddev
->degraded
> conf
->max_degraded
;
632 degraded
= raid5_calc_degraded(conf
);
633 if (degraded
> conf
->max_degraded
)
639 raid5_get_active_stripe(struct r5conf
*conf
, sector_t sector
,
640 int previous
, int noblock
, int noquiesce
)
642 struct stripe_head
*sh
;
643 int hash
= stripe_hash_locks_hash(sector
);
644 int inc_empty_inactive_list_flag
;
646 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
648 spin_lock_irq(conf
->hash_locks
+ hash
);
651 wait_event_lock_irq(conf
->wait_for_quiescent
,
652 conf
->quiesce
== 0 || noquiesce
,
653 *(conf
->hash_locks
+ hash
));
654 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
656 if (!test_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
)) {
657 sh
= get_free_stripe(conf
, hash
);
658 if (!sh
&& !test_bit(R5_DID_ALLOC
,
660 set_bit(R5_ALLOC_MORE
,
663 if (noblock
&& sh
== NULL
)
666 r5c_check_stripe_cache_usage(conf
);
668 set_bit(R5_INACTIVE_BLOCKED
,
670 r5l_wake_reclaim(conf
->log
, 0);
672 conf
->wait_for_stripe
,
673 !list_empty(conf
->inactive_list
+ hash
) &&
674 (atomic_read(&conf
->active_stripes
)
675 < (conf
->max_nr_stripes
* 3 / 4)
676 || !test_bit(R5_INACTIVE_BLOCKED
,
677 &conf
->cache_state
)),
678 *(conf
->hash_locks
+ hash
));
679 clear_bit(R5_INACTIVE_BLOCKED
,
682 init_stripe(sh
, sector
, previous
);
683 atomic_inc(&sh
->count
);
685 } else if (!atomic_inc_not_zero(&sh
->count
)) {
686 spin_lock(&conf
->device_lock
);
687 if (!atomic_read(&sh
->count
)) {
688 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
689 atomic_inc(&conf
->active_stripes
);
690 BUG_ON(list_empty(&sh
->lru
) &&
691 !test_bit(STRIPE_EXPANDING
, &sh
->state
));
692 inc_empty_inactive_list_flag
= 0;
693 if (!list_empty(conf
->inactive_list
+ hash
))
694 inc_empty_inactive_list_flag
= 1;
695 list_del_init(&sh
->lru
);
696 if (list_empty(conf
->inactive_list
+ hash
) && inc_empty_inactive_list_flag
)
697 atomic_inc(&conf
->empty_inactive_list_nr
);
699 sh
->group
->stripes_cnt
--;
703 atomic_inc(&sh
->count
);
704 spin_unlock(&conf
->device_lock
);
706 } while (sh
== NULL
);
708 spin_unlock_irq(conf
->hash_locks
+ hash
);
712 static bool is_full_stripe_write(struct stripe_head
*sh
)
714 BUG_ON(sh
->overwrite_disks
> (sh
->disks
- sh
->raid_conf
->max_degraded
));
715 return sh
->overwrite_disks
== (sh
->disks
- sh
->raid_conf
->max_degraded
);
718 static void lock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
721 spin_lock_irq(&sh2
->stripe_lock
);
722 spin_lock_nested(&sh1
->stripe_lock
, 1);
724 spin_lock_irq(&sh1
->stripe_lock
);
725 spin_lock_nested(&sh2
->stripe_lock
, 1);
729 static void unlock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
731 spin_unlock(&sh1
->stripe_lock
);
732 spin_unlock_irq(&sh2
->stripe_lock
);
735 /* Only freshly new full stripe normal write stripe can be added to a batch list */
736 static bool stripe_can_batch(struct stripe_head
*sh
)
738 struct r5conf
*conf
= sh
->raid_conf
;
740 if (conf
->log
|| raid5_has_ppl(conf
))
742 return test_bit(STRIPE_BATCH_READY
, &sh
->state
) &&
743 !test_bit(STRIPE_BITMAP_PENDING
, &sh
->state
) &&
744 is_full_stripe_write(sh
);
747 /* we only do back search */
748 static void stripe_add_to_batch_list(struct r5conf
*conf
, struct stripe_head
*sh
)
750 struct stripe_head
*head
;
751 sector_t head_sector
, tmp_sec
;
754 int inc_empty_inactive_list_flag
;
756 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
757 tmp_sec
= sh
->sector
;
758 if (!sector_div(tmp_sec
, conf
->chunk_sectors
))
760 head_sector
= sh
->sector
- STRIPE_SECTORS
;
762 hash
= stripe_hash_locks_hash(head_sector
);
763 spin_lock_irq(conf
->hash_locks
+ hash
);
764 head
= __find_stripe(conf
, head_sector
, conf
->generation
);
765 if (head
&& !atomic_inc_not_zero(&head
->count
)) {
766 spin_lock(&conf
->device_lock
);
767 if (!atomic_read(&head
->count
)) {
768 if (!test_bit(STRIPE_HANDLE
, &head
->state
))
769 atomic_inc(&conf
->active_stripes
);
770 BUG_ON(list_empty(&head
->lru
) &&
771 !test_bit(STRIPE_EXPANDING
, &head
->state
));
772 inc_empty_inactive_list_flag
= 0;
773 if (!list_empty(conf
->inactive_list
+ hash
))
774 inc_empty_inactive_list_flag
= 1;
775 list_del_init(&head
->lru
);
776 if (list_empty(conf
->inactive_list
+ hash
) && inc_empty_inactive_list_flag
)
777 atomic_inc(&conf
->empty_inactive_list_nr
);
779 head
->group
->stripes_cnt
--;
783 atomic_inc(&head
->count
);
784 spin_unlock(&conf
->device_lock
);
786 spin_unlock_irq(conf
->hash_locks
+ hash
);
790 if (!stripe_can_batch(head
))
793 lock_two_stripes(head
, sh
);
794 /* clear_batch_ready clear the flag */
795 if (!stripe_can_batch(head
) || !stripe_can_batch(sh
))
802 while (dd_idx
== sh
->pd_idx
|| dd_idx
== sh
->qd_idx
)
804 if (head
->dev
[dd_idx
].towrite
->bi_opf
!= sh
->dev
[dd_idx
].towrite
->bi_opf
||
805 bio_op(head
->dev
[dd_idx
].towrite
) != bio_op(sh
->dev
[dd_idx
].towrite
))
808 if (head
->batch_head
) {
809 spin_lock(&head
->batch_head
->batch_lock
);
810 /* This batch list is already running */
811 if (!stripe_can_batch(head
)) {
812 spin_unlock(&head
->batch_head
->batch_lock
);
817 * at this point, head's BATCH_READY could be cleared, but we
818 * can still add the stripe to batch list
820 list_add(&sh
->batch_list
, &head
->batch_list
);
821 spin_unlock(&head
->batch_head
->batch_lock
);
823 sh
->batch_head
= head
->batch_head
;
825 head
->batch_head
= head
;
826 sh
->batch_head
= head
->batch_head
;
827 spin_lock(&head
->batch_lock
);
828 list_add_tail(&sh
->batch_list
, &head
->batch_list
);
829 spin_unlock(&head
->batch_lock
);
832 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
833 if (atomic_dec_return(&conf
->preread_active_stripes
)
835 md_wakeup_thread(conf
->mddev
->thread
);
837 if (test_and_clear_bit(STRIPE_BIT_DELAY
, &sh
->state
)) {
838 int seq
= sh
->bm_seq
;
839 if (test_bit(STRIPE_BIT_DELAY
, &sh
->batch_head
->state
) &&
840 sh
->batch_head
->bm_seq
> seq
)
841 seq
= sh
->batch_head
->bm_seq
;
842 set_bit(STRIPE_BIT_DELAY
, &sh
->batch_head
->state
);
843 sh
->batch_head
->bm_seq
= seq
;
846 atomic_inc(&sh
->count
);
848 unlock_two_stripes(head
, sh
);
850 raid5_release_stripe(head
);
853 /* Determine if 'data_offset' or 'new_data_offset' should be used
854 * in this stripe_head.
856 static int use_new_offset(struct r5conf
*conf
, struct stripe_head
*sh
)
858 sector_t progress
= conf
->reshape_progress
;
859 /* Need a memory barrier to make sure we see the value
860 * of conf->generation, or ->data_offset that was set before
861 * reshape_progress was updated.
864 if (progress
== MaxSector
)
866 if (sh
->generation
== conf
->generation
- 1)
868 /* We are in a reshape, and this is a new-generation stripe,
869 * so use new_data_offset.
874 static void dispatch_bio_list(struct bio_list
*tmp
)
878 while ((bio
= bio_list_pop(tmp
)))
879 generic_make_request(bio
);
882 static int cmp_stripe(void *priv
, struct list_head
*a
, struct list_head
*b
)
884 const struct r5pending_data
*da
= list_entry(a
,
885 struct r5pending_data
, sibling
);
886 const struct r5pending_data
*db
= list_entry(b
,
887 struct r5pending_data
, sibling
);
888 if (da
->sector
> db
->sector
)
890 if (da
->sector
< db
->sector
)
895 static void dispatch_defer_bios(struct r5conf
*conf
, int target
,
896 struct bio_list
*list
)
898 struct r5pending_data
*data
;
899 struct list_head
*first
, *next
= NULL
;
902 if (conf
->pending_data_cnt
== 0)
905 list_sort(NULL
, &conf
->pending_list
, cmp_stripe
);
907 first
= conf
->pending_list
.next
;
909 /* temporarily move the head */
910 if (conf
->next_pending_data
)
911 list_move_tail(&conf
->pending_list
,
912 &conf
->next_pending_data
->sibling
);
914 while (!list_empty(&conf
->pending_list
)) {
915 data
= list_first_entry(&conf
->pending_list
,
916 struct r5pending_data
, sibling
);
917 if (&data
->sibling
== first
)
918 first
= data
->sibling
.next
;
919 next
= data
->sibling
.next
;
921 bio_list_merge(list
, &data
->bios
);
922 list_move(&data
->sibling
, &conf
->free_list
);
927 conf
->pending_data_cnt
-= cnt
;
928 BUG_ON(conf
->pending_data_cnt
< 0 || cnt
< target
);
930 if (next
!= &conf
->pending_list
)
931 conf
->next_pending_data
= list_entry(next
,
932 struct r5pending_data
, sibling
);
934 conf
->next_pending_data
= NULL
;
935 /* list isn't empty */
936 if (first
!= &conf
->pending_list
)
937 list_move_tail(&conf
->pending_list
, first
);
940 static void flush_deferred_bios(struct r5conf
*conf
)
942 struct bio_list tmp
= BIO_EMPTY_LIST
;
944 if (conf
->pending_data_cnt
== 0)
947 spin_lock(&conf
->pending_bios_lock
);
948 dispatch_defer_bios(conf
, conf
->pending_data_cnt
, &tmp
);
949 BUG_ON(conf
->pending_data_cnt
!= 0);
950 spin_unlock(&conf
->pending_bios_lock
);
952 dispatch_bio_list(&tmp
);
955 static void defer_issue_bios(struct r5conf
*conf
, sector_t sector
,
956 struct bio_list
*bios
)
958 struct bio_list tmp
= BIO_EMPTY_LIST
;
959 struct r5pending_data
*ent
;
961 spin_lock(&conf
->pending_bios_lock
);
962 ent
= list_first_entry(&conf
->free_list
, struct r5pending_data
,
964 list_move_tail(&ent
->sibling
, &conf
->pending_list
);
965 ent
->sector
= sector
;
966 bio_list_init(&ent
->bios
);
967 bio_list_merge(&ent
->bios
, bios
);
968 conf
->pending_data_cnt
++;
969 if (conf
->pending_data_cnt
>= PENDING_IO_MAX
)
970 dispatch_defer_bios(conf
, PENDING_IO_ONE_FLUSH
, &tmp
);
972 spin_unlock(&conf
->pending_bios_lock
);
974 dispatch_bio_list(&tmp
);
978 raid5_end_read_request(struct bio
*bi
);
980 raid5_end_write_request(struct bio
*bi
);
982 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
984 struct r5conf
*conf
= sh
->raid_conf
;
985 int i
, disks
= sh
->disks
;
986 struct stripe_head
*head_sh
= sh
;
987 struct bio_list pending_bios
= BIO_EMPTY_LIST
;
992 if (log_stripe(sh
, s
) == 0)
995 should_defer
= conf
->batch_bio_dispatch
&& conf
->group_cnt
;
997 for (i
= disks
; i
--; ) {
998 int op
, op_flags
= 0;
999 int replace_only
= 0;
1000 struct bio
*bi
, *rbi
;
1001 struct md_rdev
*rdev
, *rrdev
= NULL
;
1004 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
1006 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
1008 if (test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1009 op
= REQ_OP_DISCARD
;
1010 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
1012 else if (test_and_clear_bit(R5_WantReplace
,
1013 &sh
->dev
[i
].flags
)) {
1018 if (test_and_clear_bit(R5_SyncIO
, &sh
->dev
[i
].flags
))
1019 op_flags
|= REQ_SYNC
;
1022 bi
= &sh
->dev
[i
].req
;
1023 rbi
= &sh
->dev
[i
].rreq
; /* For writing to replacement */
1026 rrdev
= rcu_dereference(conf
->disks
[i
].replacement
);
1027 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1028 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
1033 if (op_is_write(op
)) {
1037 /* We raced and saw duplicates */
1040 if (test_bit(R5_ReadRepl
, &head_sh
->dev
[i
].flags
) && rrdev
)
1045 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
1048 atomic_inc(&rdev
->nr_pending
);
1049 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
1052 atomic_inc(&rrdev
->nr_pending
);
1055 /* We have already checked bad blocks for reads. Now
1056 * need to check for writes. We never accept write errors
1057 * on the replacement, so we don't to check rrdev.
1059 while (op_is_write(op
) && rdev
&&
1060 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1063 int bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
1064 &first_bad
, &bad_sectors
);
1069 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1070 if (!conf
->mddev
->external
&&
1071 conf
->mddev
->sb_flags
) {
1072 /* It is very unlikely, but we might
1073 * still need to write out the
1074 * bad block log - better give it
1076 md_check_recovery(conf
->mddev
);
1079 * Because md_wait_for_blocked_rdev
1080 * will dec nr_pending, we must
1081 * increment it first.
1083 atomic_inc(&rdev
->nr_pending
);
1084 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
1086 /* Acknowledged bad block - skip the write */
1087 rdev_dec_pending(rdev
, conf
->mddev
);
1093 if (s
->syncing
|| s
->expanding
|| s
->expanded
1095 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
1097 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
1099 bi
->bi_bdev
= rdev
->bdev
;
1100 bio_set_op_attrs(bi
, op
, op_flags
);
1101 bi
->bi_end_io
= op_is_write(op
)
1102 ? raid5_end_write_request
1103 : raid5_end_read_request
;
1104 bi
->bi_private
= sh
;
1106 pr_debug("%s: for %llu schedule op %d on disc %d\n",
1107 __func__
, (unsigned long long)sh
->sector
,
1109 atomic_inc(&sh
->count
);
1111 atomic_inc(&head_sh
->count
);
1112 if (use_new_offset(conf
, sh
))
1113 bi
->bi_iter
.bi_sector
= (sh
->sector
1114 + rdev
->new_data_offset
);
1116 bi
->bi_iter
.bi_sector
= (sh
->sector
1117 + rdev
->data_offset
);
1118 if (test_bit(R5_ReadNoMerge
, &head_sh
->dev
[i
].flags
))
1119 bi
->bi_opf
|= REQ_NOMERGE
;
1121 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1122 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1124 if (!op_is_write(op
) &&
1125 test_bit(R5_InJournal
, &sh
->dev
[i
].flags
))
1127 * issuing read for a page in journal, this
1128 * must be preparing for prexor in rmw; read
1129 * the data into orig_page
1131 sh
->dev
[i
].vec
.bv_page
= sh
->dev
[i
].orig_page
;
1133 sh
->dev
[i
].vec
.bv_page
= sh
->dev
[i
].page
;
1135 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
1136 bi
->bi_io_vec
[0].bv_offset
= 0;
1137 bi
->bi_iter
.bi_size
= STRIPE_SIZE
;
1139 * If this is discard request, set bi_vcnt 0. We don't
1140 * want to confuse SCSI because SCSI will replace payload
1142 if (op
== REQ_OP_DISCARD
)
1145 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
1147 if (conf
->mddev
->gendisk
)
1148 trace_block_bio_remap(bdev_get_queue(bi
->bi_bdev
),
1149 bi
, disk_devt(conf
->mddev
->gendisk
),
1151 if (should_defer
&& op_is_write(op
))
1152 bio_list_add(&pending_bios
, bi
);
1154 generic_make_request(bi
);
1157 if (s
->syncing
|| s
->expanding
|| s
->expanded
1159 md_sync_acct(rrdev
->bdev
, STRIPE_SECTORS
);
1161 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
1163 rbi
->bi_bdev
= rrdev
->bdev
;
1164 bio_set_op_attrs(rbi
, op
, op_flags
);
1165 BUG_ON(!op_is_write(op
));
1166 rbi
->bi_end_io
= raid5_end_write_request
;
1167 rbi
->bi_private
= sh
;
1169 pr_debug("%s: for %llu schedule op %d on "
1170 "replacement disc %d\n",
1171 __func__
, (unsigned long long)sh
->sector
,
1173 atomic_inc(&sh
->count
);
1175 atomic_inc(&head_sh
->count
);
1176 if (use_new_offset(conf
, sh
))
1177 rbi
->bi_iter
.bi_sector
= (sh
->sector
1178 + rrdev
->new_data_offset
);
1180 rbi
->bi_iter
.bi_sector
= (sh
->sector
1181 + rrdev
->data_offset
);
1182 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1183 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1184 sh
->dev
[i
].rvec
.bv_page
= sh
->dev
[i
].page
;
1186 rbi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
1187 rbi
->bi_io_vec
[0].bv_offset
= 0;
1188 rbi
->bi_iter
.bi_size
= STRIPE_SIZE
;
1190 * If this is discard request, set bi_vcnt 0. We don't
1191 * want to confuse SCSI because SCSI will replace payload
1193 if (op
== REQ_OP_DISCARD
)
1195 if (conf
->mddev
->gendisk
)
1196 trace_block_bio_remap(bdev_get_queue(rbi
->bi_bdev
),
1197 rbi
, disk_devt(conf
->mddev
->gendisk
),
1199 if (should_defer
&& op_is_write(op
))
1200 bio_list_add(&pending_bios
, rbi
);
1202 generic_make_request(rbi
);
1204 if (!rdev
&& !rrdev
) {
1205 if (op_is_write(op
))
1206 set_bit(STRIPE_DEGRADED
, &sh
->state
);
1207 pr_debug("skip op %d on disc %d for sector %llu\n",
1208 bi
->bi_opf
, i
, (unsigned long long)sh
->sector
);
1209 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1210 set_bit(STRIPE_HANDLE
, &sh
->state
);
1213 if (!head_sh
->batch_head
)
1215 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1221 if (should_defer
&& !bio_list_empty(&pending_bios
))
1222 defer_issue_bios(conf
, head_sh
->sector
, &pending_bios
);
1225 static struct dma_async_tx_descriptor
*
1226 async_copy_data(int frombio
, struct bio
*bio
, struct page
**page
,
1227 sector_t sector
, struct dma_async_tx_descriptor
*tx
,
1228 struct stripe_head
*sh
, int no_skipcopy
)
1231 struct bvec_iter iter
;
1232 struct page
*bio_page
;
1234 struct async_submit_ctl submit
;
1235 enum async_tx_flags flags
= 0;
1237 if (bio
->bi_iter
.bi_sector
>= sector
)
1238 page_offset
= (signed)(bio
->bi_iter
.bi_sector
- sector
) * 512;
1240 page_offset
= (signed)(sector
- bio
->bi_iter
.bi_sector
) * -512;
1243 flags
|= ASYNC_TX_FENCE
;
1244 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
1246 bio_for_each_segment(bvl
, bio
, iter
) {
1247 int len
= bvl
.bv_len
;
1251 if (page_offset
< 0) {
1252 b_offset
= -page_offset
;
1253 page_offset
+= b_offset
;
1257 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
1258 clen
= STRIPE_SIZE
- page_offset
;
1263 b_offset
+= bvl
.bv_offset
;
1264 bio_page
= bvl
.bv_page
;
1266 if (sh
->raid_conf
->skip_copy
&&
1267 b_offset
== 0 && page_offset
== 0 &&
1268 clen
== STRIPE_SIZE
&&
1272 tx
= async_memcpy(*page
, bio_page
, page_offset
,
1273 b_offset
, clen
, &submit
);
1275 tx
= async_memcpy(bio_page
, *page
, b_offset
,
1276 page_offset
, clen
, &submit
);
1278 /* chain the operations */
1279 submit
.depend_tx
= tx
;
1281 if (clen
< len
) /* hit end of page */
1289 static void ops_complete_biofill(void *stripe_head_ref
)
1291 struct stripe_head
*sh
= stripe_head_ref
;
1294 pr_debug("%s: stripe %llu\n", __func__
,
1295 (unsigned long long)sh
->sector
);
1297 /* clear completed biofills */
1298 for (i
= sh
->disks
; i
--; ) {
1299 struct r5dev
*dev
= &sh
->dev
[i
];
1301 /* acknowledge completion of a biofill operation */
1302 /* and check if we need to reply to a read request,
1303 * new R5_Wantfill requests are held off until
1304 * !STRIPE_BIOFILL_RUN
1306 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
1307 struct bio
*rbi
, *rbi2
;
1312 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1313 dev
->sector
+ STRIPE_SECTORS
) {
1314 rbi2
= r5_next_bio(rbi
, dev
->sector
);
1320 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
1322 set_bit(STRIPE_HANDLE
, &sh
->state
);
1323 raid5_release_stripe(sh
);
1326 static void ops_run_biofill(struct stripe_head
*sh
)
1328 struct dma_async_tx_descriptor
*tx
= NULL
;
1329 struct async_submit_ctl submit
;
1332 BUG_ON(sh
->batch_head
);
1333 pr_debug("%s: stripe %llu\n", __func__
,
1334 (unsigned long long)sh
->sector
);
1336 for (i
= sh
->disks
; i
--; ) {
1337 struct r5dev
*dev
= &sh
->dev
[i
];
1338 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
1340 spin_lock_irq(&sh
->stripe_lock
);
1341 dev
->read
= rbi
= dev
->toread
;
1343 spin_unlock_irq(&sh
->stripe_lock
);
1344 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1345 dev
->sector
+ STRIPE_SECTORS
) {
1346 tx
= async_copy_data(0, rbi
, &dev
->page
,
1347 dev
->sector
, tx
, sh
, 0);
1348 rbi
= r5_next_bio(rbi
, dev
->sector
);
1353 atomic_inc(&sh
->count
);
1354 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
1355 async_trigger_callback(&submit
);
1358 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
1365 tgt
= &sh
->dev
[target
];
1366 set_bit(R5_UPTODATE
, &tgt
->flags
);
1367 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1368 clear_bit(R5_Wantcompute
, &tgt
->flags
);
1371 static void ops_complete_compute(void *stripe_head_ref
)
1373 struct stripe_head
*sh
= stripe_head_ref
;
1375 pr_debug("%s: stripe %llu\n", __func__
,
1376 (unsigned long long)sh
->sector
);
1378 /* mark the computed target(s) as uptodate */
1379 mark_target_uptodate(sh
, sh
->ops
.target
);
1380 mark_target_uptodate(sh
, sh
->ops
.target2
);
1382 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
1383 if (sh
->check_state
== check_state_compute_run
)
1384 sh
->check_state
= check_state_compute_result
;
1385 set_bit(STRIPE_HANDLE
, &sh
->state
);
1386 raid5_release_stripe(sh
);
1389 /* return a pointer to the address conversion region of the scribble buffer */
1390 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
1391 struct raid5_percpu
*percpu
, int i
)
1395 addr
= flex_array_get(percpu
->scribble
, i
);
1396 return addr
+ sizeof(struct page
*) * (sh
->disks
+ 2);
1399 /* return a pointer to the address conversion region of the scribble buffer */
1400 static struct page
**to_addr_page(struct raid5_percpu
*percpu
, int i
)
1404 addr
= flex_array_get(percpu
->scribble
, i
);
1408 static struct dma_async_tx_descriptor
*
1409 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1411 int disks
= sh
->disks
;
1412 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1413 int target
= sh
->ops
.target
;
1414 struct r5dev
*tgt
= &sh
->dev
[target
];
1415 struct page
*xor_dest
= tgt
->page
;
1417 struct dma_async_tx_descriptor
*tx
;
1418 struct async_submit_ctl submit
;
1421 BUG_ON(sh
->batch_head
);
1423 pr_debug("%s: stripe %llu block: %d\n",
1424 __func__
, (unsigned long long)sh
->sector
, target
);
1425 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1427 for (i
= disks
; i
--; )
1429 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1431 atomic_inc(&sh
->count
);
1433 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
1434 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
, 0));
1435 if (unlikely(count
== 1))
1436 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1438 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1443 /* set_syndrome_sources - populate source buffers for gen_syndrome
1444 * @srcs - (struct page *) array of size sh->disks
1445 * @sh - stripe_head to parse
1447 * Populates srcs in proper layout order for the stripe and returns the
1448 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1449 * destination buffer is recorded in srcs[count] and the Q destination
1450 * is recorded in srcs[count+1]].
1452 static int set_syndrome_sources(struct page
**srcs
,
1453 struct stripe_head
*sh
,
1456 int disks
= sh
->disks
;
1457 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
1458 int d0_idx
= raid6_d0(sh
);
1462 for (i
= 0; i
< disks
; i
++)
1468 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1469 struct r5dev
*dev
= &sh
->dev
[i
];
1471 if (i
== sh
->qd_idx
|| i
== sh
->pd_idx
||
1472 (srctype
== SYNDROME_SRC_ALL
) ||
1473 (srctype
== SYNDROME_SRC_WANT_DRAIN
&&
1474 (test_bit(R5_Wantdrain
, &dev
->flags
) ||
1475 test_bit(R5_InJournal
, &dev
->flags
))) ||
1476 (srctype
== SYNDROME_SRC_WRITTEN
&&
1478 test_bit(R5_InJournal
, &dev
->flags
)))) {
1479 if (test_bit(R5_InJournal
, &dev
->flags
))
1480 srcs
[slot
] = sh
->dev
[i
].orig_page
;
1482 srcs
[slot
] = sh
->dev
[i
].page
;
1484 i
= raid6_next_disk(i
, disks
);
1485 } while (i
!= d0_idx
);
1487 return syndrome_disks
;
1490 static struct dma_async_tx_descriptor
*
1491 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1493 int disks
= sh
->disks
;
1494 struct page
**blocks
= to_addr_page(percpu
, 0);
1496 int qd_idx
= sh
->qd_idx
;
1497 struct dma_async_tx_descriptor
*tx
;
1498 struct async_submit_ctl submit
;
1504 BUG_ON(sh
->batch_head
);
1505 if (sh
->ops
.target
< 0)
1506 target
= sh
->ops
.target2
;
1507 else if (sh
->ops
.target2
< 0)
1508 target
= sh
->ops
.target
;
1510 /* we should only have one valid target */
1513 pr_debug("%s: stripe %llu block: %d\n",
1514 __func__
, (unsigned long long)sh
->sector
, target
);
1516 tgt
= &sh
->dev
[target
];
1517 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1520 atomic_inc(&sh
->count
);
1522 if (target
== qd_idx
) {
1523 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1524 blocks
[count
] = NULL
; /* regenerating p is not necessary */
1525 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
1526 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1527 ops_complete_compute
, sh
,
1528 to_addr_conv(sh
, percpu
, 0));
1529 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1531 /* Compute any data- or p-drive using XOR */
1533 for (i
= disks
; i
-- ; ) {
1534 if (i
== target
|| i
== qd_idx
)
1536 blocks
[count
++] = sh
->dev
[i
].page
;
1539 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1540 NULL
, ops_complete_compute
, sh
,
1541 to_addr_conv(sh
, percpu
, 0));
1542 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
1548 static struct dma_async_tx_descriptor
*
1549 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1551 int i
, count
, disks
= sh
->disks
;
1552 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1553 int d0_idx
= raid6_d0(sh
);
1554 int faila
= -1, failb
= -1;
1555 int target
= sh
->ops
.target
;
1556 int target2
= sh
->ops
.target2
;
1557 struct r5dev
*tgt
= &sh
->dev
[target
];
1558 struct r5dev
*tgt2
= &sh
->dev
[target2
];
1559 struct dma_async_tx_descriptor
*tx
;
1560 struct page
**blocks
= to_addr_page(percpu
, 0);
1561 struct async_submit_ctl submit
;
1563 BUG_ON(sh
->batch_head
);
1564 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1565 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
1566 BUG_ON(target
< 0 || target2
< 0);
1567 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1568 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
1570 /* we need to open-code set_syndrome_sources to handle the
1571 * slot number conversion for 'faila' and 'failb'
1573 for (i
= 0; i
< disks
; i
++)
1578 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1580 blocks
[slot
] = sh
->dev
[i
].page
;
1586 i
= raid6_next_disk(i
, disks
);
1587 } while (i
!= d0_idx
);
1589 BUG_ON(faila
== failb
);
1592 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1593 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1595 atomic_inc(&sh
->count
);
1597 if (failb
== syndrome_disks
+1) {
1598 /* Q disk is one of the missing disks */
1599 if (faila
== syndrome_disks
) {
1600 /* Missing P+Q, just recompute */
1601 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1602 ops_complete_compute
, sh
,
1603 to_addr_conv(sh
, percpu
, 0));
1604 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
1605 STRIPE_SIZE
, &submit
);
1609 int qd_idx
= sh
->qd_idx
;
1611 /* Missing D+Q: recompute D from P, then recompute Q */
1612 if (target
== qd_idx
)
1613 data_target
= target2
;
1615 data_target
= target
;
1618 for (i
= disks
; i
-- ; ) {
1619 if (i
== data_target
|| i
== qd_idx
)
1621 blocks
[count
++] = sh
->dev
[i
].page
;
1623 dest
= sh
->dev
[data_target
].page
;
1624 init_async_submit(&submit
,
1625 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1627 to_addr_conv(sh
, percpu
, 0));
1628 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
1631 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_ALL
);
1632 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1633 ops_complete_compute
, sh
,
1634 to_addr_conv(sh
, percpu
, 0));
1635 return async_gen_syndrome(blocks
, 0, count
+2,
1636 STRIPE_SIZE
, &submit
);
1639 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1640 ops_complete_compute
, sh
,
1641 to_addr_conv(sh
, percpu
, 0));
1642 if (failb
== syndrome_disks
) {
1643 /* We're missing D+P. */
1644 return async_raid6_datap_recov(syndrome_disks
+2,
1648 /* We're missing D+D. */
1649 return async_raid6_2data_recov(syndrome_disks
+2,
1650 STRIPE_SIZE
, faila
, failb
,
1656 static void ops_complete_prexor(void *stripe_head_ref
)
1658 struct stripe_head
*sh
= stripe_head_ref
;
1660 pr_debug("%s: stripe %llu\n", __func__
,
1661 (unsigned long long)sh
->sector
);
1663 if (r5c_is_writeback(sh
->raid_conf
->log
))
1665 * raid5-cache write back uses orig_page during prexor.
1666 * After prexor, it is time to free orig_page
1668 r5c_release_extra_page(sh
);
1671 static struct dma_async_tx_descriptor
*
1672 ops_run_prexor5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1673 struct dma_async_tx_descriptor
*tx
)
1675 int disks
= sh
->disks
;
1676 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1677 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1678 struct async_submit_ctl submit
;
1680 /* existing parity data subtracted */
1681 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1683 BUG_ON(sh
->batch_head
);
1684 pr_debug("%s: stripe %llu\n", __func__
,
1685 (unsigned long long)sh
->sector
);
1687 for (i
= disks
; i
--; ) {
1688 struct r5dev
*dev
= &sh
->dev
[i
];
1689 /* Only process blocks that are known to be uptodate */
1690 if (test_bit(R5_InJournal
, &dev
->flags
))
1691 xor_srcs
[count
++] = dev
->orig_page
;
1692 else if (test_bit(R5_Wantdrain
, &dev
->flags
))
1693 xor_srcs
[count
++] = dev
->page
;
1696 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1697 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1698 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1703 static struct dma_async_tx_descriptor
*
1704 ops_run_prexor6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1705 struct dma_async_tx_descriptor
*tx
)
1707 struct page
**blocks
= to_addr_page(percpu
, 0);
1709 struct async_submit_ctl submit
;
1711 pr_debug("%s: stripe %llu\n", __func__
,
1712 (unsigned long long)sh
->sector
);
1714 count
= set_syndrome_sources(blocks
, sh
, SYNDROME_SRC_WANT_DRAIN
);
1716 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_PQ_XOR_DST
, tx
,
1717 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1718 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1723 static struct dma_async_tx_descriptor
*
1724 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1726 struct r5conf
*conf
= sh
->raid_conf
;
1727 int disks
= sh
->disks
;
1729 struct stripe_head
*head_sh
= sh
;
1731 pr_debug("%s: stripe %llu\n", __func__
,
1732 (unsigned long long)sh
->sector
);
1734 for (i
= disks
; i
--; ) {
1739 if (test_and_clear_bit(R5_Wantdrain
, &head_sh
->dev
[i
].flags
)) {
1745 * clear R5_InJournal, so when rewriting a page in
1746 * journal, it is not skipped by r5l_log_stripe()
1748 clear_bit(R5_InJournal
, &dev
->flags
);
1749 spin_lock_irq(&sh
->stripe_lock
);
1750 chosen
= dev
->towrite
;
1751 dev
->towrite
= NULL
;
1752 sh
->overwrite_disks
= 0;
1753 BUG_ON(dev
->written
);
1754 wbi
= dev
->written
= chosen
;
1755 spin_unlock_irq(&sh
->stripe_lock
);
1756 WARN_ON(dev
->page
!= dev
->orig_page
);
1758 while (wbi
&& wbi
->bi_iter
.bi_sector
<
1759 dev
->sector
+ STRIPE_SECTORS
) {
1760 if (wbi
->bi_opf
& REQ_FUA
)
1761 set_bit(R5_WantFUA
, &dev
->flags
);
1762 if (wbi
->bi_opf
& REQ_SYNC
)
1763 set_bit(R5_SyncIO
, &dev
->flags
);
1764 if (bio_op(wbi
) == REQ_OP_DISCARD
)
1765 set_bit(R5_Discard
, &dev
->flags
);
1767 tx
= async_copy_data(1, wbi
, &dev
->page
,
1768 dev
->sector
, tx
, sh
,
1769 r5c_is_writeback(conf
->log
));
1770 if (dev
->page
!= dev
->orig_page
&&
1771 !r5c_is_writeback(conf
->log
)) {
1772 set_bit(R5_SkipCopy
, &dev
->flags
);
1773 clear_bit(R5_UPTODATE
, &dev
->flags
);
1774 clear_bit(R5_OVERWRITE
, &dev
->flags
);
1777 wbi
= r5_next_bio(wbi
, dev
->sector
);
1780 if (head_sh
->batch_head
) {
1781 sh
= list_first_entry(&sh
->batch_list
,
1794 static void ops_complete_reconstruct(void *stripe_head_ref
)
1796 struct stripe_head
*sh
= stripe_head_ref
;
1797 int disks
= sh
->disks
;
1798 int pd_idx
= sh
->pd_idx
;
1799 int qd_idx
= sh
->qd_idx
;
1801 bool fua
= false, sync
= false, discard
= false;
1803 pr_debug("%s: stripe %llu\n", __func__
,
1804 (unsigned long long)sh
->sector
);
1806 for (i
= disks
; i
--; ) {
1807 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1808 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
1809 discard
|= test_bit(R5_Discard
, &sh
->dev
[i
].flags
);
1812 for (i
= disks
; i
--; ) {
1813 struct r5dev
*dev
= &sh
->dev
[i
];
1815 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1816 if (!discard
&& !test_bit(R5_SkipCopy
, &dev
->flags
))
1817 set_bit(R5_UPTODATE
, &dev
->flags
);
1819 set_bit(R5_WantFUA
, &dev
->flags
);
1821 set_bit(R5_SyncIO
, &dev
->flags
);
1825 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1826 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1827 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1828 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1830 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1831 sh
->reconstruct_state
= reconstruct_state_result
;
1834 set_bit(STRIPE_HANDLE
, &sh
->state
);
1835 raid5_release_stripe(sh
);
1839 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1840 struct dma_async_tx_descriptor
*tx
)
1842 int disks
= sh
->disks
;
1843 struct page
**xor_srcs
;
1844 struct async_submit_ctl submit
;
1845 int count
, pd_idx
= sh
->pd_idx
, i
;
1846 struct page
*xor_dest
;
1848 unsigned long flags
;
1850 struct stripe_head
*head_sh
= sh
;
1853 pr_debug("%s: stripe %llu\n", __func__
,
1854 (unsigned long long)sh
->sector
);
1856 for (i
= 0; i
< sh
->disks
; i
++) {
1859 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1862 if (i
>= sh
->disks
) {
1863 atomic_inc(&sh
->count
);
1864 set_bit(R5_Discard
, &sh
->dev
[pd_idx
].flags
);
1865 ops_complete_reconstruct(sh
);
1870 xor_srcs
= to_addr_page(percpu
, j
);
1871 /* check if prexor is active which means only process blocks
1872 * that are part of a read-modify-write (written)
1874 if (head_sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1876 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1877 for (i
= disks
; i
--; ) {
1878 struct r5dev
*dev
= &sh
->dev
[i
];
1879 if (head_sh
->dev
[i
].written
||
1880 test_bit(R5_InJournal
, &head_sh
->dev
[i
].flags
))
1881 xor_srcs
[count
++] = dev
->page
;
1884 xor_dest
= sh
->dev
[pd_idx
].page
;
1885 for (i
= disks
; i
--; ) {
1886 struct r5dev
*dev
= &sh
->dev
[i
];
1888 xor_srcs
[count
++] = dev
->page
;
1892 /* 1/ if we prexor'd then the dest is reused as a source
1893 * 2/ if we did not prexor then we are redoing the parity
1894 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1895 * for the synchronous xor case
1897 last_stripe
= !head_sh
->batch_head
||
1898 list_first_entry(&sh
->batch_list
,
1899 struct stripe_head
, batch_list
) == head_sh
;
1901 flags
= ASYNC_TX_ACK
|
1902 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1904 atomic_inc(&head_sh
->count
);
1905 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, head_sh
,
1906 to_addr_conv(sh
, percpu
, j
));
1908 flags
= prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
;
1909 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
,
1910 to_addr_conv(sh
, percpu
, j
));
1913 if (unlikely(count
== 1))
1914 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1916 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1919 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1926 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1927 struct dma_async_tx_descriptor
*tx
)
1929 struct async_submit_ctl submit
;
1930 struct page
**blocks
;
1931 int count
, i
, j
= 0;
1932 struct stripe_head
*head_sh
= sh
;
1935 unsigned long txflags
;
1937 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1939 for (i
= 0; i
< sh
->disks
; i
++) {
1940 if (sh
->pd_idx
== i
|| sh
->qd_idx
== i
)
1942 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1945 if (i
>= sh
->disks
) {
1946 atomic_inc(&sh
->count
);
1947 set_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
1948 set_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
1949 ops_complete_reconstruct(sh
);
1954 blocks
= to_addr_page(percpu
, j
);
1956 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1957 synflags
= SYNDROME_SRC_WRITTEN
;
1958 txflags
= ASYNC_TX_ACK
| ASYNC_TX_PQ_XOR_DST
;
1960 synflags
= SYNDROME_SRC_ALL
;
1961 txflags
= ASYNC_TX_ACK
;
1964 count
= set_syndrome_sources(blocks
, sh
, synflags
);
1965 last_stripe
= !head_sh
->batch_head
||
1966 list_first_entry(&sh
->batch_list
,
1967 struct stripe_head
, batch_list
) == head_sh
;
1970 atomic_inc(&head_sh
->count
);
1971 init_async_submit(&submit
, txflags
, tx
, ops_complete_reconstruct
,
1972 head_sh
, to_addr_conv(sh
, percpu
, j
));
1974 init_async_submit(&submit
, 0, tx
, NULL
, NULL
,
1975 to_addr_conv(sh
, percpu
, j
));
1976 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1979 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1985 static void ops_complete_check(void *stripe_head_ref
)
1987 struct stripe_head
*sh
= stripe_head_ref
;
1989 pr_debug("%s: stripe %llu\n", __func__
,
1990 (unsigned long long)sh
->sector
);
1992 sh
->check_state
= check_state_check_result
;
1993 set_bit(STRIPE_HANDLE
, &sh
->state
);
1994 raid5_release_stripe(sh
);
1997 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1999 int disks
= sh
->disks
;
2000 int pd_idx
= sh
->pd_idx
;
2001 int qd_idx
= sh
->qd_idx
;
2002 struct page
*xor_dest
;
2003 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
2004 struct dma_async_tx_descriptor
*tx
;
2005 struct async_submit_ctl submit
;
2009 pr_debug("%s: stripe %llu\n", __func__
,
2010 (unsigned long long)sh
->sector
);
2012 BUG_ON(sh
->batch_head
);
2014 xor_dest
= sh
->dev
[pd_idx
].page
;
2015 xor_srcs
[count
++] = xor_dest
;
2016 for (i
= disks
; i
--; ) {
2017 if (i
== pd_idx
|| i
== qd_idx
)
2019 xor_srcs
[count
++] = sh
->dev
[i
].page
;
2022 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
2023 to_addr_conv(sh
, percpu
, 0));
2024 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
2025 &sh
->ops
.zero_sum_result
, &submit
);
2027 atomic_inc(&sh
->count
);
2028 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
2029 tx
= async_trigger_callback(&submit
);
2032 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
2034 struct page
**srcs
= to_addr_page(percpu
, 0);
2035 struct async_submit_ctl submit
;
2038 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
2039 (unsigned long long)sh
->sector
, checkp
);
2041 BUG_ON(sh
->batch_head
);
2042 count
= set_syndrome_sources(srcs
, sh
, SYNDROME_SRC_ALL
);
2046 atomic_inc(&sh
->count
);
2047 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
2048 sh
, to_addr_conv(sh
, percpu
, 0));
2049 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
2050 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
2053 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
2055 int overlap_clear
= 0, i
, disks
= sh
->disks
;
2056 struct dma_async_tx_descriptor
*tx
= NULL
;
2057 struct r5conf
*conf
= sh
->raid_conf
;
2058 int level
= conf
->level
;
2059 struct raid5_percpu
*percpu
;
2063 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
2064 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
2065 ops_run_biofill(sh
);
2069 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
2071 tx
= ops_run_compute5(sh
, percpu
);
2073 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
2074 tx
= ops_run_compute6_1(sh
, percpu
);
2076 tx
= ops_run_compute6_2(sh
, percpu
);
2078 /* terminate the chain if reconstruct is not set to be run */
2079 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
2083 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
)) {
2085 tx
= ops_run_prexor5(sh
, percpu
, tx
);
2087 tx
= ops_run_prexor6(sh
, percpu
, tx
);
2090 if (test_bit(STRIPE_OP_PARTIAL_PARITY
, &ops_request
))
2091 tx
= ops_run_partial_parity(sh
, percpu
, tx
);
2093 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
2094 tx
= ops_run_biodrain(sh
, tx
);
2098 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
2100 ops_run_reconstruct5(sh
, percpu
, tx
);
2102 ops_run_reconstruct6(sh
, percpu
, tx
);
2105 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
2106 if (sh
->check_state
== check_state_run
)
2107 ops_run_check_p(sh
, percpu
);
2108 else if (sh
->check_state
== check_state_run_q
)
2109 ops_run_check_pq(sh
, percpu
, 0);
2110 else if (sh
->check_state
== check_state_run_pq
)
2111 ops_run_check_pq(sh
, percpu
, 1);
2116 if (overlap_clear
&& !sh
->batch_head
)
2117 for (i
= disks
; i
--; ) {
2118 struct r5dev
*dev
= &sh
->dev
[i
];
2119 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
2120 wake_up(&sh
->raid_conf
->wait_for_overlap
);
2125 static void free_stripe(struct kmem_cache
*sc
, struct stripe_head
*sh
)
2128 __free_page(sh
->ppl_page
);
2129 kmem_cache_free(sc
, sh
);
2132 static struct stripe_head
*alloc_stripe(struct kmem_cache
*sc
, gfp_t gfp
,
2133 int disks
, struct r5conf
*conf
)
2135 struct stripe_head
*sh
;
2138 sh
= kmem_cache_zalloc(sc
, gfp
);
2140 spin_lock_init(&sh
->stripe_lock
);
2141 spin_lock_init(&sh
->batch_lock
);
2142 INIT_LIST_HEAD(&sh
->batch_list
);
2143 INIT_LIST_HEAD(&sh
->lru
);
2144 INIT_LIST_HEAD(&sh
->r5c
);
2145 INIT_LIST_HEAD(&sh
->log_list
);
2146 atomic_set(&sh
->count
, 1);
2147 sh
->raid_conf
= conf
;
2148 sh
->log_start
= MaxSector
;
2149 for (i
= 0; i
< disks
; i
++) {
2150 struct r5dev
*dev
= &sh
->dev
[i
];
2152 bio_init(&dev
->req
, &dev
->vec
, 1);
2153 bio_init(&dev
->rreq
, &dev
->rvec
, 1);
2156 if (raid5_has_ppl(conf
)) {
2157 sh
->ppl_page
= alloc_page(gfp
);
2158 if (!sh
->ppl_page
) {
2159 free_stripe(sc
, sh
);
2166 static int grow_one_stripe(struct r5conf
*conf
, gfp_t gfp
)
2168 struct stripe_head
*sh
;
2170 sh
= alloc_stripe(conf
->slab_cache
, gfp
, conf
->pool_size
, conf
);
2174 if (grow_buffers(sh
, gfp
)) {
2176 free_stripe(conf
->slab_cache
, sh
);
2179 sh
->hash_lock_index
=
2180 conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
;
2181 /* we just created an active stripe so... */
2182 atomic_inc(&conf
->active_stripes
);
2184 raid5_release_stripe(sh
);
2185 conf
->max_nr_stripes
++;
2189 static int grow_stripes(struct r5conf
*conf
, int num
)
2191 struct kmem_cache
*sc
;
2192 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
2194 if (conf
->mddev
->gendisk
)
2195 sprintf(conf
->cache_name
[0],
2196 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
2198 sprintf(conf
->cache_name
[0],
2199 "raid%d-%p", conf
->level
, conf
->mddev
);
2200 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
2202 conf
->active_name
= 0;
2203 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
2204 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
2208 conf
->slab_cache
= sc
;
2209 conf
->pool_size
= devs
;
2211 if (!grow_one_stripe(conf
, GFP_KERNEL
))
2218 * scribble_len - return the required size of the scribble region
2219 * @num - total number of disks in the array
2221 * The size must be enough to contain:
2222 * 1/ a struct page pointer for each device in the array +2
2223 * 2/ room to convert each entry in (1) to its corresponding dma
2224 * (dma_map_page()) or page (page_address()) address.
2226 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2227 * calculate over all devices (not just the data blocks), using zeros in place
2228 * of the P and Q blocks.
2230 static struct flex_array
*scribble_alloc(int num
, int cnt
, gfp_t flags
)
2232 struct flex_array
*ret
;
2235 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
2236 ret
= flex_array_alloc(len
, cnt
, flags
);
2239 /* always prealloc all elements, so no locking is required */
2240 if (flex_array_prealloc(ret
, 0, cnt
, flags
)) {
2241 flex_array_free(ret
);
2247 static int resize_chunks(struct r5conf
*conf
, int new_disks
, int new_sectors
)
2253 * Never shrink. And mddev_suspend() could deadlock if this is called
2254 * from raid5d. In that case, scribble_disks and scribble_sectors
2255 * should equal to new_disks and new_sectors
2257 if (conf
->scribble_disks
>= new_disks
&&
2258 conf
->scribble_sectors
>= new_sectors
)
2260 mddev_suspend(conf
->mddev
);
2262 for_each_present_cpu(cpu
) {
2263 struct raid5_percpu
*percpu
;
2264 struct flex_array
*scribble
;
2266 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
2267 scribble
= scribble_alloc(new_disks
,
2268 new_sectors
/ STRIPE_SECTORS
,
2272 flex_array_free(percpu
->scribble
);
2273 percpu
->scribble
= scribble
;
2280 mddev_resume(conf
->mddev
);
2282 conf
->scribble_disks
= new_disks
;
2283 conf
->scribble_sectors
= new_sectors
;
2288 static int resize_stripes(struct r5conf
*conf
, int newsize
)
2290 /* Make all the stripes able to hold 'newsize' devices.
2291 * New slots in each stripe get 'page' set to a new page.
2293 * This happens in stages:
2294 * 1/ create a new kmem_cache and allocate the required number of
2296 * 2/ gather all the old stripe_heads and transfer the pages across
2297 * to the new stripe_heads. This will have the side effect of
2298 * freezing the array as once all stripe_heads have been collected,
2299 * no IO will be possible. Old stripe heads are freed once their
2300 * pages have been transferred over, and the old kmem_cache is
2301 * freed when all stripes are done.
2302 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2303 * we simple return a failure status - no need to clean anything up.
2304 * 4/ allocate new pages for the new slots in the new stripe_heads.
2305 * If this fails, we don't bother trying the shrink the
2306 * stripe_heads down again, we just leave them as they are.
2307 * As each stripe_head is processed the new one is released into
2310 * Once step2 is started, we cannot afford to wait for a write,
2311 * so we use GFP_NOIO allocations.
2313 struct stripe_head
*osh
, *nsh
;
2314 LIST_HEAD(newstripes
);
2315 struct disk_info
*ndisks
;
2317 struct kmem_cache
*sc
;
2321 md_allow_write(conf
->mddev
);
2324 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
2325 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
2330 /* Need to ensure auto-resizing doesn't interfere */
2331 mutex_lock(&conf
->cache_size_mutex
);
2333 for (i
= conf
->max_nr_stripes
; i
; i
--) {
2334 nsh
= alloc_stripe(sc
, GFP_KERNEL
, newsize
, conf
);
2338 list_add(&nsh
->lru
, &newstripes
);
2341 /* didn't get enough, give up */
2342 while (!list_empty(&newstripes
)) {
2343 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2344 list_del(&nsh
->lru
);
2345 free_stripe(sc
, nsh
);
2347 kmem_cache_destroy(sc
);
2348 mutex_unlock(&conf
->cache_size_mutex
);
2351 /* Step 2 - Must use GFP_NOIO now.
2352 * OK, we have enough stripes, start collecting inactive
2353 * stripes and copying them over
2357 list_for_each_entry(nsh
, &newstripes
, lru
) {
2358 lock_device_hash_lock(conf
, hash
);
2359 wait_event_cmd(conf
->wait_for_stripe
,
2360 !list_empty(conf
->inactive_list
+ hash
),
2361 unlock_device_hash_lock(conf
, hash
),
2362 lock_device_hash_lock(conf
, hash
));
2363 osh
= get_free_stripe(conf
, hash
);
2364 unlock_device_hash_lock(conf
, hash
);
2366 for(i
=0; i
<conf
->pool_size
; i
++) {
2367 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
2368 nsh
->dev
[i
].orig_page
= osh
->dev
[i
].page
;
2370 nsh
->hash_lock_index
= hash
;
2371 free_stripe(conf
->slab_cache
, osh
);
2373 if (cnt
>= conf
->max_nr_stripes
/ NR_STRIPE_HASH_LOCKS
+
2374 !!((conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
) > hash
)) {
2379 kmem_cache_destroy(conf
->slab_cache
);
2382 * At this point, we are holding all the stripes so the array
2383 * is completely stalled, so now is a good time to resize
2384 * conf->disks and the scribble region
2386 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
2388 for (i
= 0; i
< conf
->pool_size
; i
++)
2389 ndisks
[i
] = conf
->disks
[i
];
2391 for (i
= conf
->pool_size
; i
< newsize
; i
++) {
2392 ndisks
[i
].extra_page
= alloc_page(GFP_NOIO
);
2393 if (!ndisks
[i
].extra_page
)
2398 for (i
= conf
->pool_size
; i
< newsize
; i
++)
2399 if (ndisks
[i
].extra_page
)
2400 put_page(ndisks
[i
].extra_page
);
2404 conf
->disks
= ndisks
;
2409 mutex_unlock(&conf
->cache_size_mutex
);
2411 conf
->slab_cache
= sc
;
2412 conf
->active_name
= 1-conf
->active_name
;
2414 /* Step 4, return new stripes to service */
2415 while(!list_empty(&newstripes
)) {
2416 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2417 list_del_init(&nsh
->lru
);
2419 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
2420 if (nsh
->dev
[i
].page
== NULL
) {
2421 struct page
*p
= alloc_page(GFP_NOIO
);
2422 nsh
->dev
[i
].page
= p
;
2423 nsh
->dev
[i
].orig_page
= p
;
2427 raid5_release_stripe(nsh
);
2429 /* critical section pass, GFP_NOIO no longer needed */
2432 conf
->pool_size
= newsize
;
2436 static int drop_one_stripe(struct r5conf
*conf
)
2438 struct stripe_head
*sh
;
2439 int hash
= (conf
->max_nr_stripes
- 1) & STRIPE_HASH_LOCKS_MASK
;
2441 spin_lock_irq(conf
->hash_locks
+ hash
);
2442 sh
= get_free_stripe(conf
, hash
);
2443 spin_unlock_irq(conf
->hash_locks
+ hash
);
2446 BUG_ON(atomic_read(&sh
->count
));
2448 free_stripe(conf
->slab_cache
, sh
);
2449 atomic_dec(&conf
->active_stripes
);
2450 conf
->max_nr_stripes
--;
2454 static void shrink_stripes(struct r5conf
*conf
)
2456 while (conf
->max_nr_stripes
&&
2457 drop_one_stripe(conf
))
2460 kmem_cache_destroy(conf
->slab_cache
);
2461 conf
->slab_cache
= NULL
;
2464 static void raid5_end_read_request(struct bio
* bi
)
2466 struct stripe_head
*sh
= bi
->bi_private
;
2467 struct r5conf
*conf
= sh
->raid_conf
;
2468 int disks
= sh
->disks
, i
;
2469 char b
[BDEVNAME_SIZE
];
2470 struct md_rdev
*rdev
= NULL
;
2473 for (i
=0 ; i
<disks
; i
++)
2474 if (bi
== &sh
->dev
[i
].req
)
2477 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2478 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2485 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2486 /* If replacement finished while this request was outstanding,
2487 * 'replacement' might be NULL already.
2488 * In that case it moved down to 'rdev'.
2489 * rdev is not removed until all requests are finished.
2491 rdev
= conf
->disks
[i
].replacement
;
2493 rdev
= conf
->disks
[i
].rdev
;
2495 if (use_new_offset(conf
, sh
))
2496 s
= sh
->sector
+ rdev
->new_data_offset
;
2498 s
= sh
->sector
+ rdev
->data_offset
;
2499 if (!bi
->bi_error
) {
2500 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2501 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2502 /* Note that this cannot happen on a
2503 * replacement device. We just fail those on
2506 pr_info_ratelimited(
2507 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2508 mdname(conf
->mddev
), STRIPE_SECTORS
,
2509 (unsigned long long)s
,
2510 bdevname(rdev
->bdev
, b
));
2511 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
2512 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2513 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2514 } else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2515 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2517 if (test_bit(R5_InJournal
, &sh
->dev
[i
].flags
))
2519 * end read for a page in journal, this
2520 * must be preparing for prexor in rmw
2522 set_bit(R5_OrigPageUPTDODATE
, &sh
->dev
[i
].flags
);
2524 if (atomic_read(&rdev
->read_errors
))
2525 atomic_set(&rdev
->read_errors
, 0);
2527 const char *bdn
= bdevname(rdev
->bdev
, b
);
2531 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2532 atomic_inc(&rdev
->read_errors
);
2533 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2534 pr_warn_ratelimited(
2535 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2536 mdname(conf
->mddev
),
2537 (unsigned long long)s
,
2539 else if (conf
->mddev
->degraded
>= conf
->max_degraded
) {
2541 pr_warn_ratelimited(
2542 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2543 mdname(conf
->mddev
),
2544 (unsigned long long)s
,
2546 } else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
)) {
2549 pr_warn_ratelimited(
2550 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2551 mdname(conf
->mddev
),
2552 (unsigned long long)s
,
2554 } else if (atomic_read(&rdev
->read_errors
)
2555 > conf
->max_nr_stripes
)
2556 pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2557 mdname(conf
->mddev
), bdn
);
2560 if (set_bad
&& test_bit(In_sync
, &rdev
->flags
)
2561 && !test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2564 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
)) {
2565 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2566 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2568 set_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2570 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2571 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2573 && test_bit(In_sync
, &rdev
->flags
)
2574 && rdev_set_badblocks(
2575 rdev
, sh
->sector
, STRIPE_SECTORS
, 0)))
2576 md_error(conf
->mddev
, rdev
);
2579 rdev_dec_pending(rdev
, conf
->mddev
);
2581 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2582 set_bit(STRIPE_HANDLE
, &sh
->state
);
2583 raid5_release_stripe(sh
);
2586 static void raid5_end_write_request(struct bio
*bi
)
2588 struct stripe_head
*sh
= bi
->bi_private
;
2589 struct r5conf
*conf
= sh
->raid_conf
;
2590 int disks
= sh
->disks
, i
;
2591 struct md_rdev
*uninitialized_var(rdev
);
2594 int replacement
= 0;
2596 for (i
= 0 ; i
< disks
; i
++) {
2597 if (bi
== &sh
->dev
[i
].req
) {
2598 rdev
= conf
->disks
[i
].rdev
;
2601 if (bi
== &sh
->dev
[i
].rreq
) {
2602 rdev
= conf
->disks
[i
].replacement
;
2606 /* rdev was removed and 'replacement'
2607 * replaced it. rdev is not removed
2608 * until all requests are finished.
2610 rdev
= conf
->disks
[i
].rdev
;
2614 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2615 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2625 md_error(conf
->mddev
, rdev
);
2626 else if (is_badblock(rdev
, sh
->sector
,
2628 &first_bad
, &bad_sectors
))
2629 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
2632 set_bit(STRIPE_DEGRADED
, &sh
->state
);
2633 set_bit(WriteErrorSeen
, &rdev
->flags
);
2634 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
2635 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2636 set_bit(MD_RECOVERY_NEEDED
,
2637 &rdev
->mddev
->recovery
);
2638 } else if (is_badblock(rdev
, sh
->sector
,
2640 &first_bad
, &bad_sectors
)) {
2641 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
2642 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))
2643 /* That was a successful write so make
2644 * sure it looks like we already did
2647 set_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2650 rdev_dec_pending(rdev
, conf
->mddev
);
2652 if (sh
->batch_head
&& bi
->bi_error
&& !replacement
)
2653 set_bit(STRIPE_BATCH_ERR
, &sh
->batch_head
->state
);
2656 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
2657 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2658 set_bit(STRIPE_HANDLE
, &sh
->state
);
2659 raid5_release_stripe(sh
);
2661 if (sh
->batch_head
&& sh
!= sh
->batch_head
)
2662 raid5_release_stripe(sh
->batch_head
);
2665 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
2667 struct r5dev
*dev
= &sh
->dev
[i
];
2670 dev
->sector
= raid5_compute_blocknr(sh
, i
, previous
);
2673 static void raid5_error(struct mddev
*mddev
, struct md_rdev
*rdev
)
2675 char b
[BDEVNAME_SIZE
];
2676 struct r5conf
*conf
= mddev
->private;
2677 unsigned long flags
;
2678 pr_debug("raid456: error called\n");
2680 spin_lock_irqsave(&conf
->device_lock
, 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_bit(Faulty
, &rdev
->flags
);
2688 set_mask_bits(&mddev
->sb_flags
, 0,
2689 BIT(MD_SB_CHANGE_DEVS
) | BIT(MD_SB_CHANGE_PENDING
));
2690 pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2691 "md/raid:%s: Operation continuing on %d devices.\n",
2693 bdevname(rdev
->bdev
, b
),
2695 conf
->raid_disks
- mddev
->degraded
);
2696 r5c_update_on_rdev_error(mddev
, rdev
);
2700 * Input: a 'big' sector number,
2701 * Output: index of the data and parity disk, and the sector # in them.
2703 sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
2704 int previous
, int *dd_idx
,
2705 struct stripe_head
*sh
)
2707 sector_t stripe
, stripe2
;
2708 sector_t chunk_number
;
2709 unsigned int chunk_offset
;
2712 sector_t new_sector
;
2713 int algorithm
= previous
? conf
->prev_algo
2715 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2716 : conf
->chunk_sectors
;
2717 int raid_disks
= previous
? conf
->previous_raid_disks
2719 int data_disks
= raid_disks
- conf
->max_degraded
;
2721 /* First compute the information on this sector */
2724 * Compute the chunk number and the sector offset inside the chunk
2726 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
2727 chunk_number
= r_sector
;
2730 * Compute the stripe number
2732 stripe
= chunk_number
;
2733 *dd_idx
= sector_div(stripe
, data_disks
);
2736 * Select the parity disk based on the user selected algorithm.
2738 pd_idx
= qd_idx
= -1;
2739 switch(conf
->level
) {
2741 pd_idx
= data_disks
;
2744 switch (algorithm
) {
2745 case ALGORITHM_LEFT_ASYMMETRIC
:
2746 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2747 if (*dd_idx
>= pd_idx
)
2750 case ALGORITHM_RIGHT_ASYMMETRIC
:
2751 pd_idx
= sector_div(stripe2
, raid_disks
);
2752 if (*dd_idx
>= pd_idx
)
2755 case ALGORITHM_LEFT_SYMMETRIC
:
2756 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2757 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2759 case ALGORITHM_RIGHT_SYMMETRIC
:
2760 pd_idx
= sector_div(stripe2
, raid_disks
);
2761 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2763 case ALGORITHM_PARITY_0
:
2767 case ALGORITHM_PARITY_N
:
2768 pd_idx
= data_disks
;
2776 switch (algorithm
) {
2777 case ALGORITHM_LEFT_ASYMMETRIC
:
2778 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2779 qd_idx
= pd_idx
+ 1;
2780 if (pd_idx
== raid_disks
-1) {
2781 (*dd_idx
)++; /* Q D D D P */
2783 } else if (*dd_idx
>= pd_idx
)
2784 (*dd_idx
) += 2; /* D D P Q D */
2786 case ALGORITHM_RIGHT_ASYMMETRIC
:
2787 pd_idx
= sector_div(stripe2
, raid_disks
);
2788 qd_idx
= pd_idx
+ 1;
2789 if (pd_idx
== raid_disks
-1) {
2790 (*dd_idx
)++; /* Q D D D P */
2792 } else if (*dd_idx
>= pd_idx
)
2793 (*dd_idx
) += 2; /* D D P Q D */
2795 case ALGORITHM_LEFT_SYMMETRIC
:
2796 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2797 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2798 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2800 case ALGORITHM_RIGHT_SYMMETRIC
:
2801 pd_idx
= sector_div(stripe2
, raid_disks
);
2802 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2803 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2806 case ALGORITHM_PARITY_0
:
2811 case ALGORITHM_PARITY_N
:
2812 pd_idx
= data_disks
;
2813 qd_idx
= data_disks
+ 1;
2816 case ALGORITHM_ROTATING_ZERO_RESTART
:
2817 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2818 * of blocks for computing Q is different.
2820 pd_idx
= sector_div(stripe2
, raid_disks
);
2821 qd_idx
= pd_idx
+ 1;
2822 if (pd_idx
== raid_disks
-1) {
2823 (*dd_idx
)++; /* Q D D D P */
2825 } else if (*dd_idx
>= pd_idx
)
2826 (*dd_idx
) += 2; /* D D P Q D */
2830 case ALGORITHM_ROTATING_N_RESTART
:
2831 /* Same a left_asymmetric, by first stripe is
2832 * D D D P Q rather than
2836 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2837 qd_idx
= pd_idx
+ 1;
2838 if (pd_idx
== raid_disks
-1) {
2839 (*dd_idx
)++; /* Q D D D P */
2841 } else if (*dd_idx
>= pd_idx
)
2842 (*dd_idx
) += 2; /* D D P Q D */
2846 case ALGORITHM_ROTATING_N_CONTINUE
:
2847 /* Same as left_symmetric but Q is before P */
2848 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2849 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
2850 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2854 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2855 /* RAID5 left_asymmetric, with Q on last device */
2856 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2857 if (*dd_idx
>= pd_idx
)
2859 qd_idx
= raid_disks
- 1;
2862 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2863 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2864 if (*dd_idx
>= pd_idx
)
2866 qd_idx
= raid_disks
- 1;
2869 case ALGORITHM_LEFT_SYMMETRIC_6
:
2870 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2871 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2872 qd_idx
= raid_disks
- 1;
2875 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2876 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2877 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2878 qd_idx
= raid_disks
- 1;
2881 case ALGORITHM_PARITY_0_6
:
2884 qd_idx
= raid_disks
- 1;
2894 sh
->pd_idx
= pd_idx
;
2895 sh
->qd_idx
= qd_idx
;
2896 sh
->ddf_layout
= ddf_layout
;
2899 * Finally, compute the new sector number
2901 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
2905 sector_t
raid5_compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
2907 struct r5conf
*conf
= sh
->raid_conf
;
2908 int raid_disks
= sh
->disks
;
2909 int data_disks
= raid_disks
- conf
->max_degraded
;
2910 sector_t new_sector
= sh
->sector
, check
;
2911 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2912 : conf
->chunk_sectors
;
2913 int algorithm
= previous
? conf
->prev_algo
2917 sector_t chunk_number
;
2918 int dummy1
, dd_idx
= i
;
2920 struct stripe_head sh2
;
2922 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
2923 stripe
= new_sector
;
2925 if (i
== sh
->pd_idx
)
2927 switch(conf
->level
) {
2930 switch (algorithm
) {
2931 case ALGORITHM_LEFT_ASYMMETRIC
:
2932 case ALGORITHM_RIGHT_ASYMMETRIC
:
2936 case ALGORITHM_LEFT_SYMMETRIC
:
2937 case ALGORITHM_RIGHT_SYMMETRIC
:
2940 i
-= (sh
->pd_idx
+ 1);
2942 case ALGORITHM_PARITY_0
:
2945 case ALGORITHM_PARITY_N
:
2952 if (i
== sh
->qd_idx
)
2953 return 0; /* It is the Q disk */
2954 switch (algorithm
) {
2955 case ALGORITHM_LEFT_ASYMMETRIC
:
2956 case ALGORITHM_RIGHT_ASYMMETRIC
:
2957 case ALGORITHM_ROTATING_ZERO_RESTART
:
2958 case ALGORITHM_ROTATING_N_RESTART
:
2959 if (sh
->pd_idx
== raid_disks
-1)
2960 i
--; /* Q D D D P */
2961 else if (i
> sh
->pd_idx
)
2962 i
-= 2; /* D D P Q D */
2964 case ALGORITHM_LEFT_SYMMETRIC
:
2965 case ALGORITHM_RIGHT_SYMMETRIC
:
2966 if (sh
->pd_idx
== raid_disks
-1)
2967 i
--; /* Q D D D P */
2972 i
-= (sh
->pd_idx
+ 2);
2975 case ALGORITHM_PARITY_0
:
2978 case ALGORITHM_PARITY_N
:
2980 case ALGORITHM_ROTATING_N_CONTINUE
:
2981 /* Like left_symmetric, but P is before Q */
2982 if (sh
->pd_idx
== 0)
2983 i
--; /* P D D D Q */
2988 i
-= (sh
->pd_idx
+ 1);
2991 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2992 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2996 case ALGORITHM_LEFT_SYMMETRIC_6
:
2997 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2999 i
+= data_disks
+ 1;
3000 i
-= (sh
->pd_idx
+ 1);
3002 case ALGORITHM_PARITY_0_6
:
3011 chunk_number
= stripe
* data_disks
+ i
;
3012 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
3014 check
= raid5_compute_sector(conf
, r_sector
,
3015 previous
, &dummy1
, &sh2
);
3016 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
3017 || sh2
.qd_idx
!= sh
->qd_idx
) {
3018 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3019 mdname(conf
->mddev
));
3026 * There are cases where we want handle_stripe_dirtying() and
3027 * schedule_reconstruction() to delay towrite to some dev of a stripe.
3029 * This function checks whether we want to delay the towrite. Specifically,
3030 * we delay the towrite when:
3032 * 1. degraded stripe has a non-overwrite to the missing dev, AND this
3033 * stripe has data in journal (for other devices).
3035 * In this case, when reading data for the non-overwrite dev, it is
3036 * necessary to handle complex rmw of write back cache (prexor with
3037 * orig_page, and xor with page). To keep read path simple, we would
3038 * like to flush data in journal to RAID disks first, so complex rmw
3039 * is handled in the write patch (handle_stripe_dirtying).
3041 * 2. when journal space is critical (R5C_LOG_CRITICAL=1)
3043 * It is important to be able to flush all stripes in raid5-cache.
3044 * Therefore, we need reserve some space on the journal device for
3045 * these flushes. If flush operation includes pending writes to the
3046 * stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3047 * for the flush out. If we exclude these pending writes from flush
3048 * operation, we only need (conf->max_degraded + 1) pages per stripe.
3049 * Therefore, excluding pending writes in these cases enables more
3050 * efficient use of the journal device.
3052 * Note: To make sure the stripe makes progress, we only delay
3053 * towrite for stripes with data already in journal (injournal > 0).
3054 * When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3055 * no_space_stripes list.
3057 * 3. during journal failure
3058 * In journal failure, we try to flush all cached data to raid disks
3059 * based on data in stripe cache. The array is read-only to upper
3060 * layers, so we would skip all pending writes.
3063 static inline bool delay_towrite(struct r5conf
*conf
,
3065 struct stripe_head_state
*s
)
3068 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3069 !test_bit(R5_Insync
, &dev
->flags
) && s
->injournal
)
3072 if (test_bit(R5C_LOG_CRITICAL
, &conf
->cache_state
) &&
3076 if (s
->log_failed
&& s
->injournal
)
3082 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3083 int rcw
, int expand
)
3085 int i
, pd_idx
= sh
->pd_idx
, qd_idx
= sh
->qd_idx
, disks
= sh
->disks
;
3086 struct r5conf
*conf
= sh
->raid_conf
;
3087 int level
= conf
->level
;
3091 * In some cases, handle_stripe_dirtying initially decided to
3092 * run rmw and allocates extra page for prexor. However, rcw is
3093 * cheaper later on. We need to free the extra page now,
3094 * because we won't be able to do that in ops_complete_prexor().
3096 r5c_release_extra_page(sh
);
3098 for (i
= disks
; i
--; ) {
3099 struct r5dev
*dev
= &sh
->dev
[i
];
3101 if (dev
->towrite
&& !delay_towrite(conf
, dev
, s
)) {
3102 set_bit(R5_LOCKED
, &dev
->flags
);
3103 set_bit(R5_Wantdrain
, &dev
->flags
);
3105 clear_bit(R5_UPTODATE
, &dev
->flags
);
3107 } else if (test_bit(R5_InJournal
, &dev
->flags
)) {
3108 set_bit(R5_LOCKED
, &dev
->flags
);
3112 /* if we are not expanding this is a proper write request, and
3113 * there will be bios with new data to be drained into the
3118 /* False alarm, nothing to do */
3120 sh
->reconstruct_state
= reconstruct_state_drain_run
;
3121 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
3123 sh
->reconstruct_state
= reconstruct_state_run
;
3125 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
3127 if (s
->locked
+ conf
->max_degraded
== disks
)
3128 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3129 atomic_inc(&conf
->pending_full_writes
);
3131 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
3132 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
3133 BUG_ON(level
== 6 &&
3134 (!(test_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
) ||
3135 test_bit(R5_Wantcompute
, &sh
->dev
[qd_idx
].flags
))));
3137 for (i
= disks
; i
--; ) {
3138 struct r5dev
*dev
= &sh
->dev
[i
];
3139 if (i
== pd_idx
|| i
== qd_idx
)
3143 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
3144 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3145 set_bit(R5_Wantdrain
, &dev
->flags
);
3146 set_bit(R5_LOCKED
, &dev
->flags
);
3147 clear_bit(R5_UPTODATE
, &dev
->flags
);
3149 } else if (test_bit(R5_InJournal
, &dev
->flags
)) {
3150 set_bit(R5_LOCKED
, &dev
->flags
);
3155 /* False alarm - nothing to do */
3157 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
3158 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
3159 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
3160 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
3163 /* keep the parity disk(s) locked while asynchronous operations
3166 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
3167 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
3171 int qd_idx
= sh
->qd_idx
;
3172 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
3174 set_bit(R5_LOCKED
, &dev
->flags
);
3175 clear_bit(R5_UPTODATE
, &dev
->flags
);
3179 if (raid5_has_ppl(sh
->raid_conf
) && sh
->ppl_page
&&
3180 test_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
) &&
3181 !test_bit(STRIPE_FULL_WRITE
, &sh
->state
) &&
3182 test_bit(R5_Insync
, &sh
->dev
[pd_idx
].flags
))
3183 set_bit(STRIPE_OP_PARTIAL_PARITY
, &s
->ops_request
);
3185 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3186 __func__
, (unsigned long long)sh
->sector
,
3187 s
->locked
, s
->ops_request
);
3191 * Each stripe/dev can have one or more bion attached.
3192 * toread/towrite point to the first in a chain.
3193 * The bi_next chain must be in order.
3195 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
,
3196 int forwrite
, int previous
)
3199 struct r5conf
*conf
= sh
->raid_conf
;
3202 pr_debug("adding bi b#%llu to stripe s#%llu\n",
3203 (unsigned long long)bi
->bi_iter
.bi_sector
,
3204 (unsigned long long)sh
->sector
);
3206 spin_lock_irq(&sh
->stripe_lock
);
3207 /* Don't allow new IO added to stripes in batch list */
3211 bip
= &sh
->dev
[dd_idx
].towrite
;
3215 bip
= &sh
->dev
[dd_idx
].toread
;
3216 while (*bip
&& (*bip
)->bi_iter
.bi_sector
< bi
->bi_iter
.bi_sector
) {
3217 if (bio_end_sector(*bip
) > bi
->bi_iter
.bi_sector
)
3219 bip
= & (*bip
)->bi_next
;
3221 if (*bip
&& (*bip
)->bi_iter
.bi_sector
< bio_end_sector(bi
))
3224 if (forwrite
&& raid5_has_ppl(conf
)) {
3226 * With PPL only writes to consecutive data chunks within a
3227 * stripe are allowed because for a single stripe_head we can
3228 * only have one PPL entry at a time, which describes one data
3229 * range. Not really an overlap, but wait_for_overlap can be
3230 * used to handle this.
3238 for (i
= 0; i
< sh
->disks
; i
++) {
3239 if (i
!= sh
->pd_idx
&&
3240 (i
== dd_idx
|| sh
->dev
[i
].towrite
)) {
3241 sector
= sh
->dev
[i
].sector
;
3242 if (count
== 0 || sector
< first
)
3250 if (first
+ conf
->chunk_sectors
* (count
- 1) != last
)
3254 if (!forwrite
|| previous
)
3255 clear_bit(STRIPE_BATCH_READY
, &sh
->state
);
3257 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
3261 bio_inc_remaining(bi
);
3262 md_write_inc(conf
->mddev
, bi
);
3265 /* check if page is covered */
3266 sector_t sector
= sh
->dev
[dd_idx
].sector
;
3267 for (bi
=sh
->dev
[dd_idx
].towrite
;
3268 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
3269 bi
&& bi
->bi_iter
.bi_sector
<= sector
;
3270 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
3271 if (bio_end_sector(bi
) >= sector
)
3272 sector
= bio_end_sector(bi
);
3274 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
3275 if (!test_and_set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
))
3276 sh
->overwrite_disks
++;
3279 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3280 (unsigned long long)(*bip
)->bi_iter
.bi_sector
,
3281 (unsigned long long)sh
->sector
, dd_idx
);
3283 if (conf
->mddev
->bitmap
&& firstwrite
) {
3284 /* Cannot hold spinlock over bitmap_startwrite,
3285 * but must ensure this isn't added to a batch until
3286 * we have added to the bitmap and set bm_seq.
3287 * So set STRIPE_BITMAP_PENDING to prevent
3289 * If multiple add_stripe_bio() calls race here they
3290 * much all set STRIPE_BITMAP_PENDING. So only the first one
3291 * to complete "bitmap_startwrite" gets to set
3292 * STRIPE_BIT_DELAY. This is important as once a stripe
3293 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3296 set_bit(STRIPE_BITMAP_PENDING
, &sh
->state
);
3297 spin_unlock_irq(&sh
->stripe_lock
);
3298 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
3300 spin_lock_irq(&sh
->stripe_lock
);
3301 clear_bit(STRIPE_BITMAP_PENDING
, &sh
->state
);
3302 if (!sh
->batch_head
) {
3303 sh
->bm_seq
= conf
->seq_flush
+1;
3304 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
3307 spin_unlock_irq(&sh
->stripe_lock
);
3309 if (stripe_can_batch(sh
))
3310 stripe_add_to_batch_list(conf
, sh
);
3314 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
3315 spin_unlock_irq(&sh
->stripe_lock
);
3319 static void end_reshape(struct r5conf
*conf
);
3321 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
3322 struct stripe_head
*sh
)
3324 int sectors_per_chunk
=
3325 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
3327 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
3328 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
3330 raid5_compute_sector(conf
,
3331 stripe
* (disks
- conf
->max_degraded
)
3332 *sectors_per_chunk
+ chunk_offset
,
3338 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
3339 struct stripe_head_state
*s
, int disks
)
3342 BUG_ON(sh
->batch_head
);
3343 for (i
= disks
; i
--; ) {
3347 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
3348 struct md_rdev
*rdev
;
3350 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3351 if (rdev
&& test_bit(In_sync
, &rdev
->flags
) &&
3352 !test_bit(Faulty
, &rdev
->flags
))
3353 atomic_inc(&rdev
->nr_pending
);
3358 if (!rdev_set_badblocks(
3362 md_error(conf
->mddev
, rdev
);
3363 rdev_dec_pending(rdev
, conf
->mddev
);
3366 spin_lock_irq(&sh
->stripe_lock
);
3367 /* fail all writes first */
3368 bi
= sh
->dev
[i
].towrite
;
3369 sh
->dev
[i
].towrite
= NULL
;
3370 sh
->overwrite_disks
= 0;
3371 spin_unlock_irq(&sh
->stripe_lock
);
3375 log_stripe_write_finished(sh
);
3377 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3378 wake_up(&conf
->wait_for_overlap
);
3380 while (bi
&& bi
->bi_iter
.bi_sector
<
3381 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3382 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3384 bi
->bi_error
= -EIO
;
3385 md_write_end(conf
->mddev
);
3390 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3391 STRIPE_SECTORS
, 0, 0);
3393 /* and fail all 'written' */
3394 bi
= sh
->dev
[i
].written
;
3395 sh
->dev
[i
].written
= NULL
;
3396 if (test_and_clear_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
)) {
3397 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
3398 sh
->dev
[i
].page
= sh
->dev
[i
].orig_page
;
3401 if (bi
) bitmap_end
= 1;
3402 while (bi
&& bi
->bi_iter
.bi_sector
<
3403 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3404 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
3406 bi
->bi_error
= -EIO
;
3407 md_write_end(conf
->mddev
);
3412 /* fail any reads if this device is non-operational and
3413 * the data has not reached the cache yet.
3415 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
3416 s
->failed
> conf
->max_degraded
&&
3417 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
3418 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
3419 spin_lock_irq(&sh
->stripe_lock
);
3420 bi
= sh
->dev
[i
].toread
;
3421 sh
->dev
[i
].toread
= NULL
;
3422 spin_unlock_irq(&sh
->stripe_lock
);
3423 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3424 wake_up(&conf
->wait_for_overlap
);
3427 while (bi
&& bi
->bi_iter
.bi_sector
<
3428 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
3429 struct bio
*nextbi
=
3430 r5_next_bio(bi
, sh
->dev
[i
].sector
);
3432 bi
->bi_error
= -EIO
;
3438 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3439 STRIPE_SECTORS
, 0, 0);
3440 /* If we were in the middle of a write the parity block might
3441 * still be locked - so just clear all R5_LOCKED flags
3443 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3448 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3449 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3450 md_wakeup_thread(conf
->mddev
->thread
);
3454 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
3455 struct stripe_head_state
*s
)
3460 BUG_ON(sh
->batch_head
);
3461 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3462 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
3463 wake_up(&conf
->wait_for_overlap
);
3466 /* There is nothing more to do for sync/check/repair.
3467 * Don't even need to abort as that is handled elsewhere
3468 * if needed, and not always wanted e.g. if there is a known
3470 * For recover/replace we need to record a bad block on all
3471 * non-sync devices, or abort the recovery
3473 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
3474 /* During recovery devices cannot be removed, so
3475 * locking and refcounting of rdevs is not needed
3478 for (i
= 0; i
< conf
->raid_disks
; i
++) {
3479 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3481 && !test_bit(Faulty
, &rdev
->flags
)
3482 && !test_bit(In_sync
, &rdev
->flags
)
3483 && !rdev_set_badblocks(rdev
, sh
->sector
,
3486 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
3488 && !test_bit(Faulty
, &rdev
->flags
)
3489 && !test_bit(In_sync
, &rdev
->flags
)
3490 && !rdev_set_badblocks(rdev
, sh
->sector
,
3496 conf
->recovery_disabled
=
3497 conf
->mddev
->recovery_disabled
;
3499 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, !abort
);
3502 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
3504 struct md_rdev
*rdev
;
3508 rdev
= rcu_dereference(sh
->raid_conf
->disks
[disk_idx
].replacement
);
3510 && !test_bit(Faulty
, &rdev
->flags
)
3511 && !test_bit(In_sync
, &rdev
->flags
)
3512 && (rdev
->recovery_offset
<= sh
->sector
3513 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
3519 static int need_this_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3520 int disk_idx
, int disks
)
3522 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3523 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
3524 &sh
->dev
[s
->failed_num
[1]] };
3528 if (test_bit(R5_LOCKED
, &dev
->flags
) ||
3529 test_bit(R5_UPTODATE
, &dev
->flags
))
3530 /* No point reading this as we already have it or have
3531 * decided to get it.
3536 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)))
3537 /* We need this block to directly satisfy a request */
3540 if (s
->syncing
|| s
->expanding
||
3541 (s
->replacing
&& want_replace(sh
, disk_idx
)))
3542 /* When syncing, or expanding we read everything.
3543 * When replacing, we need the replaced block.
3547 if ((s
->failed
>= 1 && fdev
[0]->toread
) ||
3548 (s
->failed
>= 2 && fdev
[1]->toread
))
3549 /* If we want to read from a failed device, then
3550 * we need to actually read every other device.
3554 /* Sometimes neither read-modify-write nor reconstruct-write
3555 * cycles can work. In those cases we read every block we
3556 * can. Then the parity-update is certain to have enough to
3558 * This can only be a problem when we need to write something,
3559 * and some device has failed. If either of those tests
3560 * fail we need look no further.
3562 if (!s
->failed
|| !s
->to_write
)
3565 if (test_bit(R5_Insync
, &dev
->flags
) &&
3566 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3567 /* Pre-reads at not permitted until after short delay
3568 * to gather multiple requests. However if this
3569 * device is no Insync, the block could only be computed
3570 * and there is no need to delay that.
3574 for (i
= 0; i
< s
->failed
&& i
< 2; i
++) {
3575 if (fdev
[i
]->towrite
&&
3576 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3577 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3578 /* If we have a partial write to a failed
3579 * device, then we will need to reconstruct
3580 * the content of that device, so all other
3581 * devices must be read.
3586 /* If we are forced to do a reconstruct-write, either because
3587 * the current RAID6 implementation only supports that, or
3588 * because parity cannot be trusted and we are currently
3589 * recovering it, there is extra need to be careful.
3590 * If one of the devices that we would need to read, because
3591 * it is not being overwritten (and maybe not written at all)
3592 * is missing/faulty, then we need to read everything we can.
3594 if (sh
->raid_conf
->level
!= 6 &&
3595 sh
->sector
< sh
->raid_conf
->mddev
->recovery_cp
)
3596 /* reconstruct-write isn't being forced */
3598 for (i
= 0; i
< s
->failed
&& i
< 2; i
++) {
3599 if (s
->failed_num
[i
] != sh
->pd_idx
&&
3600 s
->failed_num
[i
] != sh
->qd_idx
&&
3601 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3602 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3609 /* fetch_block - checks the given member device to see if its data needs
3610 * to be read or computed to satisfy a request.
3612 * Returns 1 when no more member devices need to be checked, otherwise returns
3613 * 0 to tell the loop in handle_stripe_fill to continue
3615 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3616 int disk_idx
, int disks
)
3618 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3620 /* is the data in this block needed, and can we get it? */
3621 if (need_this_block(sh
, s
, disk_idx
, disks
)) {
3622 /* we would like to get this block, possibly by computing it,
3623 * otherwise read it if the backing disk is insync
3625 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
3626 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
3627 BUG_ON(sh
->batch_head
);
3630 * In the raid6 case if the only non-uptodate disk is P
3631 * then we already trusted P to compute the other failed
3632 * drives. It is safe to compute rather than re-read P.
3633 * In other cases we only compute blocks from failed
3634 * devices, otherwise check/repair might fail to detect
3635 * a real inconsistency.
3638 if ((s
->uptodate
== disks
- 1) &&
3639 ((sh
->qd_idx
>= 0 && sh
->pd_idx
== disk_idx
) ||
3640 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
3641 disk_idx
== s
->failed_num
[1])))) {
3642 /* have disk failed, and we're requested to fetch it;
3645 pr_debug("Computing stripe %llu block %d\n",
3646 (unsigned long long)sh
->sector
, disk_idx
);
3647 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3648 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3649 set_bit(R5_Wantcompute
, &dev
->flags
);
3650 sh
->ops
.target
= disk_idx
;
3651 sh
->ops
.target2
= -1; /* no 2nd target */
3653 /* Careful: from this point on 'uptodate' is in the eye
3654 * of raid_run_ops which services 'compute' operations
3655 * before writes. R5_Wantcompute flags a block that will
3656 * be R5_UPTODATE by the time it is needed for a
3657 * subsequent operation.
3661 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
3662 /* Computing 2-failure is *very* expensive; only
3663 * do it if failed >= 2
3666 for (other
= disks
; other
--; ) {
3667 if (other
== disk_idx
)
3669 if (!test_bit(R5_UPTODATE
,
3670 &sh
->dev
[other
].flags
))
3674 pr_debug("Computing stripe %llu blocks %d,%d\n",
3675 (unsigned long long)sh
->sector
,
3677 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3678 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3679 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
3680 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
3681 sh
->ops
.target
= disk_idx
;
3682 sh
->ops
.target2
= other
;
3686 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
3687 set_bit(R5_LOCKED
, &dev
->flags
);
3688 set_bit(R5_Wantread
, &dev
->flags
);
3690 pr_debug("Reading block %d (sync=%d)\n",
3691 disk_idx
, s
->syncing
);
3699 * handle_stripe_fill - read or compute data to satisfy pending requests.
3701 static void handle_stripe_fill(struct stripe_head
*sh
,
3702 struct stripe_head_state
*s
,
3707 /* look for blocks to read/compute, skip this if a compute
3708 * is already in flight, or if the stripe contents are in the
3709 * midst of changing due to a write
3711 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
3712 !sh
->reconstruct_state
) {
3715 * For degraded stripe with data in journal, do not handle
3716 * read requests yet, instead, flush the stripe to raid
3717 * disks first, this avoids handling complex rmw of write
3718 * back cache (prexor with orig_page, and then xor with
3719 * page) in the read path
3721 if (s
->injournal
&& s
->failed
) {
3722 if (test_bit(STRIPE_R5C_CACHING
, &sh
->state
))
3723 r5c_make_stripe_write_out(sh
);
3727 for (i
= disks
; i
--; )
3728 if (fetch_block(sh
, s
, i
, disks
))
3732 set_bit(STRIPE_HANDLE
, &sh
->state
);
3735 static void break_stripe_batch_list(struct stripe_head
*head_sh
,
3736 unsigned long handle_flags
);
3737 /* handle_stripe_clean_event
3738 * any written block on an uptodate or failed drive can be returned.
3739 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3740 * never LOCKED, so we don't need to test 'failed' directly.
3742 static void handle_stripe_clean_event(struct r5conf
*conf
,
3743 struct stripe_head
*sh
, int disks
)
3747 int discard_pending
= 0;
3748 struct stripe_head
*head_sh
= sh
;
3749 bool do_endio
= false;
3751 for (i
= disks
; i
--; )
3752 if (sh
->dev
[i
].written
) {
3754 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
3755 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
3756 test_bit(R5_Discard
, &dev
->flags
) ||
3757 test_bit(R5_SkipCopy
, &dev
->flags
))) {
3758 /* We can return any write requests */
3759 struct bio
*wbi
, *wbi2
;
3760 pr_debug("Return write for disc %d\n", i
);
3761 if (test_and_clear_bit(R5_Discard
, &dev
->flags
))
3762 clear_bit(R5_UPTODATE
, &dev
->flags
);
3763 if (test_and_clear_bit(R5_SkipCopy
, &dev
->flags
)) {
3764 WARN_ON(test_bit(R5_UPTODATE
, &dev
->flags
));
3769 dev
->page
= dev
->orig_page
;
3771 dev
->written
= NULL
;
3772 while (wbi
&& wbi
->bi_iter
.bi_sector
<
3773 dev
->sector
+ STRIPE_SECTORS
) {
3774 wbi2
= r5_next_bio(wbi
, dev
->sector
);
3775 md_write_end(conf
->mddev
);
3779 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3781 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
3783 if (head_sh
->batch_head
) {
3784 sh
= list_first_entry(&sh
->batch_list
,
3787 if (sh
!= head_sh
) {
3794 } else if (test_bit(R5_Discard
, &dev
->flags
))
3795 discard_pending
= 1;
3798 log_stripe_write_finished(sh
);
3800 if (!discard_pending
&&
3801 test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
)) {
3803 clear_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
3804 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
3805 if (sh
->qd_idx
>= 0) {
3806 clear_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
3807 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
);
3809 /* now that discard is done we can proceed with any sync */
3810 clear_bit(STRIPE_DISCARD
, &sh
->state
);
3812 * SCSI discard will change some bio fields and the stripe has
3813 * no updated data, so remove it from hash list and the stripe
3814 * will be reinitialized
3817 hash
= sh
->hash_lock_index
;
3818 spin_lock_irq(conf
->hash_locks
+ hash
);
3820 spin_unlock_irq(conf
->hash_locks
+ hash
);
3821 if (head_sh
->batch_head
) {
3822 sh
= list_first_entry(&sh
->batch_list
,
3823 struct stripe_head
, batch_list
);
3829 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
))
3830 set_bit(STRIPE_HANDLE
, &sh
->state
);
3834 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3835 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3836 md_wakeup_thread(conf
->mddev
->thread
);
3838 if (head_sh
->batch_head
&& do_endio
)
3839 break_stripe_batch_list(head_sh
, STRIPE_EXPAND_SYNC_FLAGS
);
3843 * For RMW in write back cache, we need extra page in prexor to store the
3844 * old data. This page is stored in dev->orig_page.
3846 * This function checks whether we have data for prexor. The exact logic
3848 * R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3850 static inline bool uptodate_for_rmw(struct r5dev
*dev
)
3852 return (test_bit(R5_UPTODATE
, &dev
->flags
)) &&
3853 (!test_bit(R5_InJournal
, &dev
->flags
) ||
3854 test_bit(R5_OrigPageUPTDODATE
, &dev
->flags
));
3857 static int handle_stripe_dirtying(struct r5conf
*conf
,
3858 struct stripe_head
*sh
,
3859 struct stripe_head_state
*s
,
3862 int rmw
= 0, rcw
= 0, i
;
3863 sector_t recovery_cp
= conf
->mddev
->recovery_cp
;
3865 /* Check whether resync is now happening or should start.
3866 * If yes, then the array is dirty (after unclean shutdown or
3867 * initial creation), so parity in some stripes might be inconsistent.
3868 * In this case, we need to always do reconstruct-write, to ensure
3869 * that in case of drive failure or read-error correction, we
3870 * generate correct data from the parity.
3872 if (conf
->rmw_level
== PARITY_DISABLE_RMW
||
3873 (recovery_cp
< MaxSector
&& sh
->sector
>= recovery_cp
&&
3875 /* Calculate the real rcw later - for now make it
3876 * look like rcw is cheaper
3879 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3880 conf
->rmw_level
, (unsigned long long)recovery_cp
,
3881 (unsigned long long)sh
->sector
);
3882 } else for (i
= disks
; i
--; ) {
3883 /* would I have to read this buffer for read_modify_write */
3884 struct r5dev
*dev
= &sh
->dev
[i
];
3885 if (((dev
->towrite
&& !delay_towrite(conf
, dev
, s
)) ||
3886 i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3887 test_bit(R5_InJournal
, &dev
->flags
)) &&
3888 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3889 !(uptodate_for_rmw(dev
) ||
3890 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3891 if (test_bit(R5_Insync
, &dev
->flags
))
3894 rmw
+= 2*disks
; /* cannot read it */
3896 /* Would I have to read this buffer for reconstruct_write */
3897 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3898 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3899 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3900 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3901 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3902 if (test_bit(R5_Insync
, &dev
->flags
))
3909 pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3910 (unsigned long long)sh
->sector
, sh
->state
, rmw
, rcw
);
3911 set_bit(STRIPE_HANDLE
, &sh
->state
);
3912 if ((rmw
< rcw
|| (rmw
== rcw
&& conf
->rmw_level
== PARITY_PREFER_RMW
)) && rmw
> 0) {
3913 /* prefer read-modify-write, but need to get some data */
3914 if (conf
->mddev
->queue
)
3915 blk_add_trace_msg(conf
->mddev
->queue
,
3916 "raid5 rmw %llu %d",
3917 (unsigned long long)sh
->sector
, rmw
);
3918 for (i
= disks
; i
--; ) {
3919 struct r5dev
*dev
= &sh
->dev
[i
];
3920 if (test_bit(R5_InJournal
, &dev
->flags
) &&
3921 dev
->page
== dev
->orig_page
&&
3922 !test_bit(R5_LOCKED
, &sh
->dev
[sh
->pd_idx
].flags
)) {
3923 /* alloc page for prexor */
3924 struct page
*p
= alloc_page(GFP_NOIO
);
3932 * alloc_page() failed, try use
3933 * disk_info->extra_page
3935 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE
,
3936 &conf
->cache_state
)) {
3937 r5c_use_extra_page(sh
);
3941 /* extra_page in use, add to delayed_list */
3942 set_bit(STRIPE_DELAYED
, &sh
->state
);
3943 s
->waiting_extra_page
= 1;
3948 for (i
= disks
; i
--; ) {
3949 struct r5dev
*dev
= &sh
->dev
[i
];
3950 if (((dev
->towrite
&& !delay_towrite(conf
, dev
, s
)) ||
3951 i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3952 test_bit(R5_InJournal
, &dev
->flags
)) &&
3953 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3954 !(uptodate_for_rmw(dev
) ||
3955 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
3956 test_bit(R5_Insync
, &dev
->flags
)) {
3957 if (test_bit(STRIPE_PREREAD_ACTIVE
,
3959 pr_debug("Read_old block %d for r-m-w\n",
3961 set_bit(R5_LOCKED
, &dev
->flags
);
3962 set_bit(R5_Wantread
, &dev
->flags
);
3965 set_bit(STRIPE_DELAYED
, &sh
->state
);
3966 set_bit(STRIPE_HANDLE
, &sh
->state
);
3971 if ((rcw
< rmw
|| (rcw
== rmw
&& conf
->rmw_level
!= PARITY_PREFER_RMW
)) && rcw
> 0) {
3972 /* want reconstruct write, but need to get some data */
3975 for (i
= disks
; i
--; ) {
3976 struct r5dev
*dev
= &sh
->dev
[i
];
3977 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3978 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
3979 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3980 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
3981 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3983 if (test_bit(R5_Insync
, &dev
->flags
) &&
3984 test_bit(STRIPE_PREREAD_ACTIVE
,
3986 pr_debug("Read_old block "
3987 "%d for Reconstruct\n", i
);
3988 set_bit(R5_LOCKED
, &dev
->flags
);
3989 set_bit(R5_Wantread
, &dev
->flags
);
3993 set_bit(STRIPE_DELAYED
, &sh
->state
);
3994 set_bit(STRIPE_HANDLE
, &sh
->state
);
3998 if (rcw
&& conf
->mddev
->queue
)
3999 blk_add_trace_msg(conf
->mddev
->queue
, "raid5 rcw %llu %d %d %d",
4000 (unsigned long long)sh
->sector
,
4001 rcw
, qread
, test_bit(STRIPE_DELAYED
, &sh
->state
));
4004 if (rcw
> disks
&& rmw
> disks
&&
4005 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4006 set_bit(STRIPE_DELAYED
, &sh
->state
);
4008 /* now if nothing is locked, and if we have enough data,
4009 * we can start a write request
4011 /* since handle_stripe can be called at any time we need to handle the
4012 * case where a compute block operation has been submitted and then a
4013 * subsequent call wants to start a write request. raid_run_ops only
4014 * handles the case where compute block and reconstruct are requested
4015 * simultaneously. If this is not the case then new writes need to be
4016 * held off until the compute completes.
4018 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
4019 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
4020 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
4021 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
4025 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
4026 struct stripe_head_state
*s
, int disks
)
4028 struct r5dev
*dev
= NULL
;
4030 BUG_ON(sh
->batch_head
);
4031 set_bit(STRIPE_HANDLE
, &sh
->state
);
4033 switch (sh
->check_state
) {
4034 case check_state_idle
:
4035 /* start a new check operation if there are no failures */
4036 if (s
->failed
== 0) {
4037 BUG_ON(s
->uptodate
!= disks
);
4038 sh
->check_state
= check_state_run
;
4039 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
4040 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
4044 dev
= &sh
->dev
[s
->failed_num
[0]];
4046 case check_state_compute_result
:
4047 sh
->check_state
= check_state_idle
;
4049 dev
= &sh
->dev
[sh
->pd_idx
];
4051 /* check that a write has not made the stripe insync */
4052 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
4055 /* either failed parity check, or recovery is happening */
4056 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
4057 BUG_ON(s
->uptodate
!= disks
);
4059 set_bit(R5_LOCKED
, &dev
->flags
);
4061 set_bit(R5_Wantwrite
, &dev
->flags
);
4063 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
4064 set_bit(STRIPE_INSYNC
, &sh
->state
);
4066 case check_state_run
:
4067 break; /* we will be called again upon completion */
4068 case check_state_check_result
:
4069 sh
->check_state
= check_state_idle
;
4071 /* if a failure occurred during the check operation, leave
4072 * STRIPE_INSYNC not set and let the stripe be handled again
4077 /* handle a successful check operation, if parity is correct
4078 * we are done. Otherwise update the mismatch count and repair
4079 * parity if !MD_RECOVERY_CHECK
4081 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
4082 /* parity is correct (on disc,
4083 * not in buffer any more)
4085 set_bit(STRIPE_INSYNC
, &sh
->state
);
4087 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
4088 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
4089 /* don't try to repair!! */
4090 set_bit(STRIPE_INSYNC
, &sh
->state
);
4092 sh
->check_state
= check_state_compute_run
;
4093 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
4094 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
4095 set_bit(R5_Wantcompute
,
4096 &sh
->dev
[sh
->pd_idx
].flags
);
4097 sh
->ops
.target
= sh
->pd_idx
;
4098 sh
->ops
.target2
= -1;
4103 case check_state_compute_run
:
4106 pr_err("%s: unknown check_state: %d sector: %llu\n",
4107 __func__
, sh
->check_state
,
4108 (unsigned long long) sh
->sector
);
4113 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
4114 struct stripe_head_state
*s
,
4117 int pd_idx
= sh
->pd_idx
;
4118 int qd_idx
= sh
->qd_idx
;
4121 BUG_ON(sh
->batch_head
);
4122 set_bit(STRIPE_HANDLE
, &sh
->state
);
4124 BUG_ON(s
->failed
> 2);
4126 /* Want to check and possibly repair P and Q.
4127 * However there could be one 'failed' device, in which
4128 * case we can only check one of them, possibly using the
4129 * other to generate missing data
4132 switch (sh
->check_state
) {
4133 case check_state_idle
:
4134 /* start a new check operation if there are < 2 failures */
4135 if (s
->failed
== s
->q_failed
) {
4136 /* The only possible failed device holds Q, so it
4137 * makes sense to check P (If anything else were failed,
4138 * we would have used P to recreate it).
4140 sh
->check_state
= check_state_run
;
4142 if (!s
->q_failed
&& s
->failed
< 2) {
4143 /* Q is not failed, and we didn't use it to generate
4144 * anything, so it makes sense to check it
4146 if (sh
->check_state
== check_state_run
)
4147 sh
->check_state
= check_state_run_pq
;
4149 sh
->check_state
= check_state_run_q
;
4152 /* discard potentially stale zero_sum_result */
4153 sh
->ops
.zero_sum_result
= 0;
4155 if (sh
->check_state
== check_state_run
) {
4156 /* async_xor_zero_sum destroys the contents of P */
4157 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
4160 if (sh
->check_state
>= check_state_run
&&
4161 sh
->check_state
<= check_state_run_pq
) {
4162 /* async_syndrome_zero_sum preserves P and Q, so
4163 * no need to mark them !uptodate here
4165 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
4169 /* we have 2-disk failure */
4170 BUG_ON(s
->failed
!= 2);
4172 case check_state_compute_result
:
4173 sh
->check_state
= check_state_idle
;
4175 /* check that a write has not made the stripe insync */
4176 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
4179 /* now write out any block on a failed drive,
4180 * or P or Q if they were recomputed
4182 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
4183 if (s
->failed
== 2) {
4184 dev
= &sh
->dev
[s
->failed_num
[1]];
4186 set_bit(R5_LOCKED
, &dev
->flags
);
4187 set_bit(R5_Wantwrite
, &dev
->flags
);
4189 if (s
->failed
>= 1) {
4190 dev
= &sh
->dev
[s
->failed_num
[0]];
4192 set_bit(R5_LOCKED
, &dev
->flags
);
4193 set_bit(R5_Wantwrite
, &dev
->flags
);
4195 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
4196 dev
= &sh
->dev
[pd_idx
];
4198 set_bit(R5_LOCKED
, &dev
->flags
);
4199 set_bit(R5_Wantwrite
, &dev
->flags
);
4201 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
4202 dev
= &sh
->dev
[qd_idx
];
4204 set_bit(R5_LOCKED
, &dev
->flags
);
4205 set_bit(R5_Wantwrite
, &dev
->flags
);
4207 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
4209 set_bit(STRIPE_INSYNC
, &sh
->state
);
4211 case check_state_run
:
4212 case check_state_run_q
:
4213 case check_state_run_pq
:
4214 break; /* we will be called again upon completion */
4215 case check_state_check_result
:
4216 sh
->check_state
= check_state_idle
;
4218 /* handle a successful check operation, if parity is correct
4219 * we are done. Otherwise update the mismatch count and repair
4220 * parity if !MD_RECOVERY_CHECK
4222 if (sh
->ops
.zero_sum_result
== 0) {
4223 /* both parities are correct */
4225 set_bit(STRIPE_INSYNC
, &sh
->state
);
4227 /* in contrast to the raid5 case we can validate
4228 * parity, but still have a failure to write
4231 sh
->check_state
= check_state_compute_result
;
4232 /* Returning at this point means that we may go
4233 * off and bring p and/or q uptodate again so
4234 * we make sure to check zero_sum_result again
4235 * to verify if p or q need writeback
4239 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
4240 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
4241 /* don't try to repair!! */
4242 set_bit(STRIPE_INSYNC
, &sh
->state
);
4244 int *target
= &sh
->ops
.target
;
4246 sh
->ops
.target
= -1;
4247 sh
->ops
.target2
= -1;
4248 sh
->check_state
= check_state_compute_run
;
4249 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
4250 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
4251 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
4252 set_bit(R5_Wantcompute
,
4253 &sh
->dev
[pd_idx
].flags
);
4255 target
= &sh
->ops
.target2
;
4258 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
4259 set_bit(R5_Wantcompute
,
4260 &sh
->dev
[qd_idx
].flags
);
4267 case check_state_compute_run
:
4270 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4271 __func__
, sh
->check_state
,
4272 (unsigned long long) sh
->sector
);
4277 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
4281 /* We have read all the blocks in this stripe and now we need to
4282 * copy some of them into a target stripe for expand.
4284 struct dma_async_tx_descriptor
*tx
= NULL
;
4285 BUG_ON(sh
->batch_head
);
4286 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
4287 for (i
= 0; i
< sh
->disks
; i
++)
4288 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
4290 struct stripe_head
*sh2
;
4291 struct async_submit_ctl submit
;
4293 sector_t bn
= raid5_compute_blocknr(sh
, i
, 1);
4294 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
4296 sh2
= raid5_get_active_stripe(conf
, s
, 0, 1, 1);
4298 /* so far only the early blocks of this stripe
4299 * have been requested. When later blocks
4300 * get requested, we will try again
4303 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
4304 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
4305 /* must have already done this block */
4306 raid5_release_stripe(sh2
);
4310 /* place all the copies on one channel */
4311 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
4312 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
4313 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
4316 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
4317 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
4318 for (j
= 0; j
< conf
->raid_disks
; j
++)
4319 if (j
!= sh2
->pd_idx
&&
4321 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
4323 if (j
== conf
->raid_disks
) {
4324 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
4325 set_bit(STRIPE_HANDLE
, &sh2
->state
);
4327 raid5_release_stripe(sh2
);
4330 /* done submitting copies, wait for them to complete */
4331 async_tx_quiesce(&tx
);
4335 * handle_stripe - do things to a stripe.
4337 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4338 * state of various bits to see what needs to be done.
4340 * return some read requests which now have data
4341 * return some write requests which are safely on storage
4342 * schedule a read on some buffers
4343 * schedule a write of some buffers
4344 * return confirmation of parity correctness
4348 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
4350 struct r5conf
*conf
= sh
->raid_conf
;
4351 int disks
= sh
->disks
;
4354 int do_recovery
= 0;
4356 memset(s
, 0, sizeof(*s
));
4358 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
) && !sh
->batch_head
;
4359 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
) && !sh
->batch_head
;
4360 s
->failed_num
[0] = -1;
4361 s
->failed_num
[1] = -1;
4362 s
->log_failed
= r5l_log_disk_error(conf
);
4364 /* Now to look around and see what can be done */
4366 for (i
=disks
; i
--; ) {
4367 struct md_rdev
*rdev
;
4374 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4376 dev
->toread
, dev
->towrite
, dev
->written
);
4377 /* maybe we can reply to a read
4379 * new wantfill requests are only permitted while
4380 * ops_complete_biofill is guaranteed to be inactive
4382 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
4383 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
4384 set_bit(R5_Wantfill
, &dev
->flags
);
4386 /* now count some things */
4387 if (test_bit(R5_LOCKED
, &dev
->flags
))
4389 if (test_bit(R5_UPTODATE
, &dev
->flags
))
4391 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
4393 BUG_ON(s
->compute
> 2);
4396 if (test_bit(R5_Wantfill
, &dev
->flags
))
4398 else if (dev
->toread
)
4402 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
4407 /* Prefer to use the replacement for reads, but only
4408 * if it is recovered enough and has no bad blocks.
4410 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
4411 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
4412 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
4413 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4414 &first_bad
, &bad_sectors
))
4415 set_bit(R5_ReadRepl
, &dev
->flags
);
4417 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4418 set_bit(R5_NeedReplace
, &dev
->flags
);
4420 clear_bit(R5_NeedReplace
, &dev
->flags
);
4421 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
4422 clear_bit(R5_ReadRepl
, &dev
->flags
);
4424 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
4427 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
4428 &first_bad
, &bad_sectors
);
4429 if (s
->blocked_rdev
== NULL
4430 && (test_bit(Blocked
, &rdev
->flags
)
4433 set_bit(BlockedBadBlocks
,
4435 s
->blocked_rdev
= rdev
;
4436 atomic_inc(&rdev
->nr_pending
);
4439 clear_bit(R5_Insync
, &dev
->flags
);
4443 /* also not in-sync */
4444 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
4445 test_bit(R5_UPTODATE
, &dev
->flags
)) {
4446 /* treat as in-sync, but with a read error
4447 * which we can now try to correct
4449 set_bit(R5_Insync
, &dev
->flags
);
4450 set_bit(R5_ReadError
, &dev
->flags
);
4452 } else if (test_bit(In_sync
, &rdev
->flags
))
4453 set_bit(R5_Insync
, &dev
->flags
);
4454 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
4455 /* in sync if before recovery_offset */
4456 set_bit(R5_Insync
, &dev
->flags
);
4457 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
4458 test_bit(R5_Expanded
, &dev
->flags
))
4459 /* If we've reshaped into here, we assume it is Insync.
4460 * We will shortly update recovery_offset to make
4463 set_bit(R5_Insync
, &dev
->flags
);
4465 if (test_bit(R5_WriteError
, &dev
->flags
)) {
4466 /* This flag does not apply to '.replacement'
4467 * only to .rdev, so make sure to check that*/
4468 struct md_rdev
*rdev2
= rcu_dereference(
4469 conf
->disks
[i
].rdev
);
4471 clear_bit(R5_Insync
, &dev
->flags
);
4472 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4473 s
->handle_bad_blocks
= 1;
4474 atomic_inc(&rdev2
->nr_pending
);
4476 clear_bit(R5_WriteError
, &dev
->flags
);
4478 if (test_bit(R5_MadeGood
, &dev
->flags
)) {
4479 /* This flag does not apply to '.replacement'
4480 * only to .rdev, so make sure to check that*/
4481 struct md_rdev
*rdev2
= rcu_dereference(
4482 conf
->disks
[i
].rdev
);
4483 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4484 s
->handle_bad_blocks
= 1;
4485 atomic_inc(&rdev2
->nr_pending
);
4487 clear_bit(R5_MadeGood
, &dev
->flags
);
4489 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
4490 struct md_rdev
*rdev2
= rcu_dereference(
4491 conf
->disks
[i
].replacement
);
4492 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4493 s
->handle_bad_blocks
= 1;
4494 atomic_inc(&rdev2
->nr_pending
);
4496 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
4498 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4499 /* The ReadError flag will just be confusing now */
4500 clear_bit(R5_ReadError
, &dev
->flags
);
4501 clear_bit(R5_ReWrite
, &dev
->flags
);
4503 if (test_bit(R5_ReadError
, &dev
->flags
))
4504 clear_bit(R5_Insync
, &dev
->flags
);
4505 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4507 s
->failed_num
[s
->failed
] = i
;
4509 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4513 if (test_bit(R5_InJournal
, &dev
->flags
))
4515 if (test_bit(R5_InJournal
, &dev
->flags
) && dev
->written
)
4518 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
4519 /* If there is a failed device being replaced,
4520 * we must be recovering.
4521 * else if we are after recovery_cp, we must be syncing
4522 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4523 * else we can only be replacing
4524 * sync and recovery both need to read all devices, and so
4525 * use the same flag.
4528 sh
->sector
>= conf
->mddev
->recovery_cp
||
4529 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
4537 static int clear_batch_ready(struct stripe_head
*sh
)
4539 /* Return '1' if this is a member of batch, or
4540 * '0' if it is a lone stripe or a head which can now be
4543 struct stripe_head
*tmp
;
4544 if (!test_and_clear_bit(STRIPE_BATCH_READY
, &sh
->state
))
4545 return (sh
->batch_head
&& sh
->batch_head
!= sh
);
4546 spin_lock(&sh
->stripe_lock
);
4547 if (!sh
->batch_head
) {
4548 spin_unlock(&sh
->stripe_lock
);
4553 * this stripe could be added to a batch list before we check
4554 * BATCH_READY, skips it
4556 if (sh
->batch_head
!= sh
) {
4557 spin_unlock(&sh
->stripe_lock
);
4560 spin_lock(&sh
->batch_lock
);
4561 list_for_each_entry(tmp
, &sh
->batch_list
, batch_list
)
4562 clear_bit(STRIPE_BATCH_READY
, &tmp
->state
);
4563 spin_unlock(&sh
->batch_lock
);
4564 spin_unlock(&sh
->stripe_lock
);
4567 * BATCH_READY is cleared, no new stripes can be added.
4568 * batch_list can be accessed without lock
4573 static void break_stripe_batch_list(struct stripe_head
*head_sh
,
4574 unsigned long handle_flags
)
4576 struct stripe_head
*sh
, *next
;
4580 list_for_each_entry_safe(sh
, next
, &head_sh
->batch_list
, batch_list
) {
4582 list_del_init(&sh
->batch_list
);
4584 WARN_ONCE(sh
->state
& ((1 << STRIPE_ACTIVE
) |
4585 (1 << STRIPE_SYNCING
) |
4586 (1 << STRIPE_REPLACED
) |
4587 (1 << STRIPE_DELAYED
) |
4588 (1 << STRIPE_BIT_DELAY
) |
4589 (1 << STRIPE_FULL_WRITE
) |
4590 (1 << STRIPE_BIOFILL_RUN
) |
4591 (1 << STRIPE_COMPUTE_RUN
) |
4592 (1 << STRIPE_OPS_REQ_PENDING
) |
4593 (1 << STRIPE_DISCARD
) |
4594 (1 << STRIPE_BATCH_READY
) |
4595 (1 << STRIPE_BATCH_ERR
) |
4596 (1 << STRIPE_BITMAP_PENDING
)),
4597 "stripe state: %lx\n", sh
->state
);
4598 WARN_ONCE(head_sh
->state
& ((1 << STRIPE_DISCARD
) |
4599 (1 << STRIPE_REPLACED
)),
4600 "head stripe state: %lx\n", head_sh
->state
);
4602 set_mask_bits(&sh
->state
, ~(STRIPE_EXPAND_SYNC_FLAGS
|
4603 (1 << STRIPE_PREREAD_ACTIVE
) |
4604 (1 << STRIPE_DEGRADED
)),
4605 head_sh
->state
& (1 << STRIPE_INSYNC
));
4607 sh
->check_state
= head_sh
->check_state
;
4608 sh
->reconstruct_state
= head_sh
->reconstruct_state
;
4609 for (i
= 0; i
< sh
->disks
; i
++) {
4610 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
4612 sh
->dev
[i
].flags
= head_sh
->dev
[i
].flags
&
4613 (~((1 << R5_WriteError
) | (1 << R5_Overlap
)));
4615 spin_lock_irq(&sh
->stripe_lock
);
4616 sh
->batch_head
= NULL
;
4617 spin_unlock_irq(&sh
->stripe_lock
);
4618 if (handle_flags
== 0 ||
4619 sh
->state
& handle_flags
)
4620 set_bit(STRIPE_HANDLE
, &sh
->state
);
4621 raid5_release_stripe(sh
);
4623 spin_lock_irq(&head_sh
->stripe_lock
);
4624 head_sh
->batch_head
= NULL
;
4625 spin_unlock_irq(&head_sh
->stripe_lock
);
4626 for (i
= 0; i
< head_sh
->disks
; i
++)
4627 if (test_and_clear_bit(R5_Overlap
, &head_sh
->dev
[i
].flags
))
4629 if (head_sh
->state
& handle_flags
)
4630 set_bit(STRIPE_HANDLE
, &head_sh
->state
);
4633 wake_up(&head_sh
->raid_conf
->wait_for_overlap
);
4636 static void handle_stripe(struct stripe_head
*sh
)
4638 struct stripe_head_state s
;
4639 struct r5conf
*conf
= sh
->raid_conf
;
4642 int disks
= sh
->disks
;
4643 struct r5dev
*pdev
, *qdev
;
4645 clear_bit(STRIPE_HANDLE
, &sh
->state
);
4646 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
4647 /* already being handled, ensure it gets handled
4648 * again when current action finishes */
4649 set_bit(STRIPE_HANDLE
, &sh
->state
);
4653 if (clear_batch_ready(sh
) ) {
4654 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
4658 if (test_and_clear_bit(STRIPE_BATCH_ERR
, &sh
->state
))
4659 break_stripe_batch_list(sh
, 0);
4661 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
) && !sh
->batch_head
) {
4662 spin_lock(&sh
->stripe_lock
);
4664 * Cannot process 'sync' concurrently with 'discard'.
4665 * Flush data in r5cache before 'sync'.
4667 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE
, &sh
->state
) &&
4668 !test_bit(STRIPE_R5C_FULL_STRIPE
, &sh
->state
) &&
4669 !test_bit(STRIPE_DISCARD
, &sh
->state
) &&
4670 test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
4671 set_bit(STRIPE_SYNCING
, &sh
->state
);
4672 clear_bit(STRIPE_INSYNC
, &sh
->state
);
4673 clear_bit(STRIPE_REPLACED
, &sh
->state
);
4675 spin_unlock(&sh
->stripe_lock
);
4677 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4679 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4680 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4681 (unsigned long long)sh
->sector
, sh
->state
,
4682 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
4683 sh
->check_state
, sh
->reconstruct_state
);
4685 analyse_stripe(sh
, &s
);
4687 if (test_bit(STRIPE_LOG_TRAPPED
, &sh
->state
))
4690 if (s
.handle_bad_blocks
||
4691 test_bit(MD_SB_CHANGE_PENDING
, &conf
->mddev
->sb_flags
)) {
4692 set_bit(STRIPE_HANDLE
, &sh
->state
);
4696 if (unlikely(s
.blocked_rdev
)) {
4697 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
4698 s
.replacing
|| s
.to_write
|| s
.written
) {
4699 set_bit(STRIPE_HANDLE
, &sh
->state
);
4702 /* There is nothing for the blocked_rdev to block */
4703 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
4704 s
.blocked_rdev
= NULL
;
4707 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
4708 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
4709 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
4712 pr_debug("locked=%d uptodate=%d to_read=%d"
4713 " to_write=%d failed=%d failed_num=%d,%d\n",
4714 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
4715 s
.failed_num
[0], s
.failed_num
[1]);
4717 * check if the array has lost more than max_degraded devices and,
4718 * if so, some requests might need to be failed.
4720 * When journal device failed (log_failed), we will only process
4721 * the stripe if there is data need write to raid disks
4723 if (s
.failed
> conf
->max_degraded
||
4724 (s
.log_failed
&& s
.injournal
== 0)) {
4725 sh
->check_state
= 0;
4726 sh
->reconstruct_state
= 0;
4727 break_stripe_batch_list(sh
, 0);
4728 if (s
.to_read
+s
.to_write
+s
.written
)
4729 handle_failed_stripe(conf
, sh
, &s
, disks
);
4730 if (s
.syncing
+ s
.replacing
)
4731 handle_failed_sync(conf
, sh
, &s
);
4734 /* Now we check to see if any write operations have recently
4738 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
4740 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
4741 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
4742 sh
->reconstruct_state
= reconstruct_state_idle
;
4744 /* All the 'written' buffers and the parity block are ready to
4745 * be written back to disk
4747 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
) &&
4748 !test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
));
4749 BUG_ON(sh
->qd_idx
>= 0 &&
4750 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
) &&
4751 !test_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
));
4752 for (i
= disks
; i
--; ) {
4753 struct r5dev
*dev
= &sh
->dev
[i
];
4754 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
4755 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
4756 dev
->written
|| test_bit(R5_InJournal
,
4758 pr_debug("Writing block %d\n", i
);
4759 set_bit(R5_Wantwrite
, &dev
->flags
);
4764 if (!test_bit(R5_Insync
, &dev
->flags
) ||
4765 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
4767 set_bit(STRIPE_INSYNC
, &sh
->state
);
4770 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4771 s
.dec_preread_active
= 1;
4775 * might be able to return some write requests if the parity blocks
4776 * are safe, or on a failed drive
4778 pdev
= &sh
->dev
[sh
->pd_idx
];
4779 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
4780 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
4781 qdev
= &sh
->dev
[sh
->qd_idx
];
4782 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
4783 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
4787 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
4788 && !test_bit(R5_LOCKED
, &pdev
->flags
)
4789 && (test_bit(R5_UPTODATE
, &pdev
->flags
) ||
4790 test_bit(R5_Discard
, &pdev
->flags
))))) &&
4791 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
4792 && !test_bit(R5_LOCKED
, &qdev
->flags
)
4793 && (test_bit(R5_UPTODATE
, &qdev
->flags
) ||
4794 test_bit(R5_Discard
, &qdev
->flags
))))))
4795 handle_stripe_clean_event(conf
, sh
, disks
);
4798 r5c_handle_cached_data_endio(conf
, sh
, disks
);
4799 log_stripe_write_finished(sh
);
4801 /* Now we might consider reading some blocks, either to check/generate
4802 * parity, or to satisfy requests
4803 * or to load a block that is being partially written.
4805 if (s
.to_read
|| s
.non_overwrite
4806 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
4807 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
4810 handle_stripe_fill(sh
, &s
, disks
);
4813 * When the stripe finishes full journal write cycle (write to journal
4814 * and raid disk), this is the clean up procedure so it is ready for
4817 r5c_finish_stripe_write_out(conf
, sh
, &s
);
4820 * Now to consider new write requests, cache write back and what else,
4821 * if anything should be read. We do not handle new writes when:
4822 * 1/ A 'write' operation (copy+xor) is already in flight.
4823 * 2/ A 'check' operation is in flight, as it may clobber the parity
4825 * 3/ A r5c cache log write is in flight.
4828 if (!sh
->reconstruct_state
&& !sh
->check_state
&& !sh
->log_io
) {
4829 if (!r5c_is_writeback(conf
->log
)) {
4831 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
4832 } else { /* write back cache */
4835 /* First, try handle writes in caching phase */
4837 ret
= r5c_try_caching_write(conf
, sh
, &s
,
4840 * If caching phase failed: ret == -EAGAIN
4842 * stripe under reclaim: !caching && injournal
4844 * fall back to handle_stripe_dirtying()
4846 if (ret
== -EAGAIN
||
4847 /* stripe under reclaim: !caching && injournal */
4848 (!test_bit(STRIPE_R5C_CACHING
, &sh
->state
) &&
4850 ret
= handle_stripe_dirtying(conf
, sh
, &s
,
4858 /* maybe we need to check and possibly fix the parity for this stripe
4859 * Any reads will already have been scheduled, so we just see if enough
4860 * data is available. The parity check is held off while parity
4861 * dependent operations are in flight.
4863 if (sh
->check_state
||
4864 (s
.syncing
&& s
.locked
== 0 &&
4865 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4866 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
4867 if (conf
->level
== 6)
4868 handle_parity_checks6(conf
, sh
, &s
, disks
);
4870 handle_parity_checks5(conf
, sh
, &s
, disks
);
4873 if ((s
.replacing
|| s
.syncing
) && s
.locked
== 0
4874 && !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)
4875 && !test_bit(STRIPE_REPLACED
, &sh
->state
)) {
4876 /* Write out to replacement devices where possible */
4877 for (i
= 0; i
< conf
->raid_disks
; i
++)
4878 if (test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
4879 WARN_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
4880 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
4881 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4885 set_bit(STRIPE_INSYNC
, &sh
->state
);
4886 set_bit(STRIPE_REPLACED
, &sh
->state
);
4888 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
4889 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
4890 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
4891 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4892 clear_bit(STRIPE_SYNCING
, &sh
->state
);
4893 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
4894 wake_up(&conf
->wait_for_overlap
);
4897 /* If the failed drives are just a ReadError, then we might need
4898 * to progress the repair/check process
4900 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
4901 for (i
= 0; i
< s
.failed
; i
++) {
4902 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
4903 if (test_bit(R5_ReadError
, &dev
->flags
)
4904 && !test_bit(R5_LOCKED
, &dev
->flags
)
4905 && test_bit(R5_UPTODATE
, &dev
->flags
)
4907 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
4908 set_bit(R5_Wantwrite
, &dev
->flags
);
4909 set_bit(R5_ReWrite
, &dev
->flags
);
4910 set_bit(R5_LOCKED
, &dev
->flags
);
4913 /* let's read it back */
4914 set_bit(R5_Wantread
, &dev
->flags
);
4915 set_bit(R5_LOCKED
, &dev
->flags
);
4921 /* Finish reconstruct operations initiated by the expansion process */
4922 if (sh
->reconstruct_state
== reconstruct_state_result
) {
4923 struct stripe_head
*sh_src
4924 = raid5_get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
4925 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
4926 /* sh cannot be written until sh_src has been read.
4927 * so arrange for sh to be delayed a little
4929 set_bit(STRIPE_DELAYED
, &sh
->state
);
4930 set_bit(STRIPE_HANDLE
, &sh
->state
);
4931 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
4933 atomic_inc(&conf
->preread_active_stripes
);
4934 raid5_release_stripe(sh_src
);
4938 raid5_release_stripe(sh_src
);
4940 sh
->reconstruct_state
= reconstruct_state_idle
;
4941 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
4942 for (i
= conf
->raid_disks
; i
--; ) {
4943 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
4944 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
4949 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
4950 !sh
->reconstruct_state
) {
4951 /* Need to write out all blocks after computing parity */
4952 sh
->disks
= conf
->raid_disks
;
4953 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
4954 schedule_reconstruction(sh
, &s
, 1, 1);
4955 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
4956 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4957 atomic_dec(&conf
->reshape_stripes
);
4958 wake_up(&conf
->wait_for_overlap
);
4959 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
4962 if (s
.expanding
&& s
.locked
== 0 &&
4963 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
4964 handle_stripe_expansion(conf
, sh
);
4967 /* wait for this device to become unblocked */
4968 if (unlikely(s
.blocked_rdev
)) {
4969 if (conf
->mddev
->external
)
4970 md_wait_for_blocked_rdev(s
.blocked_rdev
,
4973 /* Internal metadata will immediately
4974 * be written by raid5d, so we don't
4975 * need to wait here.
4977 rdev_dec_pending(s
.blocked_rdev
,
4981 if (s
.handle_bad_blocks
)
4982 for (i
= disks
; i
--; ) {
4983 struct md_rdev
*rdev
;
4984 struct r5dev
*dev
= &sh
->dev
[i
];
4985 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
4986 /* We own a safe reference to the rdev */
4987 rdev
= conf
->disks
[i
].rdev
;
4988 if (!rdev_set_badblocks(rdev
, sh
->sector
,
4990 md_error(conf
->mddev
, rdev
);
4991 rdev_dec_pending(rdev
, conf
->mddev
);
4993 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
4994 rdev
= conf
->disks
[i
].rdev
;
4995 rdev_clear_badblocks(rdev
, sh
->sector
,
4997 rdev_dec_pending(rdev
, conf
->mddev
);
4999 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
5000 rdev
= conf
->disks
[i
].replacement
;
5002 /* rdev have been moved down */
5003 rdev
= conf
->disks
[i
].rdev
;
5004 rdev_clear_badblocks(rdev
, sh
->sector
,
5006 rdev_dec_pending(rdev
, conf
->mddev
);
5011 raid_run_ops(sh
, s
.ops_request
);
5015 if (s
.dec_preread_active
) {
5016 /* We delay this until after ops_run_io so that if make_request
5017 * is waiting on a flush, it won't continue until the writes
5018 * have actually been submitted.
5020 atomic_dec(&conf
->preread_active_stripes
);
5021 if (atomic_read(&conf
->preread_active_stripes
) <
5023 md_wakeup_thread(conf
->mddev
->thread
);
5026 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
5029 static void raid5_activate_delayed(struct r5conf
*conf
)
5031 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
5032 while (!list_empty(&conf
->delayed_list
)) {
5033 struct list_head
*l
= conf
->delayed_list
.next
;
5034 struct stripe_head
*sh
;
5035 sh
= list_entry(l
, struct stripe_head
, lru
);
5037 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5038 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5039 atomic_inc(&conf
->preread_active_stripes
);
5040 list_add_tail(&sh
->lru
, &conf
->hold_list
);
5041 raid5_wakeup_stripe_thread(sh
);
5046 static void activate_bit_delay(struct r5conf
*conf
,
5047 struct list_head
*temp_inactive_list
)
5049 /* device_lock is held */
5050 struct list_head head
;
5051 list_add(&head
, &conf
->bitmap_list
);
5052 list_del_init(&conf
->bitmap_list
);
5053 while (!list_empty(&head
)) {
5054 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
5056 list_del_init(&sh
->lru
);
5057 atomic_inc(&sh
->count
);
5058 hash
= sh
->hash_lock_index
;
5059 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
5063 static int raid5_congested(struct mddev
*mddev
, int bits
)
5065 struct r5conf
*conf
= mddev
->private;
5067 /* No difference between reads and writes. Just check
5068 * how busy the stripe_cache is
5071 if (test_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
))
5074 /* Also checks whether there is pressure on r5cache log space */
5075 if (test_bit(R5C_LOG_TIGHT
, &conf
->cache_state
))
5079 if (atomic_read(&conf
->empty_inactive_list_nr
))
5085 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
5087 struct r5conf
*conf
= mddev
->private;
5088 sector_t sector
= bio
->bi_iter
.bi_sector
+ get_start_sect(bio
->bi_bdev
);
5089 unsigned int chunk_sectors
;
5090 unsigned int bio_sectors
= bio_sectors(bio
);
5092 chunk_sectors
= min(conf
->chunk_sectors
, conf
->prev_chunk_sectors
);
5093 return chunk_sectors
>=
5094 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
5098 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
5099 * later sampled by raid5d.
5101 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
5103 unsigned long flags
;
5105 spin_lock_irqsave(&conf
->device_lock
, flags
);
5107 bi
->bi_next
= conf
->retry_read_aligned_list
;
5108 conf
->retry_read_aligned_list
= bi
;
5110 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5111 md_wakeup_thread(conf
->mddev
->thread
);
5114 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
,
5115 unsigned int *offset
)
5119 bi
= conf
->retry_read_aligned
;
5121 *offset
= conf
->retry_read_offset
;
5122 conf
->retry_read_aligned
= NULL
;
5125 bi
= conf
->retry_read_aligned_list
;
5127 conf
->retry_read_aligned_list
= bi
->bi_next
;
5136 * The "raid5_align_endio" should check if the read succeeded and if it
5137 * did, call bio_endio on the original bio (having bio_put the new bio
5139 * If the read failed..
5141 static void raid5_align_endio(struct bio
*bi
)
5143 struct bio
* raid_bi
= bi
->bi_private
;
5144 struct mddev
*mddev
;
5145 struct r5conf
*conf
;
5146 struct md_rdev
*rdev
;
5147 int error
= bi
->bi_error
;
5151 rdev
= (void*)raid_bi
->bi_next
;
5152 raid_bi
->bi_next
= NULL
;
5153 mddev
= rdev
->mddev
;
5154 conf
= mddev
->private;
5156 rdev_dec_pending(rdev
, conf
->mddev
);
5160 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
5161 wake_up(&conf
->wait_for_quiescent
);
5165 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5167 add_bio_to_retry(raid_bi
, conf
);
5170 static int raid5_read_one_chunk(struct mddev
*mddev
, struct bio
*raid_bio
)
5172 struct r5conf
*conf
= mddev
->private;
5174 struct bio
* align_bi
;
5175 struct md_rdev
*rdev
;
5176 sector_t end_sector
;
5178 if (!in_chunk_boundary(mddev
, raid_bio
)) {
5179 pr_debug("%s: non aligned\n", __func__
);
5183 * use bio_clone_fast to make a copy of the bio
5185 align_bi
= bio_clone_fast(raid_bio
, GFP_NOIO
, mddev
->bio_set
);
5189 * set bi_end_io to a new function, and set bi_private to the
5192 align_bi
->bi_end_io
= raid5_align_endio
;
5193 align_bi
->bi_private
= raid_bio
;
5197 align_bi
->bi_iter
.bi_sector
=
5198 raid5_compute_sector(conf
, raid_bio
->bi_iter
.bi_sector
,
5201 end_sector
= bio_end_sector(align_bi
);
5203 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
5204 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
5205 rdev
->recovery_offset
< end_sector
) {
5206 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
5208 (test_bit(Faulty
, &rdev
->flags
) ||
5209 !(test_bit(In_sync
, &rdev
->flags
) ||
5210 rdev
->recovery_offset
>= end_sector
)))
5214 if (r5c_big_stripe_cached(conf
, align_bi
->bi_iter
.bi_sector
)) {
5224 atomic_inc(&rdev
->nr_pending
);
5226 raid_bio
->bi_next
= (void*)rdev
;
5227 align_bi
->bi_bdev
= rdev
->bdev
;
5228 bio_clear_flag(align_bi
, BIO_SEG_VALID
);
5230 if (is_badblock(rdev
, align_bi
->bi_iter
.bi_sector
,
5231 bio_sectors(align_bi
),
5232 &first_bad
, &bad_sectors
)) {
5234 rdev_dec_pending(rdev
, mddev
);
5238 /* No reshape active, so we can trust rdev->data_offset */
5239 align_bi
->bi_iter
.bi_sector
+= rdev
->data_offset
;
5241 spin_lock_irq(&conf
->device_lock
);
5242 wait_event_lock_irq(conf
->wait_for_quiescent
,
5245 atomic_inc(&conf
->active_aligned_reads
);
5246 spin_unlock_irq(&conf
->device_lock
);
5249 trace_block_bio_remap(bdev_get_queue(align_bi
->bi_bdev
),
5250 align_bi
, disk_devt(mddev
->gendisk
),
5251 raid_bio
->bi_iter
.bi_sector
);
5252 generic_make_request(align_bi
);
5261 static struct bio
*chunk_aligned_read(struct mddev
*mddev
, struct bio
*raid_bio
)
5264 sector_t sector
= raid_bio
->bi_iter
.bi_sector
;
5265 unsigned chunk_sects
= mddev
->chunk_sectors
;
5266 unsigned sectors
= chunk_sects
- (sector
& (chunk_sects
-1));
5268 if (sectors
< bio_sectors(raid_bio
)) {
5269 struct r5conf
*conf
= mddev
->private;
5270 split
= bio_split(raid_bio
, sectors
, GFP_NOIO
, conf
->bio_split
);
5271 bio_chain(split
, raid_bio
);
5272 generic_make_request(raid_bio
);
5276 if (!raid5_read_one_chunk(mddev
, raid_bio
))
5282 /* __get_priority_stripe - get the next stripe to process
5284 * Full stripe writes are allowed to pass preread active stripes up until
5285 * the bypass_threshold is exceeded. In general the bypass_count
5286 * increments when the handle_list is handled before the hold_list; however, it
5287 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5288 * stripe with in flight i/o. The bypass_count will be reset when the
5289 * head of the hold_list has changed, i.e. the head was promoted to the
5292 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
, int group
)
5294 struct stripe_head
*sh
, *tmp
;
5295 struct list_head
*handle_list
= NULL
;
5296 struct r5worker_group
*wg
;
5297 bool second_try
= !r5c_is_writeback(conf
->log
) &&
5298 !r5l_log_disk_error(conf
);
5299 bool try_loprio
= test_bit(R5C_LOG_TIGHT
, &conf
->cache_state
) ||
5300 r5l_log_disk_error(conf
);
5305 if (conf
->worker_cnt_per_group
== 0) {
5306 handle_list
= try_loprio
? &conf
->loprio_list
:
5308 } else if (group
!= ANY_GROUP
) {
5309 handle_list
= try_loprio
? &conf
->worker_groups
[group
].loprio_list
:
5310 &conf
->worker_groups
[group
].handle_list
;
5311 wg
= &conf
->worker_groups
[group
];
5314 for (i
= 0; i
< conf
->group_cnt
; i
++) {
5315 handle_list
= try_loprio
? &conf
->worker_groups
[i
].loprio_list
:
5316 &conf
->worker_groups
[i
].handle_list
;
5317 wg
= &conf
->worker_groups
[i
];
5318 if (!list_empty(handle_list
))
5323 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5325 list_empty(handle_list
) ? "empty" : "busy",
5326 list_empty(&conf
->hold_list
) ? "empty" : "busy",
5327 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
5329 if (!list_empty(handle_list
)) {
5330 sh
= list_entry(handle_list
->next
, typeof(*sh
), lru
);
5332 if (list_empty(&conf
->hold_list
))
5333 conf
->bypass_count
= 0;
5334 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
5335 if (conf
->hold_list
.next
== conf
->last_hold
)
5336 conf
->bypass_count
++;
5338 conf
->last_hold
= conf
->hold_list
.next
;
5339 conf
->bypass_count
-= conf
->bypass_threshold
;
5340 if (conf
->bypass_count
< 0)
5341 conf
->bypass_count
= 0;
5344 } else if (!list_empty(&conf
->hold_list
) &&
5345 ((conf
->bypass_threshold
&&
5346 conf
->bypass_count
> conf
->bypass_threshold
) ||
5347 atomic_read(&conf
->pending_full_writes
) == 0)) {
5349 list_for_each_entry(tmp
, &conf
->hold_list
, lru
) {
5350 if (conf
->worker_cnt_per_group
== 0 ||
5351 group
== ANY_GROUP
||
5352 !cpu_online(tmp
->cpu
) ||
5353 cpu_to_group(tmp
->cpu
) == group
) {
5360 conf
->bypass_count
-= conf
->bypass_threshold
;
5361 if (conf
->bypass_count
< 0)
5362 conf
->bypass_count
= 0;
5371 try_loprio
= !try_loprio
;
5379 list_del_init(&sh
->lru
);
5380 BUG_ON(atomic_inc_return(&sh
->count
) != 1);
5384 struct raid5_plug_cb
{
5385 struct blk_plug_cb cb
;
5386 struct list_head list
;
5387 struct list_head temp_inactive_list
[NR_STRIPE_HASH_LOCKS
];
5390 static void raid5_unplug(struct blk_plug_cb
*blk_cb
, bool from_schedule
)
5392 struct raid5_plug_cb
*cb
= container_of(
5393 blk_cb
, struct raid5_plug_cb
, cb
);
5394 struct stripe_head
*sh
;
5395 struct mddev
*mddev
= cb
->cb
.data
;
5396 struct r5conf
*conf
= mddev
->private;
5400 if (cb
->list
.next
&& !list_empty(&cb
->list
)) {
5401 spin_lock_irq(&conf
->device_lock
);
5402 while (!list_empty(&cb
->list
)) {
5403 sh
= list_first_entry(&cb
->list
, struct stripe_head
, lru
);
5404 list_del_init(&sh
->lru
);
5406 * avoid race release_stripe_plug() sees
5407 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5408 * is still in our list
5410 smp_mb__before_atomic();
5411 clear_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
);
5413 * STRIPE_ON_RELEASE_LIST could be set here. In that
5414 * case, the count is always > 1 here
5416 hash
= sh
->hash_lock_index
;
5417 __release_stripe(conf
, sh
, &cb
->temp_inactive_list
[hash
]);
5420 spin_unlock_irq(&conf
->device_lock
);
5422 release_inactive_stripe_list(conf
, cb
->temp_inactive_list
,
5423 NR_STRIPE_HASH_LOCKS
);
5425 trace_block_unplug(mddev
->queue
, cnt
, !from_schedule
);
5429 static void release_stripe_plug(struct mddev
*mddev
,
5430 struct stripe_head
*sh
)
5432 struct blk_plug_cb
*blk_cb
= blk_check_plugged(
5433 raid5_unplug
, mddev
,
5434 sizeof(struct raid5_plug_cb
));
5435 struct raid5_plug_cb
*cb
;
5438 raid5_release_stripe(sh
);
5442 cb
= container_of(blk_cb
, struct raid5_plug_cb
, cb
);
5444 if (cb
->list
.next
== NULL
) {
5446 INIT_LIST_HEAD(&cb
->list
);
5447 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5448 INIT_LIST_HEAD(cb
->temp_inactive_list
+ i
);
5451 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
))
5452 list_add_tail(&sh
->lru
, &cb
->list
);
5454 raid5_release_stripe(sh
);
5457 static void make_discard_request(struct mddev
*mddev
, struct bio
*bi
)
5459 struct r5conf
*conf
= mddev
->private;
5460 sector_t logical_sector
, last_sector
;
5461 struct stripe_head
*sh
;
5464 if (mddev
->reshape_position
!= MaxSector
)
5465 /* Skip discard while reshape is happening */
5468 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5469 last_sector
= bi
->bi_iter
.bi_sector
+ (bi
->bi_iter
.bi_size
>>9);
5472 md_write_start(mddev
, bi
);
5474 stripe_sectors
= conf
->chunk_sectors
*
5475 (conf
->raid_disks
- conf
->max_degraded
);
5476 logical_sector
= DIV_ROUND_UP_SECTOR_T(logical_sector
,
5478 sector_div(last_sector
, stripe_sectors
);
5480 logical_sector
*= conf
->chunk_sectors
;
5481 last_sector
*= conf
->chunk_sectors
;
5483 for (; logical_sector
< last_sector
;
5484 logical_sector
+= STRIPE_SECTORS
) {
5488 sh
= raid5_get_active_stripe(conf
, logical_sector
, 0, 0, 0);
5489 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5490 TASK_UNINTERRUPTIBLE
);
5491 set_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5492 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
5493 raid5_release_stripe(sh
);
5497 clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5498 spin_lock_irq(&sh
->stripe_lock
);
5499 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5500 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5502 if (sh
->dev
[d
].towrite
|| sh
->dev
[d
].toread
) {
5503 set_bit(R5_Overlap
, &sh
->dev
[d
].flags
);
5504 spin_unlock_irq(&sh
->stripe_lock
);
5505 raid5_release_stripe(sh
);
5510 set_bit(STRIPE_DISCARD
, &sh
->state
);
5511 finish_wait(&conf
->wait_for_overlap
, &w
);
5512 sh
->overwrite_disks
= 0;
5513 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5514 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5516 sh
->dev
[d
].towrite
= bi
;
5517 set_bit(R5_OVERWRITE
, &sh
->dev
[d
].flags
);
5518 bio_inc_remaining(bi
);
5519 md_write_inc(mddev
, bi
);
5520 sh
->overwrite_disks
++;
5522 spin_unlock_irq(&sh
->stripe_lock
);
5523 if (conf
->mddev
->bitmap
) {
5525 d
< conf
->raid_disks
- conf
->max_degraded
;
5527 bitmap_startwrite(mddev
->bitmap
,
5531 sh
->bm_seq
= conf
->seq_flush
+ 1;
5532 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
5535 set_bit(STRIPE_HANDLE
, &sh
->state
);
5536 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5537 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5538 atomic_inc(&conf
->preread_active_stripes
);
5539 release_stripe_plug(mddev
, sh
);
5542 md_write_end(mddev
);
5546 static void raid5_make_request(struct mddev
*mddev
, struct bio
* bi
)
5548 struct r5conf
*conf
= mddev
->private;
5550 sector_t new_sector
;
5551 sector_t logical_sector
, last_sector
;
5552 struct stripe_head
*sh
;
5553 const int rw
= bio_data_dir(bi
);
5556 bool do_flush
= false;
5558 if (unlikely(bi
->bi_opf
& REQ_PREFLUSH
)) {
5559 int ret
= r5l_handle_flush_request(conf
->log
, bi
);
5563 if (ret
== -ENODEV
) {
5564 md_flush_request(mddev
, bi
);
5567 /* ret == -EAGAIN, fallback */
5569 * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5570 * we need to flush journal device
5572 do_flush
= bi
->bi_opf
& REQ_PREFLUSH
;
5576 * If array is degraded, better not do chunk aligned read because
5577 * later we might have to read it again in order to reconstruct
5578 * data on failed drives.
5580 if (rw
== READ
&& mddev
->degraded
== 0 &&
5581 mddev
->reshape_position
== MaxSector
) {
5582 bi
= chunk_aligned_read(mddev
, bi
);
5587 if (unlikely(bio_op(bi
) == REQ_OP_DISCARD
)) {
5588 make_discard_request(mddev
, bi
);
5592 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
5593 last_sector
= bio_end_sector(bi
);
5595 md_write_start(mddev
, bi
);
5597 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
5598 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
5604 seq
= read_seqcount_begin(&conf
->gen_lock
);
5607 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5608 TASK_UNINTERRUPTIBLE
);
5609 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
5610 /* spinlock is needed as reshape_progress may be
5611 * 64bit on a 32bit platform, and so it might be
5612 * possible to see a half-updated value
5613 * Of course reshape_progress could change after
5614 * the lock is dropped, so once we get a reference
5615 * to the stripe that we think it is, we will have
5618 spin_lock_irq(&conf
->device_lock
);
5619 if (mddev
->reshape_backwards
5620 ? logical_sector
< conf
->reshape_progress
5621 : logical_sector
>= conf
->reshape_progress
) {
5624 if (mddev
->reshape_backwards
5625 ? logical_sector
< conf
->reshape_safe
5626 : logical_sector
>= conf
->reshape_safe
) {
5627 spin_unlock_irq(&conf
->device_lock
);
5633 spin_unlock_irq(&conf
->device_lock
);
5636 new_sector
= raid5_compute_sector(conf
, logical_sector
,
5639 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5640 (unsigned long long)new_sector
,
5641 (unsigned long long)logical_sector
);
5643 sh
= raid5_get_active_stripe(conf
, new_sector
, previous
,
5644 (bi
->bi_opf
& REQ_RAHEAD
), 0);
5646 if (unlikely(previous
)) {
5647 /* expansion might have moved on while waiting for a
5648 * stripe, so we must do the range check again.
5649 * Expansion could still move past after this
5650 * test, but as we are holding a reference to
5651 * 'sh', we know that if that happens,
5652 * STRIPE_EXPANDING will get set and the expansion
5653 * won't proceed until we finish with the stripe.
5656 spin_lock_irq(&conf
->device_lock
);
5657 if (mddev
->reshape_backwards
5658 ? logical_sector
>= conf
->reshape_progress
5659 : logical_sector
< conf
->reshape_progress
)
5660 /* mismatch, need to try again */
5662 spin_unlock_irq(&conf
->device_lock
);
5664 raid5_release_stripe(sh
);
5670 if (read_seqcount_retry(&conf
->gen_lock
, seq
)) {
5671 /* Might have got the wrong stripe_head
5674 raid5_release_stripe(sh
);
5679 logical_sector
>= mddev
->suspend_lo
&&
5680 logical_sector
< mddev
->suspend_hi
) {
5681 raid5_release_stripe(sh
);
5682 /* As the suspend_* range is controlled by
5683 * userspace, we want an interruptible
5686 flush_signals(current
);
5687 prepare_to_wait(&conf
->wait_for_overlap
,
5688 &w
, TASK_INTERRUPTIBLE
);
5689 if (logical_sector
>= mddev
->suspend_lo
&&
5690 logical_sector
< mddev
->suspend_hi
) {
5697 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
5698 !add_stripe_bio(sh
, bi
, dd_idx
, rw
, previous
)) {
5699 /* Stripe is busy expanding or
5700 * add failed due to overlap. Flush everything
5703 md_wakeup_thread(mddev
->thread
);
5704 raid5_release_stripe(sh
);
5710 set_bit(STRIPE_R5C_PREFLUSH
, &sh
->state
);
5711 /* we only need flush for one stripe */
5715 set_bit(STRIPE_HANDLE
, &sh
->state
);
5716 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5717 if ((!sh
->batch_head
|| sh
== sh
->batch_head
) &&
5718 (bi
->bi_opf
& REQ_SYNC
) &&
5719 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5720 atomic_inc(&conf
->preread_active_stripes
);
5721 release_stripe_plug(mddev
, sh
);
5723 /* cannot get stripe for read-ahead, just give-up */
5724 bi
->bi_error
= -EIO
;
5728 finish_wait(&conf
->wait_for_overlap
, &w
);
5731 md_write_end(mddev
);
5735 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
5737 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
5739 /* reshaping is quite different to recovery/resync so it is
5740 * handled quite separately ... here.
5742 * On each call to sync_request, we gather one chunk worth of
5743 * destination stripes and flag them as expanding.
5744 * Then we find all the source stripes and request reads.
5745 * As the reads complete, handle_stripe will copy the data
5746 * into the destination stripe and release that stripe.
5748 struct r5conf
*conf
= mddev
->private;
5749 struct stripe_head
*sh
;
5750 sector_t first_sector
, last_sector
;
5751 int raid_disks
= conf
->previous_raid_disks
;
5752 int data_disks
= raid_disks
- conf
->max_degraded
;
5753 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5756 sector_t writepos
, readpos
, safepos
;
5757 sector_t stripe_addr
;
5758 int reshape_sectors
;
5759 struct list_head stripes
;
5762 if (sector_nr
== 0) {
5763 /* If restarting in the middle, skip the initial sectors */
5764 if (mddev
->reshape_backwards
&&
5765 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
5766 sector_nr
= raid5_size(mddev
, 0, 0)
5767 - conf
->reshape_progress
;
5768 } else if (mddev
->reshape_backwards
&&
5769 conf
->reshape_progress
== MaxSector
) {
5770 /* shouldn't happen, but just in case, finish up.*/
5771 sector_nr
= MaxSector
;
5772 } else if (!mddev
->reshape_backwards
&&
5773 conf
->reshape_progress
> 0)
5774 sector_nr
= conf
->reshape_progress
;
5775 sector_div(sector_nr
, new_data_disks
);
5777 mddev
->curr_resync_completed
= sector_nr
;
5778 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5785 /* We need to process a full chunk at a time.
5786 * If old and new chunk sizes differ, we need to process the
5790 reshape_sectors
= max(conf
->chunk_sectors
, conf
->prev_chunk_sectors
);
5792 /* We update the metadata at least every 10 seconds, or when
5793 * the data about to be copied would over-write the source of
5794 * the data at the front of the range. i.e. one new_stripe
5795 * along from reshape_progress new_maps to after where
5796 * reshape_safe old_maps to
5798 writepos
= conf
->reshape_progress
;
5799 sector_div(writepos
, new_data_disks
);
5800 readpos
= conf
->reshape_progress
;
5801 sector_div(readpos
, data_disks
);
5802 safepos
= conf
->reshape_safe
;
5803 sector_div(safepos
, data_disks
);
5804 if (mddev
->reshape_backwards
) {
5805 BUG_ON(writepos
< reshape_sectors
);
5806 writepos
-= reshape_sectors
;
5807 readpos
+= reshape_sectors
;
5808 safepos
+= reshape_sectors
;
5810 writepos
+= reshape_sectors
;
5811 /* readpos and safepos are worst-case calculations.
5812 * A negative number is overly pessimistic, and causes
5813 * obvious problems for unsigned storage. So clip to 0.
5815 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
5816 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
5819 /* Having calculated the 'writepos' possibly use it
5820 * to set 'stripe_addr' which is where we will write to.
5822 if (mddev
->reshape_backwards
) {
5823 BUG_ON(conf
->reshape_progress
== 0);
5824 stripe_addr
= writepos
;
5825 BUG_ON((mddev
->dev_sectors
&
5826 ~((sector_t
)reshape_sectors
- 1))
5827 - reshape_sectors
- stripe_addr
5830 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
5831 stripe_addr
= sector_nr
;
5834 /* 'writepos' is the most advanced device address we might write.
5835 * 'readpos' is the least advanced device address we might read.
5836 * 'safepos' is the least address recorded in the metadata as having
5838 * If there is a min_offset_diff, these are adjusted either by
5839 * increasing the safepos/readpos if diff is negative, or
5840 * increasing writepos if diff is positive.
5841 * If 'readpos' is then behind 'writepos', there is no way that we can
5842 * ensure safety in the face of a crash - that must be done by userspace
5843 * making a backup of the data. So in that case there is no particular
5844 * rush to update metadata.
5845 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5846 * update the metadata to advance 'safepos' to match 'readpos' so that
5847 * we can be safe in the event of a crash.
5848 * So we insist on updating metadata if safepos is behind writepos and
5849 * readpos is beyond writepos.
5850 * In any case, update the metadata every 10 seconds.
5851 * Maybe that number should be configurable, but I'm not sure it is
5852 * worth it.... maybe it could be a multiple of safemode_delay???
5854 if (conf
->min_offset_diff
< 0) {
5855 safepos
+= -conf
->min_offset_diff
;
5856 readpos
+= -conf
->min_offset_diff
;
5858 writepos
+= conf
->min_offset_diff
;
5860 if ((mddev
->reshape_backwards
5861 ? (safepos
> writepos
&& readpos
< writepos
)
5862 : (safepos
< writepos
&& readpos
> writepos
)) ||
5863 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
5864 /* Cannot proceed until we've updated the superblock... */
5865 wait_event(conf
->wait_for_overlap
,
5866 atomic_read(&conf
->reshape_stripes
)==0
5867 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5868 if (atomic_read(&conf
->reshape_stripes
) != 0)
5870 mddev
->reshape_position
= conf
->reshape_progress
;
5871 mddev
->curr_resync_completed
= sector_nr
;
5872 conf
->reshape_checkpoint
= jiffies
;
5873 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
5874 md_wakeup_thread(mddev
->thread
);
5875 wait_event(mddev
->sb_wait
, mddev
->sb_flags
== 0 ||
5876 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5877 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5879 spin_lock_irq(&conf
->device_lock
);
5880 conf
->reshape_safe
= mddev
->reshape_position
;
5881 spin_unlock_irq(&conf
->device_lock
);
5882 wake_up(&conf
->wait_for_overlap
);
5883 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5886 INIT_LIST_HEAD(&stripes
);
5887 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
5889 int skipped_disk
= 0;
5890 sh
= raid5_get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
5891 set_bit(STRIPE_EXPANDING
, &sh
->state
);
5892 atomic_inc(&conf
->reshape_stripes
);
5893 /* If any of this stripe is beyond the end of the old
5894 * array, then we need to zero those blocks
5896 for (j
=sh
->disks
; j
--;) {
5898 if (j
== sh
->pd_idx
)
5900 if (conf
->level
== 6 &&
5903 s
= raid5_compute_blocknr(sh
, j
, 0);
5904 if (s
< raid5_size(mddev
, 0, 0)) {
5908 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
5909 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
5910 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
5912 if (!skipped_disk
) {
5913 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
5914 set_bit(STRIPE_HANDLE
, &sh
->state
);
5916 list_add(&sh
->lru
, &stripes
);
5918 spin_lock_irq(&conf
->device_lock
);
5919 if (mddev
->reshape_backwards
)
5920 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
5922 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
5923 spin_unlock_irq(&conf
->device_lock
);
5924 /* Ok, those stripe are ready. We can start scheduling
5925 * reads on the source stripes.
5926 * The source stripes are determined by mapping the first and last
5927 * block on the destination stripes.
5930 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
5933 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
5934 * new_data_disks
- 1),
5936 if (last_sector
>= mddev
->dev_sectors
)
5937 last_sector
= mddev
->dev_sectors
- 1;
5938 while (first_sector
<= last_sector
) {
5939 sh
= raid5_get_active_stripe(conf
, first_sector
, 1, 0, 1);
5940 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
5941 set_bit(STRIPE_HANDLE
, &sh
->state
);
5942 raid5_release_stripe(sh
);
5943 first_sector
+= STRIPE_SECTORS
;
5945 /* Now that the sources are clearly marked, we can release
5946 * the destination stripes
5948 while (!list_empty(&stripes
)) {
5949 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
5950 list_del_init(&sh
->lru
);
5951 raid5_release_stripe(sh
);
5953 /* If this takes us to the resync_max point where we have to pause,
5954 * then we need to write out the superblock.
5956 sector_nr
+= reshape_sectors
;
5957 retn
= reshape_sectors
;
5959 if (mddev
->curr_resync_completed
> mddev
->resync_max
||
5960 (sector_nr
- mddev
->curr_resync_completed
) * 2
5961 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
5962 /* Cannot proceed until we've updated the superblock... */
5963 wait_event(conf
->wait_for_overlap
,
5964 atomic_read(&conf
->reshape_stripes
) == 0
5965 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5966 if (atomic_read(&conf
->reshape_stripes
) != 0)
5968 mddev
->reshape_position
= conf
->reshape_progress
;
5969 mddev
->curr_resync_completed
= sector_nr
;
5970 conf
->reshape_checkpoint
= jiffies
;
5971 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
5972 md_wakeup_thread(mddev
->thread
);
5973 wait_event(mddev
->sb_wait
,
5974 !test_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
)
5975 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
5976 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
5978 spin_lock_irq(&conf
->device_lock
);
5979 conf
->reshape_safe
= mddev
->reshape_position
;
5980 spin_unlock_irq(&conf
->device_lock
);
5981 wake_up(&conf
->wait_for_overlap
);
5982 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
5988 static inline sector_t
raid5_sync_request(struct mddev
*mddev
, sector_t sector_nr
,
5991 struct r5conf
*conf
= mddev
->private;
5992 struct stripe_head
*sh
;
5993 sector_t max_sector
= mddev
->dev_sectors
;
5994 sector_t sync_blocks
;
5995 int still_degraded
= 0;
5998 if (sector_nr
>= max_sector
) {
5999 /* just being told to finish up .. nothing much to do */
6001 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
6006 if (mddev
->curr_resync
< max_sector
) /* aborted */
6007 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
6009 else /* completed sync */
6011 bitmap_close_sync(mddev
->bitmap
);
6016 /* Allow raid5_quiesce to complete */
6017 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
6019 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
6020 return reshape_request(mddev
, sector_nr
, skipped
);
6022 /* No need to check resync_max as we never do more than one
6023 * stripe, and as resync_max will always be on a chunk boundary,
6024 * if the check in md_do_sync didn't fire, there is no chance
6025 * of overstepping resync_max here
6028 /* if there is too many failed drives and we are trying
6029 * to resync, then assert that we are finished, because there is
6030 * nothing we can do.
6032 if (mddev
->degraded
>= conf
->max_degraded
&&
6033 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
6034 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
6038 if (!test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
6040 !bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
6041 sync_blocks
>= STRIPE_SECTORS
) {
6042 /* we can skip this block, and probably more */
6043 sync_blocks
/= STRIPE_SECTORS
;
6045 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
6048 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
, false);
6050 sh
= raid5_get_active_stripe(conf
, sector_nr
, 0, 1, 0);
6052 sh
= raid5_get_active_stripe(conf
, sector_nr
, 0, 0, 0);
6053 /* make sure we don't swamp the stripe cache if someone else
6054 * is trying to get access
6056 schedule_timeout_uninterruptible(1);
6058 /* Need to check if array will still be degraded after recovery/resync
6059 * Note in case of > 1 drive failures it's possible we're rebuilding
6060 * one drive while leaving another faulty drive in array.
6063 for (i
= 0; i
< conf
->raid_disks
; i
++) {
6064 struct md_rdev
*rdev
= ACCESS_ONCE(conf
->disks
[i
].rdev
);
6066 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
))
6071 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
6073 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
6074 set_bit(STRIPE_HANDLE
, &sh
->state
);
6076 raid5_release_stripe(sh
);
6078 return STRIPE_SECTORS
;
6081 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
,
6082 unsigned int offset
)
6084 /* We may not be able to submit a whole bio at once as there
6085 * may not be enough stripe_heads available.
6086 * We cannot pre-allocate enough stripe_heads as we may need
6087 * more than exist in the cache (if we allow ever large chunks).
6088 * So we do one stripe head at a time and record in
6089 * ->bi_hw_segments how many have been done.
6091 * We *know* that this entire raid_bio is in one chunk, so
6092 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6094 struct stripe_head
*sh
;
6096 sector_t sector
, logical_sector
, last_sector
;
6100 logical_sector
= raid_bio
->bi_iter
.bi_sector
&
6101 ~((sector_t
)STRIPE_SECTORS
-1);
6102 sector
= raid5_compute_sector(conf
, logical_sector
,
6104 last_sector
= bio_end_sector(raid_bio
);
6106 for (; logical_sector
< last_sector
;
6107 logical_sector
+= STRIPE_SECTORS
,
6108 sector
+= STRIPE_SECTORS
,
6112 /* already done this stripe */
6115 sh
= raid5_get_active_stripe(conf
, sector
, 0, 1, 1);
6118 /* failed to get a stripe - must wait */
6119 conf
->retry_read_aligned
= raid_bio
;
6120 conf
->retry_read_offset
= scnt
;
6124 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0, 0)) {
6125 raid5_release_stripe(sh
);
6126 conf
->retry_read_aligned
= raid_bio
;
6127 conf
->retry_read_offset
= scnt
;
6131 set_bit(R5_ReadNoMerge
, &sh
->dev
[dd_idx
].flags
);
6133 raid5_release_stripe(sh
);
6137 bio_endio(raid_bio
);
6139 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
6140 wake_up(&conf
->wait_for_quiescent
);
6144 static int handle_active_stripes(struct r5conf
*conf
, int group
,
6145 struct r5worker
*worker
,
6146 struct list_head
*temp_inactive_list
)
6148 struct stripe_head
*batch
[MAX_STRIPE_BATCH
], *sh
;
6149 int i
, batch_size
= 0, hash
;
6150 bool release_inactive
= false;
6152 while (batch_size
< MAX_STRIPE_BATCH
&&
6153 (sh
= __get_priority_stripe(conf
, group
)) != NULL
)
6154 batch
[batch_size
++] = sh
;
6156 if (batch_size
== 0) {
6157 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6158 if (!list_empty(temp_inactive_list
+ i
))
6160 if (i
== NR_STRIPE_HASH_LOCKS
) {
6161 spin_unlock_irq(&conf
->device_lock
);
6162 r5l_flush_stripe_to_raid(conf
->log
);
6163 spin_lock_irq(&conf
->device_lock
);
6166 release_inactive
= true;
6168 spin_unlock_irq(&conf
->device_lock
);
6170 release_inactive_stripe_list(conf
, temp_inactive_list
,
6171 NR_STRIPE_HASH_LOCKS
);
6173 r5l_flush_stripe_to_raid(conf
->log
);
6174 if (release_inactive
) {
6175 spin_lock_irq(&conf
->device_lock
);
6179 for (i
= 0; i
< batch_size
; i
++)
6180 handle_stripe(batch
[i
]);
6181 log_write_stripe_run(conf
);
6185 spin_lock_irq(&conf
->device_lock
);
6186 for (i
= 0; i
< batch_size
; i
++) {
6187 hash
= batch
[i
]->hash_lock_index
;
6188 __release_stripe(conf
, batch
[i
], &temp_inactive_list
[hash
]);
6193 static void raid5_do_work(struct work_struct
*work
)
6195 struct r5worker
*worker
= container_of(work
, struct r5worker
, work
);
6196 struct r5worker_group
*group
= worker
->group
;
6197 struct r5conf
*conf
= group
->conf
;
6198 struct mddev
*mddev
= conf
->mddev
;
6199 int group_id
= group
- conf
->worker_groups
;
6201 struct blk_plug plug
;
6203 pr_debug("+++ raid5worker active\n");
6205 blk_start_plug(&plug
);
6207 spin_lock_irq(&conf
->device_lock
);
6209 int batch_size
, released
;
6211 released
= release_stripe_list(conf
, worker
->temp_inactive_list
);
6213 batch_size
= handle_active_stripes(conf
, group_id
, worker
,
6214 worker
->temp_inactive_list
);
6215 worker
->working
= false;
6216 if (!batch_size
&& !released
)
6218 handled
+= batch_size
;
6219 wait_event_lock_irq(mddev
->sb_wait
,
6220 !test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
),
6223 pr_debug("%d stripes handled\n", handled
);
6225 spin_unlock_irq(&conf
->device_lock
);
6226 blk_finish_plug(&plug
);
6228 pr_debug("--- raid5worker inactive\n");
6232 * This is our raid5 kernel thread.
6234 * We scan the hash table for stripes which can be handled now.
6235 * During the scan, completed stripes are saved for us by the interrupt
6236 * handler, so that they will not have to wait for our next wakeup.
6238 static void raid5d(struct md_thread
*thread
)
6240 struct mddev
*mddev
= thread
->mddev
;
6241 struct r5conf
*conf
= mddev
->private;
6243 struct blk_plug plug
;
6245 pr_debug("+++ raid5d active\n");
6247 md_check_recovery(mddev
);
6249 blk_start_plug(&plug
);
6251 spin_lock_irq(&conf
->device_lock
);
6254 int batch_size
, released
;
6255 unsigned int offset
;
6257 released
= release_stripe_list(conf
, conf
->temp_inactive_list
);
6259 clear_bit(R5_DID_ALLOC
, &conf
->cache_state
);
6262 !list_empty(&conf
->bitmap_list
)) {
6263 /* Now is a good time to flush some bitmap updates */
6265 spin_unlock_irq(&conf
->device_lock
);
6266 bitmap_unplug(mddev
->bitmap
);
6267 spin_lock_irq(&conf
->device_lock
);
6268 conf
->seq_write
= conf
->seq_flush
;
6269 activate_bit_delay(conf
, conf
->temp_inactive_list
);
6271 raid5_activate_delayed(conf
);
6273 while ((bio
= remove_bio_from_retry(conf
, &offset
))) {
6275 spin_unlock_irq(&conf
->device_lock
);
6276 ok
= retry_aligned_read(conf
, bio
, offset
);
6277 spin_lock_irq(&conf
->device_lock
);
6283 batch_size
= handle_active_stripes(conf
, ANY_GROUP
, NULL
,
6284 conf
->temp_inactive_list
);
6285 if (!batch_size
&& !released
)
6287 handled
+= batch_size
;
6289 if (mddev
->sb_flags
& ~(1 << MD_SB_CHANGE_PENDING
)) {
6290 spin_unlock_irq(&conf
->device_lock
);
6291 md_check_recovery(mddev
);
6292 spin_lock_irq(&conf
->device_lock
);
6295 pr_debug("%d stripes handled\n", handled
);
6297 spin_unlock_irq(&conf
->device_lock
);
6298 if (test_and_clear_bit(R5_ALLOC_MORE
, &conf
->cache_state
) &&
6299 mutex_trylock(&conf
->cache_size_mutex
)) {
6300 grow_one_stripe(conf
, __GFP_NOWARN
);
6301 /* Set flag even if allocation failed. This helps
6302 * slow down allocation requests when mem is short
6304 set_bit(R5_DID_ALLOC
, &conf
->cache_state
);
6305 mutex_unlock(&conf
->cache_size_mutex
);
6308 flush_deferred_bios(conf
);
6310 r5l_flush_stripe_to_raid(conf
->log
);
6312 async_tx_issue_pending_all();
6313 blk_finish_plug(&plug
);
6315 pr_debug("--- raid5d inactive\n");
6319 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
6321 struct r5conf
*conf
;
6323 spin_lock(&mddev
->lock
);
6324 conf
= mddev
->private;
6326 ret
= sprintf(page
, "%d\n", conf
->min_nr_stripes
);
6327 spin_unlock(&mddev
->lock
);
6332 raid5_set_cache_size(struct mddev
*mddev
, int size
)
6334 struct r5conf
*conf
= mddev
->private;
6336 if (size
<= 16 || size
> 32768)
6339 conf
->min_nr_stripes
= size
;
6340 mutex_lock(&conf
->cache_size_mutex
);
6341 while (size
< conf
->max_nr_stripes
&&
6342 drop_one_stripe(conf
))
6344 mutex_unlock(&conf
->cache_size_mutex
);
6346 md_allow_write(mddev
);
6348 mutex_lock(&conf
->cache_size_mutex
);
6349 while (size
> conf
->max_nr_stripes
)
6350 if (!grow_one_stripe(conf
, GFP_KERNEL
))
6352 mutex_unlock(&conf
->cache_size_mutex
);
6356 EXPORT_SYMBOL(raid5_set_cache_size
);
6359 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
6361 struct r5conf
*conf
;
6365 if (len
>= PAGE_SIZE
)
6367 if (kstrtoul(page
, 10, &new))
6369 err
= mddev_lock(mddev
);
6372 conf
= mddev
->private;
6376 err
= raid5_set_cache_size(mddev
, new);
6377 mddev_unlock(mddev
);
6382 static struct md_sysfs_entry
6383 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
6384 raid5_show_stripe_cache_size
,
6385 raid5_store_stripe_cache_size
);
6388 raid5_show_rmw_level(struct mddev
*mddev
, char *page
)
6390 struct r5conf
*conf
= mddev
->private;
6392 return sprintf(page
, "%d\n", conf
->rmw_level
);
6398 raid5_store_rmw_level(struct mddev
*mddev
, const char *page
, size_t len
)
6400 struct r5conf
*conf
= mddev
->private;
6406 if (len
>= PAGE_SIZE
)
6409 if (kstrtoul(page
, 10, &new))
6412 if (new != PARITY_DISABLE_RMW
&& !raid6_call
.xor_syndrome
)
6415 if (new != PARITY_DISABLE_RMW
&&
6416 new != PARITY_ENABLE_RMW
&&
6417 new != PARITY_PREFER_RMW
)
6420 conf
->rmw_level
= new;
6424 static struct md_sysfs_entry
6425 raid5_rmw_level
= __ATTR(rmw_level
, S_IRUGO
| S_IWUSR
,
6426 raid5_show_rmw_level
,
6427 raid5_store_rmw_level
);
6431 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
6433 struct r5conf
*conf
;
6435 spin_lock(&mddev
->lock
);
6436 conf
= mddev
->private;
6438 ret
= sprintf(page
, "%d\n", conf
->bypass_threshold
);
6439 spin_unlock(&mddev
->lock
);
6444 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
6446 struct r5conf
*conf
;
6450 if (len
>= PAGE_SIZE
)
6452 if (kstrtoul(page
, 10, &new))
6455 err
= mddev_lock(mddev
);
6458 conf
= mddev
->private;
6461 else if (new > conf
->min_nr_stripes
)
6464 conf
->bypass_threshold
= new;
6465 mddev_unlock(mddev
);
6469 static struct md_sysfs_entry
6470 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
6472 raid5_show_preread_threshold
,
6473 raid5_store_preread_threshold
);
6476 raid5_show_skip_copy(struct mddev
*mddev
, char *page
)
6478 struct r5conf
*conf
;
6480 spin_lock(&mddev
->lock
);
6481 conf
= mddev
->private;
6483 ret
= sprintf(page
, "%d\n", conf
->skip_copy
);
6484 spin_unlock(&mddev
->lock
);
6489 raid5_store_skip_copy(struct mddev
*mddev
, const char *page
, size_t len
)
6491 struct r5conf
*conf
;
6495 if (len
>= PAGE_SIZE
)
6497 if (kstrtoul(page
, 10, &new))
6501 err
= mddev_lock(mddev
);
6504 conf
= mddev
->private;
6507 else if (new != conf
->skip_copy
) {
6508 mddev_suspend(mddev
);
6509 conf
->skip_copy
= new;
6511 mddev
->queue
->backing_dev_info
->capabilities
|=
6512 BDI_CAP_STABLE_WRITES
;
6514 mddev
->queue
->backing_dev_info
->capabilities
&=
6515 ~BDI_CAP_STABLE_WRITES
;
6516 mddev_resume(mddev
);
6518 mddev_unlock(mddev
);
6522 static struct md_sysfs_entry
6523 raid5_skip_copy
= __ATTR(skip_copy
, S_IRUGO
| S_IWUSR
,
6524 raid5_show_skip_copy
,
6525 raid5_store_skip_copy
);
6528 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
6530 struct r5conf
*conf
= mddev
->private;
6532 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
6537 static struct md_sysfs_entry
6538 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
6541 raid5_show_group_thread_cnt(struct mddev
*mddev
, char *page
)
6543 struct r5conf
*conf
;
6545 spin_lock(&mddev
->lock
);
6546 conf
= mddev
->private;
6548 ret
= sprintf(page
, "%d\n", conf
->worker_cnt_per_group
);
6549 spin_unlock(&mddev
->lock
);
6553 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6555 int *worker_cnt_per_group
,
6556 struct r5worker_group
**worker_groups
);
6558 raid5_store_group_thread_cnt(struct mddev
*mddev
, const char *page
, size_t len
)
6560 struct r5conf
*conf
;
6563 struct r5worker_group
*new_groups
, *old_groups
;
6564 int group_cnt
, worker_cnt_per_group
;
6566 if (len
>= PAGE_SIZE
)
6568 if (kstrtoul(page
, 10, &new))
6571 err
= mddev_lock(mddev
);
6574 conf
= mddev
->private;
6577 else if (new != conf
->worker_cnt_per_group
) {
6578 mddev_suspend(mddev
);
6580 old_groups
= conf
->worker_groups
;
6582 flush_workqueue(raid5_wq
);
6584 err
= alloc_thread_groups(conf
, new,
6585 &group_cnt
, &worker_cnt_per_group
,
6588 spin_lock_irq(&conf
->device_lock
);
6589 conf
->group_cnt
= group_cnt
;
6590 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6591 conf
->worker_groups
= new_groups
;
6592 spin_unlock_irq(&conf
->device_lock
);
6595 kfree(old_groups
[0].workers
);
6598 mddev_resume(mddev
);
6600 mddev_unlock(mddev
);
6605 static struct md_sysfs_entry
6606 raid5_group_thread_cnt
= __ATTR(group_thread_cnt
, S_IRUGO
| S_IWUSR
,
6607 raid5_show_group_thread_cnt
,
6608 raid5_store_group_thread_cnt
);
6610 static struct attribute
*raid5_attrs
[] = {
6611 &raid5_stripecache_size
.attr
,
6612 &raid5_stripecache_active
.attr
,
6613 &raid5_preread_bypass_threshold
.attr
,
6614 &raid5_group_thread_cnt
.attr
,
6615 &raid5_skip_copy
.attr
,
6616 &raid5_rmw_level
.attr
,
6617 &r5c_journal_mode
.attr
,
6620 static struct attribute_group raid5_attrs_group
= {
6622 .attrs
= raid5_attrs
,
6625 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
6627 int *worker_cnt_per_group
,
6628 struct r5worker_group
**worker_groups
)
6632 struct r5worker
*workers
;
6634 *worker_cnt_per_group
= cnt
;
6637 *worker_groups
= NULL
;
6640 *group_cnt
= num_possible_nodes();
6641 size
= sizeof(struct r5worker
) * cnt
;
6642 workers
= kzalloc(size
* *group_cnt
, GFP_NOIO
);
6643 *worker_groups
= kzalloc(sizeof(struct r5worker_group
) *
6644 *group_cnt
, GFP_NOIO
);
6645 if (!*worker_groups
|| !workers
) {
6647 kfree(*worker_groups
);
6651 for (i
= 0; i
< *group_cnt
; i
++) {
6652 struct r5worker_group
*group
;
6654 group
= &(*worker_groups
)[i
];
6655 INIT_LIST_HEAD(&group
->handle_list
);
6656 INIT_LIST_HEAD(&group
->loprio_list
);
6658 group
->workers
= workers
+ i
* cnt
;
6660 for (j
= 0; j
< cnt
; j
++) {
6661 struct r5worker
*worker
= group
->workers
+ j
;
6662 worker
->group
= group
;
6663 INIT_WORK(&worker
->work
, raid5_do_work
);
6665 for (k
= 0; k
< NR_STRIPE_HASH_LOCKS
; k
++)
6666 INIT_LIST_HEAD(worker
->temp_inactive_list
+ k
);
6673 static void free_thread_groups(struct r5conf
*conf
)
6675 if (conf
->worker_groups
)
6676 kfree(conf
->worker_groups
[0].workers
);
6677 kfree(conf
->worker_groups
);
6678 conf
->worker_groups
= NULL
;
6682 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
6684 struct r5conf
*conf
= mddev
->private;
6687 sectors
= mddev
->dev_sectors
;
6689 /* size is defined by the smallest of previous and new size */
6690 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
6692 sectors
&= ~((sector_t
)conf
->chunk_sectors
- 1);
6693 sectors
&= ~((sector_t
)conf
->prev_chunk_sectors
- 1);
6694 return sectors
* (raid_disks
- conf
->max_degraded
);
6697 static void free_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6699 safe_put_page(percpu
->spare_page
);
6700 if (percpu
->scribble
)
6701 flex_array_free(percpu
->scribble
);
6702 percpu
->spare_page
= NULL
;
6703 percpu
->scribble
= NULL
;
6706 static int alloc_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
6708 if (conf
->level
== 6 && !percpu
->spare_page
)
6709 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
6710 if (!percpu
->scribble
)
6711 percpu
->scribble
= scribble_alloc(max(conf
->raid_disks
,
6712 conf
->previous_raid_disks
),
6713 max(conf
->chunk_sectors
,
6714 conf
->prev_chunk_sectors
)
6718 if (!percpu
->scribble
|| (conf
->level
== 6 && !percpu
->spare_page
)) {
6719 free_scratch_buffer(conf
, percpu
);
6726 static int raid456_cpu_dead(unsigned int cpu
, struct hlist_node
*node
)
6728 struct r5conf
*conf
= hlist_entry_safe(node
, struct r5conf
, node
);
6730 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
6734 static void raid5_free_percpu(struct r5conf
*conf
)
6739 cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE
, &conf
->node
);
6740 free_percpu(conf
->percpu
);
6743 static void free_conf(struct r5conf
*conf
)
6749 if (conf
->shrinker
.nr_deferred
)
6750 unregister_shrinker(&conf
->shrinker
);
6752 free_thread_groups(conf
);
6753 shrink_stripes(conf
);
6754 raid5_free_percpu(conf
);
6755 for (i
= 0; i
< conf
->pool_size
; i
++)
6756 if (conf
->disks
[i
].extra_page
)
6757 put_page(conf
->disks
[i
].extra_page
);
6759 if (conf
->bio_split
)
6760 bioset_free(conf
->bio_split
);
6761 kfree(conf
->stripe_hashtbl
);
6762 kfree(conf
->pending_data
);
6766 static int raid456_cpu_up_prepare(unsigned int cpu
, struct hlist_node
*node
)
6768 struct r5conf
*conf
= hlist_entry_safe(node
, struct r5conf
, node
);
6769 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
6771 if (alloc_scratch_buffer(conf
, percpu
)) {
6772 pr_warn("%s: failed memory allocation for cpu%u\n",
6779 static int raid5_alloc_percpu(struct r5conf
*conf
)
6783 conf
->percpu
= alloc_percpu(struct raid5_percpu
);
6787 err
= cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE
, &conf
->node
);
6789 conf
->scribble_disks
= max(conf
->raid_disks
,
6790 conf
->previous_raid_disks
);
6791 conf
->scribble_sectors
= max(conf
->chunk_sectors
,
6792 conf
->prev_chunk_sectors
);
6797 static unsigned long raid5_cache_scan(struct shrinker
*shrink
,
6798 struct shrink_control
*sc
)
6800 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
6801 unsigned long ret
= SHRINK_STOP
;
6803 if (mutex_trylock(&conf
->cache_size_mutex
)) {
6805 while (ret
< sc
->nr_to_scan
&&
6806 conf
->max_nr_stripes
> conf
->min_nr_stripes
) {
6807 if (drop_one_stripe(conf
) == 0) {
6813 mutex_unlock(&conf
->cache_size_mutex
);
6818 static unsigned long raid5_cache_count(struct shrinker
*shrink
,
6819 struct shrink_control
*sc
)
6821 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
6823 if (conf
->max_nr_stripes
< conf
->min_nr_stripes
)
6824 /* unlikely, but not impossible */
6826 return conf
->max_nr_stripes
- conf
->min_nr_stripes
;
6829 static struct r5conf
*setup_conf(struct mddev
*mddev
)
6831 struct r5conf
*conf
;
6832 int raid_disk
, memory
, max_disks
;
6833 struct md_rdev
*rdev
;
6834 struct disk_info
*disk
;
6837 int group_cnt
, worker_cnt_per_group
;
6838 struct r5worker_group
*new_group
;
6840 if (mddev
->new_level
!= 5
6841 && mddev
->new_level
!= 4
6842 && mddev
->new_level
!= 6) {
6843 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6844 mdname(mddev
), mddev
->new_level
);
6845 return ERR_PTR(-EIO
);
6847 if ((mddev
->new_level
== 5
6848 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
6849 (mddev
->new_level
== 6
6850 && !algorithm_valid_raid6(mddev
->new_layout
))) {
6851 pr_warn("md/raid:%s: layout %d not supported\n",
6852 mdname(mddev
), mddev
->new_layout
);
6853 return ERR_PTR(-EIO
);
6855 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
6856 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6857 mdname(mddev
), mddev
->raid_disks
);
6858 return ERR_PTR(-EINVAL
);
6861 if (!mddev
->new_chunk_sectors
||
6862 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
6863 !is_power_of_2(mddev
->new_chunk_sectors
)) {
6864 pr_warn("md/raid:%s: invalid chunk size %d\n",
6865 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
6866 return ERR_PTR(-EINVAL
);
6869 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
6872 INIT_LIST_HEAD(&conf
->free_list
);
6873 INIT_LIST_HEAD(&conf
->pending_list
);
6874 conf
->pending_data
= kzalloc(sizeof(struct r5pending_data
) *
6875 PENDING_IO_MAX
, GFP_KERNEL
);
6876 if (!conf
->pending_data
)
6878 for (i
= 0; i
< PENDING_IO_MAX
; i
++)
6879 list_add(&conf
->pending_data
[i
].sibling
, &conf
->free_list
);
6880 /* Don't enable multi-threading by default*/
6881 if (!alloc_thread_groups(conf
, 0, &group_cnt
, &worker_cnt_per_group
,
6883 conf
->group_cnt
= group_cnt
;
6884 conf
->worker_cnt_per_group
= worker_cnt_per_group
;
6885 conf
->worker_groups
= new_group
;
6888 spin_lock_init(&conf
->device_lock
);
6889 seqcount_init(&conf
->gen_lock
);
6890 mutex_init(&conf
->cache_size_mutex
);
6891 init_waitqueue_head(&conf
->wait_for_quiescent
);
6892 init_waitqueue_head(&conf
->wait_for_stripe
);
6893 init_waitqueue_head(&conf
->wait_for_overlap
);
6894 INIT_LIST_HEAD(&conf
->handle_list
);
6895 INIT_LIST_HEAD(&conf
->loprio_list
);
6896 INIT_LIST_HEAD(&conf
->hold_list
);
6897 INIT_LIST_HEAD(&conf
->delayed_list
);
6898 INIT_LIST_HEAD(&conf
->bitmap_list
);
6899 init_llist_head(&conf
->released_stripes
);
6900 atomic_set(&conf
->active_stripes
, 0);
6901 atomic_set(&conf
->preread_active_stripes
, 0);
6902 atomic_set(&conf
->active_aligned_reads
, 0);
6903 spin_lock_init(&conf
->pending_bios_lock
);
6904 conf
->batch_bio_dispatch
= true;
6905 rdev_for_each(rdev
, mddev
) {
6906 if (test_bit(Journal
, &rdev
->flags
))
6908 if (blk_queue_nonrot(bdev_get_queue(rdev
->bdev
))) {
6909 conf
->batch_bio_dispatch
= false;
6914 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
6915 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
6917 conf
->raid_disks
= mddev
->raid_disks
;
6918 if (mddev
->reshape_position
== MaxSector
)
6919 conf
->previous_raid_disks
= mddev
->raid_disks
;
6921 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
6922 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
6924 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
6930 for (i
= 0; i
< max_disks
; i
++) {
6931 conf
->disks
[i
].extra_page
= alloc_page(GFP_KERNEL
);
6932 if (!conf
->disks
[i
].extra_page
)
6936 conf
->bio_split
= bioset_create(BIO_POOL_SIZE
, 0);
6937 if (!conf
->bio_split
)
6939 conf
->mddev
= mddev
;
6941 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
6944 /* We init hash_locks[0] separately to that it can be used
6945 * as the reference lock in the spin_lock_nest_lock() call
6946 * in lock_all_device_hash_locks_irq in order to convince
6947 * lockdep that we know what we are doing.
6949 spin_lock_init(conf
->hash_locks
);
6950 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6951 spin_lock_init(conf
->hash_locks
+ i
);
6953 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6954 INIT_LIST_HEAD(conf
->inactive_list
+ i
);
6956 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6957 INIT_LIST_HEAD(conf
->temp_inactive_list
+ i
);
6959 atomic_set(&conf
->r5c_cached_full_stripes
, 0);
6960 INIT_LIST_HEAD(&conf
->r5c_full_stripe_list
);
6961 atomic_set(&conf
->r5c_cached_partial_stripes
, 0);
6962 INIT_LIST_HEAD(&conf
->r5c_partial_stripe_list
);
6963 atomic_set(&conf
->r5c_flushing_full_stripes
, 0);
6964 atomic_set(&conf
->r5c_flushing_partial_stripes
, 0);
6966 conf
->level
= mddev
->new_level
;
6967 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
6968 if (raid5_alloc_percpu(conf
) != 0)
6971 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
6973 rdev_for_each(rdev
, mddev
) {
6974 raid_disk
= rdev
->raid_disk
;
6975 if (raid_disk
>= max_disks
6976 || raid_disk
< 0 || test_bit(Journal
, &rdev
->flags
))
6978 disk
= conf
->disks
+ raid_disk
;
6980 if (test_bit(Replacement
, &rdev
->flags
)) {
6981 if (disk
->replacement
)
6983 disk
->replacement
= rdev
;
6990 if (test_bit(In_sync
, &rdev
->flags
)) {
6991 char b
[BDEVNAME_SIZE
];
6992 pr_info("md/raid:%s: device %s operational as raid disk %d\n",
6993 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
6994 } else if (rdev
->saved_raid_disk
!= raid_disk
)
6995 /* Cannot rely on bitmap to complete recovery */
6999 conf
->level
= mddev
->new_level
;
7000 if (conf
->level
== 6) {
7001 conf
->max_degraded
= 2;
7002 if (raid6_call
.xor_syndrome
)
7003 conf
->rmw_level
= PARITY_ENABLE_RMW
;
7005 conf
->rmw_level
= PARITY_DISABLE_RMW
;
7007 conf
->max_degraded
= 1;
7008 conf
->rmw_level
= PARITY_ENABLE_RMW
;
7010 conf
->algorithm
= mddev
->new_layout
;
7011 conf
->reshape_progress
= mddev
->reshape_position
;
7012 if (conf
->reshape_progress
!= MaxSector
) {
7013 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
7014 conf
->prev_algo
= mddev
->layout
;
7016 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
7017 conf
->prev_algo
= conf
->algorithm
;
7020 conf
->min_nr_stripes
= NR_STRIPES
;
7021 if (mddev
->reshape_position
!= MaxSector
) {
7022 int stripes
= max_t(int,
7023 ((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4,
7024 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4);
7025 conf
->min_nr_stripes
= max(NR_STRIPES
, stripes
);
7026 if (conf
->min_nr_stripes
!= NR_STRIPES
)
7027 pr_info("md/raid:%s: force stripe size %d for reshape\n",
7028 mdname(mddev
), conf
->min_nr_stripes
);
7030 memory
= conf
->min_nr_stripes
* (sizeof(struct stripe_head
) +
7031 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
7032 atomic_set(&conf
->empty_inactive_list_nr
, NR_STRIPE_HASH_LOCKS
);
7033 if (grow_stripes(conf
, conf
->min_nr_stripes
)) {
7034 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7035 mdname(mddev
), memory
);
7038 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev
), memory
);
7040 * Losing a stripe head costs more than the time to refill it,
7041 * it reduces the queue depth and so can hurt throughput.
7042 * So set it rather large, scaled by number of devices.
7044 conf
->shrinker
.seeks
= DEFAULT_SEEKS
* conf
->raid_disks
* 4;
7045 conf
->shrinker
.scan_objects
= raid5_cache_scan
;
7046 conf
->shrinker
.count_objects
= raid5_cache_count
;
7047 conf
->shrinker
.batch
= 128;
7048 conf
->shrinker
.flags
= 0;
7049 if (register_shrinker(&conf
->shrinker
)) {
7050 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7055 sprintf(pers_name
, "raid%d", mddev
->new_level
);
7056 conf
->thread
= md_register_thread(raid5d
, mddev
, pers_name
);
7057 if (!conf
->thread
) {
7058 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7068 return ERR_PTR(-EIO
);
7070 return ERR_PTR(-ENOMEM
);
7073 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
7076 case ALGORITHM_PARITY_0
:
7077 if (raid_disk
< max_degraded
)
7080 case ALGORITHM_PARITY_N
:
7081 if (raid_disk
>= raid_disks
- max_degraded
)
7084 case ALGORITHM_PARITY_0_6
:
7085 if (raid_disk
== 0 ||
7086 raid_disk
== raid_disks
- 1)
7089 case ALGORITHM_LEFT_ASYMMETRIC_6
:
7090 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
7091 case ALGORITHM_LEFT_SYMMETRIC_6
:
7092 case ALGORITHM_RIGHT_SYMMETRIC_6
:
7093 if (raid_disk
== raid_disks
- 1)
7099 static int raid5_run(struct mddev
*mddev
)
7101 struct r5conf
*conf
;
7102 int working_disks
= 0;
7103 int dirty_parity_disks
= 0;
7104 struct md_rdev
*rdev
;
7105 struct md_rdev
*journal_dev
= NULL
;
7106 sector_t reshape_offset
= 0;
7108 long long min_offset_diff
= 0;
7111 if (mddev
->recovery_cp
!= MaxSector
)
7112 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7115 rdev_for_each(rdev
, mddev
) {
7118 if (test_bit(Journal
, &rdev
->flags
)) {
7122 if (rdev
->raid_disk
< 0)
7124 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
7126 min_offset_diff
= diff
;
7128 } else if (mddev
->reshape_backwards
&&
7129 diff
< min_offset_diff
)
7130 min_offset_diff
= diff
;
7131 else if (!mddev
->reshape_backwards
&&
7132 diff
> min_offset_diff
)
7133 min_offset_diff
= diff
;
7136 if (mddev
->reshape_position
!= MaxSector
) {
7137 /* Check that we can continue the reshape.
7138 * Difficulties arise if the stripe we would write to
7139 * next is at or after the stripe we would read from next.
7140 * For a reshape that changes the number of devices, this
7141 * is only possible for a very short time, and mdadm makes
7142 * sure that time appears to have past before assembling
7143 * the array. So we fail if that time hasn't passed.
7144 * For a reshape that keeps the number of devices the same
7145 * mdadm must be monitoring the reshape can keeping the
7146 * critical areas read-only and backed up. It will start
7147 * the array in read-only mode, so we check for that.
7149 sector_t here_new
, here_old
;
7151 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
7156 pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7161 if (mddev
->new_level
!= mddev
->level
) {
7162 pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7166 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
7167 /* reshape_position must be on a new-stripe boundary, and one
7168 * further up in new geometry must map after here in old
7170 * If the chunk sizes are different, then as we perform reshape
7171 * in units of the largest of the two, reshape_position needs
7172 * be a multiple of the largest chunk size times new data disks.
7174 here_new
= mddev
->reshape_position
;
7175 chunk_sectors
= max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
);
7176 new_data_disks
= mddev
->raid_disks
- max_degraded
;
7177 if (sector_div(here_new
, chunk_sectors
* new_data_disks
)) {
7178 pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7182 reshape_offset
= here_new
* chunk_sectors
;
7183 /* here_new is the stripe we will write to */
7184 here_old
= mddev
->reshape_position
;
7185 sector_div(here_old
, chunk_sectors
* (old_disks
-max_degraded
));
7186 /* here_old is the first stripe that we might need to read
7188 if (mddev
->delta_disks
== 0) {
7189 /* We cannot be sure it is safe to start an in-place
7190 * reshape. It is only safe if user-space is monitoring
7191 * and taking constant backups.
7192 * mdadm always starts a situation like this in
7193 * readonly mode so it can take control before
7194 * allowing any writes. So just check for that.
7196 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
7197 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
7198 /* not really in-place - so OK */;
7199 else if (mddev
->ro
== 0) {
7200 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7204 } else if (mddev
->reshape_backwards
7205 ? (here_new
* chunk_sectors
+ min_offset_diff
<=
7206 here_old
* chunk_sectors
)
7207 : (here_new
* chunk_sectors
>=
7208 here_old
* chunk_sectors
+ (-min_offset_diff
))) {
7209 /* Reading from the same stripe as writing to - bad */
7210 pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7214 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev
));
7215 /* OK, we should be able to continue; */
7217 BUG_ON(mddev
->level
!= mddev
->new_level
);
7218 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
7219 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
7220 BUG_ON(mddev
->delta_disks
!= 0);
7223 if (test_bit(MD_HAS_JOURNAL
, &mddev
->flags
) &&
7224 test_bit(MD_HAS_PPL
, &mddev
->flags
)) {
7225 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7227 clear_bit(MD_HAS_PPL
, &mddev
->flags
);
7230 if (mddev
->private == NULL
)
7231 conf
= setup_conf(mddev
);
7233 conf
= mddev
->private;
7236 return PTR_ERR(conf
);
7238 if (test_bit(MD_HAS_JOURNAL
, &mddev
->flags
)) {
7240 pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7243 set_disk_ro(mddev
->gendisk
, 1);
7244 } else if (mddev
->recovery_cp
== MaxSector
)
7245 set_bit(MD_JOURNAL_CLEAN
, &mddev
->flags
);
7248 conf
->min_offset_diff
= min_offset_diff
;
7249 mddev
->thread
= conf
->thread
;
7250 conf
->thread
= NULL
;
7251 mddev
->private = conf
;
7253 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
7255 rdev
= conf
->disks
[i
].rdev
;
7256 if (!rdev
&& conf
->disks
[i
].replacement
) {
7257 /* The replacement is all we have yet */
7258 rdev
= conf
->disks
[i
].replacement
;
7259 conf
->disks
[i
].replacement
= NULL
;
7260 clear_bit(Replacement
, &rdev
->flags
);
7261 conf
->disks
[i
].rdev
= rdev
;
7265 if (conf
->disks
[i
].replacement
&&
7266 conf
->reshape_progress
!= MaxSector
) {
7267 /* replacements and reshape simply do not mix. */
7268 pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7271 if (test_bit(In_sync
, &rdev
->flags
)) {
7275 /* This disc is not fully in-sync. However if it
7276 * just stored parity (beyond the recovery_offset),
7277 * when we don't need to be concerned about the
7278 * array being dirty.
7279 * When reshape goes 'backwards', we never have
7280 * partially completed devices, so we only need
7281 * to worry about reshape going forwards.
7283 /* Hack because v0.91 doesn't store recovery_offset properly. */
7284 if (mddev
->major_version
== 0 &&
7285 mddev
->minor_version
> 90)
7286 rdev
->recovery_offset
= reshape_offset
;
7288 if (rdev
->recovery_offset
< reshape_offset
) {
7289 /* We need to check old and new layout */
7290 if (!only_parity(rdev
->raid_disk
,
7293 conf
->max_degraded
))
7296 if (!only_parity(rdev
->raid_disk
,
7298 conf
->previous_raid_disks
,
7299 conf
->max_degraded
))
7301 dirty_parity_disks
++;
7305 * 0 for a fully functional array, 1 or 2 for a degraded array.
7307 mddev
->degraded
= raid5_calc_degraded(conf
);
7309 if (has_failed(conf
)) {
7310 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7311 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
7315 /* device size must be a multiple of chunk size */
7316 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
7317 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
7319 if (mddev
->degraded
> dirty_parity_disks
&&
7320 mddev
->recovery_cp
!= MaxSector
) {
7321 if (test_bit(MD_HAS_PPL
, &mddev
->flags
))
7322 pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7324 else if (mddev
->ok_start_degraded
)
7325 pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7328 pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7334 pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7335 mdname(mddev
), conf
->level
,
7336 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
7339 print_raid5_conf(conf
);
7341 if (conf
->reshape_progress
!= MaxSector
) {
7342 conf
->reshape_safe
= conf
->reshape_progress
;
7343 atomic_set(&conf
->reshape_stripes
, 0);
7344 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
7345 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
7346 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
7347 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
7348 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
7352 /* Ok, everything is just fine now */
7353 if (mddev
->to_remove
== &raid5_attrs_group
)
7354 mddev
->to_remove
= NULL
;
7355 else if (mddev
->kobj
.sd
&&
7356 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
7357 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7359 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
7363 /* read-ahead size must cover two whole stripes, which
7364 * is 2 * (datadisks) * chunksize where 'n' is the
7365 * number of raid devices
7367 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
7368 int stripe
= data_disks
*
7369 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
7370 if (mddev
->queue
->backing_dev_info
->ra_pages
< 2 * stripe
)
7371 mddev
->queue
->backing_dev_info
->ra_pages
= 2 * stripe
;
7373 chunk_size
= mddev
->chunk_sectors
<< 9;
7374 blk_queue_io_min(mddev
->queue
, chunk_size
);
7375 blk_queue_io_opt(mddev
->queue
, chunk_size
*
7376 (conf
->raid_disks
- conf
->max_degraded
));
7377 mddev
->queue
->limits
.raid_partial_stripes_expensive
= 1;
7379 * We can only discard a whole stripe. It doesn't make sense to
7380 * discard data disk but write parity disk
7382 stripe
= stripe
* PAGE_SIZE
;
7383 /* Round up to power of 2, as discard handling
7384 * currently assumes that */
7385 while ((stripe
-1) & stripe
)
7386 stripe
= (stripe
| (stripe
-1)) + 1;
7387 mddev
->queue
->limits
.discard_alignment
= stripe
;
7388 mddev
->queue
->limits
.discard_granularity
= stripe
;
7390 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
7391 blk_queue_max_write_zeroes_sectors(mddev
->queue
, 0);
7393 rdev_for_each(rdev
, mddev
) {
7394 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
7395 rdev
->data_offset
<< 9);
7396 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
7397 rdev
->new_data_offset
<< 9);
7401 * zeroing is required, otherwise data
7402 * could be lost. Consider a scenario: discard a stripe
7403 * (the stripe could be inconsistent if
7404 * discard_zeroes_data is 0); write one disk of the
7405 * stripe (the stripe could be inconsistent again
7406 * depending on which disks are used to calculate
7407 * parity); the disk is broken; The stripe data of this
7410 * We only allow DISCARD if the sysadmin has confirmed that
7411 * only safe devices are in use by setting a module parameter.
7412 * A better idea might be to turn DISCARD into WRITE_ZEROES
7413 * requests, as that is required to be safe.
7415 if (devices_handle_discard_safely
&&
7416 mddev
->queue
->limits
.max_discard_sectors
>= (stripe
>> 9) &&
7417 mddev
->queue
->limits
.discard_granularity
>= stripe
)
7418 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
7421 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
7424 blk_queue_max_hw_sectors(mddev
->queue
, UINT_MAX
);
7427 if (log_init(conf
, journal_dev
, raid5_has_ppl(conf
)))
7432 md_unregister_thread(&mddev
->thread
);
7433 print_raid5_conf(conf
);
7435 mddev
->private = NULL
;
7436 pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev
));
7440 static void raid5_free(struct mddev
*mddev
, void *priv
)
7442 struct r5conf
*conf
= priv
;
7445 mddev
->to_remove
= &raid5_attrs_group
;
7448 static void raid5_status(struct seq_file
*seq
, struct mddev
*mddev
)
7450 struct r5conf
*conf
= mddev
->private;
7453 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
7454 conf
->chunk_sectors
/ 2, mddev
->layout
);
7455 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
7457 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7458 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
7459 seq_printf (seq
, "%s", rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
7462 seq_printf (seq
, "]");
7465 static void print_raid5_conf (struct r5conf
*conf
)
7468 struct disk_info
*tmp
;
7470 pr_debug("RAID conf printout:\n");
7472 pr_debug("(conf==NULL)\n");
7475 pr_debug(" --- level:%d rd:%d wd:%d\n", conf
->level
,
7477 conf
->raid_disks
- conf
->mddev
->degraded
);
7479 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7480 char b
[BDEVNAME_SIZE
];
7481 tmp
= conf
->disks
+ i
;
7483 pr_debug(" disk %d, o:%d, dev:%s\n",
7484 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
7485 bdevname(tmp
->rdev
->bdev
, b
));
7489 static int raid5_spare_active(struct mddev
*mddev
)
7492 struct r5conf
*conf
= mddev
->private;
7493 struct disk_info
*tmp
;
7495 unsigned long flags
;
7497 for (i
= 0; i
< conf
->raid_disks
; i
++) {
7498 tmp
= conf
->disks
+ i
;
7499 if (tmp
->replacement
7500 && tmp
->replacement
->recovery_offset
== MaxSector
7501 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
7502 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
7503 /* Replacement has just become active. */
7505 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
7508 /* Replaced device not technically faulty,
7509 * but we need to be sure it gets removed
7510 * and never re-added.
7512 set_bit(Faulty
, &tmp
->rdev
->flags
);
7513 sysfs_notify_dirent_safe(
7514 tmp
->rdev
->sysfs_state
);
7516 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
7517 } else if (tmp
->rdev
7518 && tmp
->rdev
->recovery_offset
== MaxSector
7519 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
7520 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
7522 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
7525 spin_lock_irqsave(&conf
->device_lock
, flags
);
7526 mddev
->degraded
= raid5_calc_degraded(conf
);
7527 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7528 print_raid5_conf(conf
);
7532 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7534 struct r5conf
*conf
= mddev
->private;
7536 int number
= rdev
->raid_disk
;
7537 struct md_rdev
**rdevp
;
7538 struct disk_info
*p
= conf
->disks
+ number
;
7540 print_raid5_conf(conf
);
7541 if (test_bit(Journal
, &rdev
->flags
) && conf
->log
) {
7543 * we can't wait pending write here, as this is called in
7544 * raid5d, wait will deadlock.
7545 * neilb: there is no locking about new writes here,
7546 * so this cannot be safe.
7548 if (atomic_read(&conf
->active_stripes
) ||
7549 atomic_read(&conf
->r5c_cached_full_stripes
) ||
7550 atomic_read(&conf
->r5c_cached_partial_stripes
)) {
7556 if (rdev
== p
->rdev
)
7558 else if (rdev
== p
->replacement
)
7559 rdevp
= &p
->replacement
;
7563 if (number
>= conf
->raid_disks
&&
7564 conf
->reshape_progress
== MaxSector
)
7565 clear_bit(In_sync
, &rdev
->flags
);
7567 if (test_bit(In_sync
, &rdev
->flags
) ||
7568 atomic_read(&rdev
->nr_pending
)) {
7572 /* Only remove non-faulty devices if recovery
7575 if (!test_bit(Faulty
, &rdev
->flags
) &&
7576 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
7577 !has_failed(conf
) &&
7578 (!p
->replacement
|| p
->replacement
== rdev
) &&
7579 number
< conf
->raid_disks
) {
7584 if (!test_bit(RemoveSynchronized
, &rdev
->flags
)) {
7586 if (atomic_read(&rdev
->nr_pending
)) {
7587 /* lost the race, try later */
7593 err
= log_modify(conf
, rdev
, false);
7597 if (p
->replacement
) {
7598 /* We must have just cleared 'rdev' */
7599 p
->rdev
= p
->replacement
;
7600 clear_bit(Replacement
, &p
->replacement
->flags
);
7601 smp_mb(); /* Make sure other CPUs may see both as identical
7602 * but will never see neither - if they are careful
7604 p
->replacement
= NULL
;
7607 err
= log_modify(conf
, p
->rdev
, true);
7610 clear_bit(WantReplacement
, &rdev
->flags
);
7613 print_raid5_conf(conf
);
7617 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
7619 struct r5conf
*conf
= mddev
->private;
7622 struct disk_info
*p
;
7624 int last
= conf
->raid_disks
- 1;
7626 if (test_bit(Journal
, &rdev
->flags
)) {
7630 rdev
->raid_disk
= 0;
7632 * The array is in readonly mode if journal is missing, so no
7633 * write requests running. We should be safe
7635 log_init(conf
, rdev
, false);
7638 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
7641 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
7642 /* no point adding a device */
7645 if (rdev
->raid_disk
>= 0)
7646 first
= last
= rdev
->raid_disk
;
7649 * find the disk ... but prefer rdev->saved_raid_disk
7652 if (rdev
->saved_raid_disk
>= 0 &&
7653 rdev
->saved_raid_disk
>= first
&&
7654 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
7655 first
= rdev
->saved_raid_disk
;
7657 for (disk
= first
; disk
<= last
; disk
++) {
7658 p
= conf
->disks
+ disk
;
7659 if (p
->rdev
== NULL
) {
7660 clear_bit(In_sync
, &rdev
->flags
);
7661 rdev
->raid_disk
= disk
;
7662 if (rdev
->saved_raid_disk
!= disk
)
7664 rcu_assign_pointer(p
->rdev
, rdev
);
7666 err
= log_modify(conf
, rdev
, true);
7671 for (disk
= first
; disk
<= last
; disk
++) {
7672 p
= conf
->disks
+ disk
;
7673 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
7674 p
->replacement
== NULL
) {
7675 clear_bit(In_sync
, &rdev
->flags
);
7676 set_bit(Replacement
, &rdev
->flags
);
7677 rdev
->raid_disk
= disk
;
7680 rcu_assign_pointer(p
->replacement
, rdev
);
7685 print_raid5_conf(conf
);
7689 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
7691 /* no resync is happening, and there is enough space
7692 * on all devices, so we can resize.
7693 * We need to make sure resync covers any new space.
7694 * If the array is shrinking we should possibly wait until
7695 * any io in the removed space completes, but it hardly seems
7699 struct r5conf
*conf
= mddev
->private;
7701 if (conf
->log
|| raid5_has_ppl(conf
))
7703 sectors
&= ~((sector_t
)conf
->chunk_sectors
- 1);
7704 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
7705 if (mddev
->external_size
&&
7706 mddev
->array_sectors
> newsize
)
7708 if (mddev
->bitmap
) {
7709 int ret
= bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
7713 md_set_array_sectors(mddev
, newsize
);
7714 if (sectors
> mddev
->dev_sectors
&&
7715 mddev
->recovery_cp
> mddev
->dev_sectors
) {
7716 mddev
->recovery_cp
= mddev
->dev_sectors
;
7717 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
7719 mddev
->dev_sectors
= sectors
;
7720 mddev
->resync_max_sectors
= sectors
;
7724 static int check_stripe_cache(struct mddev
*mddev
)
7726 /* Can only proceed if there are plenty of stripe_heads.
7727 * We need a minimum of one full stripe,, and for sensible progress
7728 * it is best to have about 4 times that.
7729 * If we require 4 times, then the default 256 4K stripe_heads will
7730 * allow for chunk sizes up to 256K, which is probably OK.
7731 * If the chunk size is greater, user-space should request more
7732 * stripe_heads first.
7734 struct r5conf
*conf
= mddev
->private;
7735 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7736 > conf
->min_nr_stripes
||
7737 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
7738 > conf
->min_nr_stripes
) {
7739 pr_warn("md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7741 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
7748 static int check_reshape(struct mddev
*mddev
)
7750 struct r5conf
*conf
= mddev
->private;
7752 if (conf
->log
|| raid5_has_ppl(conf
))
7754 if (mddev
->delta_disks
== 0 &&
7755 mddev
->new_layout
== mddev
->layout
&&
7756 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
7757 return 0; /* nothing to do */
7758 if (has_failed(conf
))
7760 if (mddev
->delta_disks
< 0 && mddev
->reshape_position
== MaxSector
) {
7761 /* We might be able to shrink, but the devices must
7762 * be made bigger first.
7763 * For raid6, 4 is the minimum size.
7764 * Otherwise 2 is the minimum
7767 if (mddev
->level
== 6)
7769 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
7773 if (!check_stripe_cache(mddev
))
7776 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
||
7777 mddev
->delta_disks
> 0)
7778 if (resize_chunks(conf
,
7779 conf
->previous_raid_disks
7780 + max(0, mddev
->delta_disks
),
7781 max(mddev
->new_chunk_sectors
,
7782 mddev
->chunk_sectors
)
7786 if (conf
->previous_raid_disks
+ mddev
->delta_disks
<= conf
->pool_size
)
7787 return 0; /* never bother to shrink */
7788 return resize_stripes(conf
, (conf
->previous_raid_disks
7789 + mddev
->delta_disks
));
7792 static int raid5_start_reshape(struct mddev
*mddev
)
7794 struct r5conf
*conf
= mddev
->private;
7795 struct md_rdev
*rdev
;
7797 unsigned long flags
;
7799 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
7802 if (!check_stripe_cache(mddev
))
7805 if (has_failed(conf
))
7808 rdev_for_each(rdev
, mddev
) {
7809 if (!test_bit(In_sync
, &rdev
->flags
)
7810 && !test_bit(Faulty
, &rdev
->flags
))
7814 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
7815 /* Not enough devices even to make a degraded array
7820 /* Refuse to reduce size of the array. Any reductions in
7821 * array size must be through explicit setting of array_size
7824 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
7825 < mddev
->array_sectors
) {
7826 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7831 atomic_set(&conf
->reshape_stripes
, 0);
7832 spin_lock_irq(&conf
->device_lock
);
7833 write_seqcount_begin(&conf
->gen_lock
);
7834 conf
->previous_raid_disks
= conf
->raid_disks
;
7835 conf
->raid_disks
+= mddev
->delta_disks
;
7836 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
7837 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
7838 conf
->prev_algo
= conf
->algorithm
;
7839 conf
->algorithm
= mddev
->new_layout
;
7841 /* Code that selects data_offset needs to see the generation update
7842 * if reshape_progress has been set - so a memory barrier needed.
7845 if (mddev
->reshape_backwards
)
7846 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
7848 conf
->reshape_progress
= 0;
7849 conf
->reshape_safe
= conf
->reshape_progress
;
7850 write_seqcount_end(&conf
->gen_lock
);
7851 spin_unlock_irq(&conf
->device_lock
);
7853 /* Now make sure any requests that proceeded on the assumption
7854 * the reshape wasn't running - like Discard or Read - have
7857 mddev_suspend(mddev
);
7858 mddev_resume(mddev
);
7860 /* Add some new drives, as many as will fit.
7861 * We know there are enough to make the newly sized array work.
7862 * Don't add devices if we are reducing the number of
7863 * devices in the array. This is because it is not possible
7864 * to correctly record the "partially reconstructed" state of
7865 * such devices during the reshape and confusion could result.
7867 if (mddev
->delta_disks
>= 0) {
7868 rdev_for_each(rdev
, mddev
)
7869 if (rdev
->raid_disk
< 0 &&
7870 !test_bit(Faulty
, &rdev
->flags
)) {
7871 if (raid5_add_disk(mddev
, rdev
) == 0) {
7873 >= conf
->previous_raid_disks
)
7874 set_bit(In_sync
, &rdev
->flags
);
7876 rdev
->recovery_offset
= 0;
7878 if (sysfs_link_rdev(mddev
, rdev
))
7879 /* Failure here is OK */;
7881 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
7882 && !test_bit(Faulty
, &rdev
->flags
)) {
7883 /* This is a spare that was manually added */
7884 set_bit(In_sync
, &rdev
->flags
);
7887 /* When a reshape changes the number of devices,
7888 * ->degraded is measured against the larger of the
7889 * pre and post number of devices.
7891 spin_lock_irqsave(&conf
->device_lock
, flags
);
7892 mddev
->degraded
= raid5_calc_degraded(conf
);
7893 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
7895 mddev
->raid_disks
= conf
->raid_disks
;
7896 mddev
->reshape_position
= conf
->reshape_progress
;
7897 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
7899 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
7900 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
7901 clear_bit(MD_RECOVERY_DONE
, &mddev
->recovery
);
7902 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
7903 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
7904 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
7906 if (!mddev
->sync_thread
) {
7907 mddev
->recovery
= 0;
7908 spin_lock_irq(&conf
->device_lock
);
7909 write_seqcount_begin(&conf
->gen_lock
);
7910 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
7911 mddev
->new_chunk_sectors
=
7912 conf
->chunk_sectors
= conf
->prev_chunk_sectors
;
7913 mddev
->new_layout
= conf
->algorithm
= conf
->prev_algo
;
7914 rdev_for_each(rdev
, mddev
)
7915 rdev
->new_data_offset
= rdev
->data_offset
;
7917 conf
->generation
--;
7918 conf
->reshape_progress
= MaxSector
;
7919 mddev
->reshape_position
= MaxSector
;
7920 write_seqcount_end(&conf
->gen_lock
);
7921 spin_unlock_irq(&conf
->device_lock
);
7924 conf
->reshape_checkpoint
= jiffies
;
7925 md_wakeup_thread(mddev
->sync_thread
);
7926 md_new_event(mddev
);
7930 /* This is called from the reshape thread and should make any
7931 * changes needed in 'conf'
7933 static void end_reshape(struct r5conf
*conf
)
7936 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
7937 struct md_rdev
*rdev
;
7939 spin_lock_irq(&conf
->device_lock
);
7940 conf
->previous_raid_disks
= conf
->raid_disks
;
7941 rdev_for_each(rdev
, conf
->mddev
)
7942 rdev
->data_offset
= rdev
->new_data_offset
;
7944 conf
->reshape_progress
= MaxSector
;
7945 conf
->mddev
->reshape_position
= MaxSector
;
7946 spin_unlock_irq(&conf
->device_lock
);
7947 wake_up(&conf
->wait_for_overlap
);
7949 /* read-ahead size must cover two whole stripes, which is
7950 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7952 if (conf
->mddev
->queue
) {
7953 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
7954 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
7956 if (conf
->mddev
->queue
->backing_dev_info
->ra_pages
< 2 * stripe
)
7957 conf
->mddev
->queue
->backing_dev_info
->ra_pages
= 2 * stripe
;
7962 /* This is called from the raid5d thread with mddev_lock held.
7963 * It makes config changes to the device.
7965 static void raid5_finish_reshape(struct mddev
*mddev
)
7967 struct r5conf
*conf
= mddev
->private;
7969 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
7971 if (mddev
->delta_disks
> 0) {
7972 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
7974 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
7975 revalidate_disk(mddev
->gendisk
);
7979 spin_lock_irq(&conf
->device_lock
);
7980 mddev
->degraded
= raid5_calc_degraded(conf
);
7981 spin_unlock_irq(&conf
->device_lock
);
7982 for (d
= conf
->raid_disks
;
7983 d
< conf
->raid_disks
- mddev
->delta_disks
;
7985 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
7987 clear_bit(In_sync
, &rdev
->flags
);
7988 rdev
= conf
->disks
[d
].replacement
;
7990 clear_bit(In_sync
, &rdev
->flags
);
7993 mddev
->layout
= conf
->algorithm
;
7994 mddev
->chunk_sectors
= conf
->chunk_sectors
;
7995 mddev
->reshape_position
= MaxSector
;
7996 mddev
->delta_disks
= 0;
7997 mddev
->reshape_backwards
= 0;
8001 static void raid5_quiesce(struct mddev
*mddev
, int state
)
8003 struct r5conf
*conf
= mddev
->private;
8006 case 2: /* resume for a suspend */
8007 wake_up(&conf
->wait_for_overlap
);
8010 case 1: /* stop all writes */
8011 lock_all_device_hash_locks_irq(conf
);
8012 /* '2' tells resync/reshape to pause so that all
8013 * active stripes can drain
8015 r5c_flush_cache(conf
, INT_MAX
);
8017 wait_event_cmd(conf
->wait_for_quiescent
,
8018 atomic_read(&conf
->active_stripes
) == 0 &&
8019 atomic_read(&conf
->active_aligned_reads
) == 0,
8020 unlock_all_device_hash_locks_irq(conf
),
8021 lock_all_device_hash_locks_irq(conf
));
8023 unlock_all_device_hash_locks_irq(conf
);
8024 /* allow reshape to continue */
8025 wake_up(&conf
->wait_for_overlap
);
8028 case 0: /* re-enable writes */
8029 lock_all_device_hash_locks_irq(conf
);
8031 wake_up(&conf
->wait_for_quiescent
);
8032 wake_up(&conf
->wait_for_overlap
);
8033 unlock_all_device_hash_locks_irq(conf
);
8036 r5l_quiesce(conf
->log
, state
);
8039 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
8041 struct r0conf
*raid0_conf
= mddev
->private;
8044 /* for raid0 takeover only one zone is supported */
8045 if (raid0_conf
->nr_strip_zones
> 1) {
8046 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8048 return ERR_PTR(-EINVAL
);
8051 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
8052 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
8053 mddev
->dev_sectors
= sectors
;
8054 mddev
->new_level
= level
;
8055 mddev
->new_layout
= ALGORITHM_PARITY_N
;
8056 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
8057 mddev
->raid_disks
+= 1;
8058 mddev
->delta_disks
= 1;
8059 /* make sure it will be not marked as dirty */
8060 mddev
->recovery_cp
= MaxSector
;
8062 return setup_conf(mddev
);
8065 static void *raid5_takeover_raid1(struct mddev
*mddev
)
8070 if (mddev
->raid_disks
!= 2 ||
8071 mddev
->degraded
> 1)
8072 return ERR_PTR(-EINVAL
);
8074 /* Should check if there are write-behind devices? */
8076 chunksect
= 64*2; /* 64K by default */
8078 /* The array must be an exact multiple of chunksize */
8079 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
8082 if ((chunksect
<<9) < STRIPE_SIZE
)
8083 /* array size does not allow a suitable chunk size */
8084 return ERR_PTR(-EINVAL
);
8086 mddev
->new_level
= 5;
8087 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
8088 mddev
->new_chunk_sectors
= chunksect
;
8090 ret
= setup_conf(mddev
);
8092 mddev_clear_unsupported_flags(mddev
,
8093 UNSUPPORTED_MDDEV_FLAGS
);
8097 static void *raid5_takeover_raid6(struct mddev
*mddev
)
8101 switch (mddev
->layout
) {
8102 case ALGORITHM_LEFT_ASYMMETRIC_6
:
8103 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
8105 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
8106 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
8108 case ALGORITHM_LEFT_SYMMETRIC_6
:
8109 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
8111 case ALGORITHM_RIGHT_SYMMETRIC_6
:
8112 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
8114 case ALGORITHM_PARITY_0_6
:
8115 new_layout
= ALGORITHM_PARITY_0
;
8117 case ALGORITHM_PARITY_N
:
8118 new_layout
= ALGORITHM_PARITY_N
;
8121 return ERR_PTR(-EINVAL
);
8123 mddev
->new_level
= 5;
8124 mddev
->new_layout
= new_layout
;
8125 mddev
->delta_disks
= -1;
8126 mddev
->raid_disks
-= 1;
8127 return setup_conf(mddev
);
8130 static int raid5_check_reshape(struct mddev
*mddev
)
8132 /* For a 2-drive array, the layout and chunk size can be changed
8133 * immediately as not restriping is needed.
8134 * For larger arrays we record the new value - after validation
8135 * to be used by a reshape pass.
8137 struct r5conf
*conf
= mddev
->private;
8138 int new_chunk
= mddev
->new_chunk_sectors
;
8140 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
8142 if (new_chunk
> 0) {
8143 if (!is_power_of_2(new_chunk
))
8145 if (new_chunk
< (PAGE_SIZE
>>9))
8147 if (mddev
->array_sectors
& (new_chunk
-1))
8148 /* not factor of array size */
8152 /* They look valid */
8154 if (mddev
->raid_disks
== 2) {
8155 /* can make the change immediately */
8156 if (mddev
->new_layout
>= 0) {
8157 conf
->algorithm
= mddev
->new_layout
;
8158 mddev
->layout
= mddev
->new_layout
;
8160 if (new_chunk
> 0) {
8161 conf
->chunk_sectors
= new_chunk
;
8162 mddev
->chunk_sectors
= new_chunk
;
8164 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
8165 md_wakeup_thread(mddev
->thread
);
8167 return check_reshape(mddev
);
8170 static int raid6_check_reshape(struct mddev
*mddev
)
8172 int new_chunk
= mddev
->new_chunk_sectors
;
8174 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
8176 if (new_chunk
> 0) {
8177 if (!is_power_of_2(new_chunk
))
8179 if (new_chunk
< (PAGE_SIZE
>> 9))
8181 if (mddev
->array_sectors
& (new_chunk
-1))
8182 /* not factor of array size */
8186 /* They look valid */
8187 return check_reshape(mddev
);
8190 static void *raid5_takeover(struct mddev
*mddev
)
8192 /* raid5 can take over:
8193 * raid0 - if there is only one strip zone - make it a raid4 layout
8194 * raid1 - if there are two drives. We need to know the chunk size
8195 * raid4 - trivial - just use a raid4 layout.
8196 * raid6 - Providing it is a *_6 layout
8198 if (mddev
->level
== 0)
8199 return raid45_takeover_raid0(mddev
, 5);
8200 if (mddev
->level
== 1)
8201 return raid5_takeover_raid1(mddev
);
8202 if (mddev
->level
== 4) {
8203 mddev
->new_layout
= ALGORITHM_PARITY_N
;
8204 mddev
->new_level
= 5;
8205 return setup_conf(mddev
);
8207 if (mddev
->level
== 6)
8208 return raid5_takeover_raid6(mddev
);
8210 return ERR_PTR(-EINVAL
);
8213 static void *raid4_takeover(struct mddev
*mddev
)
8215 /* raid4 can take over:
8216 * raid0 - if there is only one strip zone
8217 * raid5 - if layout is right
8219 if (mddev
->level
== 0)
8220 return raid45_takeover_raid0(mddev
, 4);
8221 if (mddev
->level
== 5 &&
8222 mddev
->layout
== ALGORITHM_PARITY_N
) {
8223 mddev
->new_layout
= 0;
8224 mddev
->new_level
= 4;
8225 return setup_conf(mddev
);
8227 return ERR_PTR(-EINVAL
);
8230 static struct md_personality raid5_personality
;
8232 static void *raid6_takeover(struct mddev
*mddev
)
8234 /* Currently can only take over a raid5. We map the
8235 * personality to an equivalent raid6 personality
8236 * with the Q block at the end.
8240 if (mddev
->pers
!= &raid5_personality
)
8241 return ERR_PTR(-EINVAL
);
8242 if (mddev
->degraded
> 1)
8243 return ERR_PTR(-EINVAL
);
8244 if (mddev
->raid_disks
> 253)
8245 return ERR_PTR(-EINVAL
);
8246 if (mddev
->raid_disks
< 3)
8247 return ERR_PTR(-EINVAL
);
8249 switch (mddev
->layout
) {
8250 case ALGORITHM_LEFT_ASYMMETRIC
:
8251 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
8253 case ALGORITHM_RIGHT_ASYMMETRIC
:
8254 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
8256 case ALGORITHM_LEFT_SYMMETRIC
:
8257 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
8259 case ALGORITHM_RIGHT_SYMMETRIC
:
8260 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
8262 case ALGORITHM_PARITY_0
:
8263 new_layout
= ALGORITHM_PARITY_0_6
;
8265 case ALGORITHM_PARITY_N
:
8266 new_layout
= ALGORITHM_PARITY_N
;
8269 return ERR_PTR(-EINVAL
);
8271 mddev
->new_level
= 6;
8272 mddev
->new_layout
= new_layout
;
8273 mddev
->delta_disks
= 1;
8274 mddev
->raid_disks
+= 1;
8275 return setup_conf(mddev
);
8278 static int raid5_change_consistency_policy(struct mddev
*mddev
, const char *buf
)
8280 struct r5conf
*conf
;
8283 err
= mddev_lock(mddev
);
8286 conf
= mddev
->private;
8288 mddev_unlock(mddev
);
8292 if (strncmp(buf
, "ppl", 3) == 0) {
8293 /* ppl only works with RAID 5 */
8294 if (!raid5_has_ppl(conf
) && conf
->level
== 5) {
8295 err
= log_init(conf
, NULL
, true);
8297 err
= resize_stripes(conf
, conf
->pool_size
);
8303 } else if (strncmp(buf
, "resync", 6) == 0) {
8304 if (raid5_has_ppl(conf
)) {
8305 mddev_suspend(mddev
);
8307 mddev_resume(mddev
);
8308 err
= resize_stripes(conf
, conf
->pool_size
);
8309 } else if (test_bit(MD_HAS_JOURNAL
, &conf
->mddev
->flags
) &&
8310 r5l_log_disk_error(conf
)) {
8311 bool journal_dev_exists
= false;
8312 struct md_rdev
*rdev
;
8314 rdev_for_each(rdev
, mddev
)
8315 if (test_bit(Journal
, &rdev
->flags
)) {
8316 journal_dev_exists
= true;
8320 if (!journal_dev_exists
) {
8321 mddev_suspend(mddev
);
8322 clear_bit(MD_HAS_JOURNAL
, &mddev
->flags
);
8323 mddev_resume(mddev
);
8324 } else /* need remove journal device first */
8333 md_update_sb(mddev
, 1);
8335 mddev_unlock(mddev
);
8340 static struct md_personality raid6_personality
=
8344 .owner
= THIS_MODULE
,
8345 .make_request
= raid5_make_request
,
8348 .status
= raid5_status
,
8349 .error_handler
= raid5_error
,
8350 .hot_add_disk
= raid5_add_disk
,
8351 .hot_remove_disk
= raid5_remove_disk
,
8352 .spare_active
= raid5_spare_active
,
8353 .sync_request
= raid5_sync_request
,
8354 .resize
= raid5_resize
,
8356 .check_reshape
= raid6_check_reshape
,
8357 .start_reshape
= raid5_start_reshape
,
8358 .finish_reshape
= raid5_finish_reshape
,
8359 .quiesce
= raid5_quiesce
,
8360 .takeover
= raid6_takeover
,
8361 .congested
= raid5_congested
,
8362 .change_consistency_policy
= raid5_change_consistency_policy
,
8364 static struct md_personality raid5_personality
=
8368 .owner
= THIS_MODULE
,
8369 .make_request
= raid5_make_request
,
8372 .status
= raid5_status
,
8373 .error_handler
= raid5_error
,
8374 .hot_add_disk
= raid5_add_disk
,
8375 .hot_remove_disk
= raid5_remove_disk
,
8376 .spare_active
= raid5_spare_active
,
8377 .sync_request
= raid5_sync_request
,
8378 .resize
= raid5_resize
,
8380 .check_reshape
= raid5_check_reshape
,
8381 .start_reshape
= raid5_start_reshape
,
8382 .finish_reshape
= raid5_finish_reshape
,
8383 .quiesce
= raid5_quiesce
,
8384 .takeover
= raid5_takeover
,
8385 .congested
= raid5_congested
,
8386 .change_consistency_policy
= raid5_change_consistency_policy
,
8389 static struct md_personality raid4_personality
=
8393 .owner
= THIS_MODULE
,
8394 .make_request
= raid5_make_request
,
8397 .status
= raid5_status
,
8398 .error_handler
= raid5_error
,
8399 .hot_add_disk
= raid5_add_disk
,
8400 .hot_remove_disk
= raid5_remove_disk
,
8401 .spare_active
= raid5_spare_active
,
8402 .sync_request
= raid5_sync_request
,
8403 .resize
= raid5_resize
,
8405 .check_reshape
= raid5_check_reshape
,
8406 .start_reshape
= raid5_start_reshape
,
8407 .finish_reshape
= raid5_finish_reshape
,
8408 .quiesce
= raid5_quiesce
,
8409 .takeover
= raid4_takeover
,
8410 .congested
= raid5_congested
,
8411 .change_consistency_policy
= raid5_change_consistency_policy
,
8414 static int __init
raid5_init(void)
8418 raid5_wq
= alloc_workqueue("raid5wq",
8419 WQ_UNBOUND
|WQ_MEM_RECLAIM
|WQ_CPU_INTENSIVE
|WQ_SYSFS
, 0);
8423 ret
= cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE
,
8425 raid456_cpu_up_prepare
,
8428 destroy_workqueue(raid5_wq
);
8431 register_md_personality(&raid6_personality
);
8432 register_md_personality(&raid5_personality
);
8433 register_md_personality(&raid4_personality
);
8437 static void raid5_exit(void)
8439 unregister_md_personality(&raid6_personality
);
8440 unregister_md_personality(&raid5_personality
);
8441 unregister_md_personality(&raid4_personality
);
8442 cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE
);
8443 destroy_workqueue(raid5_wq
);
8446 module_init(raid5_init
);
8447 module_exit(raid5_exit
);
8448 MODULE_LICENSE("GPL");
8449 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8450 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8451 MODULE_ALIAS("md-raid5");
8452 MODULE_ALIAS("md-raid4");
8453 MODULE_ALIAS("md-level-5");
8454 MODULE_ALIAS("md-level-4");
8455 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8456 MODULE_ALIAS("md-raid6");
8457 MODULE_ALIAS("md-level-6");
8459 /* This used to be two separate modules, they were: */
8460 MODULE_ALIAS("raid5");
8461 MODULE_ALIAS("raid6");