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 <trace/events/block.h>
67 #define NR_STRIPES 256
68 #define STRIPE_SIZE PAGE_SIZE
69 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
70 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
71 #define IO_THRESHOLD 1
72 #define BYPASS_THRESHOLD 1
73 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
74 #define HASH_MASK (NR_HASH - 1)
76 static inline struct hlist_head
*stripe_hash(struct r5conf
*conf
, sector_t sect
)
78 int hash
= (sect
>> STRIPE_SHIFT
) & HASH_MASK
;
79 return &conf
->stripe_hashtbl
[hash
];
82 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
83 * order without overlap. There may be several bio's per stripe+device, and
84 * a bio could span several devices.
85 * When walking this list for a particular stripe+device, we must never proceed
86 * beyond a bio that extends past this device, as the next bio might no longer
88 * This function is used to determine the 'next' bio in the list, given the sector
89 * of the current stripe+device
91 static inline struct bio
*r5_next_bio(struct bio
*bio
, sector_t sector
)
93 int sectors
= bio
->bi_size
>> 9;
94 if (bio
->bi_sector
+ sectors
< sector
+ STRIPE_SECTORS
)
101 * We maintain a biased count of active stripes in the bottom 16 bits of
102 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
104 static inline int raid5_bi_processed_stripes(struct bio
*bio
)
106 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
107 return (atomic_read(segments
) >> 16) & 0xffff;
110 static inline int raid5_dec_bi_active_stripes(struct bio
*bio
)
112 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
113 return atomic_sub_return(1, segments
) & 0xffff;
116 static inline void raid5_inc_bi_active_stripes(struct bio
*bio
)
118 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
119 atomic_inc(segments
);
122 static inline void raid5_set_bi_processed_stripes(struct bio
*bio
,
125 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
129 old
= atomic_read(segments
);
130 new = (old
& 0xffff) | (cnt
<< 16);
131 } while (atomic_cmpxchg(segments
, old
, new) != old
);
134 static inline void raid5_set_bi_stripes(struct bio
*bio
, unsigned int cnt
)
136 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
137 atomic_set(segments
, cnt
);
140 /* Find first data disk in a raid6 stripe */
141 static inline int raid6_d0(struct stripe_head
*sh
)
144 /* ddf always start from first device */
146 /* md starts just after Q block */
147 if (sh
->qd_idx
== sh
->disks
- 1)
150 return sh
->qd_idx
+ 1;
152 static inline int raid6_next_disk(int disk
, int raid_disks
)
155 return (disk
< raid_disks
) ? disk
: 0;
158 /* When walking through the disks in a raid5, starting at raid6_d0,
159 * We need to map each disk to a 'slot', where the data disks are slot
160 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
161 * is raid_disks-1. This help does that mapping.
163 static int raid6_idx_to_slot(int idx
, struct stripe_head
*sh
,
164 int *count
, int syndrome_disks
)
170 if (idx
== sh
->pd_idx
)
171 return syndrome_disks
;
172 if (idx
== sh
->qd_idx
)
173 return syndrome_disks
+ 1;
179 static void return_io(struct bio
*return_bi
)
181 struct bio
*bi
= return_bi
;
184 return_bi
= bi
->bi_next
;
187 trace_block_bio_complete(bdev_get_queue(bi
->bi_bdev
),
194 static void print_raid5_conf (struct r5conf
*conf
);
196 static int stripe_operations_active(struct stripe_head
*sh
)
198 return sh
->check_state
|| sh
->reconstruct_state
||
199 test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
) ||
200 test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
203 static void do_release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
)
205 BUG_ON(!list_empty(&sh
->lru
));
206 BUG_ON(atomic_read(&conf
->active_stripes
)==0);
207 if (test_bit(STRIPE_HANDLE
, &sh
->state
)) {
208 if (test_bit(STRIPE_DELAYED
, &sh
->state
) &&
209 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
210 list_add_tail(&sh
->lru
, &conf
->delayed_list
);
211 else if (test_bit(STRIPE_BIT_DELAY
, &sh
->state
) &&
212 sh
->bm_seq
- conf
->seq_write
> 0)
213 list_add_tail(&sh
->lru
, &conf
->bitmap_list
);
215 clear_bit(STRIPE_DELAYED
, &sh
->state
);
216 clear_bit(STRIPE_BIT_DELAY
, &sh
->state
);
217 list_add_tail(&sh
->lru
, &conf
->handle_list
);
219 md_wakeup_thread(conf
->mddev
->thread
);
221 BUG_ON(stripe_operations_active(sh
));
222 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
223 if (atomic_dec_return(&conf
->preread_active_stripes
)
225 md_wakeup_thread(conf
->mddev
->thread
);
226 atomic_dec(&conf
->active_stripes
);
227 if (!test_bit(STRIPE_EXPANDING
, &sh
->state
)) {
228 list_add_tail(&sh
->lru
, &conf
->inactive_list
);
229 wake_up(&conf
->wait_for_stripe
);
230 if (conf
->retry_read_aligned
)
231 md_wakeup_thread(conf
->mddev
->thread
);
236 static void __release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
)
238 if (atomic_dec_and_test(&sh
->count
))
239 do_release_stripe(conf
, sh
);
242 static void release_stripe(struct stripe_head
*sh
)
244 struct r5conf
*conf
= sh
->raid_conf
;
247 local_irq_save(flags
);
248 if (atomic_dec_and_lock(&sh
->count
, &conf
->device_lock
)) {
249 do_release_stripe(conf
, sh
);
250 spin_unlock(&conf
->device_lock
);
252 local_irq_restore(flags
);
255 static inline void remove_hash(struct stripe_head
*sh
)
257 pr_debug("remove_hash(), stripe %llu\n",
258 (unsigned long long)sh
->sector
);
260 hlist_del_init(&sh
->hash
);
263 static inline void insert_hash(struct r5conf
*conf
, struct stripe_head
*sh
)
265 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
267 pr_debug("insert_hash(), stripe %llu\n",
268 (unsigned long long)sh
->sector
);
270 hlist_add_head(&sh
->hash
, hp
);
274 /* find an idle stripe, make sure it is unhashed, and return it. */
275 static struct stripe_head
*get_free_stripe(struct r5conf
*conf
)
277 struct stripe_head
*sh
= NULL
;
278 struct list_head
*first
;
280 if (list_empty(&conf
->inactive_list
))
282 first
= conf
->inactive_list
.next
;
283 sh
= list_entry(first
, struct stripe_head
, lru
);
284 list_del_init(first
);
286 atomic_inc(&conf
->active_stripes
);
291 static void shrink_buffers(struct stripe_head
*sh
)
295 int num
= sh
->raid_conf
->pool_size
;
297 for (i
= 0; i
< num
; i
++) {
301 sh
->dev
[i
].page
= NULL
;
306 static int grow_buffers(struct stripe_head
*sh
)
309 int num
= sh
->raid_conf
->pool_size
;
311 for (i
= 0; i
< num
; i
++) {
314 if (!(page
= alloc_page(GFP_KERNEL
))) {
317 sh
->dev
[i
].page
= page
;
322 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
);
323 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
324 struct stripe_head
*sh
);
326 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int previous
)
328 struct r5conf
*conf
= sh
->raid_conf
;
331 BUG_ON(atomic_read(&sh
->count
) != 0);
332 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
333 BUG_ON(stripe_operations_active(sh
));
335 pr_debug("init_stripe called, stripe %llu\n",
336 (unsigned long long)sh
->sector
);
340 sh
->generation
= conf
->generation
- previous
;
341 sh
->disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
343 stripe_set_idx(sector
, conf
, previous
, sh
);
347 for (i
= sh
->disks
; i
--; ) {
348 struct r5dev
*dev
= &sh
->dev
[i
];
350 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
351 test_bit(R5_LOCKED
, &dev
->flags
)) {
352 printk(KERN_ERR
"sector=%llx i=%d %p %p %p %p %d\n",
353 (unsigned long long)sh
->sector
, i
, dev
->toread
,
354 dev
->read
, dev
->towrite
, dev
->written
,
355 test_bit(R5_LOCKED
, &dev
->flags
));
359 raid5_build_block(sh
, i
, previous
);
361 insert_hash(conf
, sh
);
364 static struct stripe_head
*__find_stripe(struct r5conf
*conf
, sector_t sector
,
367 struct stripe_head
*sh
;
369 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
370 hlist_for_each_entry(sh
, stripe_hash(conf
, sector
), hash
)
371 if (sh
->sector
== sector
&& sh
->generation
== generation
)
373 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
378 * Need to check if array has failed when deciding whether to:
380 * - remove non-faulty devices
383 * This determination is simple when no reshape is happening.
384 * However if there is a reshape, we need to carefully check
385 * both the before and after sections.
386 * This is because some failed devices may only affect one
387 * of the two sections, and some non-in_sync devices may
388 * be insync in the section most affected by failed devices.
390 static int calc_degraded(struct r5conf
*conf
)
392 int degraded
, degraded2
;
397 for (i
= 0; i
< conf
->previous_raid_disks
; i
++) {
398 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
399 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
400 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
401 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
403 else if (test_bit(In_sync
, &rdev
->flags
))
406 /* not in-sync or faulty.
407 * If the reshape increases the number of devices,
408 * this is being recovered by the reshape, so
409 * this 'previous' section is not in_sync.
410 * If the number of devices is being reduced however,
411 * the device can only be part of the array if
412 * we are reverting a reshape, so this section will
415 if (conf
->raid_disks
>= conf
->previous_raid_disks
)
419 if (conf
->raid_disks
== conf
->previous_raid_disks
)
423 for (i
= 0; i
< conf
->raid_disks
; i
++) {
424 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
425 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
426 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
427 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
429 else if (test_bit(In_sync
, &rdev
->flags
))
432 /* not in-sync or faulty.
433 * If reshape increases the number of devices, this
434 * section has already been recovered, else it
435 * almost certainly hasn't.
437 if (conf
->raid_disks
<= conf
->previous_raid_disks
)
441 if (degraded2
> degraded
)
446 static int has_failed(struct r5conf
*conf
)
450 if (conf
->mddev
->reshape_position
== MaxSector
)
451 return conf
->mddev
->degraded
> conf
->max_degraded
;
453 degraded
= calc_degraded(conf
);
454 if (degraded
> conf
->max_degraded
)
459 static struct stripe_head
*
460 get_active_stripe(struct r5conf
*conf
, sector_t sector
,
461 int previous
, int noblock
, int noquiesce
)
463 struct stripe_head
*sh
;
465 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
467 spin_lock_irq(&conf
->device_lock
);
470 wait_event_lock_irq(conf
->wait_for_stripe
,
471 conf
->quiesce
== 0 || noquiesce
,
473 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
475 if (!conf
->inactive_blocked
)
476 sh
= get_free_stripe(conf
);
477 if (noblock
&& sh
== NULL
)
480 conf
->inactive_blocked
= 1;
481 wait_event_lock_irq(conf
->wait_for_stripe
,
482 !list_empty(&conf
->inactive_list
) &&
483 (atomic_read(&conf
->active_stripes
)
484 < (conf
->max_nr_stripes
*3/4)
485 || !conf
->inactive_blocked
),
487 conf
->inactive_blocked
= 0;
489 init_stripe(sh
, sector
, previous
);
491 if (atomic_read(&sh
->count
)) {
492 BUG_ON(!list_empty(&sh
->lru
)
493 && !test_bit(STRIPE_EXPANDING
, &sh
->state
)
494 && !test_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
));
496 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
497 atomic_inc(&conf
->active_stripes
);
498 if (list_empty(&sh
->lru
) &&
499 !test_bit(STRIPE_EXPANDING
, &sh
->state
))
501 list_del_init(&sh
->lru
);
504 } while (sh
== NULL
);
507 atomic_inc(&sh
->count
);
509 spin_unlock_irq(&conf
->device_lock
);
513 /* Determine if 'data_offset' or 'new_data_offset' should be used
514 * in this stripe_head.
516 static int use_new_offset(struct r5conf
*conf
, struct stripe_head
*sh
)
518 sector_t progress
= conf
->reshape_progress
;
519 /* Need a memory barrier to make sure we see the value
520 * of conf->generation, or ->data_offset that was set before
521 * reshape_progress was updated.
524 if (progress
== MaxSector
)
526 if (sh
->generation
== conf
->generation
- 1)
528 /* We are in a reshape, and this is a new-generation stripe,
529 * so use new_data_offset.
535 raid5_end_read_request(struct bio
*bi
, int error
);
537 raid5_end_write_request(struct bio
*bi
, int error
);
539 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
541 struct r5conf
*conf
= sh
->raid_conf
;
542 int i
, disks
= sh
->disks
;
546 for (i
= disks
; i
--; ) {
548 int replace_only
= 0;
549 struct bio
*bi
, *rbi
;
550 struct md_rdev
*rdev
, *rrdev
= NULL
;
551 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
552 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
556 if (test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
558 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
560 else if (test_and_clear_bit(R5_WantReplace
,
561 &sh
->dev
[i
].flags
)) {
566 if (test_and_clear_bit(R5_SyncIO
, &sh
->dev
[i
].flags
))
569 bi
= &sh
->dev
[i
].req
;
570 rbi
= &sh
->dev
[i
].rreq
; /* For writing to replacement */
575 bi
->bi_end_io
= raid5_end_write_request
;
576 rbi
->bi_end_io
= raid5_end_write_request
;
578 bi
->bi_end_io
= raid5_end_read_request
;
581 rrdev
= rcu_dereference(conf
->disks
[i
].replacement
);
582 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
583 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
592 /* We raced and saw duplicates */
595 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
) && rrdev
)
600 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
603 atomic_inc(&rdev
->nr_pending
);
604 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
607 atomic_inc(&rrdev
->nr_pending
);
610 /* We have already checked bad blocks for reads. Now
611 * need to check for writes. We never accept write errors
612 * on the replacement, so we don't to check rrdev.
614 while ((rw
& WRITE
) && rdev
&&
615 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
618 int bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
619 &first_bad
, &bad_sectors
);
624 set_bit(BlockedBadBlocks
, &rdev
->flags
);
625 if (!conf
->mddev
->external
&&
626 conf
->mddev
->flags
) {
627 /* It is very unlikely, but we might
628 * still need to write out the
629 * bad block log - better give it
631 md_check_recovery(conf
->mddev
);
634 * Because md_wait_for_blocked_rdev
635 * will dec nr_pending, we must
636 * increment it first.
638 atomic_inc(&rdev
->nr_pending
);
639 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
641 /* Acknowledged bad block - skip the write */
642 rdev_dec_pending(rdev
, conf
->mddev
);
648 if (s
->syncing
|| s
->expanding
|| s
->expanded
650 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
652 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
654 bi
->bi_bdev
= rdev
->bdev
;
655 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
656 __func__
, (unsigned long long)sh
->sector
,
658 atomic_inc(&sh
->count
);
659 if (use_new_offset(conf
, sh
))
660 bi
->bi_sector
= (sh
->sector
661 + rdev
->new_data_offset
);
663 bi
->bi_sector
= (sh
->sector
664 + rdev
->data_offset
);
665 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
666 bi
->bi_rw
|= REQ_FLUSH
;
668 bi
->bi_flags
= 1 << BIO_UPTODATE
;
670 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
671 bi
->bi_io_vec
[0].bv_offset
= 0;
672 bi
->bi_size
= STRIPE_SIZE
;
675 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
677 if (conf
->mddev
->gendisk
)
678 trace_block_bio_remap(bdev_get_queue(bi
->bi_bdev
),
679 bi
, disk_devt(conf
->mddev
->gendisk
),
681 generic_make_request(bi
);
684 if (s
->syncing
|| s
->expanding
|| s
->expanded
686 md_sync_acct(rrdev
->bdev
, STRIPE_SECTORS
);
688 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
690 rbi
->bi_bdev
= rrdev
->bdev
;
691 pr_debug("%s: for %llu schedule op %ld on "
692 "replacement disc %d\n",
693 __func__
, (unsigned long long)sh
->sector
,
695 atomic_inc(&sh
->count
);
696 if (use_new_offset(conf
, sh
))
697 rbi
->bi_sector
= (sh
->sector
698 + rrdev
->new_data_offset
);
700 rbi
->bi_sector
= (sh
->sector
701 + rrdev
->data_offset
);
702 rbi
->bi_flags
= 1 << BIO_UPTODATE
;
704 rbi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
705 rbi
->bi_io_vec
[0].bv_offset
= 0;
706 rbi
->bi_size
= STRIPE_SIZE
;
708 if (conf
->mddev
->gendisk
)
709 trace_block_bio_remap(bdev_get_queue(rbi
->bi_bdev
),
710 rbi
, disk_devt(conf
->mddev
->gendisk
),
712 generic_make_request(rbi
);
714 if (!rdev
&& !rrdev
) {
716 set_bit(STRIPE_DEGRADED
, &sh
->state
);
717 pr_debug("skip op %ld on disc %d for sector %llu\n",
718 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
719 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
720 set_bit(STRIPE_HANDLE
, &sh
->state
);
725 static struct dma_async_tx_descriptor
*
726 async_copy_data(int frombio
, struct bio
*bio
, struct page
*page
,
727 sector_t sector
, struct dma_async_tx_descriptor
*tx
)
730 struct page
*bio_page
;
733 struct async_submit_ctl submit
;
734 enum async_tx_flags flags
= 0;
736 if (bio
->bi_sector
>= sector
)
737 page_offset
= (signed)(bio
->bi_sector
- sector
) * 512;
739 page_offset
= (signed)(sector
- bio
->bi_sector
) * -512;
742 flags
|= ASYNC_TX_FENCE
;
743 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
745 bio_for_each_segment(bvl
, bio
, i
) {
746 int len
= bvl
->bv_len
;
750 if (page_offset
< 0) {
751 b_offset
= -page_offset
;
752 page_offset
+= b_offset
;
756 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
757 clen
= STRIPE_SIZE
- page_offset
;
762 b_offset
+= bvl
->bv_offset
;
763 bio_page
= bvl
->bv_page
;
765 tx
= async_memcpy(page
, bio_page
, page_offset
,
766 b_offset
, clen
, &submit
);
768 tx
= async_memcpy(bio_page
, page
, b_offset
,
769 page_offset
, clen
, &submit
);
771 /* chain the operations */
772 submit
.depend_tx
= tx
;
774 if (clen
< len
) /* hit end of page */
782 static void ops_complete_biofill(void *stripe_head_ref
)
784 struct stripe_head
*sh
= stripe_head_ref
;
785 struct bio
*return_bi
= NULL
;
788 pr_debug("%s: stripe %llu\n", __func__
,
789 (unsigned long long)sh
->sector
);
791 /* clear completed biofills */
792 for (i
= sh
->disks
; i
--; ) {
793 struct r5dev
*dev
= &sh
->dev
[i
];
795 /* acknowledge completion of a biofill operation */
796 /* and check if we need to reply to a read request,
797 * new R5_Wantfill requests are held off until
798 * !STRIPE_BIOFILL_RUN
800 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
801 struct bio
*rbi
, *rbi2
;
806 while (rbi
&& rbi
->bi_sector
<
807 dev
->sector
+ STRIPE_SECTORS
) {
808 rbi2
= r5_next_bio(rbi
, dev
->sector
);
809 if (!raid5_dec_bi_active_stripes(rbi
)) {
810 rbi
->bi_next
= return_bi
;
817 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
819 return_io(return_bi
);
821 set_bit(STRIPE_HANDLE
, &sh
->state
);
825 static void ops_run_biofill(struct stripe_head
*sh
)
827 struct dma_async_tx_descriptor
*tx
= NULL
;
828 struct async_submit_ctl submit
;
831 pr_debug("%s: stripe %llu\n", __func__
,
832 (unsigned long long)sh
->sector
);
834 for (i
= sh
->disks
; i
--; ) {
835 struct r5dev
*dev
= &sh
->dev
[i
];
836 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
838 spin_lock_irq(&sh
->stripe_lock
);
839 dev
->read
= rbi
= dev
->toread
;
841 spin_unlock_irq(&sh
->stripe_lock
);
842 while (rbi
&& rbi
->bi_sector
<
843 dev
->sector
+ STRIPE_SECTORS
) {
844 tx
= async_copy_data(0, rbi
, dev
->page
,
846 rbi
= r5_next_bio(rbi
, dev
->sector
);
851 atomic_inc(&sh
->count
);
852 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
853 async_trigger_callback(&submit
);
856 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
863 tgt
= &sh
->dev
[target
];
864 set_bit(R5_UPTODATE
, &tgt
->flags
);
865 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
866 clear_bit(R5_Wantcompute
, &tgt
->flags
);
869 static void ops_complete_compute(void *stripe_head_ref
)
871 struct stripe_head
*sh
= stripe_head_ref
;
873 pr_debug("%s: stripe %llu\n", __func__
,
874 (unsigned long long)sh
->sector
);
876 /* mark the computed target(s) as uptodate */
877 mark_target_uptodate(sh
, sh
->ops
.target
);
878 mark_target_uptodate(sh
, sh
->ops
.target2
);
880 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
881 if (sh
->check_state
== check_state_compute_run
)
882 sh
->check_state
= check_state_compute_result
;
883 set_bit(STRIPE_HANDLE
, &sh
->state
);
887 /* return a pointer to the address conversion region of the scribble buffer */
888 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
889 struct raid5_percpu
*percpu
)
891 return percpu
->scribble
+ sizeof(struct page
*) * (sh
->disks
+ 2);
894 static struct dma_async_tx_descriptor
*
895 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
897 int disks
= sh
->disks
;
898 struct page
**xor_srcs
= percpu
->scribble
;
899 int target
= sh
->ops
.target
;
900 struct r5dev
*tgt
= &sh
->dev
[target
];
901 struct page
*xor_dest
= tgt
->page
;
903 struct dma_async_tx_descriptor
*tx
;
904 struct async_submit_ctl submit
;
907 pr_debug("%s: stripe %llu block: %d\n",
908 __func__
, (unsigned long long)sh
->sector
, target
);
909 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
911 for (i
= disks
; i
--; )
913 xor_srcs
[count
++] = sh
->dev
[i
].page
;
915 atomic_inc(&sh
->count
);
917 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
918 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
));
919 if (unlikely(count
== 1))
920 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
922 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
927 /* set_syndrome_sources - populate source buffers for gen_syndrome
928 * @srcs - (struct page *) array of size sh->disks
929 * @sh - stripe_head to parse
931 * Populates srcs in proper layout order for the stripe and returns the
932 * 'count' of sources to be used in a call to async_gen_syndrome. The P
933 * destination buffer is recorded in srcs[count] and the Q destination
934 * is recorded in srcs[count+1]].
936 static int set_syndrome_sources(struct page
**srcs
, struct stripe_head
*sh
)
938 int disks
= sh
->disks
;
939 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
940 int d0_idx
= raid6_d0(sh
);
944 for (i
= 0; i
< disks
; i
++)
950 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
952 srcs
[slot
] = sh
->dev
[i
].page
;
953 i
= raid6_next_disk(i
, disks
);
954 } while (i
!= d0_idx
);
956 return syndrome_disks
;
959 static struct dma_async_tx_descriptor
*
960 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
962 int disks
= sh
->disks
;
963 struct page
**blocks
= percpu
->scribble
;
965 int qd_idx
= sh
->qd_idx
;
966 struct dma_async_tx_descriptor
*tx
;
967 struct async_submit_ctl submit
;
973 if (sh
->ops
.target
< 0)
974 target
= sh
->ops
.target2
;
975 else if (sh
->ops
.target2
< 0)
976 target
= sh
->ops
.target
;
978 /* we should only have one valid target */
981 pr_debug("%s: stripe %llu block: %d\n",
982 __func__
, (unsigned long long)sh
->sector
, target
);
984 tgt
= &sh
->dev
[target
];
985 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
988 atomic_inc(&sh
->count
);
990 if (target
== qd_idx
) {
991 count
= set_syndrome_sources(blocks
, sh
);
992 blocks
[count
] = NULL
; /* regenerating p is not necessary */
993 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
994 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
995 ops_complete_compute
, sh
,
996 to_addr_conv(sh
, percpu
));
997 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
999 /* Compute any data- or p-drive using XOR */
1001 for (i
= disks
; i
-- ; ) {
1002 if (i
== target
|| i
== qd_idx
)
1004 blocks
[count
++] = sh
->dev
[i
].page
;
1007 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1008 NULL
, ops_complete_compute
, sh
,
1009 to_addr_conv(sh
, percpu
));
1010 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
1016 static struct dma_async_tx_descriptor
*
1017 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1019 int i
, count
, disks
= sh
->disks
;
1020 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1021 int d0_idx
= raid6_d0(sh
);
1022 int faila
= -1, failb
= -1;
1023 int target
= sh
->ops
.target
;
1024 int target2
= sh
->ops
.target2
;
1025 struct r5dev
*tgt
= &sh
->dev
[target
];
1026 struct r5dev
*tgt2
= &sh
->dev
[target2
];
1027 struct dma_async_tx_descriptor
*tx
;
1028 struct page
**blocks
= percpu
->scribble
;
1029 struct async_submit_ctl submit
;
1031 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1032 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
1033 BUG_ON(target
< 0 || target2
< 0);
1034 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1035 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
1037 /* we need to open-code set_syndrome_sources to handle the
1038 * slot number conversion for 'faila' and 'failb'
1040 for (i
= 0; i
< disks
; i
++)
1045 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1047 blocks
[slot
] = sh
->dev
[i
].page
;
1053 i
= raid6_next_disk(i
, disks
);
1054 } while (i
!= d0_idx
);
1056 BUG_ON(faila
== failb
);
1059 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1060 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1062 atomic_inc(&sh
->count
);
1064 if (failb
== syndrome_disks
+1) {
1065 /* Q disk is one of the missing disks */
1066 if (faila
== syndrome_disks
) {
1067 /* Missing P+Q, just recompute */
1068 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1069 ops_complete_compute
, sh
,
1070 to_addr_conv(sh
, percpu
));
1071 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
1072 STRIPE_SIZE
, &submit
);
1076 int qd_idx
= sh
->qd_idx
;
1078 /* Missing D+Q: recompute D from P, then recompute Q */
1079 if (target
== qd_idx
)
1080 data_target
= target2
;
1082 data_target
= target
;
1085 for (i
= disks
; i
-- ; ) {
1086 if (i
== data_target
|| i
== qd_idx
)
1088 blocks
[count
++] = sh
->dev
[i
].page
;
1090 dest
= sh
->dev
[data_target
].page
;
1091 init_async_submit(&submit
,
1092 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1094 to_addr_conv(sh
, percpu
));
1095 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
1098 count
= set_syndrome_sources(blocks
, sh
);
1099 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1100 ops_complete_compute
, sh
,
1101 to_addr_conv(sh
, percpu
));
1102 return async_gen_syndrome(blocks
, 0, count
+2,
1103 STRIPE_SIZE
, &submit
);
1106 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1107 ops_complete_compute
, sh
,
1108 to_addr_conv(sh
, percpu
));
1109 if (failb
== syndrome_disks
) {
1110 /* We're missing D+P. */
1111 return async_raid6_datap_recov(syndrome_disks
+2,
1115 /* We're missing D+D. */
1116 return async_raid6_2data_recov(syndrome_disks
+2,
1117 STRIPE_SIZE
, faila
, failb
,
1124 static void ops_complete_prexor(void *stripe_head_ref
)
1126 struct stripe_head
*sh
= stripe_head_ref
;
1128 pr_debug("%s: stripe %llu\n", __func__
,
1129 (unsigned long long)sh
->sector
);
1132 static struct dma_async_tx_descriptor
*
1133 ops_run_prexor(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1134 struct dma_async_tx_descriptor
*tx
)
1136 int disks
= sh
->disks
;
1137 struct page
**xor_srcs
= percpu
->scribble
;
1138 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1139 struct async_submit_ctl submit
;
1141 /* existing parity data subtracted */
1142 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1144 pr_debug("%s: stripe %llu\n", __func__
,
1145 (unsigned long long)sh
->sector
);
1147 for (i
= disks
; i
--; ) {
1148 struct r5dev
*dev
= &sh
->dev
[i
];
1149 /* Only process blocks that are known to be uptodate */
1150 if (test_bit(R5_Wantdrain
, &dev
->flags
))
1151 xor_srcs
[count
++] = dev
->page
;
1154 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1155 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
));
1156 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1161 static struct dma_async_tx_descriptor
*
1162 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1164 int disks
= sh
->disks
;
1167 pr_debug("%s: stripe %llu\n", __func__
,
1168 (unsigned long long)sh
->sector
);
1170 for (i
= disks
; i
--; ) {
1171 struct r5dev
*dev
= &sh
->dev
[i
];
1174 if (test_and_clear_bit(R5_Wantdrain
, &dev
->flags
)) {
1177 spin_lock_irq(&sh
->stripe_lock
);
1178 chosen
= dev
->towrite
;
1179 dev
->towrite
= NULL
;
1180 BUG_ON(dev
->written
);
1181 wbi
= dev
->written
= chosen
;
1182 spin_unlock_irq(&sh
->stripe_lock
);
1184 while (wbi
&& wbi
->bi_sector
<
1185 dev
->sector
+ STRIPE_SECTORS
) {
1186 if (wbi
->bi_rw
& REQ_FUA
)
1187 set_bit(R5_WantFUA
, &dev
->flags
);
1188 if (wbi
->bi_rw
& REQ_SYNC
)
1189 set_bit(R5_SyncIO
, &dev
->flags
);
1190 if (wbi
->bi_rw
& REQ_DISCARD
)
1191 set_bit(R5_Discard
, &dev
->flags
);
1193 tx
= async_copy_data(1, wbi
, dev
->page
,
1195 wbi
= r5_next_bio(wbi
, dev
->sector
);
1203 static void ops_complete_reconstruct(void *stripe_head_ref
)
1205 struct stripe_head
*sh
= stripe_head_ref
;
1206 int disks
= sh
->disks
;
1207 int pd_idx
= sh
->pd_idx
;
1208 int qd_idx
= sh
->qd_idx
;
1210 bool fua
= false, sync
= false, discard
= false;
1212 pr_debug("%s: stripe %llu\n", __func__
,
1213 (unsigned long long)sh
->sector
);
1215 for (i
= disks
; i
--; ) {
1216 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1217 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
1218 discard
|= test_bit(R5_Discard
, &sh
->dev
[i
].flags
);
1221 for (i
= disks
; i
--; ) {
1222 struct r5dev
*dev
= &sh
->dev
[i
];
1224 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1226 set_bit(R5_UPTODATE
, &dev
->flags
);
1228 set_bit(R5_WantFUA
, &dev
->flags
);
1230 set_bit(R5_SyncIO
, &dev
->flags
);
1234 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1235 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1236 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1237 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1239 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1240 sh
->reconstruct_state
= reconstruct_state_result
;
1243 set_bit(STRIPE_HANDLE
, &sh
->state
);
1248 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1249 struct dma_async_tx_descriptor
*tx
)
1251 int disks
= sh
->disks
;
1252 struct page
**xor_srcs
= percpu
->scribble
;
1253 struct async_submit_ctl submit
;
1254 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1255 struct page
*xor_dest
;
1257 unsigned long flags
;
1259 pr_debug("%s: stripe %llu\n", __func__
,
1260 (unsigned long long)sh
->sector
);
1262 for (i
= 0; i
< sh
->disks
; i
++) {
1265 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1268 if (i
>= sh
->disks
) {
1269 atomic_inc(&sh
->count
);
1270 set_bit(R5_Discard
, &sh
->dev
[pd_idx
].flags
);
1271 ops_complete_reconstruct(sh
);
1274 /* check if prexor is active which means only process blocks
1275 * that are part of a read-modify-write (written)
1277 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1279 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1280 for (i
= disks
; i
--; ) {
1281 struct r5dev
*dev
= &sh
->dev
[i
];
1283 xor_srcs
[count
++] = dev
->page
;
1286 xor_dest
= sh
->dev
[pd_idx
].page
;
1287 for (i
= disks
; i
--; ) {
1288 struct r5dev
*dev
= &sh
->dev
[i
];
1290 xor_srcs
[count
++] = dev
->page
;
1294 /* 1/ if we prexor'd then the dest is reused as a source
1295 * 2/ if we did not prexor then we are redoing the parity
1296 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1297 * for the synchronous xor case
1299 flags
= ASYNC_TX_ACK
|
1300 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1302 atomic_inc(&sh
->count
);
1304 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, sh
,
1305 to_addr_conv(sh
, percpu
));
1306 if (unlikely(count
== 1))
1307 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1309 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1313 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1314 struct dma_async_tx_descriptor
*tx
)
1316 struct async_submit_ctl submit
;
1317 struct page
**blocks
= percpu
->scribble
;
1320 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1322 for (i
= 0; i
< sh
->disks
; i
++) {
1323 if (sh
->pd_idx
== i
|| sh
->qd_idx
== i
)
1325 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1328 if (i
>= sh
->disks
) {
1329 atomic_inc(&sh
->count
);
1330 set_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
1331 set_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
1332 ops_complete_reconstruct(sh
);
1336 count
= set_syndrome_sources(blocks
, sh
);
1338 atomic_inc(&sh
->count
);
1340 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_reconstruct
,
1341 sh
, to_addr_conv(sh
, percpu
));
1342 async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1345 static void ops_complete_check(void *stripe_head_ref
)
1347 struct stripe_head
*sh
= stripe_head_ref
;
1349 pr_debug("%s: stripe %llu\n", __func__
,
1350 (unsigned long long)sh
->sector
);
1352 sh
->check_state
= check_state_check_result
;
1353 set_bit(STRIPE_HANDLE
, &sh
->state
);
1357 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1359 int disks
= sh
->disks
;
1360 int pd_idx
= sh
->pd_idx
;
1361 int qd_idx
= sh
->qd_idx
;
1362 struct page
*xor_dest
;
1363 struct page
**xor_srcs
= percpu
->scribble
;
1364 struct dma_async_tx_descriptor
*tx
;
1365 struct async_submit_ctl submit
;
1369 pr_debug("%s: stripe %llu\n", __func__
,
1370 (unsigned long long)sh
->sector
);
1373 xor_dest
= sh
->dev
[pd_idx
].page
;
1374 xor_srcs
[count
++] = xor_dest
;
1375 for (i
= disks
; i
--; ) {
1376 if (i
== pd_idx
|| i
== qd_idx
)
1378 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1381 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1382 to_addr_conv(sh
, percpu
));
1383 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1384 &sh
->ops
.zero_sum_result
, &submit
);
1386 atomic_inc(&sh
->count
);
1387 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1388 tx
= async_trigger_callback(&submit
);
1391 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1393 struct page
**srcs
= percpu
->scribble
;
1394 struct async_submit_ctl submit
;
1397 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1398 (unsigned long long)sh
->sector
, checkp
);
1400 count
= set_syndrome_sources(srcs
, sh
);
1404 atomic_inc(&sh
->count
);
1405 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1406 sh
, to_addr_conv(sh
, percpu
));
1407 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1408 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1411 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1413 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1414 struct dma_async_tx_descriptor
*tx
= NULL
;
1415 struct r5conf
*conf
= sh
->raid_conf
;
1416 int level
= conf
->level
;
1417 struct raid5_percpu
*percpu
;
1421 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1422 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1423 ops_run_biofill(sh
);
1427 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1429 tx
= ops_run_compute5(sh
, percpu
);
1431 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
1432 tx
= ops_run_compute6_1(sh
, percpu
);
1434 tx
= ops_run_compute6_2(sh
, percpu
);
1436 /* terminate the chain if reconstruct is not set to be run */
1437 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
1441 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
))
1442 tx
= ops_run_prexor(sh
, percpu
, tx
);
1444 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
1445 tx
= ops_run_biodrain(sh
, tx
);
1449 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
1451 ops_run_reconstruct5(sh
, percpu
, tx
);
1453 ops_run_reconstruct6(sh
, percpu
, tx
);
1456 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
1457 if (sh
->check_state
== check_state_run
)
1458 ops_run_check_p(sh
, percpu
);
1459 else if (sh
->check_state
== check_state_run_q
)
1460 ops_run_check_pq(sh
, percpu
, 0);
1461 else if (sh
->check_state
== check_state_run_pq
)
1462 ops_run_check_pq(sh
, percpu
, 1);
1468 for (i
= disks
; i
--; ) {
1469 struct r5dev
*dev
= &sh
->dev
[i
];
1470 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
1471 wake_up(&sh
->raid_conf
->wait_for_overlap
);
1476 static int grow_one_stripe(struct r5conf
*conf
)
1478 struct stripe_head
*sh
;
1479 sh
= kmem_cache_zalloc(conf
->slab_cache
, GFP_KERNEL
);
1483 sh
->raid_conf
= conf
;
1485 spin_lock_init(&sh
->stripe_lock
);
1487 if (grow_buffers(sh
)) {
1489 kmem_cache_free(conf
->slab_cache
, sh
);
1492 /* we just created an active stripe so... */
1493 atomic_set(&sh
->count
, 1);
1494 atomic_inc(&conf
->active_stripes
);
1495 INIT_LIST_HEAD(&sh
->lru
);
1500 static int grow_stripes(struct r5conf
*conf
, int num
)
1502 struct kmem_cache
*sc
;
1503 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
1505 if (conf
->mddev
->gendisk
)
1506 sprintf(conf
->cache_name
[0],
1507 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
1509 sprintf(conf
->cache_name
[0],
1510 "raid%d-%p", conf
->level
, conf
->mddev
);
1511 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
1513 conf
->active_name
= 0;
1514 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
1515 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
1519 conf
->slab_cache
= sc
;
1520 conf
->pool_size
= devs
;
1522 if (!grow_one_stripe(conf
))
1528 * scribble_len - return the required size of the scribble region
1529 * @num - total number of disks in the array
1531 * The size must be enough to contain:
1532 * 1/ a struct page pointer for each device in the array +2
1533 * 2/ room to convert each entry in (1) to its corresponding dma
1534 * (dma_map_page()) or page (page_address()) address.
1536 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1537 * calculate over all devices (not just the data blocks), using zeros in place
1538 * of the P and Q blocks.
1540 static size_t scribble_len(int num
)
1544 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
1549 static int resize_stripes(struct r5conf
*conf
, int newsize
)
1551 /* Make all the stripes able to hold 'newsize' devices.
1552 * New slots in each stripe get 'page' set to a new page.
1554 * This happens in stages:
1555 * 1/ create a new kmem_cache and allocate the required number of
1557 * 2/ gather all the old stripe_heads and transfer the pages across
1558 * to the new stripe_heads. This will have the side effect of
1559 * freezing the array as once all stripe_heads have been collected,
1560 * no IO will be possible. Old stripe heads are freed once their
1561 * pages have been transferred over, and the old kmem_cache is
1562 * freed when all stripes are done.
1563 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1564 * we simple return a failre status - no need to clean anything up.
1565 * 4/ allocate new pages for the new slots in the new stripe_heads.
1566 * If this fails, we don't bother trying the shrink the
1567 * stripe_heads down again, we just leave them as they are.
1568 * As each stripe_head is processed the new one is released into
1571 * Once step2 is started, we cannot afford to wait for a write,
1572 * so we use GFP_NOIO allocations.
1574 struct stripe_head
*osh
, *nsh
;
1575 LIST_HEAD(newstripes
);
1576 struct disk_info
*ndisks
;
1579 struct kmem_cache
*sc
;
1582 if (newsize
<= conf
->pool_size
)
1583 return 0; /* never bother to shrink */
1585 err
= md_allow_write(conf
->mddev
);
1590 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
1591 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
1596 for (i
= conf
->max_nr_stripes
; i
; i
--) {
1597 nsh
= kmem_cache_zalloc(sc
, GFP_KERNEL
);
1601 nsh
->raid_conf
= conf
;
1602 spin_lock_init(&nsh
->stripe_lock
);
1604 list_add(&nsh
->lru
, &newstripes
);
1607 /* didn't get enough, give up */
1608 while (!list_empty(&newstripes
)) {
1609 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1610 list_del(&nsh
->lru
);
1611 kmem_cache_free(sc
, nsh
);
1613 kmem_cache_destroy(sc
);
1616 /* Step 2 - Must use GFP_NOIO now.
1617 * OK, we have enough stripes, start collecting inactive
1618 * stripes and copying them over
1620 list_for_each_entry(nsh
, &newstripes
, lru
) {
1621 spin_lock_irq(&conf
->device_lock
);
1622 wait_event_lock_irq(conf
->wait_for_stripe
,
1623 !list_empty(&conf
->inactive_list
),
1625 osh
= get_free_stripe(conf
);
1626 spin_unlock_irq(&conf
->device_lock
);
1627 atomic_set(&nsh
->count
, 1);
1628 for(i
=0; i
<conf
->pool_size
; i
++)
1629 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
1630 for( ; i
<newsize
; i
++)
1631 nsh
->dev
[i
].page
= NULL
;
1632 kmem_cache_free(conf
->slab_cache
, osh
);
1634 kmem_cache_destroy(conf
->slab_cache
);
1637 * At this point, we are holding all the stripes so the array
1638 * is completely stalled, so now is a good time to resize
1639 * conf->disks and the scribble region
1641 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
1643 for (i
=0; i
<conf
->raid_disks
; i
++)
1644 ndisks
[i
] = conf
->disks
[i
];
1646 conf
->disks
= ndisks
;
1651 conf
->scribble_len
= scribble_len(newsize
);
1652 for_each_present_cpu(cpu
) {
1653 struct raid5_percpu
*percpu
;
1656 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1657 scribble
= kmalloc(conf
->scribble_len
, GFP_NOIO
);
1660 kfree(percpu
->scribble
);
1661 percpu
->scribble
= scribble
;
1669 /* Step 4, return new stripes to service */
1670 while(!list_empty(&newstripes
)) {
1671 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1672 list_del_init(&nsh
->lru
);
1674 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
1675 if (nsh
->dev
[i
].page
== NULL
) {
1676 struct page
*p
= alloc_page(GFP_NOIO
);
1677 nsh
->dev
[i
].page
= p
;
1681 release_stripe(nsh
);
1683 /* critical section pass, GFP_NOIO no longer needed */
1685 conf
->slab_cache
= sc
;
1686 conf
->active_name
= 1-conf
->active_name
;
1687 conf
->pool_size
= newsize
;
1691 static int drop_one_stripe(struct r5conf
*conf
)
1693 struct stripe_head
*sh
;
1695 spin_lock_irq(&conf
->device_lock
);
1696 sh
= get_free_stripe(conf
);
1697 spin_unlock_irq(&conf
->device_lock
);
1700 BUG_ON(atomic_read(&sh
->count
));
1702 kmem_cache_free(conf
->slab_cache
, sh
);
1703 atomic_dec(&conf
->active_stripes
);
1707 static void shrink_stripes(struct r5conf
*conf
)
1709 while (drop_one_stripe(conf
))
1712 if (conf
->slab_cache
)
1713 kmem_cache_destroy(conf
->slab_cache
);
1714 conf
->slab_cache
= NULL
;
1717 static void raid5_end_read_request(struct bio
* bi
, int error
)
1719 struct stripe_head
*sh
= bi
->bi_private
;
1720 struct r5conf
*conf
= sh
->raid_conf
;
1721 int disks
= sh
->disks
, i
;
1722 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1723 char b
[BDEVNAME_SIZE
];
1724 struct md_rdev
*rdev
= NULL
;
1727 for (i
=0 ; i
<disks
; i
++)
1728 if (bi
== &sh
->dev
[i
].req
)
1731 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1732 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1738 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
1739 /* If replacement finished while this request was outstanding,
1740 * 'replacement' might be NULL already.
1741 * In that case it moved down to 'rdev'.
1742 * rdev is not removed until all requests are finished.
1744 rdev
= conf
->disks
[i
].replacement
;
1746 rdev
= conf
->disks
[i
].rdev
;
1748 if (use_new_offset(conf
, sh
))
1749 s
= sh
->sector
+ rdev
->new_data_offset
;
1751 s
= sh
->sector
+ rdev
->data_offset
;
1753 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1754 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
1755 /* Note that this cannot happen on a
1756 * replacement device. We just fail those on
1761 "md/raid:%s: read error corrected"
1762 " (%lu sectors at %llu on %s)\n",
1763 mdname(conf
->mddev
), STRIPE_SECTORS
,
1764 (unsigned long long)s
,
1765 bdevname(rdev
->bdev
, b
));
1766 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
1767 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1768 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1769 } else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
1770 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
1772 if (atomic_read(&rdev
->read_errors
))
1773 atomic_set(&rdev
->read_errors
, 0);
1775 const char *bdn
= bdevname(rdev
->bdev
, b
);
1779 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1780 atomic_inc(&rdev
->read_errors
);
1781 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
1784 "md/raid:%s: read error on replacement device "
1785 "(sector %llu on %s).\n",
1786 mdname(conf
->mddev
),
1787 (unsigned long long)s
,
1789 else if (conf
->mddev
->degraded
>= conf
->max_degraded
) {
1793 "md/raid:%s: read error not correctable "
1794 "(sector %llu on %s).\n",
1795 mdname(conf
->mddev
),
1796 (unsigned long long)s
,
1798 } else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
)) {
1803 "md/raid:%s: read error NOT corrected!! "
1804 "(sector %llu on %s).\n",
1805 mdname(conf
->mddev
),
1806 (unsigned long long)s
,
1808 } else if (atomic_read(&rdev
->read_errors
)
1809 > conf
->max_nr_stripes
)
1811 "md/raid:%s: Too many read errors, failing device %s.\n",
1812 mdname(conf
->mddev
), bdn
);
1816 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
)) {
1817 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1818 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
1820 set_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
1822 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1823 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1825 && test_bit(In_sync
, &rdev
->flags
)
1826 && rdev_set_badblocks(
1827 rdev
, sh
->sector
, STRIPE_SECTORS
, 0)))
1828 md_error(conf
->mddev
, rdev
);
1831 rdev_dec_pending(rdev
, conf
->mddev
);
1832 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1833 set_bit(STRIPE_HANDLE
, &sh
->state
);
1837 static void raid5_end_write_request(struct bio
*bi
, int error
)
1839 struct stripe_head
*sh
= bi
->bi_private
;
1840 struct r5conf
*conf
= sh
->raid_conf
;
1841 int disks
= sh
->disks
, i
;
1842 struct md_rdev
*uninitialized_var(rdev
);
1843 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1846 int replacement
= 0;
1848 for (i
= 0 ; i
< disks
; i
++) {
1849 if (bi
== &sh
->dev
[i
].req
) {
1850 rdev
= conf
->disks
[i
].rdev
;
1853 if (bi
== &sh
->dev
[i
].rreq
) {
1854 rdev
= conf
->disks
[i
].replacement
;
1858 /* rdev was removed and 'replacement'
1859 * replaced it. rdev is not removed
1860 * until all requests are finished.
1862 rdev
= conf
->disks
[i
].rdev
;
1866 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1867 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1876 md_error(conf
->mddev
, rdev
);
1877 else if (is_badblock(rdev
, sh
->sector
,
1879 &first_bad
, &bad_sectors
))
1880 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
1883 set_bit(WriteErrorSeen
, &rdev
->flags
);
1884 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
1885 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1886 set_bit(MD_RECOVERY_NEEDED
,
1887 &rdev
->mddev
->recovery
);
1888 } else if (is_badblock(rdev
, sh
->sector
,
1890 &first_bad
, &bad_sectors
)) {
1891 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
1892 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))
1893 /* That was a successful write so make
1894 * sure it looks like we already did
1897 set_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1900 rdev_dec_pending(rdev
, conf
->mddev
);
1902 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
1903 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1904 set_bit(STRIPE_HANDLE
, &sh
->state
);
1908 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
);
1910 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
1912 struct r5dev
*dev
= &sh
->dev
[i
];
1914 bio_init(&dev
->req
);
1915 dev
->req
.bi_io_vec
= &dev
->vec
;
1917 dev
->req
.bi_max_vecs
++;
1918 dev
->req
.bi_private
= sh
;
1919 dev
->vec
.bv_page
= dev
->page
;
1921 bio_init(&dev
->rreq
);
1922 dev
->rreq
.bi_io_vec
= &dev
->rvec
;
1923 dev
->rreq
.bi_vcnt
++;
1924 dev
->rreq
.bi_max_vecs
++;
1925 dev
->rreq
.bi_private
= sh
;
1926 dev
->rvec
.bv_page
= dev
->page
;
1929 dev
->sector
= compute_blocknr(sh
, i
, previous
);
1932 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1934 char b
[BDEVNAME_SIZE
];
1935 struct r5conf
*conf
= mddev
->private;
1936 unsigned long flags
;
1937 pr_debug("raid456: error called\n");
1939 spin_lock_irqsave(&conf
->device_lock
, flags
);
1940 clear_bit(In_sync
, &rdev
->flags
);
1941 mddev
->degraded
= calc_degraded(conf
);
1942 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1943 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1945 set_bit(Blocked
, &rdev
->flags
);
1946 set_bit(Faulty
, &rdev
->flags
);
1947 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1949 "md/raid:%s: Disk failure on %s, disabling device.\n"
1950 "md/raid:%s: Operation continuing on %d devices.\n",
1952 bdevname(rdev
->bdev
, b
),
1954 conf
->raid_disks
- mddev
->degraded
);
1958 * Input: a 'big' sector number,
1959 * Output: index of the data and parity disk, and the sector # in them.
1961 static sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
1962 int previous
, int *dd_idx
,
1963 struct stripe_head
*sh
)
1965 sector_t stripe
, stripe2
;
1966 sector_t chunk_number
;
1967 unsigned int chunk_offset
;
1970 sector_t new_sector
;
1971 int algorithm
= previous
? conf
->prev_algo
1973 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
1974 : conf
->chunk_sectors
;
1975 int raid_disks
= previous
? conf
->previous_raid_disks
1977 int data_disks
= raid_disks
- conf
->max_degraded
;
1979 /* First compute the information on this sector */
1982 * Compute the chunk number and the sector offset inside the chunk
1984 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
1985 chunk_number
= r_sector
;
1988 * Compute the stripe number
1990 stripe
= chunk_number
;
1991 *dd_idx
= sector_div(stripe
, data_disks
);
1994 * Select the parity disk based on the user selected algorithm.
1996 pd_idx
= qd_idx
= -1;
1997 switch(conf
->level
) {
1999 pd_idx
= data_disks
;
2002 switch (algorithm
) {
2003 case ALGORITHM_LEFT_ASYMMETRIC
:
2004 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2005 if (*dd_idx
>= pd_idx
)
2008 case ALGORITHM_RIGHT_ASYMMETRIC
:
2009 pd_idx
= sector_div(stripe2
, raid_disks
);
2010 if (*dd_idx
>= pd_idx
)
2013 case ALGORITHM_LEFT_SYMMETRIC
:
2014 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2015 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2017 case ALGORITHM_RIGHT_SYMMETRIC
:
2018 pd_idx
= sector_div(stripe2
, raid_disks
);
2019 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2021 case ALGORITHM_PARITY_0
:
2025 case ALGORITHM_PARITY_N
:
2026 pd_idx
= data_disks
;
2034 switch (algorithm
) {
2035 case ALGORITHM_LEFT_ASYMMETRIC
:
2036 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2037 qd_idx
= pd_idx
+ 1;
2038 if (pd_idx
== raid_disks
-1) {
2039 (*dd_idx
)++; /* Q D D D P */
2041 } else if (*dd_idx
>= pd_idx
)
2042 (*dd_idx
) += 2; /* D D P Q D */
2044 case ALGORITHM_RIGHT_ASYMMETRIC
:
2045 pd_idx
= sector_div(stripe2
, raid_disks
);
2046 qd_idx
= pd_idx
+ 1;
2047 if (pd_idx
== raid_disks
-1) {
2048 (*dd_idx
)++; /* Q D D D P */
2050 } else if (*dd_idx
>= pd_idx
)
2051 (*dd_idx
) += 2; /* D D P Q D */
2053 case ALGORITHM_LEFT_SYMMETRIC
:
2054 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2055 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2056 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2058 case ALGORITHM_RIGHT_SYMMETRIC
:
2059 pd_idx
= sector_div(stripe2
, raid_disks
);
2060 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2061 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2064 case ALGORITHM_PARITY_0
:
2069 case ALGORITHM_PARITY_N
:
2070 pd_idx
= data_disks
;
2071 qd_idx
= data_disks
+ 1;
2074 case ALGORITHM_ROTATING_ZERO_RESTART
:
2075 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2076 * of blocks for computing Q is different.
2078 pd_idx
= sector_div(stripe2
, raid_disks
);
2079 qd_idx
= pd_idx
+ 1;
2080 if (pd_idx
== raid_disks
-1) {
2081 (*dd_idx
)++; /* Q D D D P */
2083 } else if (*dd_idx
>= pd_idx
)
2084 (*dd_idx
) += 2; /* D D P Q D */
2088 case ALGORITHM_ROTATING_N_RESTART
:
2089 /* Same a left_asymmetric, by first stripe is
2090 * D D D P Q rather than
2094 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2095 qd_idx
= pd_idx
+ 1;
2096 if (pd_idx
== raid_disks
-1) {
2097 (*dd_idx
)++; /* Q D D D P */
2099 } else if (*dd_idx
>= pd_idx
)
2100 (*dd_idx
) += 2; /* D D P Q D */
2104 case ALGORITHM_ROTATING_N_CONTINUE
:
2105 /* Same as left_symmetric but Q is before P */
2106 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2107 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
2108 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2112 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2113 /* RAID5 left_asymmetric, with Q on last device */
2114 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2115 if (*dd_idx
>= pd_idx
)
2117 qd_idx
= raid_disks
- 1;
2120 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2121 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2122 if (*dd_idx
>= pd_idx
)
2124 qd_idx
= raid_disks
- 1;
2127 case ALGORITHM_LEFT_SYMMETRIC_6
:
2128 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2129 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2130 qd_idx
= raid_disks
- 1;
2133 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2134 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2135 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2136 qd_idx
= raid_disks
- 1;
2139 case ALGORITHM_PARITY_0_6
:
2142 qd_idx
= raid_disks
- 1;
2152 sh
->pd_idx
= pd_idx
;
2153 sh
->qd_idx
= qd_idx
;
2154 sh
->ddf_layout
= ddf_layout
;
2157 * Finally, compute the new sector number
2159 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
2164 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
2166 struct r5conf
*conf
= sh
->raid_conf
;
2167 int raid_disks
= sh
->disks
;
2168 int data_disks
= raid_disks
- conf
->max_degraded
;
2169 sector_t new_sector
= sh
->sector
, check
;
2170 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2171 : conf
->chunk_sectors
;
2172 int algorithm
= previous
? conf
->prev_algo
2176 sector_t chunk_number
;
2177 int dummy1
, dd_idx
= i
;
2179 struct stripe_head sh2
;
2182 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
2183 stripe
= new_sector
;
2185 if (i
== sh
->pd_idx
)
2187 switch(conf
->level
) {
2190 switch (algorithm
) {
2191 case ALGORITHM_LEFT_ASYMMETRIC
:
2192 case ALGORITHM_RIGHT_ASYMMETRIC
:
2196 case ALGORITHM_LEFT_SYMMETRIC
:
2197 case ALGORITHM_RIGHT_SYMMETRIC
:
2200 i
-= (sh
->pd_idx
+ 1);
2202 case ALGORITHM_PARITY_0
:
2205 case ALGORITHM_PARITY_N
:
2212 if (i
== sh
->qd_idx
)
2213 return 0; /* It is the Q disk */
2214 switch (algorithm
) {
2215 case ALGORITHM_LEFT_ASYMMETRIC
:
2216 case ALGORITHM_RIGHT_ASYMMETRIC
:
2217 case ALGORITHM_ROTATING_ZERO_RESTART
:
2218 case ALGORITHM_ROTATING_N_RESTART
:
2219 if (sh
->pd_idx
== raid_disks
-1)
2220 i
--; /* Q D D D P */
2221 else if (i
> sh
->pd_idx
)
2222 i
-= 2; /* D D P Q D */
2224 case ALGORITHM_LEFT_SYMMETRIC
:
2225 case ALGORITHM_RIGHT_SYMMETRIC
:
2226 if (sh
->pd_idx
== raid_disks
-1)
2227 i
--; /* Q D D D P */
2232 i
-= (sh
->pd_idx
+ 2);
2235 case ALGORITHM_PARITY_0
:
2238 case ALGORITHM_PARITY_N
:
2240 case ALGORITHM_ROTATING_N_CONTINUE
:
2241 /* Like left_symmetric, but P is before Q */
2242 if (sh
->pd_idx
== 0)
2243 i
--; /* P D D D Q */
2248 i
-= (sh
->pd_idx
+ 1);
2251 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2252 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2256 case ALGORITHM_LEFT_SYMMETRIC_6
:
2257 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2259 i
+= data_disks
+ 1;
2260 i
-= (sh
->pd_idx
+ 1);
2262 case ALGORITHM_PARITY_0_6
:
2271 chunk_number
= stripe
* data_disks
+ i
;
2272 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2274 check
= raid5_compute_sector(conf
, r_sector
,
2275 previous
, &dummy1
, &sh2
);
2276 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2277 || sh2
.qd_idx
!= sh
->qd_idx
) {
2278 printk(KERN_ERR
"md/raid:%s: compute_blocknr: map not correct\n",
2279 mdname(conf
->mddev
));
2287 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2288 int rcw
, int expand
)
2290 int i
, pd_idx
= sh
->pd_idx
, disks
= sh
->disks
;
2291 struct r5conf
*conf
= sh
->raid_conf
;
2292 int level
= conf
->level
;
2296 for (i
= disks
; i
--; ) {
2297 struct r5dev
*dev
= &sh
->dev
[i
];
2300 set_bit(R5_LOCKED
, &dev
->flags
);
2301 set_bit(R5_Wantdrain
, &dev
->flags
);
2303 clear_bit(R5_UPTODATE
, &dev
->flags
);
2307 /* if we are not expanding this is a proper write request, and
2308 * there will be bios with new data to be drained into the
2313 /* False alarm, nothing to do */
2315 sh
->reconstruct_state
= reconstruct_state_drain_run
;
2316 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2318 sh
->reconstruct_state
= reconstruct_state_run
;
2320 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2322 if (s
->locked
+ conf
->max_degraded
== disks
)
2323 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2324 atomic_inc(&conf
->pending_full_writes
);
2327 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
2328 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
2330 for (i
= disks
; i
--; ) {
2331 struct r5dev
*dev
= &sh
->dev
[i
];
2336 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2337 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2338 set_bit(R5_Wantdrain
, &dev
->flags
);
2339 set_bit(R5_LOCKED
, &dev
->flags
);
2340 clear_bit(R5_UPTODATE
, &dev
->flags
);
2345 /* False alarm - nothing to do */
2347 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
2348 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
2349 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2350 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2353 /* keep the parity disk(s) locked while asynchronous operations
2356 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
2357 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2361 int qd_idx
= sh
->qd_idx
;
2362 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
2364 set_bit(R5_LOCKED
, &dev
->flags
);
2365 clear_bit(R5_UPTODATE
, &dev
->flags
);
2369 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2370 __func__
, (unsigned long long)sh
->sector
,
2371 s
->locked
, s
->ops_request
);
2375 * Each stripe/dev can have one or more bion attached.
2376 * toread/towrite point to the first in a chain.
2377 * The bi_next chain must be in order.
2379 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
, int forwrite
)
2382 struct r5conf
*conf
= sh
->raid_conf
;
2385 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2386 (unsigned long long)bi
->bi_sector
,
2387 (unsigned long long)sh
->sector
);
2390 * If several bio share a stripe. The bio bi_phys_segments acts as a
2391 * reference count to avoid race. The reference count should already be
2392 * increased before this function is called (for example, in
2393 * make_request()), so other bio sharing this stripe will not free the
2394 * stripe. If a stripe is owned by one stripe, the stripe lock will
2397 spin_lock_irq(&sh
->stripe_lock
);
2399 bip
= &sh
->dev
[dd_idx
].towrite
;
2403 bip
= &sh
->dev
[dd_idx
].toread
;
2404 while (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
) {
2405 if ((*bip
)->bi_sector
+ ((*bip
)->bi_size
>> 9) > bi
->bi_sector
)
2407 bip
= & (*bip
)->bi_next
;
2409 if (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
+ ((bi
->bi_size
)>>9))
2412 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
2416 raid5_inc_bi_active_stripes(bi
);
2419 /* check if page is covered */
2420 sector_t sector
= sh
->dev
[dd_idx
].sector
;
2421 for (bi
=sh
->dev
[dd_idx
].towrite
;
2422 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
2423 bi
&& bi
->bi_sector
<= sector
;
2424 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
2425 if (bi
->bi_sector
+ (bi
->bi_size
>>9) >= sector
)
2426 sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
2428 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
2429 set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
);
2432 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2433 (unsigned long long)(*bip
)->bi_sector
,
2434 (unsigned long long)sh
->sector
, dd_idx
);
2435 spin_unlock_irq(&sh
->stripe_lock
);
2437 if (conf
->mddev
->bitmap
&& firstwrite
) {
2438 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
2440 sh
->bm_seq
= conf
->seq_flush
+1;
2441 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
2446 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
2447 spin_unlock_irq(&sh
->stripe_lock
);
2451 static void end_reshape(struct r5conf
*conf
);
2453 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
2454 struct stripe_head
*sh
)
2456 int sectors_per_chunk
=
2457 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
2459 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
2460 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
2462 raid5_compute_sector(conf
,
2463 stripe
* (disks
- conf
->max_degraded
)
2464 *sectors_per_chunk
+ chunk_offset
,
2470 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
2471 struct stripe_head_state
*s
, int disks
,
2472 struct bio
**return_bi
)
2475 for (i
= disks
; i
--; ) {
2479 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2480 struct md_rdev
*rdev
;
2482 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2483 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
2484 atomic_inc(&rdev
->nr_pending
);
2489 if (!rdev_set_badblocks(
2493 md_error(conf
->mddev
, rdev
);
2494 rdev_dec_pending(rdev
, conf
->mddev
);
2497 spin_lock_irq(&sh
->stripe_lock
);
2498 /* fail all writes first */
2499 bi
= sh
->dev
[i
].towrite
;
2500 sh
->dev
[i
].towrite
= NULL
;
2501 spin_unlock_irq(&sh
->stripe_lock
);
2505 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2506 wake_up(&conf
->wait_for_overlap
);
2508 while (bi
&& bi
->bi_sector
<
2509 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2510 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2511 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2512 if (!raid5_dec_bi_active_stripes(bi
)) {
2513 md_write_end(conf
->mddev
);
2514 bi
->bi_next
= *return_bi
;
2520 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2521 STRIPE_SECTORS
, 0, 0);
2523 /* and fail all 'written' */
2524 bi
= sh
->dev
[i
].written
;
2525 sh
->dev
[i
].written
= NULL
;
2526 if (bi
) bitmap_end
= 1;
2527 while (bi
&& bi
->bi_sector
<
2528 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2529 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2530 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2531 if (!raid5_dec_bi_active_stripes(bi
)) {
2532 md_write_end(conf
->mddev
);
2533 bi
->bi_next
= *return_bi
;
2539 /* fail any reads if this device is non-operational and
2540 * the data has not reached the cache yet.
2542 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
2543 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
2544 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
2545 spin_lock_irq(&sh
->stripe_lock
);
2546 bi
= sh
->dev
[i
].toread
;
2547 sh
->dev
[i
].toread
= NULL
;
2548 spin_unlock_irq(&sh
->stripe_lock
);
2549 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2550 wake_up(&conf
->wait_for_overlap
);
2551 while (bi
&& bi
->bi_sector
<
2552 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2553 struct bio
*nextbi
=
2554 r5_next_bio(bi
, sh
->dev
[i
].sector
);
2555 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2556 if (!raid5_dec_bi_active_stripes(bi
)) {
2557 bi
->bi_next
= *return_bi
;
2564 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2565 STRIPE_SECTORS
, 0, 0);
2566 /* If we were in the middle of a write the parity block might
2567 * still be locked - so just clear all R5_LOCKED flags
2569 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2572 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2573 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2574 md_wakeup_thread(conf
->mddev
->thread
);
2578 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
2579 struct stripe_head_state
*s
)
2584 clear_bit(STRIPE_SYNCING
, &sh
->state
);
2585 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
2586 wake_up(&conf
->wait_for_overlap
);
2589 /* There is nothing more to do for sync/check/repair.
2590 * Don't even need to abort as that is handled elsewhere
2591 * if needed, and not always wanted e.g. if there is a known
2593 * For recover/replace we need to record a bad block on all
2594 * non-sync devices, or abort the recovery
2596 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
2597 /* During recovery devices cannot be removed, so
2598 * locking and refcounting of rdevs is not needed
2600 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2601 struct md_rdev
*rdev
= conf
->disks
[i
].rdev
;
2603 && !test_bit(Faulty
, &rdev
->flags
)
2604 && !test_bit(In_sync
, &rdev
->flags
)
2605 && !rdev_set_badblocks(rdev
, sh
->sector
,
2608 rdev
= conf
->disks
[i
].replacement
;
2610 && !test_bit(Faulty
, &rdev
->flags
)
2611 && !test_bit(In_sync
, &rdev
->flags
)
2612 && !rdev_set_badblocks(rdev
, sh
->sector
,
2617 conf
->recovery_disabled
=
2618 conf
->mddev
->recovery_disabled
;
2620 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, !abort
);
2623 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
2625 struct md_rdev
*rdev
;
2627 /* Doing recovery so rcu locking not required */
2628 rdev
= sh
->raid_conf
->disks
[disk_idx
].replacement
;
2630 && !test_bit(Faulty
, &rdev
->flags
)
2631 && !test_bit(In_sync
, &rdev
->flags
)
2632 && (rdev
->recovery_offset
<= sh
->sector
2633 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
2639 /* fetch_block - checks the given member device to see if its data needs
2640 * to be read or computed to satisfy a request.
2642 * Returns 1 when no more member devices need to be checked, otherwise returns
2643 * 0 to tell the loop in handle_stripe_fill to continue
2645 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2646 int disk_idx
, int disks
)
2648 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
2649 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
2650 &sh
->dev
[s
->failed_num
[1]] };
2652 /* is the data in this block needed, and can we get it? */
2653 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2654 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2656 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2657 s
->syncing
|| s
->expanding
||
2658 (s
->replacing
&& want_replace(sh
, disk_idx
)) ||
2659 (s
->failed
>= 1 && fdev
[0]->toread
) ||
2660 (s
->failed
>= 2 && fdev
[1]->toread
) ||
2661 (sh
->raid_conf
->level
<= 5 && s
->failed
&& fdev
[0]->towrite
&&
2662 !test_bit(R5_OVERWRITE
, &fdev
[0]->flags
)) ||
2663 (sh
->raid_conf
->level
== 6 && s
->failed
&& s
->to_write
))) {
2664 /* we would like to get this block, possibly by computing it,
2665 * otherwise read it if the backing disk is insync
2667 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
2668 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
2669 if ((s
->uptodate
== disks
- 1) &&
2670 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
2671 disk_idx
== s
->failed_num
[1]))) {
2672 /* have disk failed, and we're requested to fetch it;
2675 pr_debug("Computing stripe %llu block %d\n",
2676 (unsigned long long)sh
->sector
, disk_idx
);
2677 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2678 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2679 set_bit(R5_Wantcompute
, &dev
->flags
);
2680 sh
->ops
.target
= disk_idx
;
2681 sh
->ops
.target2
= -1; /* no 2nd target */
2683 /* Careful: from this point on 'uptodate' is in the eye
2684 * of raid_run_ops which services 'compute' operations
2685 * before writes. R5_Wantcompute flags a block that will
2686 * be R5_UPTODATE by the time it is needed for a
2687 * subsequent operation.
2691 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
2692 /* Computing 2-failure is *very* expensive; only
2693 * do it if failed >= 2
2696 for (other
= disks
; other
--; ) {
2697 if (other
== disk_idx
)
2699 if (!test_bit(R5_UPTODATE
,
2700 &sh
->dev
[other
].flags
))
2704 pr_debug("Computing stripe %llu blocks %d,%d\n",
2705 (unsigned long long)sh
->sector
,
2707 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2708 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2709 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
2710 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
2711 sh
->ops
.target
= disk_idx
;
2712 sh
->ops
.target2
= other
;
2716 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2717 set_bit(R5_LOCKED
, &dev
->flags
);
2718 set_bit(R5_Wantread
, &dev
->flags
);
2720 pr_debug("Reading block %d (sync=%d)\n",
2721 disk_idx
, s
->syncing
);
2729 * handle_stripe_fill - read or compute data to satisfy pending requests.
2731 static void handle_stripe_fill(struct stripe_head
*sh
,
2732 struct stripe_head_state
*s
,
2737 /* look for blocks to read/compute, skip this if a compute
2738 * is already in flight, or if the stripe contents are in the
2739 * midst of changing due to a write
2741 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2742 !sh
->reconstruct_state
)
2743 for (i
= disks
; i
--; )
2744 if (fetch_block(sh
, s
, i
, disks
))
2746 set_bit(STRIPE_HANDLE
, &sh
->state
);
2750 /* handle_stripe_clean_event
2751 * any written block on an uptodate or failed drive can be returned.
2752 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2753 * never LOCKED, so we don't need to test 'failed' directly.
2755 static void handle_stripe_clean_event(struct r5conf
*conf
,
2756 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
2760 int discard_pending
= 0;
2762 for (i
= disks
; i
--; )
2763 if (sh
->dev
[i
].written
) {
2765 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2766 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2767 test_bit(R5_Discard
, &dev
->flags
))) {
2768 /* We can return any write requests */
2769 struct bio
*wbi
, *wbi2
;
2770 pr_debug("Return write for disc %d\n", i
);
2771 if (test_and_clear_bit(R5_Discard
, &dev
->flags
))
2772 clear_bit(R5_UPTODATE
, &dev
->flags
);
2774 dev
->written
= NULL
;
2775 while (wbi
&& wbi
->bi_sector
<
2776 dev
->sector
+ STRIPE_SECTORS
) {
2777 wbi2
= r5_next_bio(wbi
, dev
->sector
);
2778 if (!raid5_dec_bi_active_stripes(wbi
)) {
2779 md_write_end(conf
->mddev
);
2780 wbi
->bi_next
= *return_bi
;
2785 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2787 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
2789 } else if (test_bit(R5_Discard
, &dev
->flags
))
2790 discard_pending
= 1;
2792 if (!discard_pending
&&
2793 test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
)) {
2794 clear_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
2795 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2796 if (sh
->qd_idx
>= 0) {
2797 clear_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
2798 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
);
2800 /* now that discard is done we can proceed with any sync */
2801 clear_bit(STRIPE_DISCARD
, &sh
->state
);
2802 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
))
2803 set_bit(STRIPE_HANDLE
, &sh
->state
);
2807 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2808 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2809 md_wakeup_thread(conf
->mddev
->thread
);
2812 static void handle_stripe_dirtying(struct r5conf
*conf
,
2813 struct stripe_head
*sh
,
2814 struct stripe_head_state
*s
,
2817 int rmw
= 0, rcw
= 0, i
;
2818 sector_t recovery_cp
= conf
->mddev
->recovery_cp
;
2820 /* RAID6 requires 'rcw' in current implementation.
2821 * Otherwise, check whether resync is now happening or should start.
2822 * If yes, then the array is dirty (after unclean shutdown or
2823 * initial creation), so parity in some stripes might be inconsistent.
2824 * In this case, we need to always do reconstruct-write, to ensure
2825 * that in case of drive failure or read-error correction, we
2826 * generate correct data from the parity.
2828 if (conf
->max_degraded
== 2 ||
2829 (recovery_cp
< MaxSector
&& sh
->sector
>= recovery_cp
)) {
2830 /* Calculate the real rcw later - for now make it
2831 * look like rcw is cheaper
2834 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
2835 conf
->max_degraded
, (unsigned long long)recovery_cp
,
2836 (unsigned long long)sh
->sector
);
2837 } else for (i
= disks
; i
--; ) {
2838 /* would I have to read this buffer for read_modify_write */
2839 struct r5dev
*dev
= &sh
->dev
[i
];
2840 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2841 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2842 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2843 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2844 if (test_bit(R5_Insync
, &dev
->flags
))
2847 rmw
+= 2*disks
; /* cannot read it */
2849 /* Would I have to read this buffer for reconstruct_write */
2850 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) && i
!= sh
->pd_idx
&&
2851 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2852 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2853 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2854 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2859 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2860 (unsigned long long)sh
->sector
, rmw
, rcw
);
2861 set_bit(STRIPE_HANDLE
, &sh
->state
);
2862 if (rmw
< rcw
&& rmw
> 0) {
2863 /* prefer read-modify-write, but need to get some data */
2864 if (conf
->mddev
->queue
)
2865 blk_add_trace_msg(conf
->mddev
->queue
,
2866 "raid5 rmw %llu %d",
2867 (unsigned long long)sh
->sector
, rmw
);
2868 for (i
= disks
; i
--; ) {
2869 struct r5dev
*dev
= &sh
->dev
[i
];
2870 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2871 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2872 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2873 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2874 test_bit(R5_Insync
, &dev
->flags
)) {
2876 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2877 pr_debug("Read_old block "
2878 "%d for r-m-w\n", i
);
2879 set_bit(R5_LOCKED
, &dev
->flags
);
2880 set_bit(R5_Wantread
, &dev
->flags
);
2883 set_bit(STRIPE_DELAYED
, &sh
->state
);
2884 set_bit(STRIPE_HANDLE
, &sh
->state
);
2889 if (rcw
<= rmw
&& rcw
> 0) {
2890 /* want reconstruct write, but need to get some data */
2893 for (i
= disks
; i
--; ) {
2894 struct r5dev
*dev
= &sh
->dev
[i
];
2895 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2896 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
2897 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2898 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2899 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2901 if (!test_bit(R5_Insync
, &dev
->flags
))
2902 continue; /* it's a failed drive */
2904 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2905 pr_debug("Read_old block "
2906 "%d for Reconstruct\n", i
);
2907 set_bit(R5_LOCKED
, &dev
->flags
);
2908 set_bit(R5_Wantread
, &dev
->flags
);
2912 set_bit(STRIPE_DELAYED
, &sh
->state
);
2913 set_bit(STRIPE_HANDLE
, &sh
->state
);
2917 if (rcw
&& conf
->mddev
->queue
)
2918 blk_add_trace_msg(conf
->mddev
->queue
, "raid5 rcw %llu %d %d %d",
2919 (unsigned long long)sh
->sector
,
2920 rcw
, qread
, test_bit(STRIPE_DELAYED
, &sh
->state
));
2922 /* now if nothing is locked, and if we have enough data,
2923 * we can start a write request
2925 /* since handle_stripe can be called at any time we need to handle the
2926 * case where a compute block operation has been submitted and then a
2927 * subsequent call wants to start a write request. raid_run_ops only
2928 * handles the case where compute block and reconstruct are requested
2929 * simultaneously. If this is not the case then new writes need to be
2930 * held off until the compute completes.
2932 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
2933 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
2934 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
2935 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
2938 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
2939 struct stripe_head_state
*s
, int disks
)
2941 struct r5dev
*dev
= NULL
;
2943 set_bit(STRIPE_HANDLE
, &sh
->state
);
2945 switch (sh
->check_state
) {
2946 case check_state_idle
:
2947 /* start a new check operation if there are no failures */
2948 if (s
->failed
== 0) {
2949 BUG_ON(s
->uptodate
!= disks
);
2950 sh
->check_state
= check_state_run
;
2951 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2952 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2956 dev
= &sh
->dev
[s
->failed_num
[0]];
2958 case check_state_compute_result
:
2959 sh
->check_state
= check_state_idle
;
2961 dev
= &sh
->dev
[sh
->pd_idx
];
2963 /* check that a write has not made the stripe insync */
2964 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2967 /* either failed parity check, or recovery is happening */
2968 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
2969 BUG_ON(s
->uptodate
!= disks
);
2971 set_bit(R5_LOCKED
, &dev
->flags
);
2973 set_bit(R5_Wantwrite
, &dev
->flags
);
2975 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2976 set_bit(STRIPE_INSYNC
, &sh
->state
);
2978 case check_state_run
:
2979 break; /* we will be called again upon completion */
2980 case check_state_check_result
:
2981 sh
->check_state
= check_state_idle
;
2983 /* if a failure occurred during the check operation, leave
2984 * STRIPE_INSYNC not set and let the stripe be handled again
2989 /* handle a successful check operation, if parity is correct
2990 * we are done. Otherwise update the mismatch count and repair
2991 * parity if !MD_RECOVERY_CHECK
2993 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
2994 /* parity is correct (on disc,
2995 * not in buffer any more)
2997 set_bit(STRIPE_INSYNC
, &sh
->state
);
2999 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3000 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3001 /* don't try to repair!! */
3002 set_bit(STRIPE_INSYNC
, &sh
->state
);
3004 sh
->check_state
= check_state_compute_run
;
3005 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3006 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3007 set_bit(R5_Wantcompute
,
3008 &sh
->dev
[sh
->pd_idx
].flags
);
3009 sh
->ops
.target
= sh
->pd_idx
;
3010 sh
->ops
.target2
= -1;
3015 case check_state_compute_run
:
3018 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3019 __func__
, sh
->check_state
,
3020 (unsigned long long) sh
->sector
);
3026 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
3027 struct stripe_head_state
*s
,
3030 int pd_idx
= sh
->pd_idx
;
3031 int qd_idx
= sh
->qd_idx
;
3034 set_bit(STRIPE_HANDLE
, &sh
->state
);
3036 BUG_ON(s
->failed
> 2);
3038 /* Want to check and possibly repair P and Q.
3039 * However there could be one 'failed' device, in which
3040 * case we can only check one of them, possibly using the
3041 * other to generate missing data
3044 switch (sh
->check_state
) {
3045 case check_state_idle
:
3046 /* start a new check operation if there are < 2 failures */
3047 if (s
->failed
== s
->q_failed
) {
3048 /* The only possible failed device holds Q, so it
3049 * makes sense to check P (If anything else were failed,
3050 * we would have used P to recreate it).
3052 sh
->check_state
= check_state_run
;
3054 if (!s
->q_failed
&& s
->failed
< 2) {
3055 /* Q is not failed, and we didn't use it to generate
3056 * anything, so it makes sense to check it
3058 if (sh
->check_state
== check_state_run
)
3059 sh
->check_state
= check_state_run_pq
;
3061 sh
->check_state
= check_state_run_q
;
3064 /* discard potentially stale zero_sum_result */
3065 sh
->ops
.zero_sum_result
= 0;
3067 if (sh
->check_state
== check_state_run
) {
3068 /* async_xor_zero_sum destroys the contents of P */
3069 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
3072 if (sh
->check_state
>= check_state_run
&&
3073 sh
->check_state
<= check_state_run_pq
) {
3074 /* async_syndrome_zero_sum preserves P and Q, so
3075 * no need to mark them !uptodate here
3077 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3081 /* we have 2-disk failure */
3082 BUG_ON(s
->failed
!= 2);
3084 case check_state_compute_result
:
3085 sh
->check_state
= check_state_idle
;
3087 /* check that a write has not made the stripe insync */
3088 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3091 /* now write out any block on a failed drive,
3092 * or P or Q if they were recomputed
3094 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
3095 if (s
->failed
== 2) {
3096 dev
= &sh
->dev
[s
->failed_num
[1]];
3098 set_bit(R5_LOCKED
, &dev
->flags
);
3099 set_bit(R5_Wantwrite
, &dev
->flags
);
3101 if (s
->failed
>= 1) {
3102 dev
= &sh
->dev
[s
->failed_num
[0]];
3104 set_bit(R5_LOCKED
, &dev
->flags
);
3105 set_bit(R5_Wantwrite
, &dev
->flags
);
3107 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3108 dev
= &sh
->dev
[pd_idx
];
3110 set_bit(R5_LOCKED
, &dev
->flags
);
3111 set_bit(R5_Wantwrite
, &dev
->flags
);
3113 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3114 dev
= &sh
->dev
[qd_idx
];
3116 set_bit(R5_LOCKED
, &dev
->flags
);
3117 set_bit(R5_Wantwrite
, &dev
->flags
);
3119 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3121 set_bit(STRIPE_INSYNC
, &sh
->state
);
3123 case check_state_run
:
3124 case check_state_run_q
:
3125 case check_state_run_pq
:
3126 break; /* we will be called again upon completion */
3127 case check_state_check_result
:
3128 sh
->check_state
= check_state_idle
;
3130 /* handle a successful check operation, if parity is correct
3131 * we are done. Otherwise update the mismatch count and repair
3132 * parity if !MD_RECOVERY_CHECK
3134 if (sh
->ops
.zero_sum_result
== 0) {
3135 /* both parities are correct */
3137 set_bit(STRIPE_INSYNC
, &sh
->state
);
3139 /* in contrast to the raid5 case we can validate
3140 * parity, but still have a failure to write
3143 sh
->check_state
= check_state_compute_result
;
3144 /* Returning at this point means that we may go
3145 * off and bring p and/or q uptodate again so
3146 * we make sure to check zero_sum_result again
3147 * to verify if p or q need writeback
3151 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3152 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3153 /* don't try to repair!! */
3154 set_bit(STRIPE_INSYNC
, &sh
->state
);
3156 int *target
= &sh
->ops
.target
;
3158 sh
->ops
.target
= -1;
3159 sh
->ops
.target2
= -1;
3160 sh
->check_state
= check_state_compute_run
;
3161 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3162 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3163 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3164 set_bit(R5_Wantcompute
,
3165 &sh
->dev
[pd_idx
].flags
);
3167 target
= &sh
->ops
.target2
;
3170 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3171 set_bit(R5_Wantcompute
,
3172 &sh
->dev
[qd_idx
].flags
);
3179 case check_state_compute_run
:
3182 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3183 __func__
, sh
->check_state
,
3184 (unsigned long long) sh
->sector
);
3189 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
3193 /* We have read all the blocks in this stripe and now we need to
3194 * copy some of them into a target stripe for expand.
3196 struct dma_async_tx_descriptor
*tx
= NULL
;
3197 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3198 for (i
= 0; i
< sh
->disks
; i
++)
3199 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
3201 struct stripe_head
*sh2
;
3202 struct async_submit_ctl submit
;
3204 sector_t bn
= compute_blocknr(sh
, i
, 1);
3205 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
3207 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
3209 /* so far only the early blocks of this stripe
3210 * have been requested. When later blocks
3211 * get requested, we will try again
3214 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
3215 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
3216 /* must have already done this block */
3217 release_stripe(sh2
);
3221 /* place all the copies on one channel */
3222 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
3223 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
3224 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
3227 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
3228 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
3229 for (j
= 0; j
< conf
->raid_disks
; j
++)
3230 if (j
!= sh2
->pd_idx
&&
3232 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
3234 if (j
== conf
->raid_disks
) {
3235 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
3236 set_bit(STRIPE_HANDLE
, &sh2
->state
);
3238 release_stripe(sh2
);
3241 /* done submitting copies, wait for them to complete */
3242 async_tx_quiesce(&tx
);
3246 * handle_stripe - do things to a stripe.
3248 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3249 * state of various bits to see what needs to be done.
3251 * return some read requests which now have data
3252 * return some write requests which are safely on storage
3253 * schedule a read on some buffers
3254 * schedule a write of some buffers
3255 * return confirmation of parity correctness
3259 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
3261 struct r5conf
*conf
= sh
->raid_conf
;
3262 int disks
= sh
->disks
;
3265 int do_recovery
= 0;
3267 memset(s
, 0, sizeof(*s
));
3269 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3270 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3271 s
->failed_num
[0] = -1;
3272 s
->failed_num
[1] = -1;
3274 /* Now to look around and see what can be done */
3276 for (i
=disks
; i
--; ) {
3277 struct md_rdev
*rdev
;
3284 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3286 dev
->toread
, dev
->towrite
, dev
->written
);
3287 /* maybe we can reply to a read
3289 * new wantfill requests are only permitted while
3290 * ops_complete_biofill is guaranteed to be inactive
3292 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
3293 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
3294 set_bit(R5_Wantfill
, &dev
->flags
);
3296 /* now count some things */
3297 if (test_bit(R5_LOCKED
, &dev
->flags
))
3299 if (test_bit(R5_UPTODATE
, &dev
->flags
))
3301 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
3303 BUG_ON(s
->compute
> 2);
3306 if (test_bit(R5_Wantfill
, &dev
->flags
))
3308 else if (dev
->toread
)
3312 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3317 /* Prefer to use the replacement for reads, but only
3318 * if it is recovered enough and has no bad blocks.
3320 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
3321 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
3322 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
3323 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3324 &first_bad
, &bad_sectors
))
3325 set_bit(R5_ReadRepl
, &dev
->flags
);
3328 set_bit(R5_NeedReplace
, &dev
->flags
);
3329 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3330 clear_bit(R5_ReadRepl
, &dev
->flags
);
3332 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
3335 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3336 &first_bad
, &bad_sectors
);
3337 if (s
->blocked_rdev
== NULL
3338 && (test_bit(Blocked
, &rdev
->flags
)
3341 set_bit(BlockedBadBlocks
,
3343 s
->blocked_rdev
= rdev
;
3344 atomic_inc(&rdev
->nr_pending
);
3347 clear_bit(R5_Insync
, &dev
->flags
);
3351 /* also not in-sync */
3352 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
3353 test_bit(R5_UPTODATE
, &dev
->flags
)) {
3354 /* treat as in-sync, but with a read error
3355 * which we can now try to correct
3357 set_bit(R5_Insync
, &dev
->flags
);
3358 set_bit(R5_ReadError
, &dev
->flags
);
3360 } else if (test_bit(In_sync
, &rdev
->flags
))
3361 set_bit(R5_Insync
, &dev
->flags
);
3362 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
3363 /* in sync if before recovery_offset */
3364 set_bit(R5_Insync
, &dev
->flags
);
3365 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
3366 test_bit(R5_Expanded
, &dev
->flags
))
3367 /* If we've reshaped into here, we assume it is Insync.
3368 * We will shortly update recovery_offset to make
3371 set_bit(R5_Insync
, &dev
->flags
);
3373 if (rdev
&& test_bit(R5_WriteError
, &dev
->flags
)) {
3374 /* This flag does not apply to '.replacement'
3375 * only to .rdev, so make sure to check that*/
3376 struct md_rdev
*rdev2
= rcu_dereference(
3377 conf
->disks
[i
].rdev
);
3379 clear_bit(R5_Insync
, &dev
->flags
);
3380 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3381 s
->handle_bad_blocks
= 1;
3382 atomic_inc(&rdev2
->nr_pending
);
3384 clear_bit(R5_WriteError
, &dev
->flags
);
3386 if (rdev
&& test_bit(R5_MadeGood
, &dev
->flags
)) {
3387 /* This flag does not apply to '.replacement'
3388 * only to .rdev, so make sure to check that*/
3389 struct md_rdev
*rdev2
= rcu_dereference(
3390 conf
->disks
[i
].rdev
);
3391 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3392 s
->handle_bad_blocks
= 1;
3393 atomic_inc(&rdev2
->nr_pending
);
3395 clear_bit(R5_MadeGood
, &dev
->flags
);
3397 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
3398 struct md_rdev
*rdev2
= rcu_dereference(
3399 conf
->disks
[i
].replacement
);
3400 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3401 s
->handle_bad_blocks
= 1;
3402 atomic_inc(&rdev2
->nr_pending
);
3404 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
3406 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3407 /* The ReadError flag will just be confusing now */
3408 clear_bit(R5_ReadError
, &dev
->flags
);
3409 clear_bit(R5_ReWrite
, &dev
->flags
);
3411 if (test_bit(R5_ReadError
, &dev
->flags
))
3412 clear_bit(R5_Insync
, &dev
->flags
);
3413 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3415 s
->failed_num
[s
->failed
] = i
;
3417 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
3421 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
3422 /* If there is a failed device being replaced,
3423 * we must be recovering.
3424 * else if we are after recovery_cp, we must be syncing
3425 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3426 * else we can only be replacing
3427 * sync and recovery both need to read all devices, and so
3428 * use the same flag.
3431 sh
->sector
>= conf
->mddev
->recovery_cp
||
3432 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
3440 static void handle_stripe(struct stripe_head
*sh
)
3442 struct stripe_head_state s
;
3443 struct r5conf
*conf
= sh
->raid_conf
;
3446 int disks
= sh
->disks
;
3447 struct r5dev
*pdev
, *qdev
;
3449 clear_bit(STRIPE_HANDLE
, &sh
->state
);
3450 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
3451 /* already being handled, ensure it gets handled
3452 * again when current action finishes */
3453 set_bit(STRIPE_HANDLE
, &sh
->state
);
3457 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
3458 spin_lock(&sh
->stripe_lock
);
3459 /* Cannot process 'sync' concurrently with 'discard' */
3460 if (!test_bit(STRIPE_DISCARD
, &sh
->state
) &&
3461 test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
3462 set_bit(STRIPE_SYNCING
, &sh
->state
);
3463 clear_bit(STRIPE_INSYNC
, &sh
->state
);
3465 spin_unlock(&sh
->stripe_lock
);
3467 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3469 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3470 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3471 (unsigned long long)sh
->sector
, sh
->state
,
3472 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
3473 sh
->check_state
, sh
->reconstruct_state
);
3475 analyse_stripe(sh
, &s
);
3477 if (s
.handle_bad_blocks
) {
3478 set_bit(STRIPE_HANDLE
, &sh
->state
);
3482 if (unlikely(s
.blocked_rdev
)) {
3483 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
3484 s
.replacing
|| s
.to_write
|| s
.written
) {
3485 set_bit(STRIPE_HANDLE
, &sh
->state
);
3488 /* There is nothing for the blocked_rdev to block */
3489 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
3490 s
.blocked_rdev
= NULL
;
3493 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
3494 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
3495 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
3498 pr_debug("locked=%d uptodate=%d to_read=%d"
3499 " to_write=%d failed=%d failed_num=%d,%d\n",
3500 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
3501 s
.failed_num
[0], s
.failed_num
[1]);
3502 /* check if the array has lost more than max_degraded devices and,
3503 * if so, some requests might need to be failed.
3505 if (s
.failed
> conf
->max_degraded
) {
3506 sh
->check_state
= 0;
3507 sh
->reconstruct_state
= 0;
3508 if (s
.to_read
+s
.to_write
+s
.written
)
3509 handle_failed_stripe(conf
, sh
, &s
, disks
, &s
.return_bi
);
3510 if (s
.syncing
+ s
.replacing
)
3511 handle_failed_sync(conf
, sh
, &s
);
3514 /* Now we check to see if any write operations have recently
3518 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
3520 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
3521 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
3522 sh
->reconstruct_state
= reconstruct_state_idle
;
3524 /* All the 'written' buffers and the parity block are ready to
3525 * be written back to disk
3527 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
) &&
3528 !test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
));
3529 BUG_ON(sh
->qd_idx
>= 0 &&
3530 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
) &&
3531 !test_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
));
3532 for (i
= disks
; i
--; ) {
3533 struct r5dev
*dev
= &sh
->dev
[i
];
3534 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
3535 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3537 pr_debug("Writing block %d\n", i
);
3538 set_bit(R5_Wantwrite
, &dev
->flags
);
3541 if (!test_bit(R5_Insync
, &dev
->flags
) ||
3542 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3544 set_bit(STRIPE_INSYNC
, &sh
->state
);
3547 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3548 s
.dec_preread_active
= 1;
3552 * might be able to return some write requests if the parity blocks
3553 * are safe, or on a failed drive
3555 pdev
= &sh
->dev
[sh
->pd_idx
];
3556 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
3557 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
3558 qdev
= &sh
->dev
[sh
->qd_idx
];
3559 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
3560 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
3564 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
3565 && !test_bit(R5_LOCKED
, &pdev
->flags
)
3566 && (test_bit(R5_UPTODATE
, &pdev
->flags
) ||
3567 test_bit(R5_Discard
, &pdev
->flags
))))) &&
3568 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
3569 && !test_bit(R5_LOCKED
, &qdev
->flags
)
3570 && (test_bit(R5_UPTODATE
, &qdev
->flags
) ||
3571 test_bit(R5_Discard
, &qdev
->flags
))))))
3572 handle_stripe_clean_event(conf
, sh
, disks
, &s
.return_bi
);
3574 /* Now we might consider reading some blocks, either to check/generate
3575 * parity, or to satisfy requests
3576 * or to load a block that is being partially written.
3578 if (s
.to_read
|| s
.non_overwrite
3579 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
3580 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
3583 handle_stripe_fill(sh
, &s
, disks
);
3585 /* Now to consider new write requests and what else, if anything
3586 * should be read. We do not handle new writes when:
3587 * 1/ A 'write' operation (copy+xor) is already in flight.
3588 * 2/ A 'check' operation is in flight, as it may clobber the parity
3591 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
3592 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
3594 /* maybe we need to check and possibly fix the parity for this stripe
3595 * Any reads will already have been scheduled, so we just see if enough
3596 * data is available. The parity check is held off while parity
3597 * dependent operations are in flight.
3599 if (sh
->check_state
||
3600 (s
.syncing
&& s
.locked
== 0 &&
3601 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
3602 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
3603 if (conf
->level
== 6)
3604 handle_parity_checks6(conf
, sh
, &s
, disks
);
3606 handle_parity_checks5(conf
, sh
, &s
, disks
);
3609 if (s
.replacing
&& s
.locked
== 0
3610 && !test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3611 /* Write out to replacement devices where possible */
3612 for (i
= 0; i
< conf
->raid_disks
; i
++)
3613 if (test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
) &&
3614 test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
3615 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
3616 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3619 set_bit(STRIPE_INSYNC
, &sh
->state
);
3621 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
3622 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3623 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3624 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3625 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
3626 wake_up(&conf
->wait_for_overlap
);
3629 /* If the failed drives are just a ReadError, then we might need
3630 * to progress the repair/check process
3632 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
3633 for (i
= 0; i
< s
.failed
; i
++) {
3634 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
3635 if (test_bit(R5_ReadError
, &dev
->flags
)
3636 && !test_bit(R5_LOCKED
, &dev
->flags
)
3637 && test_bit(R5_UPTODATE
, &dev
->flags
)
3639 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3640 set_bit(R5_Wantwrite
, &dev
->flags
);
3641 set_bit(R5_ReWrite
, &dev
->flags
);
3642 set_bit(R5_LOCKED
, &dev
->flags
);
3645 /* let's read it back */
3646 set_bit(R5_Wantread
, &dev
->flags
);
3647 set_bit(R5_LOCKED
, &dev
->flags
);
3654 /* Finish reconstruct operations initiated by the expansion process */
3655 if (sh
->reconstruct_state
== reconstruct_state_result
) {
3656 struct stripe_head
*sh_src
3657 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
3658 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
3659 /* sh cannot be written until sh_src has been read.
3660 * so arrange for sh to be delayed a little
3662 set_bit(STRIPE_DELAYED
, &sh
->state
);
3663 set_bit(STRIPE_HANDLE
, &sh
->state
);
3664 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
3666 atomic_inc(&conf
->preread_active_stripes
);
3667 release_stripe(sh_src
);
3671 release_stripe(sh_src
);
3673 sh
->reconstruct_state
= reconstruct_state_idle
;
3674 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3675 for (i
= conf
->raid_disks
; i
--; ) {
3676 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3677 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3682 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
3683 !sh
->reconstruct_state
) {
3684 /* Need to write out all blocks after computing parity */
3685 sh
->disks
= conf
->raid_disks
;
3686 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
3687 schedule_reconstruction(sh
, &s
, 1, 1);
3688 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
3689 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3690 atomic_dec(&conf
->reshape_stripes
);
3691 wake_up(&conf
->wait_for_overlap
);
3692 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3695 if (s
.expanding
&& s
.locked
== 0 &&
3696 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
3697 handle_stripe_expansion(conf
, sh
);
3700 /* wait for this device to become unblocked */
3701 if (unlikely(s
.blocked_rdev
)) {
3702 if (conf
->mddev
->external
)
3703 md_wait_for_blocked_rdev(s
.blocked_rdev
,
3706 /* Internal metadata will immediately
3707 * be written by raid5d, so we don't
3708 * need to wait here.
3710 rdev_dec_pending(s
.blocked_rdev
,
3714 if (s
.handle_bad_blocks
)
3715 for (i
= disks
; i
--; ) {
3716 struct md_rdev
*rdev
;
3717 struct r5dev
*dev
= &sh
->dev
[i
];
3718 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
3719 /* We own a safe reference to the rdev */
3720 rdev
= conf
->disks
[i
].rdev
;
3721 if (!rdev_set_badblocks(rdev
, sh
->sector
,
3723 md_error(conf
->mddev
, rdev
);
3724 rdev_dec_pending(rdev
, conf
->mddev
);
3726 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
3727 rdev
= conf
->disks
[i
].rdev
;
3728 rdev_clear_badblocks(rdev
, sh
->sector
,
3730 rdev_dec_pending(rdev
, conf
->mddev
);
3732 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
3733 rdev
= conf
->disks
[i
].replacement
;
3735 /* rdev have been moved down */
3736 rdev
= conf
->disks
[i
].rdev
;
3737 rdev_clear_badblocks(rdev
, sh
->sector
,
3739 rdev_dec_pending(rdev
, conf
->mddev
);
3744 raid_run_ops(sh
, s
.ops_request
);
3748 if (s
.dec_preread_active
) {
3749 /* We delay this until after ops_run_io so that if make_request
3750 * is waiting on a flush, it won't continue until the writes
3751 * have actually been submitted.
3753 atomic_dec(&conf
->preread_active_stripes
);
3754 if (atomic_read(&conf
->preread_active_stripes
) <
3756 md_wakeup_thread(conf
->mddev
->thread
);
3759 return_io(s
.return_bi
);
3761 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
3764 static void raid5_activate_delayed(struct r5conf
*conf
)
3766 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
3767 while (!list_empty(&conf
->delayed_list
)) {
3768 struct list_head
*l
= conf
->delayed_list
.next
;
3769 struct stripe_head
*sh
;
3770 sh
= list_entry(l
, struct stripe_head
, lru
);
3772 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3773 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3774 atomic_inc(&conf
->preread_active_stripes
);
3775 list_add_tail(&sh
->lru
, &conf
->hold_list
);
3780 static void activate_bit_delay(struct r5conf
*conf
)
3782 /* device_lock is held */
3783 struct list_head head
;
3784 list_add(&head
, &conf
->bitmap_list
);
3785 list_del_init(&conf
->bitmap_list
);
3786 while (!list_empty(&head
)) {
3787 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
3788 list_del_init(&sh
->lru
);
3789 atomic_inc(&sh
->count
);
3790 __release_stripe(conf
, sh
);
3794 int md_raid5_congested(struct mddev
*mddev
, int bits
)
3796 struct r5conf
*conf
= mddev
->private;
3798 /* No difference between reads and writes. Just check
3799 * how busy the stripe_cache is
3802 if (conf
->inactive_blocked
)
3806 if (list_empty_careful(&conf
->inactive_list
))
3811 EXPORT_SYMBOL_GPL(md_raid5_congested
);
3813 static int raid5_congested(void *data
, int bits
)
3815 struct mddev
*mddev
= data
;
3817 return mddev_congested(mddev
, bits
) ||
3818 md_raid5_congested(mddev
, bits
);
3821 /* We want read requests to align with chunks where possible,
3822 * but write requests don't need to.
3824 static int raid5_mergeable_bvec(struct request_queue
*q
,
3825 struct bvec_merge_data
*bvm
,
3826 struct bio_vec
*biovec
)
3828 struct mddev
*mddev
= q
->queuedata
;
3829 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
3831 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3832 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
3834 if ((bvm
->bi_rw
& 1) == WRITE
)
3835 return biovec
->bv_len
; /* always allow writes to be mergeable */
3837 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3838 chunk_sectors
= mddev
->new_chunk_sectors
;
3839 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
3840 if (max
< 0) max
= 0;
3841 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
3842 return biovec
->bv_len
;
3848 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
3850 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
3851 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3852 unsigned int bio_sectors
= bio
->bi_size
>> 9;
3854 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3855 chunk_sectors
= mddev
->new_chunk_sectors
;
3856 return chunk_sectors
>=
3857 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
3861 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3862 * later sampled by raid5d.
3864 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
3866 unsigned long flags
;
3868 spin_lock_irqsave(&conf
->device_lock
, flags
);
3870 bi
->bi_next
= conf
->retry_read_aligned_list
;
3871 conf
->retry_read_aligned_list
= bi
;
3873 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3874 md_wakeup_thread(conf
->mddev
->thread
);
3878 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
)
3882 bi
= conf
->retry_read_aligned
;
3884 conf
->retry_read_aligned
= NULL
;
3887 bi
= conf
->retry_read_aligned_list
;
3889 conf
->retry_read_aligned_list
= bi
->bi_next
;
3892 * this sets the active strip count to 1 and the processed
3893 * strip count to zero (upper 8 bits)
3895 raid5_set_bi_stripes(bi
, 1); /* biased count of active stripes */
3903 * The "raid5_align_endio" should check if the read succeeded and if it
3904 * did, call bio_endio on the original bio (having bio_put the new bio
3906 * If the read failed..
3908 static void raid5_align_endio(struct bio
*bi
, int error
)
3910 struct bio
* raid_bi
= bi
->bi_private
;
3911 struct mddev
*mddev
;
3912 struct r5conf
*conf
;
3913 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3914 struct md_rdev
*rdev
;
3918 rdev
= (void*)raid_bi
->bi_next
;
3919 raid_bi
->bi_next
= NULL
;
3920 mddev
= rdev
->mddev
;
3921 conf
= mddev
->private;
3923 rdev_dec_pending(rdev
, conf
->mddev
);
3925 if (!error
&& uptodate
) {
3926 trace_block_bio_complete(bdev_get_queue(raid_bi
->bi_bdev
),
3928 bio_endio(raid_bi
, 0);
3929 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
3930 wake_up(&conf
->wait_for_stripe
);
3935 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3937 add_bio_to_retry(raid_bi
, conf
);
3940 static int bio_fits_rdev(struct bio
*bi
)
3942 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
3944 if ((bi
->bi_size
>>9) > queue_max_sectors(q
))
3946 blk_recount_segments(q
, bi
);
3947 if (bi
->bi_phys_segments
> queue_max_segments(q
))
3950 if (q
->merge_bvec_fn
)
3951 /* it's too hard to apply the merge_bvec_fn at this stage,
3960 static int chunk_aligned_read(struct mddev
*mddev
, struct bio
* raid_bio
)
3962 struct r5conf
*conf
= mddev
->private;
3964 struct bio
* align_bi
;
3965 struct md_rdev
*rdev
;
3966 sector_t end_sector
;
3968 if (!in_chunk_boundary(mddev
, raid_bio
)) {
3969 pr_debug("chunk_aligned_read : non aligned\n");
3973 * use bio_clone_mddev to make a copy of the bio
3975 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
3979 * set bi_end_io to a new function, and set bi_private to the
3982 align_bi
->bi_end_io
= raid5_align_endio
;
3983 align_bi
->bi_private
= raid_bio
;
3987 align_bi
->bi_sector
= raid5_compute_sector(conf
, raid_bio
->bi_sector
,
3991 end_sector
= align_bi
->bi_sector
+ (align_bi
->bi_size
>> 9);
3993 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
3994 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
3995 rdev
->recovery_offset
< end_sector
) {
3996 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
3998 (test_bit(Faulty
, &rdev
->flags
) ||
3999 !(test_bit(In_sync
, &rdev
->flags
) ||
4000 rdev
->recovery_offset
>= end_sector
)))
4007 atomic_inc(&rdev
->nr_pending
);
4009 raid_bio
->bi_next
= (void*)rdev
;
4010 align_bi
->bi_bdev
= rdev
->bdev
;
4011 align_bi
->bi_flags
&= ~(1 << BIO_SEG_VALID
);
4013 if (!bio_fits_rdev(align_bi
) ||
4014 is_badblock(rdev
, align_bi
->bi_sector
, align_bi
->bi_size
>>9,
4015 &first_bad
, &bad_sectors
)) {
4016 /* too big in some way, or has a known bad block */
4018 rdev_dec_pending(rdev
, mddev
);
4022 /* No reshape active, so we can trust rdev->data_offset */
4023 align_bi
->bi_sector
+= rdev
->data_offset
;
4025 spin_lock_irq(&conf
->device_lock
);
4026 wait_event_lock_irq(conf
->wait_for_stripe
,
4029 atomic_inc(&conf
->active_aligned_reads
);
4030 spin_unlock_irq(&conf
->device_lock
);
4033 trace_block_bio_remap(bdev_get_queue(align_bi
->bi_bdev
),
4034 align_bi
, disk_devt(mddev
->gendisk
),
4035 raid_bio
->bi_sector
);
4036 generic_make_request(align_bi
);
4045 /* __get_priority_stripe - get the next stripe to process
4047 * Full stripe writes are allowed to pass preread active stripes up until
4048 * the bypass_threshold is exceeded. In general the bypass_count
4049 * increments when the handle_list is handled before the hold_list; however, it
4050 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4051 * stripe with in flight i/o. The bypass_count will be reset when the
4052 * head of the hold_list has changed, i.e. the head was promoted to the
4055 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
)
4057 struct stripe_head
*sh
;
4059 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4061 list_empty(&conf
->handle_list
) ? "empty" : "busy",
4062 list_empty(&conf
->hold_list
) ? "empty" : "busy",
4063 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
4065 if (!list_empty(&conf
->handle_list
)) {
4066 sh
= list_entry(conf
->handle_list
.next
, typeof(*sh
), lru
);
4068 if (list_empty(&conf
->hold_list
))
4069 conf
->bypass_count
= 0;
4070 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
4071 if (conf
->hold_list
.next
== conf
->last_hold
)
4072 conf
->bypass_count
++;
4074 conf
->last_hold
= conf
->hold_list
.next
;
4075 conf
->bypass_count
-= conf
->bypass_threshold
;
4076 if (conf
->bypass_count
< 0)
4077 conf
->bypass_count
= 0;
4080 } else if (!list_empty(&conf
->hold_list
) &&
4081 ((conf
->bypass_threshold
&&
4082 conf
->bypass_count
> conf
->bypass_threshold
) ||
4083 atomic_read(&conf
->pending_full_writes
) == 0)) {
4084 sh
= list_entry(conf
->hold_list
.next
,
4086 conf
->bypass_count
-= conf
->bypass_threshold
;
4087 if (conf
->bypass_count
< 0)
4088 conf
->bypass_count
= 0;
4092 list_del_init(&sh
->lru
);
4093 atomic_inc(&sh
->count
);
4094 BUG_ON(atomic_read(&sh
->count
) != 1);
4098 struct raid5_plug_cb
{
4099 struct blk_plug_cb cb
;
4100 struct list_head list
;
4103 static void raid5_unplug(struct blk_plug_cb
*blk_cb
, bool from_schedule
)
4105 struct raid5_plug_cb
*cb
= container_of(
4106 blk_cb
, struct raid5_plug_cb
, cb
);
4107 struct stripe_head
*sh
;
4108 struct mddev
*mddev
= cb
->cb
.data
;
4109 struct r5conf
*conf
= mddev
->private;
4112 if (cb
->list
.next
&& !list_empty(&cb
->list
)) {
4113 spin_lock_irq(&conf
->device_lock
);
4114 while (!list_empty(&cb
->list
)) {
4115 sh
= list_first_entry(&cb
->list
, struct stripe_head
, lru
);
4116 list_del_init(&sh
->lru
);
4118 * avoid race release_stripe_plug() sees
4119 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4120 * is still in our list
4122 smp_mb__before_clear_bit();
4123 clear_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
);
4124 __release_stripe(conf
, sh
);
4127 spin_unlock_irq(&conf
->device_lock
);
4130 trace_block_unplug(mddev
->queue
, cnt
, !from_schedule
);
4134 static void release_stripe_plug(struct mddev
*mddev
,
4135 struct stripe_head
*sh
)
4137 struct blk_plug_cb
*blk_cb
= blk_check_plugged(
4138 raid5_unplug
, mddev
,
4139 sizeof(struct raid5_plug_cb
));
4140 struct raid5_plug_cb
*cb
;
4147 cb
= container_of(blk_cb
, struct raid5_plug_cb
, cb
);
4149 if (cb
->list
.next
== NULL
)
4150 INIT_LIST_HEAD(&cb
->list
);
4152 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
))
4153 list_add_tail(&sh
->lru
, &cb
->list
);
4158 static void make_discard_request(struct mddev
*mddev
, struct bio
*bi
)
4160 struct r5conf
*conf
= mddev
->private;
4161 sector_t logical_sector
, last_sector
;
4162 struct stripe_head
*sh
;
4166 if (mddev
->reshape_position
!= MaxSector
)
4167 /* Skip discard while reshape is happening */
4170 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4171 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
4174 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
4176 stripe_sectors
= conf
->chunk_sectors
*
4177 (conf
->raid_disks
- conf
->max_degraded
);
4178 logical_sector
= DIV_ROUND_UP_SECTOR_T(logical_sector
,
4180 sector_div(last_sector
, stripe_sectors
);
4182 logical_sector
*= conf
->chunk_sectors
;
4183 last_sector
*= conf
->chunk_sectors
;
4185 for (; logical_sector
< last_sector
;
4186 logical_sector
+= STRIPE_SECTORS
) {
4190 sh
= get_active_stripe(conf
, logical_sector
, 0, 0, 0);
4191 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
4192 TASK_UNINTERRUPTIBLE
);
4193 set_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
4194 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
4199 clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
4200 spin_lock_irq(&sh
->stripe_lock
);
4201 for (d
= 0; d
< conf
->raid_disks
; d
++) {
4202 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
4204 if (sh
->dev
[d
].towrite
|| sh
->dev
[d
].toread
) {
4205 set_bit(R5_Overlap
, &sh
->dev
[d
].flags
);
4206 spin_unlock_irq(&sh
->stripe_lock
);
4212 set_bit(STRIPE_DISCARD
, &sh
->state
);
4213 finish_wait(&conf
->wait_for_overlap
, &w
);
4214 for (d
= 0; d
< conf
->raid_disks
; d
++) {
4215 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
4217 sh
->dev
[d
].towrite
= bi
;
4218 set_bit(R5_OVERWRITE
, &sh
->dev
[d
].flags
);
4219 raid5_inc_bi_active_stripes(bi
);
4221 spin_unlock_irq(&sh
->stripe_lock
);
4222 if (conf
->mddev
->bitmap
) {
4224 d
< conf
->raid_disks
- conf
->max_degraded
;
4226 bitmap_startwrite(mddev
->bitmap
,
4230 sh
->bm_seq
= conf
->seq_flush
+ 1;
4231 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
4234 set_bit(STRIPE_HANDLE
, &sh
->state
);
4235 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4236 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4237 atomic_inc(&conf
->preread_active_stripes
);
4238 release_stripe_plug(mddev
, sh
);
4241 remaining
= raid5_dec_bi_active_stripes(bi
);
4242 if (remaining
== 0) {
4243 md_write_end(mddev
);
4248 static void make_request(struct mddev
*mddev
, struct bio
* bi
)
4250 struct r5conf
*conf
= mddev
->private;
4252 sector_t new_sector
;
4253 sector_t logical_sector
, last_sector
;
4254 struct stripe_head
*sh
;
4255 const int rw
= bio_data_dir(bi
);
4258 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
4259 md_flush_request(mddev
, bi
);
4263 md_write_start(mddev
, bi
);
4266 mddev
->reshape_position
== MaxSector
&&
4267 chunk_aligned_read(mddev
,bi
))
4270 if (unlikely(bi
->bi_rw
& REQ_DISCARD
)) {
4271 make_discard_request(mddev
, bi
);
4275 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4276 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
4278 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
4280 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
4286 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
4287 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
4288 /* spinlock is needed as reshape_progress may be
4289 * 64bit on a 32bit platform, and so it might be
4290 * possible to see a half-updated value
4291 * Of course reshape_progress could change after
4292 * the lock is dropped, so once we get a reference
4293 * to the stripe that we think it is, we will have
4296 spin_lock_irq(&conf
->device_lock
);
4297 if (mddev
->reshape_backwards
4298 ? logical_sector
< conf
->reshape_progress
4299 : logical_sector
>= conf
->reshape_progress
) {
4302 if (mddev
->reshape_backwards
4303 ? logical_sector
< conf
->reshape_safe
4304 : logical_sector
>= conf
->reshape_safe
) {
4305 spin_unlock_irq(&conf
->device_lock
);
4310 spin_unlock_irq(&conf
->device_lock
);
4313 new_sector
= raid5_compute_sector(conf
, logical_sector
,
4316 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4317 (unsigned long long)new_sector
,
4318 (unsigned long long)logical_sector
);
4320 sh
= get_active_stripe(conf
, new_sector
, previous
,
4321 (bi
->bi_rw
&RWA_MASK
), 0);
4323 if (unlikely(previous
)) {
4324 /* expansion might have moved on while waiting for a
4325 * stripe, so we must do the range check again.
4326 * Expansion could still move past after this
4327 * test, but as we are holding a reference to
4328 * 'sh', we know that if that happens,
4329 * STRIPE_EXPANDING will get set and the expansion
4330 * won't proceed until we finish with the stripe.
4333 spin_lock_irq(&conf
->device_lock
);
4334 if (mddev
->reshape_backwards
4335 ? logical_sector
>= conf
->reshape_progress
4336 : logical_sector
< conf
->reshape_progress
)
4337 /* mismatch, need to try again */
4339 spin_unlock_irq(&conf
->device_lock
);
4348 logical_sector
>= mddev
->suspend_lo
&&
4349 logical_sector
< mddev
->suspend_hi
) {
4351 /* As the suspend_* range is controlled by
4352 * userspace, we want an interruptible
4355 flush_signals(current
);
4356 prepare_to_wait(&conf
->wait_for_overlap
,
4357 &w
, TASK_INTERRUPTIBLE
);
4358 if (logical_sector
>= mddev
->suspend_lo
&&
4359 logical_sector
< mddev
->suspend_hi
)
4364 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
4365 !add_stripe_bio(sh
, bi
, dd_idx
, rw
)) {
4366 /* Stripe is busy expanding or
4367 * add failed due to overlap. Flush everything
4370 md_wakeup_thread(mddev
->thread
);
4375 finish_wait(&conf
->wait_for_overlap
, &w
);
4376 set_bit(STRIPE_HANDLE
, &sh
->state
);
4377 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4378 if ((bi
->bi_rw
& REQ_SYNC
) &&
4379 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4380 atomic_inc(&conf
->preread_active_stripes
);
4381 release_stripe_plug(mddev
, sh
);
4383 /* cannot get stripe for read-ahead, just give-up */
4384 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
4385 finish_wait(&conf
->wait_for_overlap
, &w
);
4390 remaining
= raid5_dec_bi_active_stripes(bi
);
4391 if (remaining
== 0) {
4394 md_write_end(mddev
);
4396 trace_block_bio_complete(bdev_get_queue(bi
->bi_bdev
),
4402 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
4404 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
4406 /* reshaping is quite different to recovery/resync so it is
4407 * handled quite separately ... here.
4409 * On each call to sync_request, we gather one chunk worth of
4410 * destination stripes and flag them as expanding.
4411 * Then we find all the source stripes and request reads.
4412 * As the reads complete, handle_stripe will copy the data
4413 * into the destination stripe and release that stripe.
4415 struct r5conf
*conf
= mddev
->private;
4416 struct stripe_head
*sh
;
4417 sector_t first_sector
, last_sector
;
4418 int raid_disks
= conf
->previous_raid_disks
;
4419 int data_disks
= raid_disks
- conf
->max_degraded
;
4420 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
4423 sector_t writepos
, readpos
, safepos
;
4424 sector_t stripe_addr
;
4425 int reshape_sectors
;
4426 struct list_head stripes
;
4428 if (sector_nr
== 0) {
4429 /* If restarting in the middle, skip the initial sectors */
4430 if (mddev
->reshape_backwards
&&
4431 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
4432 sector_nr
= raid5_size(mddev
, 0, 0)
4433 - conf
->reshape_progress
;
4434 } else if (!mddev
->reshape_backwards
&&
4435 conf
->reshape_progress
> 0)
4436 sector_nr
= conf
->reshape_progress
;
4437 sector_div(sector_nr
, new_data_disks
);
4439 mddev
->curr_resync_completed
= sector_nr
;
4440 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4446 /* We need to process a full chunk at a time.
4447 * If old and new chunk sizes differ, we need to process the
4450 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
)
4451 reshape_sectors
= mddev
->new_chunk_sectors
;
4453 reshape_sectors
= mddev
->chunk_sectors
;
4455 /* We update the metadata at least every 10 seconds, or when
4456 * the data about to be copied would over-write the source of
4457 * the data at the front of the range. i.e. one new_stripe
4458 * along from reshape_progress new_maps to after where
4459 * reshape_safe old_maps to
4461 writepos
= conf
->reshape_progress
;
4462 sector_div(writepos
, new_data_disks
);
4463 readpos
= conf
->reshape_progress
;
4464 sector_div(readpos
, data_disks
);
4465 safepos
= conf
->reshape_safe
;
4466 sector_div(safepos
, data_disks
);
4467 if (mddev
->reshape_backwards
) {
4468 writepos
-= min_t(sector_t
, reshape_sectors
, writepos
);
4469 readpos
+= reshape_sectors
;
4470 safepos
+= reshape_sectors
;
4472 writepos
+= reshape_sectors
;
4473 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
4474 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
4477 /* Having calculated the 'writepos' possibly use it
4478 * to set 'stripe_addr' which is where we will write to.
4480 if (mddev
->reshape_backwards
) {
4481 BUG_ON(conf
->reshape_progress
== 0);
4482 stripe_addr
= writepos
;
4483 BUG_ON((mddev
->dev_sectors
&
4484 ~((sector_t
)reshape_sectors
- 1))
4485 - reshape_sectors
- stripe_addr
4488 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
4489 stripe_addr
= sector_nr
;
4492 /* 'writepos' is the most advanced device address we might write.
4493 * 'readpos' is the least advanced device address we might read.
4494 * 'safepos' is the least address recorded in the metadata as having
4496 * If there is a min_offset_diff, these are adjusted either by
4497 * increasing the safepos/readpos if diff is negative, or
4498 * increasing writepos if diff is positive.
4499 * If 'readpos' is then behind 'writepos', there is no way that we can
4500 * ensure safety in the face of a crash - that must be done by userspace
4501 * making a backup of the data. So in that case there is no particular
4502 * rush to update metadata.
4503 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4504 * update the metadata to advance 'safepos' to match 'readpos' so that
4505 * we can be safe in the event of a crash.
4506 * So we insist on updating metadata if safepos is behind writepos and
4507 * readpos is beyond writepos.
4508 * In any case, update the metadata every 10 seconds.
4509 * Maybe that number should be configurable, but I'm not sure it is
4510 * worth it.... maybe it could be a multiple of safemode_delay???
4512 if (conf
->min_offset_diff
< 0) {
4513 safepos
+= -conf
->min_offset_diff
;
4514 readpos
+= -conf
->min_offset_diff
;
4516 writepos
+= conf
->min_offset_diff
;
4518 if ((mddev
->reshape_backwards
4519 ? (safepos
> writepos
&& readpos
< writepos
)
4520 : (safepos
< writepos
&& readpos
> writepos
)) ||
4521 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4522 /* Cannot proceed until we've updated the superblock... */
4523 wait_event(conf
->wait_for_overlap
,
4524 atomic_read(&conf
->reshape_stripes
)==0);
4525 mddev
->reshape_position
= conf
->reshape_progress
;
4526 mddev
->curr_resync_completed
= sector_nr
;
4527 conf
->reshape_checkpoint
= jiffies
;
4528 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4529 md_wakeup_thread(mddev
->thread
);
4530 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4531 kthread_should_stop());
4532 spin_lock_irq(&conf
->device_lock
);
4533 conf
->reshape_safe
= mddev
->reshape_position
;
4534 spin_unlock_irq(&conf
->device_lock
);
4535 wake_up(&conf
->wait_for_overlap
);
4536 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4539 INIT_LIST_HEAD(&stripes
);
4540 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
4542 int skipped_disk
= 0;
4543 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
4544 set_bit(STRIPE_EXPANDING
, &sh
->state
);
4545 atomic_inc(&conf
->reshape_stripes
);
4546 /* If any of this stripe is beyond the end of the old
4547 * array, then we need to zero those blocks
4549 for (j
=sh
->disks
; j
--;) {
4551 if (j
== sh
->pd_idx
)
4553 if (conf
->level
== 6 &&
4556 s
= compute_blocknr(sh
, j
, 0);
4557 if (s
< raid5_size(mddev
, 0, 0)) {
4561 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
4562 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
4563 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
4565 if (!skipped_disk
) {
4566 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4567 set_bit(STRIPE_HANDLE
, &sh
->state
);
4569 list_add(&sh
->lru
, &stripes
);
4571 spin_lock_irq(&conf
->device_lock
);
4572 if (mddev
->reshape_backwards
)
4573 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
4575 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
4576 spin_unlock_irq(&conf
->device_lock
);
4577 /* Ok, those stripe are ready. We can start scheduling
4578 * reads on the source stripes.
4579 * The source stripes are determined by mapping the first and last
4580 * block on the destination stripes.
4583 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
4586 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
4587 * new_data_disks
- 1),
4589 if (last_sector
>= mddev
->dev_sectors
)
4590 last_sector
= mddev
->dev_sectors
- 1;
4591 while (first_sector
<= last_sector
) {
4592 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
4593 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
4594 set_bit(STRIPE_HANDLE
, &sh
->state
);
4596 first_sector
+= STRIPE_SECTORS
;
4598 /* Now that the sources are clearly marked, we can release
4599 * the destination stripes
4601 while (!list_empty(&stripes
)) {
4602 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
4603 list_del_init(&sh
->lru
);
4606 /* If this takes us to the resync_max point where we have to pause,
4607 * then we need to write out the superblock.
4609 sector_nr
+= reshape_sectors
;
4610 if ((sector_nr
- mddev
->curr_resync_completed
) * 2
4611 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
4612 /* Cannot proceed until we've updated the superblock... */
4613 wait_event(conf
->wait_for_overlap
,
4614 atomic_read(&conf
->reshape_stripes
) == 0);
4615 mddev
->reshape_position
= conf
->reshape_progress
;
4616 mddev
->curr_resync_completed
= sector_nr
;
4617 conf
->reshape_checkpoint
= jiffies
;
4618 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4619 md_wakeup_thread(mddev
->thread
);
4620 wait_event(mddev
->sb_wait
,
4621 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
4622 || kthread_should_stop());
4623 spin_lock_irq(&conf
->device_lock
);
4624 conf
->reshape_safe
= mddev
->reshape_position
;
4625 spin_unlock_irq(&conf
->device_lock
);
4626 wake_up(&conf
->wait_for_overlap
);
4627 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4629 return reshape_sectors
;
4632 /* FIXME go_faster isn't used */
4633 static inline sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
4635 struct r5conf
*conf
= mddev
->private;
4636 struct stripe_head
*sh
;
4637 sector_t max_sector
= mddev
->dev_sectors
;
4638 sector_t sync_blocks
;
4639 int still_degraded
= 0;
4642 if (sector_nr
>= max_sector
) {
4643 /* just being told to finish up .. nothing much to do */
4645 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
4650 if (mddev
->curr_resync
< max_sector
) /* aborted */
4651 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
4653 else /* completed sync */
4655 bitmap_close_sync(mddev
->bitmap
);
4660 /* Allow raid5_quiesce to complete */
4661 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
4663 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
4664 return reshape_request(mddev
, sector_nr
, skipped
);
4666 /* No need to check resync_max as we never do more than one
4667 * stripe, and as resync_max will always be on a chunk boundary,
4668 * if the check in md_do_sync didn't fire, there is no chance
4669 * of overstepping resync_max here
4672 /* if there is too many failed drives and we are trying
4673 * to resync, then assert that we are finished, because there is
4674 * nothing we can do.
4676 if (mddev
->degraded
>= conf
->max_degraded
&&
4677 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
4678 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
4682 if (!test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
4684 !bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
4685 sync_blocks
>= STRIPE_SECTORS
) {
4686 /* we can skip this block, and probably more */
4687 sync_blocks
/= STRIPE_SECTORS
;
4689 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
4692 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
4694 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
4696 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
4697 /* make sure we don't swamp the stripe cache if someone else
4698 * is trying to get access
4700 schedule_timeout_uninterruptible(1);
4702 /* Need to check if array will still be degraded after recovery/resync
4703 * We don't need to check the 'failed' flag as when that gets set,
4706 for (i
= 0; i
< conf
->raid_disks
; i
++)
4707 if (conf
->disks
[i
].rdev
== NULL
)
4710 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
4712 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
4717 return STRIPE_SECTORS
;
4720 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
)
4722 /* We may not be able to submit a whole bio at once as there
4723 * may not be enough stripe_heads available.
4724 * We cannot pre-allocate enough stripe_heads as we may need
4725 * more than exist in the cache (if we allow ever large chunks).
4726 * So we do one stripe head at a time and record in
4727 * ->bi_hw_segments how many have been done.
4729 * We *know* that this entire raid_bio is in one chunk, so
4730 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4732 struct stripe_head
*sh
;
4734 sector_t sector
, logical_sector
, last_sector
;
4739 logical_sector
= raid_bio
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4740 sector
= raid5_compute_sector(conf
, logical_sector
,
4742 last_sector
= raid_bio
->bi_sector
+ (raid_bio
->bi_size
>>9);
4744 for (; logical_sector
< last_sector
;
4745 logical_sector
+= STRIPE_SECTORS
,
4746 sector
+= STRIPE_SECTORS
,
4749 if (scnt
< raid5_bi_processed_stripes(raid_bio
))
4750 /* already done this stripe */
4753 sh
= get_active_stripe(conf
, sector
, 0, 1, 0);
4756 /* failed to get a stripe - must wait */
4757 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
4758 conf
->retry_read_aligned
= raid_bio
;
4762 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0)) {
4764 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
4765 conf
->retry_read_aligned
= raid_bio
;
4769 set_bit(R5_ReadNoMerge
, &sh
->dev
[dd_idx
].flags
);
4774 remaining
= raid5_dec_bi_active_stripes(raid_bio
);
4775 if (remaining
== 0) {
4776 trace_block_bio_complete(bdev_get_queue(raid_bio
->bi_bdev
),
4778 bio_endio(raid_bio
, 0);
4780 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4781 wake_up(&conf
->wait_for_stripe
);
4785 #define MAX_STRIPE_BATCH 8
4786 static int handle_active_stripes(struct r5conf
*conf
)
4788 struct stripe_head
*batch
[MAX_STRIPE_BATCH
], *sh
;
4789 int i
, batch_size
= 0;
4791 while (batch_size
< MAX_STRIPE_BATCH
&&
4792 (sh
= __get_priority_stripe(conf
)) != NULL
)
4793 batch
[batch_size
++] = sh
;
4795 if (batch_size
== 0)
4797 spin_unlock_irq(&conf
->device_lock
);
4799 for (i
= 0; i
< batch_size
; i
++)
4800 handle_stripe(batch
[i
]);
4804 spin_lock_irq(&conf
->device_lock
);
4805 for (i
= 0; i
< batch_size
; i
++)
4806 __release_stripe(conf
, batch
[i
]);
4811 * This is our raid5 kernel thread.
4813 * We scan the hash table for stripes which can be handled now.
4814 * During the scan, completed stripes are saved for us by the interrupt
4815 * handler, so that they will not have to wait for our next wakeup.
4817 static void raid5d(struct md_thread
*thread
)
4819 struct mddev
*mddev
= thread
->mddev
;
4820 struct r5conf
*conf
= mddev
->private;
4822 struct blk_plug plug
;
4824 pr_debug("+++ raid5d active\n");
4826 md_check_recovery(mddev
);
4828 blk_start_plug(&plug
);
4830 spin_lock_irq(&conf
->device_lock
);
4836 !list_empty(&conf
->bitmap_list
)) {
4837 /* Now is a good time to flush some bitmap updates */
4839 spin_unlock_irq(&conf
->device_lock
);
4840 bitmap_unplug(mddev
->bitmap
);
4841 spin_lock_irq(&conf
->device_lock
);
4842 conf
->seq_write
= conf
->seq_flush
;
4843 activate_bit_delay(conf
);
4845 raid5_activate_delayed(conf
);
4847 while ((bio
= remove_bio_from_retry(conf
))) {
4849 spin_unlock_irq(&conf
->device_lock
);
4850 ok
= retry_aligned_read(conf
, bio
);
4851 spin_lock_irq(&conf
->device_lock
);
4857 batch_size
= handle_active_stripes(conf
);
4860 handled
+= batch_size
;
4862 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
)) {
4863 spin_unlock_irq(&conf
->device_lock
);
4864 md_check_recovery(mddev
);
4865 spin_lock_irq(&conf
->device_lock
);
4868 pr_debug("%d stripes handled\n", handled
);
4870 spin_unlock_irq(&conf
->device_lock
);
4872 async_tx_issue_pending_all();
4873 blk_finish_plug(&plug
);
4875 pr_debug("--- raid5d inactive\n");
4879 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
4881 struct r5conf
*conf
= mddev
->private;
4883 return sprintf(page
, "%d\n", conf
->max_nr_stripes
);
4889 raid5_set_cache_size(struct mddev
*mddev
, int size
)
4891 struct r5conf
*conf
= mddev
->private;
4894 if (size
<= 16 || size
> 32768)
4896 while (size
< conf
->max_nr_stripes
) {
4897 if (drop_one_stripe(conf
))
4898 conf
->max_nr_stripes
--;
4902 err
= md_allow_write(mddev
);
4905 while (size
> conf
->max_nr_stripes
) {
4906 if (grow_one_stripe(conf
))
4907 conf
->max_nr_stripes
++;
4912 EXPORT_SYMBOL(raid5_set_cache_size
);
4915 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
4917 struct r5conf
*conf
= mddev
->private;
4921 if (len
>= PAGE_SIZE
)
4926 if (strict_strtoul(page
, 10, &new))
4928 err
= raid5_set_cache_size(mddev
, new);
4934 static struct md_sysfs_entry
4935 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
4936 raid5_show_stripe_cache_size
,
4937 raid5_store_stripe_cache_size
);
4940 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
4942 struct r5conf
*conf
= mddev
->private;
4944 return sprintf(page
, "%d\n", conf
->bypass_threshold
);
4950 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
4952 struct r5conf
*conf
= mddev
->private;
4954 if (len
>= PAGE_SIZE
)
4959 if (strict_strtoul(page
, 10, &new))
4961 if (new > conf
->max_nr_stripes
)
4963 conf
->bypass_threshold
= new;
4967 static struct md_sysfs_entry
4968 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
4970 raid5_show_preread_threshold
,
4971 raid5_store_preread_threshold
);
4974 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
4976 struct r5conf
*conf
= mddev
->private;
4978 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
4983 static struct md_sysfs_entry
4984 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
4986 static struct attribute
*raid5_attrs
[] = {
4987 &raid5_stripecache_size
.attr
,
4988 &raid5_stripecache_active
.attr
,
4989 &raid5_preread_bypass_threshold
.attr
,
4992 static struct attribute_group raid5_attrs_group
= {
4994 .attrs
= raid5_attrs
,
4998 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
5000 struct r5conf
*conf
= mddev
->private;
5003 sectors
= mddev
->dev_sectors
;
5005 /* size is defined by the smallest of previous and new size */
5006 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
5008 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
5009 sectors
&= ~((sector_t
)mddev
->new_chunk_sectors
- 1);
5010 return sectors
* (raid_disks
- conf
->max_degraded
);
5013 static void raid5_free_percpu(struct r5conf
*conf
)
5015 struct raid5_percpu
*percpu
;
5022 for_each_possible_cpu(cpu
) {
5023 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
5024 safe_put_page(percpu
->spare_page
);
5025 kfree(percpu
->scribble
);
5027 #ifdef CONFIG_HOTPLUG_CPU
5028 unregister_cpu_notifier(&conf
->cpu_notify
);
5032 free_percpu(conf
->percpu
);
5035 static void free_conf(struct r5conf
*conf
)
5037 shrink_stripes(conf
);
5038 raid5_free_percpu(conf
);
5040 kfree(conf
->stripe_hashtbl
);
5044 #ifdef CONFIG_HOTPLUG_CPU
5045 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
5048 struct r5conf
*conf
= container_of(nfb
, struct r5conf
, cpu_notify
);
5049 long cpu
= (long)hcpu
;
5050 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
5053 case CPU_UP_PREPARE
:
5054 case CPU_UP_PREPARE_FROZEN
:
5055 if (conf
->level
== 6 && !percpu
->spare_page
)
5056 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
5057 if (!percpu
->scribble
)
5058 percpu
->scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
5060 if (!percpu
->scribble
||
5061 (conf
->level
== 6 && !percpu
->spare_page
)) {
5062 safe_put_page(percpu
->spare_page
);
5063 kfree(percpu
->scribble
);
5064 pr_err("%s: failed memory allocation for cpu%ld\n",
5066 return notifier_from_errno(-ENOMEM
);
5070 case CPU_DEAD_FROZEN
:
5071 safe_put_page(percpu
->spare_page
);
5072 kfree(percpu
->scribble
);
5073 percpu
->spare_page
= NULL
;
5074 percpu
->scribble
= NULL
;
5083 static int raid5_alloc_percpu(struct r5conf
*conf
)
5086 struct page
*spare_page
;
5087 struct raid5_percpu __percpu
*allcpus
;
5091 allcpus
= alloc_percpu(struct raid5_percpu
);
5094 conf
->percpu
= allcpus
;
5098 for_each_present_cpu(cpu
) {
5099 if (conf
->level
== 6) {
5100 spare_page
= alloc_page(GFP_KERNEL
);
5105 per_cpu_ptr(conf
->percpu
, cpu
)->spare_page
= spare_page
;
5107 scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
5112 per_cpu_ptr(conf
->percpu
, cpu
)->scribble
= scribble
;
5114 #ifdef CONFIG_HOTPLUG_CPU
5115 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
5116 conf
->cpu_notify
.priority
= 0;
5118 err
= register_cpu_notifier(&conf
->cpu_notify
);
5125 static struct r5conf
*setup_conf(struct mddev
*mddev
)
5127 struct r5conf
*conf
;
5128 int raid_disk
, memory
, max_disks
;
5129 struct md_rdev
*rdev
;
5130 struct disk_info
*disk
;
5133 if (mddev
->new_level
!= 5
5134 && mddev
->new_level
!= 4
5135 && mddev
->new_level
!= 6) {
5136 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5137 mdname(mddev
), mddev
->new_level
);
5138 return ERR_PTR(-EIO
);
5140 if ((mddev
->new_level
== 5
5141 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
5142 (mddev
->new_level
== 6
5143 && !algorithm_valid_raid6(mddev
->new_layout
))) {
5144 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
5145 mdname(mddev
), mddev
->new_layout
);
5146 return ERR_PTR(-EIO
);
5148 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
5149 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5150 mdname(mddev
), mddev
->raid_disks
);
5151 return ERR_PTR(-EINVAL
);
5154 if (!mddev
->new_chunk_sectors
||
5155 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
5156 !is_power_of_2(mddev
->new_chunk_sectors
)) {
5157 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
5158 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
5159 return ERR_PTR(-EINVAL
);
5162 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
5165 spin_lock_init(&conf
->device_lock
);
5166 init_waitqueue_head(&conf
->wait_for_stripe
);
5167 init_waitqueue_head(&conf
->wait_for_overlap
);
5168 INIT_LIST_HEAD(&conf
->handle_list
);
5169 INIT_LIST_HEAD(&conf
->hold_list
);
5170 INIT_LIST_HEAD(&conf
->delayed_list
);
5171 INIT_LIST_HEAD(&conf
->bitmap_list
);
5172 INIT_LIST_HEAD(&conf
->inactive_list
);
5173 atomic_set(&conf
->active_stripes
, 0);
5174 atomic_set(&conf
->preread_active_stripes
, 0);
5175 atomic_set(&conf
->active_aligned_reads
, 0);
5176 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
5177 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
5179 conf
->raid_disks
= mddev
->raid_disks
;
5180 if (mddev
->reshape_position
== MaxSector
)
5181 conf
->previous_raid_disks
= mddev
->raid_disks
;
5183 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
5184 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
5185 conf
->scribble_len
= scribble_len(max_disks
);
5187 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
5192 conf
->mddev
= mddev
;
5194 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
5197 conf
->level
= mddev
->new_level
;
5198 if (raid5_alloc_percpu(conf
) != 0)
5201 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
5203 rdev_for_each(rdev
, mddev
) {
5204 raid_disk
= rdev
->raid_disk
;
5205 if (raid_disk
>= max_disks
5208 disk
= conf
->disks
+ raid_disk
;
5210 if (test_bit(Replacement
, &rdev
->flags
)) {
5211 if (disk
->replacement
)
5213 disk
->replacement
= rdev
;
5220 if (test_bit(In_sync
, &rdev
->flags
)) {
5221 char b
[BDEVNAME_SIZE
];
5222 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
5224 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
5225 } else if (rdev
->saved_raid_disk
!= raid_disk
)
5226 /* Cannot rely on bitmap to complete recovery */
5230 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
5231 conf
->level
= mddev
->new_level
;
5232 if (conf
->level
== 6)
5233 conf
->max_degraded
= 2;
5235 conf
->max_degraded
= 1;
5236 conf
->algorithm
= mddev
->new_layout
;
5237 conf
->max_nr_stripes
= NR_STRIPES
;
5238 conf
->reshape_progress
= mddev
->reshape_position
;
5239 if (conf
->reshape_progress
!= MaxSector
) {
5240 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
5241 conf
->prev_algo
= mddev
->layout
;
5244 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
5245 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
5246 if (grow_stripes(conf
, conf
->max_nr_stripes
)) {
5248 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5249 mdname(mddev
), memory
);
5252 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
5253 mdname(mddev
), memory
);
5255 sprintf(pers_name
, "raid%d", mddev
->new_level
);
5256 conf
->thread
= md_register_thread(raid5d
, mddev
, pers_name
);
5257 if (!conf
->thread
) {
5259 "md/raid:%s: couldn't allocate thread.\n",
5269 return ERR_PTR(-EIO
);
5271 return ERR_PTR(-ENOMEM
);
5275 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
5278 case ALGORITHM_PARITY_0
:
5279 if (raid_disk
< max_degraded
)
5282 case ALGORITHM_PARITY_N
:
5283 if (raid_disk
>= raid_disks
- max_degraded
)
5286 case ALGORITHM_PARITY_0_6
:
5287 if (raid_disk
== 0 ||
5288 raid_disk
== raid_disks
- 1)
5291 case ALGORITHM_LEFT_ASYMMETRIC_6
:
5292 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
5293 case ALGORITHM_LEFT_SYMMETRIC_6
:
5294 case ALGORITHM_RIGHT_SYMMETRIC_6
:
5295 if (raid_disk
== raid_disks
- 1)
5301 static int run(struct mddev
*mddev
)
5303 struct r5conf
*conf
;
5304 int working_disks
= 0;
5305 int dirty_parity_disks
= 0;
5306 struct md_rdev
*rdev
;
5307 sector_t reshape_offset
= 0;
5309 long long min_offset_diff
= 0;
5312 if (mddev
->recovery_cp
!= MaxSector
)
5313 printk(KERN_NOTICE
"md/raid:%s: not clean"
5314 " -- starting background reconstruction\n",
5317 rdev_for_each(rdev
, mddev
) {
5319 if (rdev
->raid_disk
< 0)
5321 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
5323 min_offset_diff
= diff
;
5325 } else if (mddev
->reshape_backwards
&&
5326 diff
< min_offset_diff
)
5327 min_offset_diff
= diff
;
5328 else if (!mddev
->reshape_backwards
&&
5329 diff
> min_offset_diff
)
5330 min_offset_diff
= diff
;
5333 if (mddev
->reshape_position
!= MaxSector
) {
5334 /* Check that we can continue the reshape.
5335 * Difficulties arise if the stripe we would write to
5336 * next is at or after the stripe we would read from next.
5337 * For a reshape that changes the number of devices, this
5338 * is only possible for a very short time, and mdadm makes
5339 * sure that time appears to have past before assembling
5340 * the array. So we fail if that time hasn't passed.
5341 * For a reshape that keeps the number of devices the same
5342 * mdadm must be monitoring the reshape can keeping the
5343 * critical areas read-only and backed up. It will start
5344 * the array in read-only mode, so we check for that.
5346 sector_t here_new
, here_old
;
5348 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
5350 if (mddev
->new_level
!= mddev
->level
) {
5351 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
5352 "required - aborting.\n",
5356 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
5357 /* reshape_position must be on a new-stripe boundary, and one
5358 * further up in new geometry must map after here in old
5361 here_new
= mddev
->reshape_position
;
5362 if (sector_div(here_new
, mddev
->new_chunk_sectors
*
5363 (mddev
->raid_disks
- max_degraded
))) {
5364 printk(KERN_ERR
"md/raid:%s: reshape_position not "
5365 "on a stripe boundary\n", mdname(mddev
));
5368 reshape_offset
= here_new
* mddev
->new_chunk_sectors
;
5369 /* here_new is the stripe we will write to */
5370 here_old
= mddev
->reshape_position
;
5371 sector_div(here_old
, mddev
->chunk_sectors
*
5372 (old_disks
-max_degraded
));
5373 /* here_old is the first stripe that we might need to read
5375 if (mddev
->delta_disks
== 0) {
5376 if ((here_new
* mddev
->new_chunk_sectors
!=
5377 here_old
* mddev
->chunk_sectors
)) {
5378 printk(KERN_ERR
"md/raid:%s: reshape position is"
5379 " confused - aborting\n", mdname(mddev
));
5382 /* We cannot be sure it is safe to start an in-place
5383 * reshape. It is only safe if user-space is monitoring
5384 * and taking constant backups.
5385 * mdadm always starts a situation like this in
5386 * readonly mode so it can take control before
5387 * allowing any writes. So just check for that.
5389 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
5390 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
5391 /* not really in-place - so OK */;
5392 else if (mddev
->ro
== 0) {
5393 printk(KERN_ERR
"md/raid:%s: in-place reshape "
5394 "must be started in read-only mode "
5399 } else if (mddev
->reshape_backwards
5400 ? (here_new
* mddev
->new_chunk_sectors
+ min_offset_diff
<=
5401 here_old
* mddev
->chunk_sectors
)
5402 : (here_new
* mddev
->new_chunk_sectors
>=
5403 here_old
* mddev
->chunk_sectors
+ (-min_offset_diff
))) {
5404 /* Reading from the same stripe as writing to - bad */
5405 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
5406 "auto-recovery - aborting.\n",
5410 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
5412 /* OK, we should be able to continue; */
5414 BUG_ON(mddev
->level
!= mddev
->new_level
);
5415 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
5416 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
5417 BUG_ON(mddev
->delta_disks
!= 0);
5420 if (mddev
->private == NULL
)
5421 conf
= setup_conf(mddev
);
5423 conf
= mddev
->private;
5426 return PTR_ERR(conf
);
5428 conf
->min_offset_diff
= min_offset_diff
;
5429 mddev
->thread
= conf
->thread
;
5430 conf
->thread
= NULL
;
5431 mddev
->private = conf
;
5433 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
5435 rdev
= conf
->disks
[i
].rdev
;
5436 if (!rdev
&& conf
->disks
[i
].replacement
) {
5437 /* The replacement is all we have yet */
5438 rdev
= conf
->disks
[i
].replacement
;
5439 conf
->disks
[i
].replacement
= NULL
;
5440 clear_bit(Replacement
, &rdev
->flags
);
5441 conf
->disks
[i
].rdev
= rdev
;
5445 if (conf
->disks
[i
].replacement
&&
5446 conf
->reshape_progress
!= MaxSector
) {
5447 /* replacements and reshape simply do not mix. */
5448 printk(KERN_ERR
"md: cannot handle concurrent "
5449 "replacement and reshape.\n");
5452 if (test_bit(In_sync
, &rdev
->flags
)) {
5456 /* This disc is not fully in-sync. However if it
5457 * just stored parity (beyond the recovery_offset),
5458 * when we don't need to be concerned about the
5459 * array being dirty.
5460 * When reshape goes 'backwards', we never have
5461 * partially completed devices, so we only need
5462 * to worry about reshape going forwards.
5464 /* Hack because v0.91 doesn't store recovery_offset properly. */
5465 if (mddev
->major_version
== 0 &&
5466 mddev
->minor_version
> 90)
5467 rdev
->recovery_offset
= reshape_offset
;
5469 if (rdev
->recovery_offset
< reshape_offset
) {
5470 /* We need to check old and new layout */
5471 if (!only_parity(rdev
->raid_disk
,
5474 conf
->max_degraded
))
5477 if (!only_parity(rdev
->raid_disk
,
5479 conf
->previous_raid_disks
,
5480 conf
->max_degraded
))
5482 dirty_parity_disks
++;
5486 * 0 for a fully functional array, 1 or 2 for a degraded array.
5488 mddev
->degraded
= calc_degraded(conf
);
5490 if (has_failed(conf
)) {
5491 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
5492 " (%d/%d failed)\n",
5493 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
5497 /* device size must be a multiple of chunk size */
5498 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
5499 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
5501 if (mddev
->degraded
> dirty_parity_disks
&&
5502 mddev
->recovery_cp
!= MaxSector
) {
5503 if (mddev
->ok_start_degraded
)
5505 "md/raid:%s: starting dirty degraded array"
5506 " - data corruption possible.\n",
5510 "md/raid:%s: cannot start dirty degraded array.\n",
5516 if (mddev
->degraded
== 0)
5517 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
5518 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
5519 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
5522 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
5523 " out of %d devices, algorithm %d\n",
5524 mdname(mddev
), conf
->level
,
5525 mddev
->raid_disks
- mddev
->degraded
,
5526 mddev
->raid_disks
, mddev
->new_layout
);
5528 print_raid5_conf(conf
);
5530 if (conf
->reshape_progress
!= MaxSector
) {
5531 conf
->reshape_safe
= conf
->reshape_progress
;
5532 atomic_set(&conf
->reshape_stripes
, 0);
5533 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5534 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5535 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5536 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5537 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5542 /* Ok, everything is just fine now */
5543 if (mddev
->to_remove
== &raid5_attrs_group
)
5544 mddev
->to_remove
= NULL
;
5545 else if (mddev
->kobj
.sd
&&
5546 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
5548 "raid5: failed to create sysfs attributes for %s\n",
5550 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5554 bool discard_supported
= true;
5555 /* read-ahead size must cover two whole stripes, which
5556 * is 2 * (datadisks) * chunksize where 'n' is the
5557 * number of raid devices
5559 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
5560 int stripe
= data_disks
*
5561 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
5562 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
5563 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
5565 blk_queue_merge_bvec(mddev
->queue
, raid5_mergeable_bvec
);
5567 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
5568 mddev
->queue
->backing_dev_info
.congested_fn
= raid5_congested
;
5570 chunk_size
= mddev
->chunk_sectors
<< 9;
5571 blk_queue_io_min(mddev
->queue
, chunk_size
);
5572 blk_queue_io_opt(mddev
->queue
, chunk_size
*
5573 (conf
->raid_disks
- conf
->max_degraded
));
5575 * We can only discard a whole stripe. It doesn't make sense to
5576 * discard data disk but write parity disk
5578 stripe
= stripe
* PAGE_SIZE
;
5579 /* Round up to power of 2, as discard handling
5580 * currently assumes that */
5581 while ((stripe
-1) & stripe
)
5582 stripe
= (stripe
| (stripe
-1)) + 1;
5583 mddev
->queue
->limits
.discard_alignment
= stripe
;
5584 mddev
->queue
->limits
.discard_granularity
= stripe
;
5586 * unaligned part of discard request will be ignored, so can't
5587 * guarantee discard_zerors_data
5589 mddev
->queue
->limits
.discard_zeroes_data
= 0;
5591 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
5593 rdev_for_each(rdev
, mddev
) {
5594 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
5595 rdev
->data_offset
<< 9);
5596 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
5597 rdev
->new_data_offset
<< 9);
5599 * discard_zeroes_data is required, otherwise data
5600 * could be lost. Consider a scenario: discard a stripe
5601 * (the stripe could be inconsistent if
5602 * discard_zeroes_data is 0); write one disk of the
5603 * stripe (the stripe could be inconsistent again
5604 * depending on which disks are used to calculate
5605 * parity); the disk is broken; The stripe data of this
5608 if (!blk_queue_discard(bdev_get_queue(rdev
->bdev
)) ||
5609 !bdev_get_queue(rdev
->bdev
)->
5610 limits
.discard_zeroes_data
)
5611 discard_supported
= false;
5614 if (discard_supported
&&
5615 mddev
->queue
->limits
.max_discard_sectors
>= stripe
&&
5616 mddev
->queue
->limits
.discard_granularity
>= stripe
)
5617 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
5620 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
5626 md_unregister_thread(&mddev
->thread
);
5627 print_raid5_conf(conf
);
5629 mddev
->private = NULL
;
5630 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
5634 static int stop(struct mddev
*mddev
)
5636 struct r5conf
*conf
= mddev
->private;
5638 md_unregister_thread(&mddev
->thread
);
5640 mddev
->queue
->backing_dev_info
.congested_fn
= NULL
;
5642 mddev
->private = NULL
;
5643 mddev
->to_remove
= &raid5_attrs_group
;
5647 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
5649 struct r5conf
*conf
= mddev
->private;
5652 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
5653 mddev
->chunk_sectors
/ 2, mddev
->layout
);
5654 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
5655 for (i
= 0; i
< conf
->raid_disks
; i
++)
5656 seq_printf (seq
, "%s",
5657 conf
->disks
[i
].rdev
&&
5658 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
5659 seq_printf (seq
, "]");
5662 static void print_raid5_conf (struct r5conf
*conf
)
5665 struct disk_info
*tmp
;
5667 printk(KERN_DEBUG
"RAID conf printout:\n");
5669 printk("(conf==NULL)\n");
5672 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
5674 conf
->raid_disks
- conf
->mddev
->degraded
);
5676 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5677 char b
[BDEVNAME_SIZE
];
5678 tmp
= conf
->disks
+ i
;
5680 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
5681 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
5682 bdevname(tmp
->rdev
->bdev
, b
));
5686 static int raid5_spare_active(struct mddev
*mddev
)
5689 struct r5conf
*conf
= mddev
->private;
5690 struct disk_info
*tmp
;
5692 unsigned long flags
;
5694 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5695 tmp
= conf
->disks
+ i
;
5696 if (tmp
->replacement
5697 && tmp
->replacement
->recovery_offset
== MaxSector
5698 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
5699 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
5700 /* Replacement has just become active. */
5702 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
5705 /* Replaced device not technically faulty,
5706 * but we need to be sure it gets removed
5707 * and never re-added.
5709 set_bit(Faulty
, &tmp
->rdev
->flags
);
5710 sysfs_notify_dirent_safe(
5711 tmp
->rdev
->sysfs_state
);
5713 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
5714 } else if (tmp
->rdev
5715 && tmp
->rdev
->recovery_offset
== MaxSector
5716 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
5717 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
5719 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
5722 spin_lock_irqsave(&conf
->device_lock
, flags
);
5723 mddev
->degraded
= calc_degraded(conf
);
5724 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5725 print_raid5_conf(conf
);
5729 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
5731 struct r5conf
*conf
= mddev
->private;
5733 int number
= rdev
->raid_disk
;
5734 struct md_rdev
**rdevp
;
5735 struct disk_info
*p
= conf
->disks
+ number
;
5737 print_raid5_conf(conf
);
5738 if (rdev
== p
->rdev
)
5740 else if (rdev
== p
->replacement
)
5741 rdevp
= &p
->replacement
;
5745 if (number
>= conf
->raid_disks
&&
5746 conf
->reshape_progress
== MaxSector
)
5747 clear_bit(In_sync
, &rdev
->flags
);
5749 if (test_bit(In_sync
, &rdev
->flags
) ||
5750 atomic_read(&rdev
->nr_pending
)) {
5754 /* Only remove non-faulty devices if recovery
5757 if (!test_bit(Faulty
, &rdev
->flags
) &&
5758 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
5759 !has_failed(conf
) &&
5760 (!p
->replacement
|| p
->replacement
== rdev
) &&
5761 number
< conf
->raid_disks
) {
5767 if (atomic_read(&rdev
->nr_pending
)) {
5768 /* lost the race, try later */
5771 } else if (p
->replacement
) {
5772 /* We must have just cleared 'rdev' */
5773 p
->rdev
= p
->replacement
;
5774 clear_bit(Replacement
, &p
->replacement
->flags
);
5775 smp_mb(); /* Make sure other CPUs may see both as identical
5776 * but will never see neither - if they are careful
5778 p
->replacement
= NULL
;
5779 clear_bit(WantReplacement
, &rdev
->flags
);
5781 /* We might have just removed the Replacement as faulty-
5782 * clear the bit just in case
5784 clear_bit(WantReplacement
, &rdev
->flags
);
5787 print_raid5_conf(conf
);
5791 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
5793 struct r5conf
*conf
= mddev
->private;
5796 struct disk_info
*p
;
5798 int last
= conf
->raid_disks
- 1;
5800 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
5803 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
5804 /* no point adding a device */
5807 if (rdev
->raid_disk
>= 0)
5808 first
= last
= rdev
->raid_disk
;
5811 * find the disk ... but prefer rdev->saved_raid_disk
5814 if (rdev
->saved_raid_disk
>= 0 &&
5815 rdev
->saved_raid_disk
>= first
&&
5816 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
5817 first
= rdev
->saved_raid_disk
;
5819 for (disk
= first
; disk
<= last
; disk
++) {
5820 p
= conf
->disks
+ disk
;
5821 if (p
->rdev
== NULL
) {
5822 clear_bit(In_sync
, &rdev
->flags
);
5823 rdev
->raid_disk
= disk
;
5825 if (rdev
->saved_raid_disk
!= disk
)
5827 rcu_assign_pointer(p
->rdev
, rdev
);
5831 for (disk
= first
; disk
<= last
; disk
++) {
5832 p
= conf
->disks
+ disk
;
5833 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
5834 p
->replacement
== NULL
) {
5835 clear_bit(In_sync
, &rdev
->flags
);
5836 set_bit(Replacement
, &rdev
->flags
);
5837 rdev
->raid_disk
= disk
;
5840 rcu_assign_pointer(p
->replacement
, rdev
);
5845 print_raid5_conf(conf
);
5849 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
5851 /* no resync is happening, and there is enough space
5852 * on all devices, so we can resize.
5853 * We need to make sure resync covers any new space.
5854 * If the array is shrinking we should possibly wait until
5855 * any io in the removed space completes, but it hardly seems
5859 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
5860 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
5861 if (mddev
->external_size
&&
5862 mddev
->array_sectors
> newsize
)
5864 if (mddev
->bitmap
) {
5865 int ret
= bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
5869 md_set_array_sectors(mddev
, newsize
);
5870 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5871 revalidate_disk(mddev
->gendisk
);
5872 if (sectors
> mddev
->dev_sectors
&&
5873 mddev
->recovery_cp
> mddev
->dev_sectors
) {
5874 mddev
->recovery_cp
= mddev
->dev_sectors
;
5875 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
5877 mddev
->dev_sectors
= sectors
;
5878 mddev
->resync_max_sectors
= sectors
;
5882 static int check_stripe_cache(struct mddev
*mddev
)
5884 /* Can only proceed if there are plenty of stripe_heads.
5885 * We need a minimum of one full stripe,, and for sensible progress
5886 * it is best to have about 4 times that.
5887 * If we require 4 times, then the default 256 4K stripe_heads will
5888 * allow for chunk sizes up to 256K, which is probably OK.
5889 * If the chunk size is greater, user-space should request more
5890 * stripe_heads first.
5892 struct r5conf
*conf
= mddev
->private;
5893 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5894 > conf
->max_nr_stripes
||
5895 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5896 > conf
->max_nr_stripes
) {
5897 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5899 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
5906 static int check_reshape(struct mddev
*mddev
)
5908 struct r5conf
*conf
= mddev
->private;
5910 if (mddev
->delta_disks
== 0 &&
5911 mddev
->new_layout
== mddev
->layout
&&
5912 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
5913 return 0; /* nothing to do */
5914 if (has_failed(conf
))
5916 if (mddev
->delta_disks
< 0) {
5917 /* We might be able to shrink, but the devices must
5918 * be made bigger first.
5919 * For raid6, 4 is the minimum size.
5920 * Otherwise 2 is the minimum
5923 if (mddev
->level
== 6)
5925 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
5929 if (!check_stripe_cache(mddev
))
5932 return resize_stripes(conf
, (conf
->previous_raid_disks
5933 + mddev
->delta_disks
));
5936 static int raid5_start_reshape(struct mddev
*mddev
)
5938 struct r5conf
*conf
= mddev
->private;
5939 struct md_rdev
*rdev
;
5941 unsigned long flags
;
5943 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
5946 if (!check_stripe_cache(mddev
))
5949 if (has_failed(conf
))
5952 rdev_for_each(rdev
, mddev
) {
5953 if (!test_bit(In_sync
, &rdev
->flags
)
5954 && !test_bit(Faulty
, &rdev
->flags
))
5958 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
5959 /* Not enough devices even to make a degraded array
5964 /* Refuse to reduce size of the array. Any reductions in
5965 * array size must be through explicit setting of array_size
5968 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
5969 < mddev
->array_sectors
) {
5970 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
5971 "before number of disks\n", mdname(mddev
));
5975 atomic_set(&conf
->reshape_stripes
, 0);
5976 spin_lock_irq(&conf
->device_lock
);
5977 conf
->previous_raid_disks
= conf
->raid_disks
;
5978 conf
->raid_disks
+= mddev
->delta_disks
;
5979 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
5980 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
5981 conf
->prev_algo
= conf
->algorithm
;
5982 conf
->algorithm
= mddev
->new_layout
;
5984 /* Code that selects data_offset needs to see the generation update
5985 * if reshape_progress has been set - so a memory barrier needed.
5988 if (mddev
->reshape_backwards
)
5989 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
5991 conf
->reshape_progress
= 0;
5992 conf
->reshape_safe
= conf
->reshape_progress
;
5993 spin_unlock_irq(&conf
->device_lock
);
5995 /* Add some new drives, as many as will fit.
5996 * We know there are enough to make the newly sized array work.
5997 * Don't add devices if we are reducing the number of
5998 * devices in the array. This is because it is not possible
5999 * to correctly record the "partially reconstructed" state of
6000 * such devices during the reshape and confusion could result.
6002 if (mddev
->delta_disks
>= 0) {
6003 rdev_for_each(rdev
, mddev
)
6004 if (rdev
->raid_disk
< 0 &&
6005 !test_bit(Faulty
, &rdev
->flags
)) {
6006 if (raid5_add_disk(mddev
, rdev
) == 0) {
6008 >= conf
->previous_raid_disks
)
6009 set_bit(In_sync
, &rdev
->flags
);
6011 rdev
->recovery_offset
= 0;
6013 if (sysfs_link_rdev(mddev
, rdev
))
6014 /* Failure here is OK */;
6016 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
6017 && !test_bit(Faulty
, &rdev
->flags
)) {
6018 /* This is a spare that was manually added */
6019 set_bit(In_sync
, &rdev
->flags
);
6022 /* When a reshape changes the number of devices,
6023 * ->degraded is measured against the larger of the
6024 * pre and post number of devices.
6026 spin_lock_irqsave(&conf
->device_lock
, flags
);
6027 mddev
->degraded
= calc_degraded(conf
);
6028 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
6030 mddev
->raid_disks
= conf
->raid_disks
;
6031 mddev
->reshape_position
= conf
->reshape_progress
;
6032 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
6034 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
6035 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
6036 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
6037 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
6038 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
6040 if (!mddev
->sync_thread
) {
6041 mddev
->recovery
= 0;
6042 spin_lock_irq(&conf
->device_lock
);
6043 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
6044 rdev_for_each(rdev
, mddev
)
6045 rdev
->new_data_offset
= rdev
->data_offset
;
6047 conf
->reshape_progress
= MaxSector
;
6048 mddev
->reshape_position
= MaxSector
;
6049 spin_unlock_irq(&conf
->device_lock
);
6052 conf
->reshape_checkpoint
= jiffies
;
6053 md_wakeup_thread(mddev
->sync_thread
);
6054 md_new_event(mddev
);
6058 /* This is called from the reshape thread and should make any
6059 * changes needed in 'conf'
6061 static void end_reshape(struct r5conf
*conf
)
6064 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
6065 struct md_rdev
*rdev
;
6067 spin_lock_irq(&conf
->device_lock
);
6068 conf
->previous_raid_disks
= conf
->raid_disks
;
6069 rdev_for_each(rdev
, conf
->mddev
)
6070 rdev
->data_offset
= rdev
->new_data_offset
;
6072 conf
->reshape_progress
= MaxSector
;
6073 spin_unlock_irq(&conf
->device_lock
);
6074 wake_up(&conf
->wait_for_overlap
);
6076 /* read-ahead size must cover two whole stripes, which is
6077 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6079 if (conf
->mddev
->queue
) {
6080 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
6081 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
6083 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
6084 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
6089 /* This is called from the raid5d thread with mddev_lock held.
6090 * It makes config changes to the device.
6092 static void raid5_finish_reshape(struct mddev
*mddev
)
6094 struct r5conf
*conf
= mddev
->private;
6096 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
6098 if (mddev
->delta_disks
> 0) {
6099 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
6100 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
6101 revalidate_disk(mddev
->gendisk
);
6104 spin_lock_irq(&conf
->device_lock
);
6105 mddev
->degraded
= calc_degraded(conf
);
6106 spin_unlock_irq(&conf
->device_lock
);
6107 for (d
= conf
->raid_disks
;
6108 d
< conf
->raid_disks
- mddev
->delta_disks
;
6110 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
6112 clear_bit(In_sync
, &rdev
->flags
);
6113 rdev
= conf
->disks
[d
].replacement
;
6115 clear_bit(In_sync
, &rdev
->flags
);
6118 mddev
->layout
= conf
->algorithm
;
6119 mddev
->chunk_sectors
= conf
->chunk_sectors
;
6120 mddev
->reshape_position
= MaxSector
;
6121 mddev
->delta_disks
= 0;
6122 mddev
->reshape_backwards
= 0;
6126 static void raid5_quiesce(struct mddev
*mddev
, int state
)
6128 struct r5conf
*conf
= mddev
->private;
6131 case 2: /* resume for a suspend */
6132 wake_up(&conf
->wait_for_overlap
);
6135 case 1: /* stop all writes */
6136 spin_lock_irq(&conf
->device_lock
);
6137 /* '2' tells resync/reshape to pause so that all
6138 * active stripes can drain
6141 wait_event_lock_irq(conf
->wait_for_stripe
,
6142 atomic_read(&conf
->active_stripes
) == 0 &&
6143 atomic_read(&conf
->active_aligned_reads
) == 0,
6146 spin_unlock_irq(&conf
->device_lock
);
6147 /* allow reshape to continue */
6148 wake_up(&conf
->wait_for_overlap
);
6151 case 0: /* re-enable writes */
6152 spin_lock_irq(&conf
->device_lock
);
6154 wake_up(&conf
->wait_for_stripe
);
6155 wake_up(&conf
->wait_for_overlap
);
6156 spin_unlock_irq(&conf
->device_lock
);
6162 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
6164 struct r0conf
*raid0_conf
= mddev
->private;
6167 /* for raid0 takeover only one zone is supported */
6168 if (raid0_conf
->nr_strip_zones
> 1) {
6169 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6171 return ERR_PTR(-EINVAL
);
6174 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
6175 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
6176 mddev
->dev_sectors
= sectors
;
6177 mddev
->new_level
= level
;
6178 mddev
->new_layout
= ALGORITHM_PARITY_N
;
6179 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
6180 mddev
->raid_disks
+= 1;
6181 mddev
->delta_disks
= 1;
6182 /* make sure it will be not marked as dirty */
6183 mddev
->recovery_cp
= MaxSector
;
6185 return setup_conf(mddev
);
6189 static void *raid5_takeover_raid1(struct mddev
*mddev
)
6193 if (mddev
->raid_disks
!= 2 ||
6194 mddev
->degraded
> 1)
6195 return ERR_PTR(-EINVAL
);
6197 /* Should check if there are write-behind devices? */
6199 chunksect
= 64*2; /* 64K by default */
6201 /* The array must be an exact multiple of chunksize */
6202 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
6205 if ((chunksect
<<9) < STRIPE_SIZE
)
6206 /* array size does not allow a suitable chunk size */
6207 return ERR_PTR(-EINVAL
);
6209 mddev
->new_level
= 5;
6210 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
6211 mddev
->new_chunk_sectors
= chunksect
;
6213 return setup_conf(mddev
);
6216 static void *raid5_takeover_raid6(struct mddev
*mddev
)
6220 switch (mddev
->layout
) {
6221 case ALGORITHM_LEFT_ASYMMETRIC_6
:
6222 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
6224 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
6225 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
6227 case ALGORITHM_LEFT_SYMMETRIC_6
:
6228 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
6230 case ALGORITHM_RIGHT_SYMMETRIC_6
:
6231 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
6233 case ALGORITHM_PARITY_0_6
:
6234 new_layout
= ALGORITHM_PARITY_0
;
6236 case ALGORITHM_PARITY_N
:
6237 new_layout
= ALGORITHM_PARITY_N
;
6240 return ERR_PTR(-EINVAL
);
6242 mddev
->new_level
= 5;
6243 mddev
->new_layout
= new_layout
;
6244 mddev
->delta_disks
= -1;
6245 mddev
->raid_disks
-= 1;
6246 return setup_conf(mddev
);
6250 static int raid5_check_reshape(struct mddev
*mddev
)
6252 /* For a 2-drive array, the layout and chunk size can be changed
6253 * immediately as not restriping is needed.
6254 * For larger arrays we record the new value - after validation
6255 * to be used by a reshape pass.
6257 struct r5conf
*conf
= mddev
->private;
6258 int new_chunk
= mddev
->new_chunk_sectors
;
6260 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
6262 if (new_chunk
> 0) {
6263 if (!is_power_of_2(new_chunk
))
6265 if (new_chunk
< (PAGE_SIZE
>>9))
6267 if (mddev
->array_sectors
& (new_chunk
-1))
6268 /* not factor of array size */
6272 /* They look valid */
6274 if (mddev
->raid_disks
== 2) {
6275 /* can make the change immediately */
6276 if (mddev
->new_layout
>= 0) {
6277 conf
->algorithm
= mddev
->new_layout
;
6278 mddev
->layout
= mddev
->new_layout
;
6280 if (new_chunk
> 0) {
6281 conf
->chunk_sectors
= new_chunk
;
6282 mddev
->chunk_sectors
= new_chunk
;
6284 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
6285 md_wakeup_thread(mddev
->thread
);
6287 return check_reshape(mddev
);
6290 static int raid6_check_reshape(struct mddev
*mddev
)
6292 int new_chunk
= mddev
->new_chunk_sectors
;
6294 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
6296 if (new_chunk
> 0) {
6297 if (!is_power_of_2(new_chunk
))
6299 if (new_chunk
< (PAGE_SIZE
>> 9))
6301 if (mddev
->array_sectors
& (new_chunk
-1))
6302 /* not factor of array size */
6306 /* They look valid */
6307 return check_reshape(mddev
);
6310 static void *raid5_takeover(struct mddev
*mddev
)
6312 /* raid5 can take over:
6313 * raid0 - if there is only one strip zone - make it a raid4 layout
6314 * raid1 - if there are two drives. We need to know the chunk size
6315 * raid4 - trivial - just use a raid4 layout.
6316 * raid6 - Providing it is a *_6 layout
6318 if (mddev
->level
== 0)
6319 return raid45_takeover_raid0(mddev
, 5);
6320 if (mddev
->level
== 1)
6321 return raid5_takeover_raid1(mddev
);
6322 if (mddev
->level
== 4) {
6323 mddev
->new_layout
= ALGORITHM_PARITY_N
;
6324 mddev
->new_level
= 5;
6325 return setup_conf(mddev
);
6327 if (mddev
->level
== 6)
6328 return raid5_takeover_raid6(mddev
);
6330 return ERR_PTR(-EINVAL
);
6333 static void *raid4_takeover(struct mddev
*mddev
)
6335 /* raid4 can take over:
6336 * raid0 - if there is only one strip zone
6337 * raid5 - if layout is right
6339 if (mddev
->level
== 0)
6340 return raid45_takeover_raid0(mddev
, 4);
6341 if (mddev
->level
== 5 &&
6342 mddev
->layout
== ALGORITHM_PARITY_N
) {
6343 mddev
->new_layout
= 0;
6344 mddev
->new_level
= 4;
6345 return setup_conf(mddev
);
6347 return ERR_PTR(-EINVAL
);
6350 static struct md_personality raid5_personality
;
6352 static void *raid6_takeover(struct mddev
*mddev
)
6354 /* Currently can only take over a raid5. We map the
6355 * personality to an equivalent raid6 personality
6356 * with the Q block at the end.
6360 if (mddev
->pers
!= &raid5_personality
)
6361 return ERR_PTR(-EINVAL
);
6362 if (mddev
->degraded
> 1)
6363 return ERR_PTR(-EINVAL
);
6364 if (mddev
->raid_disks
> 253)
6365 return ERR_PTR(-EINVAL
);
6366 if (mddev
->raid_disks
< 3)
6367 return ERR_PTR(-EINVAL
);
6369 switch (mddev
->layout
) {
6370 case ALGORITHM_LEFT_ASYMMETRIC
:
6371 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
6373 case ALGORITHM_RIGHT_ASYMMETRIC
:
6374 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
6376 case ALGORITHM_LEFT_SYMMETRIC
:
6377 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
6379 case ALGORITHM_RIGHT_SYMMETRIC
:
6380 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
6382 case ALGORITHM_PARITY_0
:
6383 new_layout
= ALGORITHM_PARITY_0_6
;
6385 case ALGORITHM_PARITY_N
:
6386 new_layout
= ALGORITHM_PARITY_N
;
6389 return ERR_PTR(-EINVAL
);
6391 mddev
->new_level
= 6;
6392 mddev
->new_layout
= new_layout
;
6393 mddev
->delta_disks
= 1;
6394 mddev
->raid_disks
+= 1;
6395 return setup_conf(mddev
);
6399 static struct md_personality raid6_personality
=
6403 .owner
= THIS_MODULE
,
6404 .make_request
= make_request
,
6408 .error_handler
= error
,
6409 .hot_add_disk
= raid5_add_disk
,
6410 .hot_remove_disk
= raid5_remove_disk
,
6411 .spare_active
= raid5_spare_active
,
6412 .sync_request
= sync_request
,
6413 .resize
= raid5_resize
,
6415 .check_reshape
= raid6_check_reshape
,
6416 .start_reshape
= raid5_start_reshape
,
6417 .finish_reshape
= raid5_finish_reshape
,
6418 .quiesce
= raid5_quiesce
,
6419 .takeover
= raid6_takeover
,
6421 static struct md_personality raid5_personality
=
6425 .owner
= THIS_MODULE
,
6426 .make_request
= make_request
,
6430 .error_handler
= error
,
6431 .hot_add_disk
= raid5_add_disk
,
6432 .hot_remove_disk
= raid5_remove_disk
,
6433 .spare_active
= raid5_spare_active
,
6434 .sync_request
= sync_request
,
6435 .resize
= raid5_resize
,
6437 .check_reshape
= raid5_check_reshape
,
6438 .start_reshape
= raid5_start_reshape
,
6439 .finish_reshape
= raid5_finish_reshape
,
6440 .quiesce
= raid5_quiesce
,
6441 .takeover
= raid5_takeover
,
6444 static struct md_personality raid4_personality
=
6448 .owner
= THIS_MODULE
,
6449 .make_request
= make_request
,
6453 .error_handler
= error
,
6454 .hot_add_disk
= raid5_add_disk
,
6455 .hot_remove_disk
= raid5_remove_disk
,
6456 .spare_active
= raid5_spare_active
,
6457 .sync_request
= sync_request
,
6458 .resize
= raid5_resize
,
6460 .check_reshape
= raid5_check_reshape
,
6461 .start_reshape
= raid5_start_reshape
,
6462 .finish_reshape
= raid5_finish_reshape
,
6463 .quiesce
= raid5_quiesce
,
6464 .takeover
= raid4_takeover
,
6467 static int __init
raid5_init(void)
6469 register_md_personality(&raid6_personality
);
6470 register_md_personality(&raid5_personality
);
6471 register_md_personality(&raid4_personality
);
6475 static void raid5_exit(void)
6477 unregister_md_personality(&raid6_personality
);
6478 unregister_md_personality(&raid5_personality
);
6479 unregister_md_personality(&raid4_personality
);
6482 module_init(raid5_init
);
6483 module_exit(raid5_exit
);
6484 MODULE_LICENSE("GPL");
6485 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6486 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6487 MODULE_ALIAS("md-raid5");
6488 MODULE_ALIAS("md-raid4");
6489 MODULE_ALIAS("md-level-5");
6490 MODULE_ALIAS("md-level-4");
6491 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6492 MODULE_ALIAS("md-raid6");
6493 MODULE_ALIAS("md-level-6");
6495 /* This used to be two separate modules, they were: */
6496 MODULE_ALIAS("raid5");
6497 MODULE_ALIAS("raid6");