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_sectors(bio
);
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 */
573 rrdev
= rcu_dereference(conf
->disks
[i
].replacement
);
574 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
575 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
584 /* We raced and saw duplicates */
587 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
) && rrdev
)
592 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
595 atomic_inc(&rdev
->nr_pending
);
596 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
599 atomic_inc(&rrdev
->nr_pending
);
602 /* We have already checked bad blocks for reads. Now
603 * need to check for writes. We never accept write errors
604 * on the replacement, so we don't to check rrdev.
606 while ((rw
& WRITE
) && rdev
&&
607 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
610 int bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
611 &first_bad
, &bad_sectors
);
616 set_bit(BlockedBadBlocks
, &rdev
->flags
);
617 if (!conf
->mddev
->external
&&
618 conf
->mddev
->flags
) {
619 /* It is very unlikely, but we might
620 * still need to write out the
621 * bad block log - better give it
623 md_check_recovery(conf
->mddev
);
626 * Because md_wait_for_blocked_rdev
627 * will dec nr_pending, we must
628 * increment it first.
630 atomic_inc(&rdev
->nr_pending
);
631 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
633 /* Acknowledged bad block - skip the write */
634 rdev_dec_pending(rdev
, conf
->mddev
);
640 if (s
->syncing
|| s
->expanding
|| s
->expanded
642 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
644 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
647 bi
->bi_bdev
= rdev
->bdev
;
649 bi
->bi_end_io
= (rw
& WRITE
)
650 ? raid5_end_write_request
651 : raid5_end_read_request
;
654 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
655 __func__
, (unsigned long long)sh
->sector
,
657 atomic_inc(&sh
->count
);
658 if (use_new_offset(conf
, sh
))
659 bi
->bi_sector
= (sh
->sector
660 + rdev
->new_data_offset
);
662 bi
->bi_sector
= (sh
->sector
663 + rdev
->data_offset
);
664 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
665 bi
->bi_rw
|= REQ_FLUSH
;
668 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
669 bi
->bi_io_vec
[0].bv_offset
= 0;
670 bi
->bi_size
= STRIPE_SIZE
;
672 * If this is discard request, set bi_vcnt 0. We don't
673 * want to confuse SCSI because SCSI will replace payload
675 if (rw
& REQ_DISCARD
)
678 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
680 if (conf
->mddev
->gendisk
)
681 trace_block_bio_remap(bdev_get_queue(bi
->bi_bdev
),
682 bi
, disk_devt(conf
->mddev
->gendisk
),
684 generic_make_request(bi
);
687 if (s
->syncing
|| s
->expanding
|| s
->expanded
689 md_sync_acct(rrdev
->bdev
, STRIPE_SECTORS
);
691 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
694 rbi
->bi_bdev
= rrdev
->bdev
;
696 BUG_ON(!(rw
& WRITE
));
697 rbi
->bi_end_io
= raid5_end_write_request
;
698 rbi
->bi_private
= sh
;
700 pr_debug("%s: for %llu schedule op %ld on "
701 "replacement disc %d\n",
702 __func__
, (unsigned long long)sh
->sector
,
704 atomic_inc(&sh
->count
);
705 if (use_new_offset(conf
, sh
))
706 rbi
->bi_sector
= (sh
->sector
707 + rrdev
->new_data_offset
);
709 rbi
->bi_sector
= (sh
->sector
710 + rrdev
->data_offset
);
712 rbi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
713 rbi
->bi_io_vec
[0].bv_offset
= 0;
714 rbi
->bi_size
= STRIPE_SIZE
;
716 * If this is discard request, set bi_vcnt 0. We don't
717 * want to confuse SCSI because SCSI will replace payload
719 if (rw
& REQ_DISCARD
)
721 if (conf
->mddev
->gendisk
)
722 trace_block_bio_remap(bdev_get_queue(rbi
->bi_bdev
),
723 rbi
, disk_devt(conf
->mddev
->gendisk
),
725 generic_make_request(rbi
);
727 if (!rdev
&& !rrdev
) {
729 set_bit(STRIPE_DEGRADED
, &sh
->state
);
730 pr_debug("skip op %ld on disc %d for sector %llu\n",
731 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
732 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
733 set_bit(STRIPE_HANDLE
, &sh
->state
);
738 static struct dma_async_tx_descriptor
*
739 async_copy_data(int frombio
, struct bio
*bio
, struct page
*page
,
740 sector_t sector
, struct dma_async_tx_descriptor
*tx
)
743 struct page
*bio_page
;
746 struct async_submit_ctl submit
;
747 enum async_tx_flags flags
= 0;
749 if (bio
->bi_sector
>= sector
)
750 page_offset
= (signed)(bio
->bi_sector
- sector
) * 512;
752 page_offset
= (signed)(sector
- bio
->bi_sector
) * -512;
755 flags
|= ASYNC_TX_FENCE
;
756 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
758 bio_for_each_segment(bvl
, bio
, i
) {
759 int len
= bvl
->bv_len
;
763 if (page_offset
< 0) {
764 b_offset
= -page_offset
;
765 page_offset
+= b_offset
;
769 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
770 clen
= STRIPE_SIZE
- page_offset
;
775 b_offset
+= bvl
->bv_offset
;
776 bio_page
= bvl
->bv_page
;
778 tx
= async_memcpy(page
, bio_page
, page_offset
,
779 b_offset
, clen
, &submit
);
781 tx
= async_memcpy(bio_page
, page
, b_offset
,
782 page_offset
, clen
, &submit
);
784 /* chain the operations */
785 submit
.depend_tx
= tx
;
787 if (clen
< len
) /* hit end of page */
795 static void ops_complete_biofill(void *stripe_head_ref
)
797 struct stripe_head
*sh
= stripe_head_ref
;
798 struct bio
*return_bi
= NULL
;
801 pr_debug("%s: stripe %llu\n", __func__
,
802 (unsigned long long)sh
->sector
);
804 /* clear completed biofills */
805 for (i
= sh
->disks
; i
--; ) {
806 struct r5dev
*dev
= &sh
->dev
[i
];
808 /* acknowledge completion of a biofill operation */
809 /* and check if we need to reply to a read request,
810 * new R5_Wantfill requests are held off until
811 * !STRIPE_BIOFILL_RUN
813 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
814 struct bio
*rbi
, *rbi2
;
819 while (rbi
&& rbi
->bi_sector
<
820 dev
->sector
+ STRIPE_SECTORS
) {
821 rbi2
= r5_next_bio(rbi
, dev
->sector
);
822 if (!raid5_dec_bi_active_stripes(rbi
)) {
823 rbi
->bi_next
= return_bi
;
830 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
832 return_io(return_bi
);
834 set_bit(STRIPE_HANDLE
, &sh
->state
);
838 static void ops_run_biofill(struct stripe_head
*sh
)
840 struct dma_async_tx_descriptor
*tx
= NULL
;
841 struct async_submit_ctl submit
;
844 pr_debug("%s: stripe %llu\n", __func__
,
845 (unsigned long long)sh
->sector
);
847 for (i
= sh
->disks
; i
--; ) {
848 struct r5dev
*dev
= &sh
->dev
[i
];
849 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
851 spin_lock_irq(&sh
->stripe_lock
);
852 dev
->read
= rbi
= dev
->toread
;
854 spin_unlock_irq(&sh
->stripe_lock
);
855 while (rbi
&& rbi
->bi_sector
<
856 dev
->sector
+ STRIPE_SECTORS
) {
857 tx
= async_copy_data(0, rbi
, dev
->page
,
859 rbi
= r5_next_bio(rbi
, dev
->sector
);
864 atomic_inc(&sh
->count
);
865 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
866 async_trigger_callback(&submit
);
869 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
876 tgt
= &sh
->dev
[target
];
877 set_bit(R5_UPTODATE
, &tgt
->flags
);
878 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
879 clear_bit(R5_Wantcompute
, &tgt
->flags
);
882 static void ops_complete_compute(void *stripe_head_ref
)
884 struct stripe_head
*sh
= stripe_head_ref
;
886 pr_debug("%s: stripe %llu\n", __func__
,
887 (unsigned long long)sh
->sector
);
889 /* mark the computed target(s) as uptodate */
890 mark_target_uptodate(sh
, sh
->ops
.target
);
891 mark_target_uptodate(sh
, sh
->ops
.target2
);
893 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
894 if (sh
->check_state
== check_state_compute_run
)
895 sh
->check_state
= check_state_compute_result
;
896 set_bit(STRIPE_HANDLE
, &sh
->state
);
900 /* return a pointer to the address conversion region of the scribble buffer */
901 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
902 struct raid5_percpu
*percpu
)
904 return percpu
->scribble
+ sizeof(struct page
*) * (sh
->disks
+ 2);
907 static struct dma_async_tx_descriptor
*
908 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
910 int disks
= sh
->disks
;
911 struct page
**xor_srcs
= percpu
->scribble
;
912 int target
= sh
->ops
.target
;
913 struct r5dev
*tgt
= &sh
->dev
[target
];
914 struct page
*xor_dest
= tgt
->page
;
916 struct dma_async_tx_descriptor
*tx
;
917 struct async_submit_ctl submit
;
920 pr_debug("%s: stripe %llu block: %d\n",
921 __func__
, (unsigned long long)sh
->sector
, target
);
922 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
924 for (i
= disks
; i
--; )
926 xor_srcs
[count
++] = sh
->dev
[i
].page
;
928 atomic_inc(&sh
->count
);
930 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
931 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
));
932 if (unlikely(count
== 1))
933 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
935 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
940 /* set_syndrome_sources - populate source buffers for gen_syndrome
941 * @srcs - (struct page *) array of size sh->disks
942 * @sh - stripe_head to parse
944 * Populates srcs in proper layout order for the stripe and returns the
945 * 'count' of sources to be used in a call to async_gen_syndrome. The P
946 * destination buffer is recorded in srcs[count] and the Q destination
947 * is recorded in srcs[count+1]].
949 static int set_syndrome_sources(struct page
**srcs
, struct stripe_head
*sh
)
951 int disks
= sh
->disks
;
952 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
953 int d0_idx
= raid6_d0(sh
);
957 for (i
= 0; i
< disks
; i
++)
963 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
965 srcs
[slot
] = sh
->dev
[i
].page
;
966 i
= raid6_next_disk(i
, disks
);
967 } while (i
!= d0_idx
);
969 return syndrome_disks
;
972 static struct dma_async_tx_descriptor
*
973 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
975 int disks
= sh
->disks
;
976 struct page
**blocks
= percpu
->scribble
;
978 int qd_idx
= sh
->qd_idx
;
979 struct dma_async_tx_descriptor
*tx
;
980 struct async_submit_ctl submit
;
986 if (sh
->ops
.target
< 0)
987 target
= sh
->ops
.target2
;
988 else if (sh
->ops
.target2
< 0)
989 target
= sh
->ops
.target
;
991 /* we should only have one valid target */
994 pr_debug("%s: stripe %llu block: %d\n",
995 __func__
, (unsigned long long)sh
->sector
, target
);
997 tgt
= &sh
->dev
[target
];
998 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1001 atomic_inc(&sh
->count
);
1003 if (target
== qd_idx
) {
1004 count
= set_syndrome_sources(blocks
, sh
);
1005 blocks
[count
] = NULL
; /* regenerating p is not necessary */
1006 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
1007 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1008 ops_complete_compute
, sh
,
1009 to_addr_conv(sh
, percpu
));
1010 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1012 /* Compute any data- or p-drive using XOR */
1014 for (i
= disks
; i
-- ; ) {
1015 if (i
== target
|| i
== qd_idx
)
1017 blocks
[count
++] = sh
->dev
[i
].page
;
1020 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1021 NULL
, ops_complete_compute
, sh
,
1022 to_addr_conv(sh
, percpu
));
1023 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
1029 static struct dma_async_tx_descriptor
*
1030 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1032 int i
, count
, disks
= sh
->disks
;
1033 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1034 int d0_idx
= raid6_d0(sh
);
1035 int faila
= -1, failb
= -1;
1036 int target
= sh
->ops
.target
;
1037 int target2
= sh
->ops
.target2
;
1038 struct r5dev
*tgt
= &sh
->dev
[target
];
1039 struct r5dev
*tgt2
= &sh
->dev
[target2
];
1040 struct dma_async_tx_descriptor
*tx
;
1041 struct page
**blocks
= percpu
->scribble
;
1042 struct async_submit_ctl submit
;
1044 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1045 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
1046 BUG_ON(target
< 0 || target2
< 0);
1047 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1048 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
1050 /* we need to open-code set_syndrome_sources to handle the
1051 * slot number conversion for 'faila' and 'failb'
1053 for (i
= 0; i
< disks
; i
++)
1058 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1060 blocks
[slot
] = sh
->dev
[i
].page
;
1066 i
= raid6_next_disk(i
, disks
);
1067 } while (i
!= d0_idx
);
1069 BUG_ON(faila
== failb
);
1072 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1073 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1075 atomic_inc(&sh
->count
);
1077 if (failb
== syndrome_disks
+1) {
1078 /* Q disk is one of the missing disks */
1079 if (faila
== syndrome_disks
) {
1080 /* Missing P+Q, just recompute */
1081 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1082 ops_complete_compute
, sh
,
1083 to_addr_conv(sh
, percpu
));
1084 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
1085 STRIPE_SIZE
, &submit
);
1089 int qd_idx
= sh
->qd_idx
;
1091 /* Missing D+Q: recompute D from P, then recompute Q */
1092 if (target
== qd_idx
)
1093 data_target
= target2
;
1095 data_target
= target
;
1098 for (i
= disks
; i
-- ; ) {
1099 if (i
== data_target
|| i
== qd_idx
)
1101 blocks
[count
++] = sh
->dev
[i
].page
;
1103 dest
= sh
->dev
[data_target
].page
;
1104 init_async_submit(&submit
,
1105 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1107 to_addr_conv(sh
, percpu
));
1108 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
1111 count
= set_syndrome_sources(blocks
, sh
);
1112 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1113 ops_complete_compute
, sh
,
1114 to_addr_conv(sh
, percpu
));
1115 return async_gen_syndrome(blocks
, 0, count
+2,
1116 STRIPE_SIZE
, &submit
);
1119 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1120 ops_complete_compute
, sh
,
1121 to_addr_conv(sh
, percpu
));
1122 if (failb
== syndrome_disks
) {
1123 /* We're missing D+P. */
1124 return async_raid6_datap_recov(syndrome_disks
+2,
1128 /* We're missing D+D. */
1129 return async_raid6_2data_recov(syndrome_disks
+2,
1130 STRIPE_SIZE
, faila
, failb
,
1137 static void ops_complete_prexor(void *stripe_head_ref
)
1139 struct stripe_head
*sh
= stripe_head_ref
;
1141 pr_debug("%s: stripe %llu\n", __func__
,
1142 (unsigned long long)sh
->sector
);
1145 static struct dma_async_tx_descriptor
*
1146 ops_run_prexor(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1147 struct dma_async_tx_descriptor
*tx
)
1149 int disks
= sh
->disks
;
1150 struct page
**xor_srcs
= percpu
->scribble
;
1151 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1152 struct async_submit_ctl submit
;
1154 /* existing parity data subtracted */
1155 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1157 pr_debug("%s: stripe %llu\n", __func__
,
1158 (unsigned long long)sh
->sector
);
1160 for (i
= disks
; i
--; ) {
1161 struct r5dev
*dev
= &sh
->dev
[i
];
1162 /* Only process blocks that are known to be uptodate */
1163 if (test_bit(R5_Wantdrain
, &dev
->flags
))
1164 xor_srcs
[count
++] = dev
->page
;
1167 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1168 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
));
1169 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1174 static struct dma_async_tx_descriptor
*
1175 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1177 int disks
= sh
->disks
;
1180 pr_debug("%s: stripe %llu\n", __func__
,
1181 (unsigned long long)sh
->sector
);
1183 for (i
= disks
; i
--; ) {
1184 struct r5dev
*dev
= &sh
->dev
[i
];
1187 if (test_and_clear_bit(R5_Wantdrain
, &dev
->flags
)) {
1190 spin_lock_irq(&sh
->stripe_lock
);
1191 chosen
= dev
->towrite
;
1192 dev
->towrite
= NULL
;
1193 BUG_ON(dev
->written
);
1194 wbi
= dev
->written
= chosen
;
1195 spin_unlock_irq(&sh
->stripe_lock
);
1197 while (wbi
&& wbi
->bi_sector
<
1198 dev
->sector
+ STRIPE_SECTORS
) {
1199 if (wbi
->bi_rw
& REQ_FUA
)
1200 set_bit(R5_WantFUA
, &dev
->flags
);
1201 if (wbi
->bi_rw
& REQ_SYNC
)
1202 set_bit(R5_SyncIO
, &dev
->flags
);
1203 if (wbi
->bi_rw
& REQ_DISCARD
)
1204 set_bit(R5_Discard
, &dev
->flags
);
1206 tx
= async_copy_data(1, wbi
, dev
->page
,
1208 wbi
= r5_next_bio(wbi
, dev
->sector
);
1216 static void ops_complete_reconstruct(void *stripe_head_ref
)
1218 struct stripe_head
*sh
= stripe_head_ref
;
1219 int disks
= sh
->disks
;
1220 int pd_idx
= sh
->pd_idx
;
1221 int qd_idx
= sh
->qd_idx
;
1223 bool fua
= false, sync
= false, discard
= false;
1225 pr_debug("%s: stripe %llu\n", __func__
,
1226 (unsigned long long)sh
->sector
);
1228 for (i
= disks
; i
--; ) {
1229 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1230 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
1231 discard
|= test_bit(R5_Discard
, &sh
->dev
[i
].flags
);
1234 for (i
= disks
; i
--; ) {
1235 struct r5dev
*dev
= &sh
->dev
[i
];
1237 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1239 set_bit(R5_UPTODATE
, &dev
->flags
);
1241 set_bit(R5_WantFUA
, &dev
->flags
);
1243 set_bit(R5_SyncIO
, &dev
->flags
);
1247 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1248 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1249 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1250 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1252 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1253 sh
->reconstruct_state
= reconstruct_state_result
;
1256 set_bit(STRIPE_HANDLE
, &sh
->state
);
1261 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1262 struct dma_async_tx_descriptor
*tx
)
1264 int disks
= sh
->disks
;
1265 struct page
**xor_srcs
= percpu
->scribble
;
1266 struct async_submit_ctl submit
;
1267 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1268 struct page
*xor_dest
;
1270 unsigned long flags
;
1272 pr_debug("%s: stripe %llu\n", __func__
,
1273 (unsigned long long)sh
->sector
);
1275 for (i
= 0; i
< sh
->disks
; i
++) {
1278 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1281 if (i
>= sh
->disks
) {
1282 atomic_inc(&sh
->count
);
1283 set_bit(R5_Discard
, &sh
->dev
[pd_idx
].flags
);
1284 ops_complete_reconstruct(sh
);
1287 /* check if prexor is active which means only process blocks
1288 * that are part of a read-modify-write (written)
1290 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1292 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1293 for (i
= disks
; i
--; ) {
1294 struct r5dev
*dev
= &sh
->dev
[i
];
1296 xor_srcs
[count
++] = dev
->page
;
1299 xor_dest
= sh
->dev
[pd_idx
].page
;
1300 for (i
= disks
; i
--; ) {
1301 struct r5dev
*dev
= &sh
->dev
[i
];
1303 xor_srcs
[count
++] = dev
->page
;
1307 /* 1/ if we prexor'd then the dest is reused as a source
1308 * 2/ if we did not prexor then we are redoing the parity
1309 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1310 * for the synchronous xor case
1312 flags
= ASYNC_TX_ACK
|
1313 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1315 atomic_inc(&sh
->count
);
1317 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, sh
,
1318 to_addr_conv(sh
, percpu
));
1319 if (unlikely(count
== 1))
1320 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1322 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1326 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1327 struct dma_async_tx_descriptor
*tx
)
1329 struct async_submit_ctl submit
;
1330 struct page
**blocks
= percpu
->scribble
;
1333 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1335 for (i
= 0; i
< sh
->disks
; i
++) {
1336 if (sh
->pd_idx
== i
|| sh
->qd_idx
== i
)
1338 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1341 if (i
>= sh
->disks
) {
1342 atomic_inc(&sh
->count
);
1343 set_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
1344 set_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
1345 ops_complete_reconstruct(sh
);
1349 count
= set_syndrome_sources(blocks
, sh
);
1351 atomic_inc(&sh
->count
);
1353 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_reconstruct
,
1354 sh
, to_addr_conv(sh
, percpu
));
1355 async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1358 static void ops_complete_check(void *stripe_head_ref
)
1360 struct stripe_head
*sh
= stripe_head_ref
;
1362 pr_debug("%s: stripe %llu\n", __func__
,
1363 (unsigned long long)sh
->sector
);
1365 sh
->check_state
= check_state_check_result
;
1366 set_bit(STRIPE_HANDLE
, &sh
->state
);
1370 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1372 int disks
= sh
->disks
;
1373 int pd_idx
= sh
->pd_idx
;
1374 int qd_idx
= sh
->qd_idx
;
1375 struct page
*xor_dest
;
1376 struct page
**xor_srcs
= percpu
->scribble
;
1377 struct dma_async_tx_descriptor
*tx
;
1378 struct async_submit_ctl submit
;
1382 pr_debug("%s: stripe %llu\n", __func__
,
1383 (unsigned long long)sh
->sector
);
1386 xor_dest
= sh
->dev
[pd_idx
].page
;
1387 xor_srcs
[count
++] = xor_dest
;
1388 for (i
= disks
; i
--; ) {
1389 if (i
== pd_idx
|| i
== qd_idx
)
1391 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1394 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1395 to_addr_conv(sh
, percpu
));
1396 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1397 &sh
->ops
.zero_sum_result
, &submit
);
1399 atomic_inc(&sh
->count
);
1400 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1401 tx
= async_trigger_callback(&submit
);
1404 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1406 struct page
**srcs
= percpu
->scribble
;
1407 struct async_submit_ctl submit
;
1410 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1411 (unsigned long long)sh
->sector
, checkp
);
1413 count
= set_syndrome_sources(srcs
, sh
);
1417 atomic_inc(&sh
->count
);
1418 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1419 sh
, to_addr_conv(sh
, percpu
));
1420 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1421 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1424 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1426 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1427 struct dma_async_tx_descriptor
*tx
= NULL
;
1428 struct r5conf
*conf
= sh
->raid_conf
;
1429 int level
= conf
->level
;
1430 struct raid5_percpu
*percpu
;
1434 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1435 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1436 ops_run_biofill(sh
);
1440 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1442 tx
= ops_run_compute5(sh
, percpu
);
1444 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
1445 tx
= ops_run_compute6_1(sh
, percpu
);
1447 tx
= ops_run_compute6_2(sh
, percpu
);
1449 /* terminate the chain if reconstruct is not set to be run */
1450 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
1454 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
))
1455 tx
= ops_run_prexor(sh
, percpu
, tx
);
1457 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
1458 tx
= ops_run_biodrain(sh
, tx
);
1462 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
1464 ops_run_reconstruct5(sh
, percpu
, tx
);
1466 ops_run_reconstruct6(sh
, percpu
, tx
);
1469 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
1470 if (sh
->check_state
== check_state_run
)
1471 ops_run_check_p(sh
, percpu
);
1472 else if (sh
->check_state
== check_state_run_q
)
1473 ops_run_check_pq(sh
, percpu
, 0);
1474 else if (sh
->check_state
== check_state_run_pq
)
1475 ops_run_check_pq(sh
, percpu
, 1);
1481 for (i
= disks
; i
--; ) {
1482 struct r5dev
*dev
= &sh
->dev
[i
];
1483 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
1484 wake_up(&sh
->raid_conf
->wait_for_overlap
);
1489 static int grow_one_stripe(struct r5conf
*conf
)
1491 struct stripe_head
*sh
;
1492 sh
= kmem_cache_zalloc(conf
->slab_cache
, GFP_KERNEL
);
1496 sh
->raid_conf
= conf
;
1498 spin_lock_init(&sh
->stripe_lock
);
1500 if (grow_buffers(sh
)) {
1502 kmem_cache_free(conf
->slab_cache
, sh
);
1505 /* we just created an active stripe so... */
1506 atomic_set(&sh
->count
, 1);
1507 atomic_inc(&conf
->active_stripes
);
1508 INIT_LIST_HEAD(&sh
->lru
);
1513 static int grow_stripes(struct r5conf
*conf
, int num
)
1515 struct kmem_cache
*sc
;
1516 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
1518 if (conf
->mddev
->gendisk
)
1519 sprintf(conf
->cache_name
[0],
1520 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
1522 sprintf(conf
->cache_name
[0],
1523 "raid%d-%p", conf
->level
, conf
->mddev
);
1524 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
1526 conf
->active_name
= 0;
1527 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
1528 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
1532 conf
->slab_cache
= sc
;
1533 conf
->pool_size
= devs
;
1535 if (!grow_one_stripe(conf
))
1541 * scribble_len - return the required size of the scribble region
1542 * @num - total number of disks in the array
1544 * The size must be enough to contain:
1545 * 1/ a struct page pointer for each device in the array +2
1546 * 2/ room to convert each entry in (1) to its corresponding dma
1547 * (dma_map_page()) or page (page_address()) address.
1549 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1550 * calculate over all devices (not just the data blocks), using zeros in place
1551 * of the P and Q blocks.
1553 static size_t scribble_len(int num
)
1557 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
1562 static int resize_stripes(struct r5conf
*conf
, int newsize
)
1564 /* Make all the stripes able to hold 'newsize' devices.
1565 * New slots in each stripe get 'page' set to a new page.
1567 * This happens in stages:
1568 * 1/ create a new kmem_cache and allocate the required number of
1570 * 2/ gather all the old stripe_heads and transfer the pages across
1571 * to the new stripe_heads. This will have the side effect of
1572 * freezing the array as once all stripe_heads have been collected,
1573 * no IO will be possible. Old stripe heads are freed once their
1574 * pages have been transferred over, and the old kmem_cache is
1575 * freed when all stripes are done.
1576 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1577 * we simple return a failre status - no need to clean anything up.
1578 * 4/ allocate new pages for the new slots in the new stripe_heads.
1579 * If this fails, we don't bother trying the shrink the
1580 * stripe_heads down again, we just leave them as they are.
1581 * As each stripe_head is processed the new one is released into
1584 * Once step2 is started, we cannot afford to wait for a write,
1585 * so we use GFP_NOIO allocations.
1587 struct stripe_head
*osh
, *nsh
;
1588 LIST_HEAD(newstripes
);
1589 struct disk_info
*ndisks
;
1592 struct kmem_cache
*sc
;
1595 if (newsize
<= conf
->pool_size
)
1596 return 0; /* never bother to shrink */
1598 err
= md_allow_write(conf
->mddev
);
1603 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
1604 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
1609 for (i
= conf
->max_nr_stripes
; i
; i
--) {
1610 nsh
= kmem_cache_zalloc(sc
, GFP_KERNEL
);
1614 nsh
->raid_conf
= conf
;
1615 spin_lock_init(&nsh
->stripe_lock
);
1617 list_add(&nsh
->lru
, &newstripes
);
1620 /* didn't get enough, give up */
1621 while (!list_empty(&newstripes
)) {
1622 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1623 list_del(&nsh
->lru
);
1624 kmem_cache_free(sc
, nsh
);
1626 kmem_cache_destroy(sc
);
1629 /* Step 2 - Must use GFP_NOIO now.
1630 * OK, we have enough stripes, start collecting inactive
1631 * stripes and copying them over
1633 list_for_each_entry(nsh
, &newstripes
, lru
) {
1634 spin_lock_irq(&conf
->device_lock
);
1635 wait_event_lock_irq(conf
->wait_for_stripe
,
1636 !list_empty(&conf
->inactive_list
),
1638 osh
= get_free_stripe(conf
);
1639 spin_unlock_irq(&conf
->device_lock
);
1640 atomic_set(&nsh
->count
, 1);
1641 for(i
=0; i
<conf
->pool_size
; i
++)
1642 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
1643 for( ; i
<newsize
; i
++)
1644 nsh
->dev
[i
].page
= NULL
;
1645 kmem_cache_free(conf
->slab_cache
, osh
);
1647 kmem_cache_destroy(conf
->slab_cache
);
1650 * At this point, we are holding all the stripes so the array
1651 * is completely stalled, so now is a good time to resize
1652 * conf->disks and the scribble region
1654 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
1656 for (i
=0; i
<conf
->raid_disks
; i
++)
1657 ndisks
[i
] = conf
->disks
[i
];
1659 conf
->disks
= ndisks
;
1664 conf
->scribble_len
= scribble_len(newsize
);
1665 for_each_present_cpu(cpu
) {
1666 struct raid5_percpu
*percpu
;
1669 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1670 scribble
= kmalloc(conf
->scribble_len
, GFP_NOIO
);
1673 kfree(percpu
->scribble
);
1674 percpu
->scribble
= scribble
;
1682 /* Step 4, return new stripes to service */
1683 while(!list_empty(&newstripes
)) {
1684 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1685 list_del_init(&nsh
->lru
);
1687 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
1688 if (nsh
->dev
[i
].page
== NULL
) {
1689 struct page
*p
= alloc_page(GFP_NOIO
);
1690 nsh
->dev
[i
].page
= p
;
1694 release_stripe(nsh
);
1696 /* critical section pass, GFP_NOIO no longer needed */
1698 conf
->slab_cache
= sc
;
1699 conf
->active_name
= 1-conf
->active_name
;
1700 conf
->pool_size
= newsize
;
1704 static int drop_one_stripe(struct r5conf
*conf
)
1706 struct stripe_head
*sh
;
1708 spin_lock_irq(&conf
->device_lock
);
1709 sh
= get_free_stripe(conf
);
1710 spin_unlock_irq(&conf
->device_lock
);
1713 BUG_ON(atomic_read(&sh
->count
));
1715 kmem_cache_free(conf
->slab_cache
, sh
);
1716 atomic_dec(&conf
->active_stripes
);
1720 static void shrink_stripes(struct r5conf
*conf
)
1722 while (drop_one_stripe(conf
))
1725 if (conf
->slab_cache
)
1726 kmem_cache_destroy(conf
->slab_cache
);
1727 conf
->slab_cache
= NULL
;
1730 static void raid5_end_read_request(struct bio
* bi
, int error
)
1732 struct stripe_head
*sh
= bi
->bi_private
;
1733 struct r5conf
*conf
= sh
->raid_conf
;
1734 int disks
= sh
->disks
, i
;
1735 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1736 char b
[BDEVNAME_SIZE
];
1737 struct md_rdev
*rdev
= NULL
;
1740 for (i
=0 ; i
<disks
; i
++)
1741 if (bi
== &sh
->dev
[i
].req
)
1744 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1745 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1751 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
1752 /* If replacement finished while this request was outstanding,
1753 * 'replacement' might be NULL already.
1754 * In that case it moved down to 'rdev'.
1755 * rdev is not removed until all requests are finished.
1757 rdev
= conf
->disks
[i
].replacement
;
1759 rdev
= conf
->disks
[i
].rdev
;
1761 if (use_new_offset(conf
, sh
))
1762 s
= sh
->sector
+ rdev
->new_data_offset
;
1764 s
= sh
->sector
+ rdev
->data_offset
;
1766 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1767 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
1768 /* Note that this cannot happen on a
1769 * replacement device. We just fail those on
1774 "md/raid:%s: read error corrected"
1775 " (%lu sectors at %llu on %s)\n",
1776 mdname(conf
->mddev
), STRIPE_SECTORS
,
1777 (unsigned long long)s
,
1778 bdevname(rdev
->bdev
, b
));
1779 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
1780 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1781 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1782 } else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
1783 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
1785 if (atomic_read(&rdev
->read_errors
))
1786 atomic_set(&rdev
->read_errors
, 0);
1788 const char *bdn
= bdevname(rdev
->bdev
, b
);
1792 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1793 atomic_inc(&rdev
->read_errors
);
1794 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
1797 "md/raid:%s: read error on replacement device "
1798 "(sector %llu on %s).\n",
1799 mdname(conf
->mddev
),
1800 (unsigned long long)s
,
1802 else if (conf
->mddev
->degraded
>= conf
->max_degraded
) {
1806 "md/raid:%s: read error not correctable "
1807 "(sector %llu on %s).\n",
1808 mdname(conf
->mddev
),
1809 (unsigned long long)s
,
1811 } else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
)) {
1816 "md/raid:%s: read error NOT corrected!! "
1817 "(sector %llu on %s).\n",
1818 mdname(conf
->mddev
),
1819 (unsigned long long)s
,
1821 } else if (atomic_read(&rdev
->read_errors
)
1822 > conf
->max_nr_stripes
)
1824 "md/raid:%s: Too many read errors, failing device %s.\n",
1825 mdname(conf
->mddev
), bdn
);
1829 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
)) {
1830 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1831 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
1833 set_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
1835 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1836 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1838 && test_bit(In_sync
, &rdev
->flags
)
1839 && rdev_set_badblocks(
1840 rdev
, sh
->sector
, STRIPE_SECTORS
, 0)))
1841 md_error(conf
->mddev
, rdev
);
1844 rdev_dec_pending(rdev
, conf
->mddev
);
1845 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1846 set_bit(STRIPE_HANDLE
, &sh
->state
);
1850 static void raid5_end_write_request(struct bio
*bi
, int error
)
1852 struct stripe_head
*sh
= bi
->bi_private
;
1853 struct r5conf
*conf
= sh
->raid_conf
;
1854 int disks
= sh
->disks
, i
;
1855 struct md_rdev
*uninitialized_var(rdev
);
1856 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1859 int replacement
= 0;
1861 for (i
= 0 ; i
< disks
; i
++) {
1862 if (bi
== &sh
->dev
[i
].req
) {
1863 rdev
= conf
->disks
[i
].rdev
;
1866 if (bi
== &sh
->dev
[i
].rreq
) {
1867 rdev
= conf
->disks
[i
].replacement
;
1871 /* rdev was removed and 'replacement'
1872 * replaced it. rdev is not removed
1873 * until all requests are finished.
1875 rdev
= conf
->disks
[i
].rdev
;
1879 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1880 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1889 md_error(conf
->mddev
, rdev
);
1890 else if (is_badblock(rdev
, sh
->sector
,
1892 &first_bad
, &bad_sectors
))
1893 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
1896 set_bit(STRIPE_DEGRADED
, &sh
->state
);
1897 set_bit(WriteErrorSeen
, &rdev
->flags
);
1898 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
1899 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1900 set_bit(MD_RECOVERY_NEEDED
,
1901 &rdev
->mddev
->recovery
);
1902 } else if (is_badblock(rdev
, sh
->sector
,
1904 &first_bad
, &bad_sectors
)) {
1905 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
1906 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))
1907 /* That was a successful write so make
1908 * sure it looks like we already did
1911 set_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1914 rdev_dec_pending(rdev
, conf
->mddev
);
1916 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
1917 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1918 set_bit(STRIPE_HANDLE
, &sh
->state
);
1922 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
);
1924 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
1926 struct r5dev
*dev
= &sh
->dev
[i
];
1928 bio_init(&dev
->req
);
1929 dev
->req
.bi_io_vec
= &dev
->vec
;
1931 dev
->req
.bi_max_vecs
++;
1932 dev
->req
.bi_private
= sh
;
1933 dev
->vec
.bv_page
= dev
->page
;
1935 bio_init(&dev
->rreq
);
1936 dev
->rreq
.bi_io_vec
= &dev
->rvec
;
1937 dev
->rreq
.bi_vcnt
++;
1938 dev
->rreq
.bi_max_vecs
++;
1939 dev
->rreq
.bi_private
= sh
;
1940 dev
->rvec
.bv_page
= dev
->page
;
1943 dev
->sector
= compute_blocknr(sh
, i
, previous
);
1946 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1948 char b
[BDEVNAME_SIZE
];
1949 struct r5conf
*conf
= mddev
->private;
1950 unsigned long flags
;
1951 pr_debug("raid456: error called\n");
1953 spin_lock_irqsave(&conf
->device_lock
, flags
);
1954 clear_bit(In_sync
, &rdev
->flags
);
1955 mddev
->degraded
= calc_degraded(conf
);
1956 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1957 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1959 set_bit(Blocked
, &rdev
->flags
);
1960 set_bit(Faulty
, &rdev
->flags
);
1961 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1963 "md/raid:%s: Disk failure on %s, disabling device.\n"
1964 "md/raid:%s: Operation continuing on %d devices.\n",
1966 bdevname(rdev
->bdev
, b
),
1968 conf
->raid_disks
- mddev
->degraded
);
1972 * Input: a 'big' sector number,
1973 * Output: index of the data and parity disk, and the sector # in them.
1975 static sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
1976 int previous
, int *dd_idx
,
1977 struct stripe_head
*sh
)
1979 sector_t stripe
, stripe2
;
1980 sector_t chunk_number
;
1981 unsigned int chunk_offset
;
1984 sector_t new_sector
;
1985 int algorithm
= previous
? conf
->prev_algo
1987 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
1988 : conf
->chunk_sectors
;
1989 int raid_disks
= previous
? conf
->previous_raid_disks
1991 int data_disks
= raid_disks
- conf
->max_degraded
;
1993 /* First compute the information on this sector */
1996 * Compute the chunk number and the sector offset inside the chunk
1998 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
1999 chunk_number
= r_sector
;
2002 * Compute the stripe number
2004 stripe
= chunk_number
;
2005 *dd_idx
= sector_div(stripe
, data_disks
);
2008 * Select the parity disk based on the user selected algorithm.
2010 pd_idx
= qd_idx
= -1;
2011 switch(conf
->level
) {
2013 pd_idx
= data_disks
;
2016 switch (algorithm
) {
2017 case ALGORITHM_LEFT_ASYMMETRIC
:
2018 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2019 if (*dd_idx
>= pd_idx
)
2022 case ALGORITHM_RIGHT_ASYMMETRIC
:
2023 pd_idx
= sector_div(stripe2
, raid_disks
);
2024 if (*dd_idx
>= pd_idx
)
2027 case ALGORITHM_LEFT_SYMMETRIC
:
2028 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2029 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2031 case ALGORITHM_RIGHT_SYMMETRIC
:
2032 pd_idx
= sector_div(stripe2
, raid_disks
);
2033 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2035 case ALGORITHM_PARITY_0
:
2039 case ALGORITHM_PARITY_N
:
2040 pd_idx
= data_disks
;
2048 switch (algorithm
) {
2049 case ALGORITHM_LEFT_ASYMMETRIC
:
2050 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2051 qd_idx
= pd_idx
+ 1;
2052 if (pd_idx
== raid_disks
-1) {
2053 (*dd_idx
)++; /* Q D D D P */
2055 } else if (*dd_idx
>= pd_idx
)
2056 (*dd_idx
) += 2; /* D D P Q D */
2058 case ALGORITHM_RIGHT_ASYMMETRIC
:
2059 pd_idx
= sector_div(stripe2
, raid_disks
);
2060 qd_idx
= pd_idx
+ 1;
2061 if (pd_idx
== raid_disks
-1) {
2062 (*dd_idx
)++; /* Q D D D P */
2064 } else if (*dd_idx
>= pd_idx
)
2065 (*dd_idx
) += 2; /* D D P Q D */
2067 case ALGORITHM_LEFT_SYMMETRIC
:
2068 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2069 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2070 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2072 case ALGORITHM_RIGHT_SYMMETRIC
:
2073 pd_idx
= sector_div(stripe2
, raid_disks
);
2074 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2075 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2078 case ALGORITHM_PARITY_0
:
2083 case ALGORITHM_PARITY_N
:
2084 pd_idx
= data_disks
;
2085 qd_idx
= data_disks
+ 1;
2088 case ALGORITHM_ROTATING_ZERO_RESTART
:
2089 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2090 * of blocks for computing Q is different.
2092 pd_idx
= sector_div(stripe2
, raid_disks
);
2093 qd_idx
= pd_idx
+ 1;
2094 if (pd_idx
== raid_disks
-1) {
2095 (*dd_idx
)++; /* Q D D D P */
2097 } else if (*dd_idx
>= pd_idx
)
2098 (*dd_idx
) += 2; /* D D P Q D */
2102 case ALGORITHM_ROTATING_N_RESTART
:
2103 /* Same a left_asymmetric, by first stripe is
2104 * D D D P Q rather than
2108 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2109 qd_idx
= pd_idx
+ 1;
2110 if (pd_idx
== raid_disks
-1) {
2111 (*dd_idx
)++; /* Q D D D P */
2113 } else if (*dd_idx
>= pd_idx
)
2114 (*dd_idx
) += 2; /* D D P Q D */
2118 case ALGORITHM_ROTATING_N_CONTINUE
:
2119 /* Same as left_symmetric but Q is before P */
2120 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2121 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
2122 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2126 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2127 /* RAID5 left_asymmetric, with Q on last device */
2128 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2129 if (*dd_idx
>= pd_idx
)
2131 qd_idx
= raid_disks
- 1;
2134 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2135 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2136 if (*dd_idx
>= pd_idx
)
2138 qd_idx
= raid_disks
- 1;
2141 case ALGORITHM_LEFT_SYMMETRIC_6
:
2142 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2143 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2144 qd_idx
= raid_disks
- 1;
2147 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2148 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2149 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2150 qd_idx
= raid_disks
- 1;
2153 case ALGORITHM_PARITY_0_6
:
2156 qd_idx
= raid_disks
- 1;
2166 sh
->pd_idx
= pd_idx
;
2167 sh
->qd_idx
= qd_idx
;
2168 sh
->ddf_layout
= ddf_layout
;
2171 * Finally, compute the new sector number
2173 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
2178 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
2180 struct r5conf
*conf
= sh
->raid_conf
;
2181 int raid_disks
= sh
->disks
;
2182 int data_disks
= raid_disks
- conf
->max_degraded
;
2183 sector_t new_sector
= sh
->sector
, check
;
2184 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2185 : conf
->chunk_sectors
;
2186 int algorithm
= previous
? conf
->prev_algo
2190 sector_t chunk_number
;
2191 int dummy1
, dd_idx
= i
;
2193 struct stripe_head sh2
;
2196 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
2197 stripe
= new_sector
;
2199 if (i
== sh
->pd_idx
)
2201 switch(conf
->level
) {
2204 switch (algorithm
) {
2205 case ALGORITHM_LEFT_ASYMMETRIC
:
2206 case ALGORITHM_RIGHT_ASYMMETRIC
:
2210 case ALGORITHM_LEFT_SYMMETRIC
:
2211 case ALGORITHM_RIGHT_SYMMETRIC
:
2214 i
-= (sh
->pd_idx
+ 1);
2216 case ALGORITHM_PARITY_0
:
2219 case ALGORITHM_PARITY_N
:
2226 if (i
== sh
->qd_idx
)
2227 return 0; /* It is the Q disk */
2228 switch (algorithm
) {
2229 case ALGORITHM_LEFT_ASYMMETRIC
:
2230 case ALGORITHM_RIGHT_ASYMMETRIC
:
2231 case ALGORITHM_ROTATING_ZERO_RESTART
:
2232 case ALGORITHM_ROTATING_N_RESTART
:
2233 if (sh
->pd_idx
== raid_disks
-1)
2234 i
--; /* Q D D D P */
2235 else if (i
> sh
->pd_idx
)
2236 i
-= 2; /* D D P Q D */
2238 case ALGORITHM_LEFT_SYMMETRIC
:
2239 case ALGORITHM_RIGHT_SYMMETRIC
:
2240 if (sh
->pd_idx
== raid_disks
-1)
2241 i
--; /* Q D D D P */
2246 i
-= (sh
->pd_idx
+ 2);
2249 case ALGORITHM_PARITY_0
:
2252 case ALGORITHM_PARITY_N
:
2254 case ALGORITHM_ROTATING_N_CONTINUE
:
2255 /* Like left_symmetric, but P is before Q */
2256 if (sh
->pd_idx
== 0)
2257 i
--; /* P D D D Q */
2262 i
-= (sh
->pd_idx
+ 1);
2265 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2266 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2270 case ALGORITHM_LEFT_SYMMETRIC_6
:
2271 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2273 i
+= data_disks
+ 1;
2274 i
-= (sh
->pd_idx
+ 1);
2276 case ALGORITHM_PARITY_0_6
:
2285 chunk_number
= stripe
* data_disks
+ i
;
2286 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2288 check
= raid5_compute_sector(conf
, r_sector
,
2289 previous
, &dummy1
, &sh2
);
2290 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2291 || sh2
.qd_idx
!= sh
->qd_idx
) {
2292 printk(KERN_ERR
"md/raid:%s: compute_blocknr: map not correct\n",
2293 mdname(conf
->mddev
));
2301 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2302 int rcw
, int expand
)
2304 int i
, pd_idx
= sh
->pd_idx
, disks
= sh
->disks
;
2305 struct r5conf
*conf
= sh
->raid_conf
;
2306 int level
= conf
->level
;
2310 for (i
= disks
; i
--; ) {
2311 struct r5dev
*dev
= &sh
->dev
[i
];
2314 set_bit(R5_LOCKED
, &dev
->flags
);
2315 set_bit(R5_Wantdrain
, &dev
->flags
);
2317 clear_bit(R5_UPTODATE
, &dev
->flags
);
2321 /* if we are not expanding this is a proper write request, and
2322 * there will be bios with new data to be drained into the
2327 /* False alarm, nothing to do */
2329 sh
->reconstruct_state
= reconstruct_state_drain_run
;
2330 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2332 sh
->reconstruct_state
= reconstruct_state_run
;
2334 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2336 if (s
->locked
+ conf
->max_degraded
== disks
)
2337 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2338 atomic_inc(&conf
->pending_full_writes
);
2341 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
2342 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
2344 for (i
= disks
; i
--; ) {
2345 struct r5dev
*dev
= &sh
->dev
[i
];
2350 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2351 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2352 set_bit(R5_Wantdrain
, &dev
->flags
);
2353 set_bit(R5_LOCKED
, &dev
->flags
);
2354 clear_bit(R5_UPTODATE
, &dev
->flags
);
2359 /* False alarm - nothing to do */
2361 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
2362 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
2363 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2364 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2367 /* keep the parity disk(s) locked while asynchronous operations
2370 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
2371 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2375 int qd_idx
= sh
->qd_idx
;
2376 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
2378 set_bit(R5_LOCKED
, &dev
->flags
);
2379 clear_bit(R5_UPTODATE
, &dev
->flags
);
2383 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2384 __func__
, (unsigned long long)sh
->sector
,
2385 s
->locked
, s
->ops_request
);
2389 * Each stripe/dev can have one or more bion attached.
2390 * toread/towrite point to the first in a chain.
2391 * The bi_next chain must be in order.
2393 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
, int forwrite
)
2396 struct r5conf
*conf
= sh
->raid_conf
;
2399 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2400 (unsigned long long)bi
->bi_sector
,
2401 (unsigned long long)sh
->sector
);
2404 * If several bio share a stripe. The bio bi_phys_segments acts as a
2405 * reference count to avoid race. The reference count should already be
2406 * increased before this function is called (for example, in
2407 * make_request()), so other bio sharing this stripe will not free the
2408 * stripe. If a stripe is owned by one stripe, the stripe lock will
2411 spin_lock_irq(&sh
->stripe_lock
);
2413 bip
= &sh
->dev
[dd_idx
].towrite
;
2417 bip
= &sh
->dev
[dd_idx
].toread
;
2418 while (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
) {
2419 if (bio_end_sector(*bip
) > bi
->bi_sector
)
2421 bip
= & (*bip
)->bi_next
;
2423 if (*bip
&& (*bip
)->bi_sector
< bio_end_sector(bi
))
2426 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
2430 raid5_inc_bi_active_stripes(bi
);
2433 /* check if page is covered */
2434 sector_t sector
= sh
->dev
[dd_idx
].sector
;
2435 for (bi
=sh
->dev
[dd_idx
].towrite
;
2436 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
2437 bi
&& bi
->bi_sector
<= sector
;
2438 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
2439 if (bio_end_sector(bi
) >= sector
)
2440 sector
= bio_end_sector(bi
);
2442 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
2443 set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
);
2446 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2447 (unsigned long long)(*bip
)->bi_sector
,
2448 (unsigned long long)sh
->sector
, dd_idx
);
2449 spin_unlock_irq(&sh
->stripe_lock
);
2451 if (conf
->mddev
->bitmap
&& firstwrite
) {
2452 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
2454 sh
->bm_seq
= conf
->seq_flush
+1;
2455 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
2460 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
2461 spin_unlock_irq(&sh
->stripe_lock
);
2465 static void end_reshape(struct r5conf
*conf
);
2467 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
2468 struct stripe_head
*sh
)
2470 int sectors_per_chunk
=
2471 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
2473 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
2474 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
2476 raid5_compute_sector(conf
,
2477 stripe
* (disks
- conf
->max_degraded
)
2478 *sectors_per_chunk
+ chunk_offset
,
2484 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
2485 struct stripe_head_state
*s
, int disks
,
2486 struct bio
**return_bi
)
2489 for (i
= disks
; i
--; ) {
2493 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2494 struct md_rdev
*rdev
;
2496 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2497 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
2498 atomic_inc(&rdev
->nr_pending
);
2503 if (!rdev_set_badblocks(
2507 md_error(conf
->mddev
, rdev
);
2508 rdev_dec_pending(rdev
, conf
->mddev
);
2511 spin_lock_irq(&sh
->stripe_lock
);
2512 /* fail all writes first */
2513 bi
= sh
->dev
[i
].towrite
;
2514 sh
->dev
[i
].towrite
= NULL
;
2515 spin_unlock_irq(&sh
->stripe_lock
);
2519 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2520 wake_up(&conf
->wait_for_overlap
);
2522 while (bi
&& bi
->bi_sector
<
2523 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2524 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2525 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2526 if (!raid5_dec_bi_active_stripes(bi
)) {
2527 md_write_end(conf
->mddev
);
2528 bi
->bi_next
= *return_bi
;
2534 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2535 STRIPE_SECTORS
, 0, 0);
2537 /* and fail all 'written' */
2538 bi
= sh
->dev
[i
].written
;
2539 sh
->dev
[i
].written
= NULL
;
2540 if (bi
) bitmap_end
= 1;
2541 while (bi
&& bi
->bi_sector
<
2542 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2543 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2544 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2545 if (!raid5_dec_bi_active_stripes(bi
)) {
2546 md_write_end(conf
->mddev
);
2547 bi
->bi_next
= *return_bi
;
2553 /* fail any reads if this device is non-operational and
2554 * the data has not reached the cache yet.
2556 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
2557 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
2558 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
2559 spin_lock_irq(&sh
->stripe_lock
);
2560 bi
= sh
->dev
[i
].toread
;
2561 sh
->dev
[i
].toread
= NULL
;
2562 spin_unlock_irq(&sh
->stripe_lock
);
2563 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2564 wake_up(&conf
->wait_for_overlap
);
2565 while (bi
&& bi
->bi_sector
<
2566 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2567 struct bio
*nextbi
=
2568 r5_next_bio(bi
, sh
->dev
[i
].sector
);
2569 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2570 if (!raid5_dec_bi_active_stripes(bi
)) {
2571 bi
->bi_next
= *return_bi
;
2578 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2579 STRIPE_SECTORS
, 0, 0);
2580 /* If we were in the middle of a write the parity block might
2581 * still be locked - so just clear all R5_LOCKED flags
2583 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2586 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2587 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2588 md_wakeup_thread(conf
->mddev
->thread
);
2592 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
2593 struct stripe_head_state
*s
)
2598 clear_bit(STRIPE_SYNCING
, &sh
->state
);
2599 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
2600 wake_up(&conf
->wait_for_overlap
);
2603 /* There is nothing more to do for sync/check/repair.
2604 * Don't even need to abort as that is handled elsewhere
2605 * if needed, and not always wanted e.g. if there is a known
2607 * For recover/replace we need to record a bad block on all
2608 * non-sync devices, or abort the recovery
2610 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
2611 /* During recovery devices cannot be removed, so
2612 * locking and refcounting of rdevs is not needed
2614 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2615 struct md_rdev
*rdev
= conf
->disks
[i
].rdev
;
2617 && !test_bit(Faulty
, &rdev
->flags
)
2618 && !test_bit(In_sync
, &rdev
->flags
)
2619 && !rdev_set_badblocks(rdev
, sh
->sector
,
2622 rdev
= conf
->disks
[i
].replacement
;
2624 && !test_bit(Faulty
, &rdev
->flags
)
2625 && !test_bit(In_sync
, &rdev
->flags
)
2626 && !rdev_set_badblocks(rdev
, sh
->sector
,
2631 conf
->recovery_disabled
=
2632 conf
->mddev
->recovery_disabled
;
2634 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, !abort
);
2637 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
2639 struct md_rdev
*rdev
;
2641 /* Doing recovery so rcu locking not required */
2642 rdev
= sh
->raid_conf
->disks
[disk_idx
].replacement
;
2644 && !test_bit(Faulty
, &rdev
->flags
)
2645 && !test_bit(In_sync
, &rdev
->flags
)
2646 && (rdev
->recovery_offset
<= sh
->sector
2647 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
2653 /* fetch_block - checks the given member device to see if its data needs
2654 * to be read or computed to satisfy a request.
2656 * Returns 1 when no more member devices need to be checked, otherwise returns
2657 * 0 to tell the loop in handle_stripe_fill to continue
2659 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2660 int disk_idx
, int disks
)
2662 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
2663 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
2664 &sh
->dev
[s
->failed_num
[1]] };
2666 /* is the data in this block needed, and can we get it? */
2667 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2668 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2670 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2671 s
->syncing
|| s
->expanding
||
2672 (s
->replacing
&& want_replace(sh
, disk_idx
)) ||
2673 (s
->failed
>= 1 && fdev
[0]->toread
) ||
2674 (s
->failed
>= 2 && fdev
[1]->toread
) ||
2675 (sh
->raid_conf
->level
<= 5 && s
->failed
&& fdev
[0]->towrite
&&
2676 !test_bit(R5_OVERWRITE
, &fdev
[0]->flags
)) ||
2677 (sh
->raid_conf
->level
== 6 && s
->failed
&& s
->to_write
))) {
2678 /* we would like to get this block, possibly by computing it,
2679 * otherwise read it if the backing disk is insync
2681 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
2682 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
2683 if ((s
->uptodate
== disks
- 1) &&
2684 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
2685 disk_idx
== s
->failed_num
[1]))) {
2686 /* have disk failed, and we're requested to fetch it;
2689 pr_debug("Computing stripe %llu block %d\n",
2690 (unsigned long long)sh
->sector
, disk_idx
);
2691 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2692 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2693 set_bit(R5_Wantcompute
, &dev
->flags
);
2694 sh
->ops
.target
= disk_idx
;
2695 sh
->ops
.target2
= -1; /* no 2nd target */
2697 /* Careful: from this point on 'uptodate' is in the eye
2698 * of raid_run_ops which services 'compute' operations
2699 * before writes. R5_Wantcompute flags a block that will
2700 * be R5_UPTODATE by the time it is needed for a
2701 * subsequent operation.
2705 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
2706 /* Computing 2-failure is *very* expensive; only
2707 * do it if failed >= 2
2710 for (other
= disks
; other
--; ) {
2711 if (other
== disk_idx
)
2713 if (!test_bit(R5_UPTODATE
,
2714 &sh
->dev
[other
].flags
))
2718 pr_debug("Computing stripe %llu blocks %d,%d\n",
2719 (unsigned long long)sh
->sector
,
2721 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2722 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2723 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
2724 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
2725 sh
->ops
.target
= disk_idx
;
2726 sh
->ops
.target2
= other
;
2730 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2731 set_bit(R5_LOCKED
, &dev
->flags
);
2732 set_bit(R5_Wantread
, &dev
->flags
);
2734 pr_debug("Reading block %d (sync=%d)\n",
2735 disk_idx
, s
->syncing
);
2743 * handle_stripe_fill - read or compute data to satisfy pending requests.
2745 static void handle_stripe_fill(struct stripe_head
*sh
,
2746 struct stripe_head_state
*s
,
2751 /* look for blocks to read/compute, skip this if a compute
2752 * is already in flight, or if the stripe contents are in the
2753 * midst of changing due to a write
2755 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2756 !sh
->reconstruct_state
)
2757 for (i
= disks
; i
--; )
2758 if (fetch_block(sh
, s
, i
, disks
))
2760 set_bit(STRIPE_HANDLE
, &sh
->state
);
2764 /* handle_stripe_clean_event
2765 * any written block on an uptodate or failed drive can be returned.
2766 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2767 * never LOCKED, so we don't need to test 'failed' directly.
2769 static void handle_stripe_clean_event(struct r5conf
*conf
,
2770 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
2774 int discard_pending
= 0;
2776 for (i
= disks
; i
--; )
2777 if (sh
->dev
[i
].written
) {
2779 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2780 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2781 test_bit(R5_Discard
, &dev
->flags
))) {
2782 /* We can return any write requests */
2783 struct bio
*wbi
, *wbi2
;
2784 pr_debug("Return write for disc %d\n", i
);
2785 if (test_and_clear_bit(R5_Discard
, &dev
->flags
))
2786 clear_bit(R5_UPTODATE
, &dev
->flags
);
2788 dev
->written
= NULL
;
2789 while (wbi
&& wbi
->bi_sector
<
2790 dev
->sector
+ STRIPE_SECTORS
) {
2791 wbi2
= r5_next_bio(wbi
, dev
->sector
);
2792 if (!raid5_dec_bi_active_stripes(wbi
)) {
2793 md_write_end(conf
->mddev
);
2794 wbi
->bi_next
= *return_bi
;
2799 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2801 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
2803 } else if (test_bit(R5_Discard
, &dev
->flags
))
2804 discard_pending
= 1;
2806 if (!discard_pending
&&
2807 test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
)) {
2808 clear_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
2809 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2810 if (sh
->qd_idx
>= 0) {
2811 clear_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
2812 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
);
2814 /* now that discard is done we can proceed with any sync */
2815 clear_bit(STRIPE_DISCARD
, &sh
->state
);
2817 * SCSI discard will change some bio fields and the stripe has
2818 * no updated data, so remove it from hash list and the stripe
2819 * will be reinitialized
2821 spin_lock_irq(&conf
->device_lock
);
2823 spin_unlock_irq(&conf
->device_lock
);
2824 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
))
2825 set_bit(STRIPE_HANDLE
, &sh
->state
);
2829 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2830 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2831 md_wakeup_thread(conf
->mddev
->thread
);
2834 static void handle_stripe_dirtying(struct r5conf
*conf
,
2835 struct stripe_head
*sh
,
2836 struct stripe_head_state
*s
,
2839 int rmw
= 0, rcw
= 0, i
;
2840 sector_t recovery_cp
= conf
->mddev
->recovery_cp
;
2842 /* RAID6 requires 'rcw' in current implementation.
2843 * Otherwise, check whether resync is now happening or should start.
2844 * If yes, then the array is dirty (after unclean shutdown or
2845 * initial creation), so parity in some stripes might be inconsistent.
2846 * In this case, we need to always do reconstruct-write, to ensure
2847 * that in case of drive failure or read-error correction, we
2848 * generate correct data from the parity.
2850 if (conf
->max_degraded
== 2 ||
2851 (recovery_cp
< MaxSector
&& sh
->sector
>= recovery_cp
)) {
2852 /* Calculate the real rcw later - for now make it
2853 * look like rcw is cheaper
2856 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
2857 conf
->max_degraded
, (unsigned long long)recovery_cp
,
2858 (unsigned long long)sh
->sector
);
2859 } else for (i
= disks
; i
--; ) {
2860 /* would I have to read this buffer for read_modify_write */
2861 struct r5dev
*dev
= &sh
->dev
[i
];
2862 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2863 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2864 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2865 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2866 if (test_bit(R5_Insync
, &dev
->flags
))
2869 rmw
+= 2*disks
; /* cannot read it */
2871 /* Would I have to read this buffer for reconstruct_write */
2872 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) && i
!= sh
->pd_idx
&&
2873 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2874 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2875 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2876 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2881 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2882 (unsigned long long)sh
->sector
, rmw
, rcw
);
2883 set_bit(STRIPE_HANDLE
, &sh
->state
);
2884 if (rmw
< rcw
&& rmw
> 0) {
2885 /* prefer read-modify-write, but need to get some data */
2886 if (conf
->mddev
->queue
)
2887 blk_add_trace_msg(conf
->mddev
->queue
,
2888 "raid5 rmw %llu %d",
2889 (unsigned long long)sh
->sector
, rmw
);
2890 for (i
= disks
; i
--; ) {
2891 struct r5dev
*dev
= &sh
->dev
[i
];
2892 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2893 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2894 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2895 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2896 test_bit(R5_Insync
, &dev
->flags
)) {
2898 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2899 pr_debug("Read_old block "
2900 "%d for r-m-w\n", i
);
2901 set_bit(R5_LOCKED
, &dev
->flags
);
2902 set_bit(R5_Wantread
, &dev
->flags
);
2905 set_bit(STRIPE_DELAYED
, &sh
->state
);
2906 set_bit(STRIPE_HANDLE
, &sh
->state
);
2911 if (rcw
<= rmw
&& rcw
> 0) {
2912 /* want reconstruct write, but need to get some data */
2915 for (i
= disks
; i
--; ) {
2916 struct r5dev
*dev
= &sh
->dev
[i
];
2917 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2918 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
2919 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2920 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2921 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2923 if (!test_bit(R5_Insync
, &dev
->flags
))
2924 continue; /* it's a failed drive */
2926 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2927 pr_debug("Read_old block "
2928 "%d for Reconstruct\n", i
);
2929 set_bit(R5_LOCKED
, &dev
->flags
);
2930 set_bit(R5_Wantread
, &dev
->flags
);
2934 set_bit(STRIPE_DELAYED
, &sh
->state
);
2935 set_bit(STRIPE_HANDLE
, &sh
->state
);
2939 if (rcw
&& conf
->mddev
->queue
)
2940 blk_add_trace_msg(conf
->mddev
->queue
, "raid5 rcw %llu %d %d %d",
2941 (unsigned long long)sh
->sector
,
2942 rcw
, qread
, test_bit(STRIPE_DELAYED
, &sh
->state
));
2944 /* now if nothing is locked, and if we have enough data,
2945 * we can start a write request
2947 /* since handle_stripe can be called at any time we need to handle the
2948 * case where a compute block operation has been submitted and then a
2949 * subsequent call wants to start a write request. raid_run_ops only
2950 * handles the case where compute block and reconstruct are requested
2951 * simultaneously. If this is not the case then new writes need to be
2952 * held off until the compute completes.
2954 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
2955 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
2956 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
2957 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
2960 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
2961 struct stripe_head_state
*s
, int disks
)
2963 struct r5dev
*dev
= NULL
;
2965 set_bit(STRIPE_HANDLE
, &sh
->state
);
2967 switch (sh
->check_state
) {
2968 case check_state_idle
:
2969 /* start a new check operation if there are no failures */
2970 if (s
->failed
== 0) {
2971 BUG_ON(s
->uptodate
!= disks
);
2972 sh
->check_state
= check_state_run
;
2973 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2974 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2978 dev
= &sh
->dev
[s
->failed_num
[0]];
2980 case check_state_compute_result
:
2981 sh
->check_state
= check_state_idle
;
2983 dev
= &sh
->dev
[sh
->pd_idx
];
2985 /* check that a write has not made the stripe insync */
2986 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2989 /* either failed parity check, or recovery is happening */
2990 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
2991 BUG_ON(s
->uptodate
!= disks
);
2993 set_bit(R5_LOCKED
, &dev
->flags
);
2995 set_bit(R5_Wantwrite
, &dev
->flags
);
2997 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2998 set_bit(STRIPE_INSYNC
, &sh
->state
);
3000 case check_state_run
:
3001 break; /* we will be called again upon completion */
3002 case check_state_check_result
:
3003 sh
->check_state
= check_state_idle
;
3005 /* if a failure occurred during the check operation, leave
3006 * STRIPE_INSYNC not set and let the stripe be handled again
3011 /* handle a successful check operation, if parity is correct
3012 * we are done. Otherwise update the mismatch count and repair
3013 * parity if !MD_RECOVERY_CHECK
3015 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
3016 /* parity is correct (on disc,
3017 * not in buffer any more)
3019 set_bit(STRIPE_INSYNC
, &sh
->state
);
3021 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3022 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3023 /* don't try to repair!! */
3024 set_bit(STRIPE_INSYNC
, &sh
->state
);
3026 sh
->check_state
= check_state_compute_run
;
3027 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3028 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3029 set_bit(R5_Wantcompute
,
3030 &sh
->dev
[sh
->pd_idx
].flags
);
3031 sh
->ops
.target
= sh
->pd_idx
;
3032 sh
->ops
.target2
= -1;
3037 case check_state_compute_run
:
3040 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3041 __func__
, sh
->check_state
,
3042 (unsigned long long) sh
->sector
);
3048 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
3049 struct stripe_head_state
*s
,
3052 int pd_idx
= sh
->pd_idx
;
3053 int qd_idx
= sh
->qd_idx
;
3056 set_bit(STRIPE_HANDLE
, &sh
->state
);
3058 BUG_ON(s
->failed
> 2);
3060 /* Want to check and possibly repair P and Q.
3061 * However there could be one 'failed' device, in which
3062 * case we can only check one of them, possibly using the
3063 * other to generate missing data
3066 switch (sh
->check_state
) {
3067 case check_state_idle
:
3068 /* start a new check operation if there are < 2 failures */
3069 if (s
->failed
== s
->q_failed
) {
3070 /* The only possible failed device holds Q, so it
3071 * makes sense to check P (If anything else were failed,
3072 * we would have used P to recreate it).
3074 sh
->check_state
= check_state_run
;
3076 if (!s
->q_failed
&& s
->failed
< 2) {
3077 /* Q is not failed, and we didn't use it to generate
3078 * anything, so it makes sense to check it
3080 if (sh
->check_state
== check_state_run
)
3081 sh
->check_state
= check_state_run_pq
;
3083 sh
->check_state
= check_state_run_q
;
3086 /* discard potentially stale zero_sum_result */
3087 sh
->ops
.zero_sum_result
= 0;
3089 if (sh
->check_state
== check_state_run
) {
3090 /* async_xor_zero_sum destroys the contents of P */
3091 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
3094 if (sh
->check_state
>= check_state_run
&&
3095 sh
->check_state
<= check_state_run_pq
) {
3096 /* async_syndrome_zero_sum preserves P and Q, so
3097 * no need to mark them !uptodate here
3099 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3103 /* we have 2-disk failure */
3104 BUG_ON(s
->failed
!= 2);
3106 case check_state_compute_result
:
3107 sh
->check_state
= check_state_idle
;
3109 /* check that a write has not made the stripe insync */
3110 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3113 /* now write out any block on a failed drive,
3114 * or P or Q if they were recomputed
3116 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
3117 if (s
->failed
== 2) {
3118 dev
= &sh
->dev
[s
->failed_num
[1]];
3120 set_bit(R5_LOCKED
, &dev
->flags
);
3121 set_bit(R5_Wantwrite
, &dev
->flags
);
3123 if (s
->failed
>= 1) {
3124 dev
= &sh
->dev
[s
->failed_num
[0]];
3126 set_bit(R5_LOCKED
, &dev
->flags
);
3127 set_bit(R5_Wantwrite
, &dev
->flags
);
3129 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3130 dev
= &sh
->dev
[pd_idx
];
3132 set_bit(R5_LOCKED
, &dev
->flags
);
3133 set_bit(R5_Wantwrite
, &dev
->flags
);
3135 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3136 dev
= &sh
->dev
[qd_idx
];
3138 set_bit(R5_LOCKED
, &dev
->flags
);
3139 set_bit(R5_Wantwrite
, &dev
->flags
);
3141 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3143 set_bit(STRIPE_INSYNC
, &sh
->state
);
3145 case check_state_run
:
3146 case check_state_run_q
:
3147 case check_state_run_pq
:
3148 break; /* we will be called again upon completion */
3149 case check_state_check_result
:
3150 sh
->check_state
= check_state_idle
;
3152 /* handle a successful check operation, if parity is correct
3153 * we are done. Otherwise update the mismatch count and repair
3154 * parity if !MD_RECOVERY_CHECK
3156 if (sh
->ops
.zero_sum_result
== 0) {
3157 /* both parities are correct */
3159 set_bit(STRIPE_INSYNC
, &sh
->state
);
3161 /* in contrast to the raid5 case we can validate
3162 * parity, but still have a failure to write
3165 sh
->check_state
= check_state_compute_result
;
3166 /* Returning at this point means that we may go
3167 * off and bring p and/or q uptodate again so
3168 * we make sure to check zero_sum_result again
3169 * to verify if p or q need writeback
3173 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3174 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3175 /* don't try to repair!! */
3176 set_bit(STRIPE_INSYNC
, &sh
->state
);
3178 int *target
= &sh
->ops
.target
;
3180 sh
->ops
.target
= -1;
3181 sh
->ops
.target2
= -1;
3182 sh
->check_state
= check_state_compute_run
;
3183 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3184 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3185 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3186 set_bit(R5_Wantcompute
,
3187 &sh
->dev
[pd_idx
].flags
);
3189 target
= &sh
->ops
.target2
;
3192 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3193 set_bit(R5_Wantcompute
,
3194 &sh
->dev
[qd_idx
].flags
);
3201 case check_state_compute_run
:
3204 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3205 __func__
, sh
->check_state
,
3206 (unsigned long long) sh
->sector
);
3211 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
3215 /* We have read all the blocks in this stripe and now we need to
3216 * copy some of them into a target stripe for expand.
3218 struct dma_async_tx_descriptor
*tx
= NULL
;
3219 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3220 for (i
= 0; i
< sh
->disks
; i
++)
3221 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
3223 struct stripe_head
*sh2
;
3224 struct async_submit_ctl submit
;
3226 sector_t bn
= compute_blocknr(sh
, i
, 1);
3227 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
3229 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
3231 /* so far only the early blocks of this stripe
3232 * have been requested. When later blocks
3233 * get requested, we will try again
3236 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
3237 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
3238 /* must have already done this block */
3239 release_stripe(sh2
);
3243 /* place all the copies on one channel */
3244 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
3245 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
3246 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
3249 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
3250 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
3251 for (j
= 0; j
< conf
->raid_disks
; j
++)
3252 if (j
!= sh2
->pd_idx
&&
3254 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
3256 if (j
== conf
->raid_disks
) {
3257 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
3258 set_bit(STRIPE_HANDLE
, &sh2
->state
);
3260 release_stripe(sh2
);
3263 /* done submitting copies, wait for them to complete */
3264 async_tx_quiesce(&tx
);
3268 * handle_stripe - do things to a stripe.
3270 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3271 * state of various bits to see what needs to be done.
3273 * return some read requests which now have data
3274 * return some write requests which are safely on storage
3275 * schedule a read on some buffers
3276 * schedule a write of some buffers
3277 * return confirmation of parity correctness
3281 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
3283 struct r5conf
*conf
= sh
->raid_conf
;
3284 int disks
= sh
->disks
;
3287 int do_recovery
= 0;
3289 memset(s
, 0, sizeof(*s
));
3291 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3292 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3293 s
->failed_num
[0] = -1;
3294 s
->failed_num
[1] = -1;
3296 /* Now to look around and see what can be done */
3298 for (i
=disks
; i
--; ) {
3299 struct md_rdev
*rdev
;
3306 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3308 dev
->toread
, dev
->towrite
, dev
->written
);
3309 /* maybe we can reply to a read
3311 * new wantfill requests are only permitted while
3312 * ops_complete_biofill is guaranteed to be inactive
3314 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
3315 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
3316 set_bit(R5_Wantfill
, &dev
->flags
);
3318 /* now count some things */
3319 if (test_bit(R5_LOCKED
, &dev
->flags
))
3321 if (test_bit(R5_UPTODATE
, &dev
->flags
))
3323 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
3325 BUG_ON(s
->compute
> 2);
3328 if (test_bit(R5_Wantfill
, &dev
->flags
))
3330 else if (dev
->toread
)
3334 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3339 /* Prefer to use the replacement for reads, but only
3340 * if it is recovered enough and has no bad blocks.
3342 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
3343 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
3344 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
3345 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3346 &first_bad
, &bad_sectors
))
3347 set_bit(R5_ReadRepl
, &dev
->flags
);
3350 set_bit(R5_NeedReplace
, &dev
->flags
);
3351 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3352 clear_bit(R5_ReadRepl
, &dev
->flags
);
3354 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
3357 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3358 &first_bad
, &bad_sectors
);
3359 if (s
->blocked_rdev
== NULL
3360 && (test_bit(Blocked
, &rdev
->flags
)
3363 set_bit(BlockedBadBlocks
,
3365 s
->blocked_rdev
= rdev
;
3366 atomic_inc(&rdev
->nr_pending
);
3369 clear_bit(R5_Insync
, &dev
->flags
);
3373 /* also not in-sync */
3374 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
3375 test_bit(R5_UPTODATE
, &dev
->flags
)) {
3376 /* treat as in-sync, but with a read error
3377 * which we can now try to correct
3379 set_bit(R5_Insync
, &dev
->flags
);
3380 set_bit(R5_ReadError
, &dev
->flags
);
3382 } else if (test_bit(In_sync
, &rdev
->flags
))
3383 set_bit(R5_Insync
, &dev
->flags
);
3384 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
3385 /* in sync if before recovery_offset */
3386 set_bit(R5_Insync
, &dev
->flags
);
3387 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
3388 test_bit(R5_Expanded
, &dev
->flags
))
3389 /* If we've reshaped into here, we assume it is Insync.
3390 * We will shortly update recovery_offset to make
3393 set_bit(R5_Insync
, &dev
->flags
);
3395 if (test_bit(R5_WriteError
, &dev
->flags
)) {
3396 /* This flag does not apply to '.replacement'
3397 * only to .rdev, so make sure to check that*/
3398 struct md_rdev
*rdev2
= rcu_dereference(
3399 conf
->disks
[i
].rdev
);
3401 clear_bit(R5_Insync
, &dev
->flags
);
3402 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3403 s
->handle_bad_blocks
= 1;
3404 atomic_inc(&rdev2
->nr_pending
);
3406 clear_bit(R5_WriteError
, &dev
->flags
);
3408 if (test_bit(R5_MadeGood
, &dev
->flags
)) {
3409 /* This flag does not apply to '.replacement'
3410 * only to .rdev, so make sure to check that*/
3411 struct md_rdev
*rdev2
= rcu_dereference(
3412 conf
->disks
[i
].rdev
);
3413 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3414 s
->handle_bad_blocks
= 1;
3415 atomic_inc(&rdev2
->nr_pending
);
3417 clear_bit(R5_MadeGood
, &dev
->flags
);
3419 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
3420 struct md_rdev
*rdev2
= rcu_dereference(
3421 conf
->disks
[i
].replacement
);
3422 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3423 s
->handle_bad_blocks
= 1;
3424 atomic_inc(&rdev2
->nr_pending
);
3426 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
3428 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3429 /* The ReadError flag will just be confusing now */
3430 clear_bit(R5_ReadError
, &dev
->flags
);
3431 clear_bit(R5_ReWrite
, &dev
->flags
);
3433 if (test_bit(R5_ReadError
, &dev
->flags
))
3434 clear_bit(R5_Insync
, &dev
->flags
);
3435 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3437 s
->failed_num
[s
->failed
] = i
;
3439 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
3443 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
3444 /* If there is a failed device being replaced,
3445 * we must be recovering.
3446 * else if we are after recovery_cp, we must be syncing
3447 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3448 * else we can only be replacing
3449 * sync and recovery both need to read all devices, and so
3450 * use the same flag.
3453 sh
->sector
>= conf
->mddev
->recovery_cp
||
3454 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
3462 static void handle_stripe(struct stripe_head
*sh
)
3464 struct stripe_head_state s
;
3465 struct r5conf
*conf
= sh
->raid_conf
;
3468 int disks
= sh
->disks
;
3469 struct r5dev
*pdev
, *qdev
;
3471 clear_bit(STRIPE_HANDLE
, &sh
->state
);
3472 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
3473 /* already being handled, ensure it gets handled
3474 * again when current action finishes */
3475 set_bit(STRIPE_HANDLE
, &sh
->state
);
3479 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
3480 spin_lock(&sh
->stripe_lock
);
3481 /* Cannot process 'sync' concurrently with 'discard' */
3482 if (!test_bit(STRIPE_DISCARD
, &sh
->state
) &&
3483 test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
3484 set_bit(STRIPE_SYNCING
, &sh
->state
);
3485 clear_bit(STRIPE_INSYNC
, &sh
->state
);
3486 clear_bit(STRIPE_REPLACED
, &sh
->state
);
3488 spin_unlock(&sh
->stripe_lock
);
3490 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3492 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3493 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3494 (unsigned long long)sh
->sector
, sh
->state
,
3495 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
3496 sh
->check_state
, sh
->reconstruct_state
);
3498 analyse_stripe(sh
, &s
);
3500 if (s
.handle_bad_blocks
) {
3501 set_bit(STRIPE_HANDLE
, &sh
->state
);
3505 if (unlikely(s
.blocked_rdev
)) {
3506 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
3507 s
.replacing
|| s
.to_write
|| s
.written
) {
3508 set_bit(STRIPE_HANDLE
, &sh
->state
);
3511 /* There is nothing for the blocked_rdev to block */
3512 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
3513 s
.blocked_rdev
= NULL
;
3516 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
3517 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
3518 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
3521 pr_debug("locked=%d uptodate=%d to_read=%d"
3522 " to_write=%d failed=%d failed_num=%d,%d\n",
3523 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
3524 s
.failed_num
[0], s
.failed_num
[1]);
3525 /* check if the array has lost more than max_degraded devices and,
3526 * if so, some requests might need to be failed.
3528 if (s
.failed
> conf
->max_degraded
) {
3529 sh
->check_state
= 0;
3530 sh
->reconstruct_state
= 0;
3531 if (s
.to_read
+s
.to_write
+s
.written
)
3532 handle_failed_stripe(conf
, sh
, &s
, disks
, &s
.return_bi
);
3533 if (s
.syncing
+ s
.replacing
)
3534 handle_failed_sync(conf
, sh
, &s
);
3537 /* Now we check to see if any write operations have recently
3541 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
3543 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
3544 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
3545 sh
->reconstruct_state
= reconstruct_state_idle
;
3547 /* All the 'written' buffers and the parity block are ready to
3548 * be written back to disk
3550 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
) &&
3551 !test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
));
3552 BUG_ON(sh
->qd_idx
>= 0 &&
3553 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
) &&
3554 !test_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
));
3555 for (i
= disks
; i
--; ) {
3556 struct r5dev
*dev
= &sh
->dev
[i
];
3557 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
3558 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3560 pr_debug("Writing block %d\n", i
);
3561 set_bit(R5_Wantwrite
, &dev
->flags
);
3564 if (!test_bit(R5_Insync
, &dev
->flags
) ||
3565 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3567 set_bit(STRIPE_INSYNC
, &sh
->state
);
3570 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3571 s
.dec_preread_active
= 1;
3575 * might be able to return some write requests if the parity blocks
3576 * are safe, or on a failed drive
3578 pdev
= &sh
->dev
[sh
->pd_idx
];
3579 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
3580 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
3581 qdev
= &sh
->dev
[sh
->qd_idx
];
3582 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
3583 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
3587 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
3588 && !test_bit(R5_LOCKED
, &pdev
->flags
)
3589 && (test_bit(R5_UPTODATE
, &pdev
->flags
) ||
3590 test_bit(R5_Discard
, &pdev
->flags
))))) &&
3591 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
3592 && !test_bit(R5_LOCKED
, &qdev
->flags
)
3593 && (test_bit(R5_UPTODATE
, &qdev
->flags
) ||
3594 test_bit(R5_Discard
, &qdev
->flags
))))))
3595 handle_stripe_clean_event(conf
, sh
, disks
, &s
.return_bi
);
3597 /* Now we might consider reading some blocks, either to check/generate
3598 * parity, or to satisfy requests
3599 * or to load a block that is being partially written.
3601 if (s
.to_read
|| s
.non_overwrite
3602 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
3603 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
3606 handle_stripe_fill(sh
, &s
, disks
);
3608 /* Now to consider new write requests and what else, if anything
3609 * should be read. We do not handle new writes when:
3610 * 1/ A 'write' operation (copy+xor) is already in flight.
3611 * 2/ A 'check' operation is in flight, as it may clobber the parity
3614 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
3615 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
3617 /* maybe we need to check and possibly fix the parity for this stripe
3618 * Any reads will already have been scheduled, so we just see if enough
3619 * data is available. The parity check is held off while parity
3620 * dependent operations are in flight.
3622 if (sh
->check_state
||
3623 (s
.syncing
&& s
.locked
== 0 &&
3624 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
3625 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
3626 if (conf
->level
== 6)
3627 handle_parity_checks6(conf
, sh
, &s
, disks
);
3629 handle_parity_checks5(conf
, sh
, &s
, disks
);
3632 if ((s
.replacing
|| s
.syncing
) && s
.locked
== 0
3633 && !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)
3634 && !test_bit(STRIPE_REPLACED
, &sh
->state
)) {
3635 /* Write out to replacement devices where possible */
3636 for (i
= 0; i
< conf
->raid_disks
; i
++)
3637 if (test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
3638 WARN_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
3639 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
3640 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3644 set_bit(STRIPE_INSYNC
, &sh
->state
);
3645 set_bit(STRIPE_REPLACED
, &sh
->state
);
3647 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
3648 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
3649 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3650 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3651 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3652 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
3653 wake_up(&conf
->wait_for_overlap
);
3656 /* If the failed drives are just a ReadError, then we might need
3657 * to progress the repair/check process
3659 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
3660 for (i
= 0; i
< s
.failed
; i
++) {
3661 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
3662 if (test_bit(R5_ReadError
, &dev
->flags
)
3663 && !test_bit(R5_LOCKED
, &dev
->flags
)
3664 && test_bit(R5_UPTODATE
, &dev
->flags
)
3666 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3667 set_bit(R5_Wantwrite
, &dev
->flags
);
3668 set_bit(R5_ReWrite
, &dev
->flags
);
3669 set_bit(R5_LOCKED
, &dev
->flags
);
3672 /* let's read it back */
3673 set_bit(R5_Wantread
, &dev
->flags
);
3674 set_bit(R5_LOCKED
, &dev
->flags
);
3681 /* Finish reconstruct operations initiated by the expansion process */
3682 if (sh
->reconstruct_state
== reconstruct_state_result
) {
3683 struct stripe_head
*sh_src
3684 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
3685 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
3686 /* sh cannot be written until sh_src has been read.
3687 * so arrange for sh to be delayed a little
3689 set_bit(STRIPE_DELAYED
, &sh
->state
);
3690 set_bit(STRIPE_HANDLE
, &sh
->state
);
3691 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
3693 atomic_inc(&conf
->preread_active_stripes
);
3694 release_stripe(sh_src
);
3698 release_stripe(sh_src
);
3700 sh
->reconstruct_state
= reconstruct_state_idle
;
3701 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3702 for (i
= conf
->raid_disks
; i
--; ) {
3703 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3704 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3709 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
3710 !sh
->reconstruct_state
) {
3711 /* Need to write out all blocks after computing parity */
3712 sh
->disks
= conf
->raid_disks
;
3713 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
3714 schedule_reconstruction(sh
, &s
, 1, 1);
3715 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
3716 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3717 atomic_dec(&conf
->reshape_stripes
);
3718 wake_up(&conf
->wait_for_overlap
);
3719 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3722 if (s
.expanding
&& s
.locked
== 0 &&
3723 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
3724 handle_stripe_expansion(conf
, sh
);
3727 /* wait for this device to become unblocked */
3728 if (unlikely(s
.blocked_rdev
)) {
3729 if (conf
->mddev
->external
)
3730 md_wait_for_blocked_rdev(s
.blocked_rdev
,
3733 /* Internal metadata will immediately
3734 * be written by raid5d, so we don't
3735 * need to wait here.
3737 rdev_dec_pending(s
.blocked_rdev
,
3741 if (s
.handle_bad_blocks
)
3742 for (i
= disks
; i
--; ) {
3743 struct md_rdev
*rdev
;
3744 struct r5dev
*dev
= &sh
->dev
[i
];
3745 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
3746 /* We own a safe reference to the rdev */
3747 rdev
= conf
->disks
[i
].rdev
;
3748 if (!rdev_set_badblocks(rdev
, sh
->sector
,
3750 md_error(conf
->mddev
, rdev
);
3751 rdev_dec_pending(rdev
, conf
->mddev
);
3753 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
3754 rdev
= conf
->disks
[i
].rdev
;
3755 rdev_clear_badblocks(rdev
, sh
->sector
,
3757 rdev_dec_pending(rdev
, conf
->mddev
);
3759 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
3760 rdev
= conf
->disks
[i
].replacement
;
3762 /* rdev have been moved down */
3763 rdev
= conf
->disks
[i
].rdev
;
3764 rdev_clear_badblocks(rdev
, sh
->sector
,
3766 rdev_dec_pending(rdev
, conf
->mddev
);
3771 raid_run_ops(sh
, s
.ops_request
);
3775 if (s
.dec_preread_active
) {
3776 /* We delay this until after ops_run_io so that if make_request
3777 * is waiting on a flush, it won't continue until the writes
3778 * have actually been submitted.
3780 atomic_dec(&conf
->preread_active_stripes
);
3781 if (atomic_read(&conf
->preread_active_stripes
) <
3783 md_wakeup_thread(conf
->mddev
->thread
);
3786 return_io(s
.return_bi
);
3788 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
3791 static void raid5_activate_delayed(struct r5conf
*conf
)
3793 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
3794 while (!list_empty(&conf
->delayed_list
)) {
3795 struct list_head
*l
= conf
->delayed_list
.next
;
3796 struct stripe_head
*sh
;
3797 sh
= list_entry(l
, struct stripe_head
, lru
);
3799 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3800 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3801 atomic_inc(&conf
->preread_active_stripes
);
3802 list_add_tail(&sh
->lru
, &conf
->hold_list
);
3807 static void activate_bit_delay(struct r5conf
*conf
)
3809 /* device_lock is held */
3810 struct list_head head
;
3811 list_add(&head
, &conf
->bitmap_list
);
3812 list_del_init(&conf
->bitmap_list
);
3813 while (!list_empty(&head
)) {
3814 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
3815 list_del_init(&sh
->lru
);
3816 atomic_inc(&sh
->count
);
3817 __release_stripe(conf
, sh
);
3821 int md_raid5_congested(struct mddev
*mddev
, int bits
)
3823 struct r5conf
*conf
= mddev
->private;
3825 /* No difference between reads and writes. Just check
3826 * how busy the stripe_cache is
3829 if (conf
->inactive_blocked
)
3833 if (list_empty_careful(&conf
->inactive_list
))
3838 EXPORT_SYMBOL_GPL(md_raid5_congested
);
3840 static int raid5_congested(void *data
, int bits
)
3842 struct mddev
*mddev
= data
;
3844 return mddev_congested(mddev
, bits
) ||
3845 md_raid5_congested(mddev
, bits
);
3848 /* We want read requests to align with chunks where possible,
3849 * but write requests don't need to.
3851 static int raid5_mergeable_bvec(struct request_queue
*q
,
3852 struct bvec_merge_data
*bvm
,
3853 struct bio_vec
*biovec
)
3855 struct mddev
*mddev
= q
->queuedata
;
3856 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
3858 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3859 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
3861 if ((bvm
->bi_rw
& 1) == WRITE
)
3862 return biovec
->bv_len
; /* always allow writes to be mergeable */
3864 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3865 chunk_sectors
= mddev
->new_chunk_sectors
;
3866 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
3867 if (max
< 0) max
= 0;
3868 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
3869 return biovec
->bv_len
;
3875 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
3877 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
3878 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3879 unsigned int bio_sectors
= bio_sectors(bio
);
3881 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3882 chunk_sectors
= mddev
->new_chunk_sectors
;
3883 return chunk_sectors
>=
3884 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
3888 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3889 * later sampled by raid5d.
3891 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
3893 unsigned long flags
;
3895 spin_lock_irqsave(&conf
->device_lock
, flags
);
3897 bi
->bi_next
= conf
->retry_read_aligned_list
;
3898 conf
->retry_read_aligned_list
= bi
;
3900 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3901 md_wakeup_thread(conf
->mddev
->thread
);
3905 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
)
3909 bi
= conf
->retry_read_aligned
;
3911 conf
->retry_read_aligned
= NULL
;
3914 bi
= conf
->retry_read_aligned_list
;
3916 conf
->retry_read_aligned_list
= bi
->bi_next
;
3919 * this sets the active strip count to 1 and the processed
3920 * strip count to zero (upper 8 bits)
3922 raid5_set_bi_stripes(bi
, 1); /* biased count of active stripes */
3930 * The "raid5_align_endio" should check if the read succeeded and if it
3931 * did, call bio_endio on the original bio (having bio_put the new bio
3933 * If the read failed..
3935 static void raid5_align_endio(struct bio
*bi
, int error
)
3937 struct bio
* raid_bi
= bi
->bi_private
;
3938 struct mddev
*mddev
;
3939 struct r5conf
*conf
;
3940 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3941 struct md_rdev
*rdev
;
3945 rdev
= (void*)raid_bi
->bi_next
;
3946 raid_bi
->bi_next
= NULL
;
3947 mddev
= rdev
->mddev
;
3948 conf
= mddev
->private;
3950 rdev_dec_pending(rdev
, conf
->mddev
);
3952 if (!error
&& uptodate
) {
3953 trace_block_bio_complete(bdev_get_queue(raid_bi
->bi_bdev
),
3955 bio_endio(raid_bi
, 0);
3956 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
3957 wake_up(&conf
->wait_for_stripe
);
3962 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3964 add_bio_to_retry(raid_bi
, conf
);
3967 static int bio_fits_rdev(struct bio
*bi
)
3969 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
3971 if (bio_sectors(bi
) > queue_max_sectors(q
))
3973 blk_recount_segments(q
, bi
);
3974 if (bi
->bi_phys_segments
> queue_max_segments(q
))
3977 if (q
->merge_bvec_fn
)
3978 /* it's too hard to apply the merge_bvec_fn at this stage,
3987 static int chunk_aligned_read(struct mddev
*mddev
, struct bio
* raid_bio
)
3989 struct r5conf
*conf
= mddev
->private;
3991 struct bio
* align_bi
;
3992 struct md_rdev
*rdev
;
3993 sector_t end_sector
;
3995 if (!in_chunk_boundary(mddev
, raid_bio
)) {
3996 pr_debug("chunk_aligned_read : non aligned\n");
4000 * use bio_clone_mddev to make a copy of the bio
4002 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
4006 * set bi_end_io to a new function, and set bi_private to the
4009 align_bi
->bi_end_io
= raid5_align_endio
;
4010 align_bi
->bi_private
= raid_bio
;
4014 align_bi
->bi_sector
= raid5_compute_sector(conf
, raid_bio
->bi_sector
,
4018 end_sector
= bio_end_sector(align_bi
);
4020 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
4021 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
4022 rdev
->recovery_offset
< end_sector
) {
4023 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
4025 (test_bit(Faulty
, &rdev
->flags
) ||
4026 !(test_bit(In_sync
, &rdev
->flags
) ||
4027 rdev
->recovery_offset
>= end_sector
)))
4034 atomic_inc(&rdev
->nr_pending
);
4036 raid_bio
->bi_next
= (void*)rdev
;
4037 align_bi
->bi_bdev
= rdev
->bdev
;
4038 align_bi
->bi_flags
&= ~(1 << BIO_SEG_VALID
);
4040 if (!bio_fits_rdev(align_bi
) ||
4041 is_badblock(rdev
, align_bi
->bi_sector
, bio_sectors(align_bi
),
4042 &first_bad
, &bad_sectors
)) {
4043 /* too big in some way, or has a known bad block */
4045 rdev_dec_pending(rdev
, mddev
);
4049 /* No reshape active, so we can trust rdev->data_offset */
4050 align_bi
->bi_sector
+= rdev
->data_offset
;
4052 spin_lock_irq(&conf
->device_lock
);
4053 wait_event_lock_irq(conf
->wait_for_stripe
,
4056 atomic_inc(&conf
->active_aligned_reads
);
4057 spin_unlock_irq(&conf
->device_lock
);
4060 trace_block_bio_remap(bdev_get_queue(align_bi
->bi_bdev
),
4061 align_bi
, disk_devt(mddev
->gendisk
),
4062 raid_bio
->bi_sector
);
4063 generic_make_request(align_bi
);
4072 /* __get_priority_stripe - get the next stripe to process
4074 * Full stripe writes are allowed to pass preread active stripes up until
4075 * the bypass_threshold is exceeded. In general the bypass_count
4076 * increments when the handle_list is handled before the hold_list; however, it
4077 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4078 * stripe with in flight i/o. The bypass_count will be reset when the
4079 * head of the hold_list has changed, i.e. the head was promoted to the
4082 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
)
4084 struct stripe_head
*sh
;
4086 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4088 list_empty(&conf
->handle_list
) ? "empty" : "busy",
4089 list_empty(&conf
->hold_list
) ? "empty" : "busy",
4090 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
4092 if (!list_empty(&conf
->handle_list
)) {
4093 sh
= list_entry(conf
->handle_list
.next
, typeof(*sh
), lru
);
4095 if (list_empty(&conf
->hold_list
))
4096 conf
->bypass_count
= 0;
4097 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
4098 if (conf
->hold_list
.next
== conf
->last_hold
)
4099 conf
->bypass_count
++;
4101 conf
->last_hold
= conf
->hold_list
.next
;
4102 conf
->bypass_count
-= conf
->bypass_threshold
;
4103 if (conf
->bypass_count
< 0)
4104 conf
->bypass_count
= 0;
4107 } else if (!list_empty(&conf
->hold_list
) &&
4108 ((conf
->bypass_threshold
&&
4109 conf
->bypass_count
> conf
->bypass_threshold
) ||
4110 atomic_read(&conf
->pending_full_writes
) == 0)) {
4111 sh
= list_entry(conf
->hold_list
.next
,
4113 conf
->bypass_count
-= conf
->bypass_threshold
;
4114 if (conf
->bypass_count
< 0)
4115 conf
->bypass_count
= 0;
4119 list_del_init(&sh
->lru
);
4120 atomic_inc(&sh
->count
);
4121 BUG_ON(atomic_read(&sh
->count
) != 1);
4125 struct raid5_plug_cb
{
4126 struct blk_plug_cb cb
;
4127 struct list_head list
;
4130 static void raid5_unplug(struct blk_plug_cb
*blk_cb
, bool from_schedule
)
4132 struct raid5_plug_cb
*cb
= container_of(
4133 blk_cb
, struct raid5_plug_cb
, cb
);
4134 struct stripe_head
*sh
;
4135 struct mddev
*mddev
= cb
->cb
.data
;
4136 struct r5conf
*conf
= mddev
->private;
4139 if (cb
->list
.next
&& !list_empty(&cb
->list
)) {
4140 spin_lock_irq(&conf
->device_lock
);
4141 while (!list_empty(&cb
->list
)) {
4142 sh
= list_first_entry(&cb
->list
, struct stripe_head
, lru
);
4143 list_del_init(&sh
->lru
);
4145 * avoid race release_stripe_plug() sees
4146 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4147 * is still in our list
4149 smp_mb__before_clear_bit();
4150 clear_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
);
4151 __release_stripe(conf
, sh
);
4154 spin_unlock_irq(&conf
->device_lock
);
4157 trace_block_unplug(mddev
->queue
, cnt
, !from_schedule
);
4161 static void release_stripe_plug(struct mddev
*mddev
,
4162 struct stripe_head
*sh
)
4164 struct blk_plug_cb
*blk_cb
= blk_check_plugged(
4165 raid5_unplug
, mddev
,
4166 sizeof(struct raid5_plug_cb
));
4167 struct raid5_plug_cb
*cb
;
4174 cb
= container_of(blk_cb
, struct raid5_plug_cb
, cb
);
4176 if (cb
->list
.next
== NULL
)
4177 INIT_LIST_HEAD(&cb
->list
);
4179 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
))
4180 list_add_tail(&sh
->lru
, &cb
->list
);
4185 static void make_discard_request(struct mddev
*mddev
, struct bio
*bi
)
4187 struct r5conf
*conf
= mddev
->private;
4188 sector_t logical_sector
, last_sector
;
4189 struct stripe_head
*sh
;
4193 if (mddev
->reshape_position
!= MaxSector
)
4194 /* Skip discard while reshape is happening */
4197 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4198 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
4201 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
4203 stripe_sectors
= conf
->chunk_sectors
*
4204 (conf
->raid_disks
- conf
->max_degraded
);
4205 logical_sector
= DIV_ROUND_UP_SECTOR_T(logical_sector
,
4207 sector_div(last_sector
, stripe_sectors
);
4209 logical_sector
*= conf
->chunk_sectors
;
4210 last_sector
*= conf
->chunk_sectors
;
4212 for (; logical_sector
< last_sector
;
4213 logical_sector
+= STRIPE_SECTORS
) {
4217 sh
= get_active_stripe(conf
, logical_sector
, 0, 0, 0);
4218 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
4219 TASK_UNINTERRUPTIBLE
);
4220 set_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
4221 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
4226 clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
4227 spin_lock_irq(&sh
->stripe_lock
);
4228 for (d
= 0; d
< conf
->raid_disks
; d
++) {
4229 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
4231 if (sh
->dev
[d
].towrite
|| sh
->dev
[d
].toread
) {
4232 set_bit(R5_Overlap
, &sh
->dev
[d
].flags
);
4233 spin_unlock_irq(&sh
->stripe_lock
);
4239 set_bit(STRIPE_DISCARD
, &sh
->state
);
4240 finish_wait(&conf
->wait_for_overlap
, &w
);
4241 for (d
= 0; d
< conf
->raid_disks
; d
++) {
4242 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
4244 sh
->dev
[d
].towrite
= bi
;
4245 set_bit(R5_OVERWRITE
, &sh
->dev
[d
].flags
);
4246 raid5_inc_bi_active_stripes(bi
);
4248 spin_unlock_irq(&sh
->stripe_lock
);
4249 if (conf
->mddev
->bitmap
) {
4251 d
< conf
->raid_disks
- conf
->max_degraded
;
4253 bitmap_startwrite(mddev
->bitmap
,
4257 sh
->bm_seq
= conf
->seq_flush
+ 1;
4258 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
4261 set_bit(STRIPE_HANDLE
, &sh
->state
);
4262 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4263 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4264 atomic_inc(&conf
->preread_active_stripes
);
4265 release_stripe_plug(mddev
, sh
);
4268 remaining
= raid5_dec_bi_active_stripes(bi
);
4269 if (remaining
== 0) {
4270 md_write_end(mddev
);
4275 static void make_request(struct mddev
*mddev
, struct bio
* bi
)
4277 struct r5conf
*conf
= mddev
->private;
4279 sector_t new_sector
;
4280 sector_t logical_sector
, last_sector
;
4281 struct stripe_head
*sh
;
4282 const int rw
= bio_data_dir(bi
);
4285 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
4286 md_flush_request(mddev
, bi
);
4290 md_write_start(mddev
, bi
);
4293 mddev
->reshape_position
== MaxSector
&&
4294 chunk_aligned_read(mddev
,bi
))
4297 if (unlikely(bi
->bi_rw
& REQ_DISCARD
)) {
4298 make_discard_request(mddev
, bi
);
4302 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4303 last_sector
= bio_end_sector(bi
);
4305 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
4307 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
4313 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
4314 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
4315 /* spinlock is needed as reshape_progress may be
4316 * 64bit on a 32bit platform, and so it might be
4317 * possible to see a half-updated value
4318 * Of course reshape_progress could change after
4319 * the lock is dropped, so once we get a reference
4320 * to the stripe that we think it is, we will have
4323 spin_lock_irq(&conf
->device_lock
);
4324 if (mddev
->reshape_backwards
4325 ? logical_sector
< conf
->reshape_progress
4326 : logical_sector
>= conf
->reshape_progress
) {
4329 if (mddev
->reshape_backwards
4330 ? logical_sector
< conf
->reshape_safe
4331 : logical_sector
>= conf
->reshape_safe
) {
4332 spin_unlock_irq(&conf
->device_lock
);
4337 spin_unlock_irq(&conf
->device_lock
);
4340 new_sector
= raid5_compute_sector(conf
, logical_sector
,
4343 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4344 (unsigned long long)new_sector
,
4345 (unsigned long long)logical_sector
);
4347 sh
= get_active_stripe(conf
, new_sector
, previous
,
4348 (bi
->bi_rw
&RWA_MASK
), 0);
4350 if (unlikely(previous
)) {
4351 /* expansion might have moved on while waiting for a
4352 * stripe, so we must do the range check again.
4353 * Expansion could still move past after this
4354 * test, but as we are holding a reference to
4355 * 'sh', we know that if that happens,
4356 * STRIPE_EXPANDING will get set and the expansion
4357 * won't proceed until we finish with the stripe.
4360 spin_lock_irq(&conf
->device_lock
);
4361 if (mddev
->reshape_backwards
4362 ? logical_sector
>= conf
->reshape_progress
4363 : logical_sector
< conf
->reshape_progress
)
4364 /* mismatch, need to try again */
4366 spin_unlock_irq(&conf
->device_lock
);
4375 logical_sector
>= mddev
->suspend_lo
&&
4376 logical_sector
< mddev
->suspend_hi
) {
4378 /* As the suspend_* range is controlled by
4379 * userspace, we want an interruptible
4382 flush_signals(current
);
4383 prepare_to_wait(&conf
->wait_for_overlap
,
4384 &w
, TASK_INTERRUPTIBLE
);
4385 if (logical_sector
>= mddev
->suspend_lo
&&
4386 logical_sector
< mddev
->suspend_hi
)
4391 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
4392 !add_stripe_bio(sh
, bi
, dd_idx
, rw
)) {
4393 /* Stripe is busy expanding or
4394 * add failed due to overlap. Flush everything
4397 md_wakeup_thread(mddev
->thread
);
4402 finish_wait(&conf
->wait_for_overlap
, &w
);
4403 set_bit(STRIPE_HANDLE
, &sh
->state
);
4404 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4405 if ((bi
->bi_rw
& REQ_SYNC
) &&
4406 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4407 atomic_inc(&conf
->preread_active_stripes
);
4408 release_stripe_plug(mddev
, sh
);
4410 /* cannot get stripe for read-ahead, just give-up */
4411 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
4412 finish_wait(&conf
->wait_for_overlap
, &w
);
4417 remaining
= raid5_dec_bi_active_stripes(bi
);
4418 if (remaining
== 0) {
4421 md_write_end(mddev
);
4423 trace_block_bio_complete(bdev_get_queue(bi
->bi_bdev
),
4429 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
4431 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
4433 /* reshaping is quite different to recovery/resync so it is
4434 * handled quite separately ... here.
4436 * On each call to sync_request, we gather one chunk worth of
4437 * destination stripes and flag them as expanding.
4438 * Then we find all the source stripes and request reads.
4439 * As the reads complete, handle_stripe will copy the data
4440 * into the destination stripe and release that stripe.
4442 struct r5conf
*conf
= mddev
->private;
4443 struct stripe_head
*sh
;
4444 sector_t first_sector
, last_sector
;
4445 int raid_disks
= conf
->previous_raid_disks
;
4446 int data_disks
= raid_disks
- conf
->max_degraded
;
4447 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
4450 sector_t writepos
, readpos
, safepos
;
4451 sector_t stripe_addr
;
4452 int reshape_sectors
;
4453 struct list_head stripes
;
4455 if (sector_nr
== 0) {
4456 /* If restarting in the middle, skip the initial sectors */
4457 if (mddev
->reshape_backwards
&&
4458 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
4459 sector_nr
= raid5_size(mddev
, 0, 0)
4460 - conf
->reshape_progress
;
4461 } else if (!mddev
->reshape_backwards
&&
4462 conf
->reshape_progress
> 0)
4463 sector_nr
= conf
->reshape_progress
;
4464 sector_div(sector_nr
, new_data_disks
);
4466 mddev
->curr_resync_completed
= sector_nr
;
4467 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4473 /* We need to process a full chunk at a time.
4474 * If old and new chunk sizes differ, we need to process the
4477 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
)
4478 reshape_sectors
= mddev
->new_chunk_sectors
;
4480 reshape_sectors
= mddev
->chunk_sectors
;
4482 /* We update the metadata at least every 10 seconds, or when
4483 * the data about to be copied would over-write the source of
4484 * the data at the front of the range. i.e. one new_stripe
4485 * along from reshape_progress new_maps to after where
4486 * reshape_safe old_maps to
4488 writepos
= conf
->reshape_progress
;
4489 sector_div(writepos
, new_data_disks
);
4490 readpos
= conf
->reshape_progress
;
4491 sector_div(readpos
, data_disks
);
4492 safepos
= conf
->reshape_safe
;
4493 sector_div(safepos
, data_disks
);
4494 if (mddev
->reshape_backwards
) {
4495 writepos
-= min_t(sector_t
, reshape_sectors
, writepos
);
4496 readpos
+= reshape_sectors
;
4497 safepos
+= reshape_sectors
;
4499 writepos
+= reshape_sectors
;
4500 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
4501 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
4504 /* Having calculated the 'writepos' possibly use it
4505 * to set 'stripe_addr' which is where we will write to.
4507 if (mddev
->reshape_backwards
) {
4508 BUG_ON(conf
->reshape_progress
== 0);
4509 stripe_addr
= writepos
;
4510 BUG_ON((mddev
->dev_sectors
&
4511 ~((sector_t
)reshape_sectors
- 1))
4512 - reshape_sectors
- stripe_addr
4515 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
4516 stripe_addr
= sector_nr
;
4519 /* 'writepos' is the most advanced device address we might write.
4520 * 'readpos' is the least advanced device address we might read.
4521 * 'safepos' is the least address recorded in the metadata as having
4523 * If there is a min_offset_diff, these are adjusted either by
4524 * increasing the safepos/readpos if diff is negative, or
4525 * increasing writepos if diff is positive.
4526 * If 'readpos' is then behind 'writepos', there is no way that we can
4527 * ensure safety in the face of a crash - that must be done by userspace
4528 * making a backup of the data. So in that case there is no particular
4529 * rush to update metadata.
4530 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4531 * update the metadata to advance 'safepos' to match 'readpos' so that
4532 * we can be safe in the event of a crash.
4533 * So we insist on updating metadata if safepos is behind writepos and
4534 * readpos is beyond writepos.
4535 * In any case, update the metadata every 10 seconds.
4536 * Maybe that number should be configurable, but I'm not sure it is
4537 * worth it.... maybe it could be a multiple of safemode_delay???
4539 if (conf
->min_offset_diff
< 0) {
4540 safepos
+= -conf
->min_offset_diff
;
4541 readpos
+= -conf
->min_offset_diff
;
4543 writepos
+= conf
->min_offset_diff
;
4545 if ((mddev
->reshape_backwards
4546 ? (safepos
> writepos
&& readpos
< writepos
)
4547 : (safepos
< writepos
&& readpos
> writepos
)) ||
4548 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4549 /* Cannot proceed until we've updated the superblock... */
4550 wait_event(conf
->wait_for_overlap
,
4551 atomic_read(&conf
->reshape_stripes
)==0);
4552 mddev
->reshape_position
= conf
->reshape_progress
;
4553 mddev
->curr_resync_completed
= sector_nr
;
4554 conf
->reshape_checkpoint
= jiffies
;
4555 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4556 md_wakeup_thread(mddev
->thread
);
4557 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4558 kthread_should_stop());
4559 spin_lock_irq(&conf
->device_lock
);
4560 conf
->reshape_safe
= mddev
->reshape_position
;
4561 spin_unlock_irq(&conf
->device_lock
);
4562 wake_up(&conf
->wait_for_overlap
);
4563 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4566 INIT_LIST_HEAD(&stripes
);
4567 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
4569 int skipped_disk
= 0;
4570 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
4571 set_bit(STRIPE_EXPANDING
, &sh
->state
);
4572 atomic_inc(&conf
->reshape_stripes
);
4573 /* If any of this stripe is beyond the end of the old
4574 * array, then we need to zero those blocks
4576 for (j
=sh
->disks
; j
--;) {
4578 if (j
== sh
->pd_idx
)
4580 if (conf
->level
== 6 &&
4583 s
= compute_blocknr(sh
, j
, 0);
4584 if (s
< raid5_size(mddev
, 0, 0)) {
4588 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
4589 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
4590 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
4592 if (!skipped_disk
) {
4593 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4594 set_bit(STRIPE_HANDLE
, &sh
->state
);
4596 list_add(&sh
->lru
, &stripes
);
4598 spin_lock_irq(&conf
->device_lock
);
4599 if (mddev
->reshape_backwards
)
4600 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
4602 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
4603 spin_unlock_irq(&conf
->device_lock
);
4604 /* Ok, those stripe are ready. We can start scheduling
4605 * reads on the source stripes.
4606 * The source stripes are determined by mapping the first and last
4607 * block on the destination stripes.
4610 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
4613 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
4614 * new_data_disks
- 1),
4616 if (last_sector
>= mddev
->dev_sectors
)
4617 last_sector
= mddev
->dev_sectors
- 1;
4618 while (first_sector
<= last_sector
) {
4619 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
4620 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
4621 set_bit(STRIPE_HANDLE
, &sh
->state
);
4623 first_sector
+= STRIPE_SECTORS
;
4625 /* Now that the sources are clearly marked, we can release
4626 * the destination stripes
4628 while (!list_empty(&stripes
)) {
4629 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
4630 list_del_init(&sh
->lru
);
4633 /* If this takes us to the resync_max point where we have to pause,
4634 * then we need to write out the superblock.
4636 sector_nr
+= reshape_sectors
;
4637 if ((sector_nr
- mddev
->curr_resync_completed
) * 2
4638 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
4639 /* Cannot proceed until we've updated the superblock... */
4640 wait_event(conf
->wait_for_overlap
,
4641 atomic_read(&conf
->reshape_stripes
) == 0);
4642 mddev
->reshape_position
= conf
->reshape_progress
;
4643 mddev
->curr_resync_completed
= sector_nr
;
4644 conf
->reshape_checkpoint
= jiffies
;
4645 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4646 md_wakeup_thread(mddev
->thread
);
4647 wait_event(mddev
->sb_wait
,
4648 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
4649 || kthread_should_stop());
4650 spin_lock_irq(&conf
->device_lock
);
4651 conf
->reshape_safe
= mddev
->reshape_position
;
4652 spin_unlock_irq(&conf
->device_lock
);
4653 wake_up(&conf
->wait_for_overlap
);
4654 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4656 return reshape_sectors
;
4659 /* FIXME go_faster isn't used */
4660 static inline sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
4662 struct r5conf
*conf
= mddev
->private;
4663 struct stripe_head
*sh
;
4664 sector_t max_sector
= mddev
->dev_sectors
;
4665 sector_t sync_blocks
;
4666 int still_degraded
= 0;
4669 if (sector_nr
>= max_sector
) {
4670 /* just being told to finish up .. nothing much to do */
4672 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
4677 if (mddev
->curr_resync
< max_sector
) /* aborted */
4678 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
4680 else /* completed sync */
4682 bitmap_close_sync(mddev
->bitmap
);
4687 /* Allow raid5_quiesce to complete */
4688 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
4690 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
4691 return reshape_request(mddev
, sector_nr
, skipped
);
4693 /* No need to check resync_max as we never do more than one
4694 * stripe, and as resync_max will always be on a chunk boundary,
4695 * if the check in md_do_sync didn't fire, there is no chance
4696 * of overstepping resync_max here
4699 /* if there is too many failed drives and we are trying
4700 * to resync, then assert that we are finished, because there is
4701 * nothing we can do.
4703 if (mddev
->degraded
>= conf
->max_degraded
&&
4704 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
4705 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
4709 if (!test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
4711 !bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
4712 sync_blocks
>= STRIPE_SECTORS
) {
4713 /* we can skip this block, and probably more */
4714 sync_blocks
/= STRIPE_SECTORS
;
4716 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
4719 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
4721 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
4723 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
4724 /* make sure we don't swamp the stripe cache if someone else
4725 * is trying to get access
4727 schedule_timeout_uninterruptible(1);
4729 /* Need to check if array will still be degraded after recovery/resync
4730 * We don't need to check the 'failed' flag as when that gets set,
4733 for (i
= 0; i
< conf
->raid_disks
; i
++)
4734 if (conf
->disks
[i
].rdev
== NULL
)
4737 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
4739 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
4744 return STRIPE_SECTORS
;
4747 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
)
4749 /* We may not be able to submit a whole bio at once as there
4750 * may not be enough stripe_heads available.
4751 * We cannot pre-allocate enough stripe_heads as we may need
4752 * more than exist in the cache (if we allow ever large chunks).
4753 * So we do one stripe head at a time and record in
4754 * ->bi_hw_segments how many have been done.
4756 * We *know* that this entire raid_bio is in one chunk, so
4757 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4759 struct stripe_head
*sh
;
4761 sector_t sector
, logical_sector
, last_sector
;
4766 logical_sector
= raid_bio
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4767 sector
= raid5_compute_sector(conf
, logical_sector
,
4769 last_sector
= bio_end_sector(raid_bio
);
4771 for (; logical_sector
< last_sector
;
4772 logical_sector
+= STRIPE_SECTORS
,
4773 sector
+= STRIPE_SECTORS
,
4776 if (scnt
< raid5_bi_processed_stripes(raid_bio
))
4777 /* already done this stripe */
4780 sh
= get_active_stripe(conf
, sector
, 0, 1, 0);
4783 /* failed to get a stripe - must wait */
4784 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
4785 conf
->retry_read_aligned
= raid_bio
;
4789 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0)) {
4791 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
4792 conf
->retry_read_aligned
= raid_bio
;
4796 set_bit(R5_ReadNoMerge
, &sh
->dev
[dd_idx
].flags
);
4801 remaining
= raid5_dec_bi_active_stripes(raid_bio
);
4802 if (remaining
== 0) {
4803 trace_block_bio_complete(bdev_get_queue(raid_bio
->bi_bdev
),
4805 bio_endio(raid_bio
, 0);
4807 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4808 wake_up(&conf
->wait_for_stripe
);
4812 #define MAX_STRIPE_BATCH 8
4813 static int handle_active_stripes(struct r5conf
*conf
)
4815 struct stripe_head
*batch
[MAX_STRIPE_BATCH
], *sh
;
4816 int i
, batch_size
= 0;
4818 while (batch_size
< MAX_STRIPE_BATCH
&&
4819 (sh
= __get_priority_stripe(conf
)) != NULL
)
4820 batch
[batch_size
++] = sh
;
4822 if (batch_size
== 0)
4824 spin_unlock_irq(&conf
->device_lock
);
4826 for (i
= 0; i
< batch_size
; i
++)
4827 handle_stripe(batch
[i
]);
4831 spin_lock_irq(&conf
->device_lock
);
4832 for (i
= 0; i
< batch_size
; i
++)
4833 __release_stripe(conf
, batch
[i
]);
4838 * This is our raid5 kernel thread.
4840 * We scan the hash table for stripes which can be handled now.
4841 * During the scan, completed stripes are saved for us by the interrupt
4842 * handler, so that they will not have to wait for our next wakeup.
4844 static void raid5d(struct md_thread
*thread
)
4846 struct mddev
*mddev
= thread
->mddev
;
4847 struct r5conf
*conf
= mddev
->private;
4849 struct blk_plug plug
;
4851 pr_debug("+++ raid5d active\n");
4853 md_check_recovery(mddev
);
4855 blk_start_plug(&plug
);
4857 spin_lock_irq(&conf
->device_lock
);
4863 !list_empty(&conf
->bitmap_list
)) {
4864 /* Now is a good time to flush some bitmap updates */
4866 spin_unlock_irq(&conf
->device_lock
);
4867 bitmap_unplug(mddev
->bitmap
);
4868 spin_lock_irq(&conf
->device_lock
);
4869 conf
->seq_write
= conf
->seq_flush
;
4870 activate_bit_delay(conf
);
4872 raid5_activate_delayed(conf
);
4874 while ((bio
= remove_bio_from_retry(conf
))) {
4876 spin_unlock_irq(&conf
->device_lock
);
4877 ok
= retry_aligned_read(conf
, bio
);
4878 spin_lock_irq(&conf
->device_lock
);
4884 batch_size
= handle_active_stripes(conf
);
4887 handled
+= batch_size
;
4889 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
)) {
4890 spin_unlock_irq(&conf
->device_lock
);
4891 md_check_recovery(mddev
);
4892 spin_lock_irq(&conf
->device_lock
);
4895 pr_debug("%d stripes handled\n", handled
);
4897 spin_unlock_irq(&conf
->device_lock
);
4899 async_tx_issue_pending_all();
4900 blk_finish_plug(&plug
);
4902 pr_debug("--- raid5d inactive\n");
4906 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
4908 struct r5conf
*conf
= mddev
->private;
4910 return sprintf(page
, "%d\n", conf
->max_nr_stripes
);
4916 raid5_set_cache_size(struct mddev
*mddev
, int size
)
4918 struct r5conf
*conf
= mddev
->private;
4921 if (size
<= 16 || size
> 32768)
4923 while (size
< conf
->max_nr_stripes
) {
4924 if (drop_one_stripe(conf
))
4925 conf
->max_nr_stripes
--;
4929 err
= md_allow_write(mddev
);
4932 while (size
> conf
->max_nr_stripes
) {
4933 if (grow_one_stripe(conf
))
4934 conf
->max_nr_stripes
++;
4939 EXPORT_SYMBOL(raid5_set_cache_size
);
4942 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
4944 struct r5conf
*conf
= mddev
->private;
4948 if (len
>= PAGE_SIZE
)
4953 if (strict_strtoul(page
, 10, &new))
4955 err
= raid5_set_cache_size(mddev
, new);
4961 static struct md_sysfs_entry
4962 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
4963 raid5_show_stripe_cache_size
,
4964 raid5_store_stripe_cache_size
);
4967 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
4969 struct r5conf
*conf
= mddev
->private;
4971 return sprintf(page
, "%d\n", conf
->bypass_threshold
);
4977 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
4979 struct r5conf
*conf
= mddev
->private;
4981 if (len
>= PAGE_SIZE
)
4986 if (strict_strtoul(page
, 10, &new))
4988 if (new > conf
->max_nr_stripes
)
4990 conf
->bypass_threshold
= new;
4994 static struct md_sysfs_entry
4995 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
4997 raid5_show_preread_threshold
,
4998 raid5_store_preread_threshold
);
5001 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
5003 struct r5conf
*conf
= mddev
->private;
5005 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
5010 static struct md_sysfs_entry
5011 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
5013 static struct attribute
*raid5_attrs
[] = {
5014 &raid5_stripecache_size
.attr
,
5015 &raid5_stripecache_active
.attr
,
5016 &raid5_preread_bypass_threshold
.attr
,
5019 static struct attribute_group raid5_attrs_group
= {
5021 .attrs
= raid5_attrs
,
5025 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
5027 struct r5conf
*conf
= mddev
->private;
5030 sectors
= mddev
->dev_sectors
;
5032 /* size is defined by the smallest of previous and new size */
5033 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
5035 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
5036 sectors
&= ~((sector_t
)mddev
->new_chunk_sectors
- 1);
5037 return sectors
* (raid_disks
- conf
->max_degraded
);
5040 static void free_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
5042 safe_put_page(percpu
->spare_page
);
5043 kfree(percpu
->scribble
);
5044 percpu
->spare_page
= NULL
;
5045 percpu
->scribble
= NULL
;
5048 static int alloc_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
5050 if (conf
->level
== 6 && !percpu
->spare_page
)
5051 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
5052 if (!percpu
->scribble
)
5053 percpu
->scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
5055 if (!percpu
->scribble
|| (conf
->level
== 6 && !percpu
->spare_page
)) {
5056 free_scratch_buffer(conf
, percpu
);
5063 static void raid5_free_percpu(struct r5conf
*conf
)
5070 #ifdef CONFIG_HOTPLUG_CPU
5071 unregister_cpu_notifier(&conf
->cpu_notify
);
5075 for_each_possible_cpu(cpu
)
5076 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
5079 free_percpu(conf
->percpu
);
5082 static void free_conf(struct r5conf
*conf
)
5084 shrink_stripes(conf
);
5085 raid5_free_percpu(conf
);
5087 kfree(conf
->stripe_hashtbl
);
5091 #ifdef CONFIG_HOTPLUG_CPU
5092 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
5095 struct r5conf
*conf
= container_of(nfb
, struct r5conf
, cpu_notify
);
5096 long cpu
= (long)hcpu
;
5097 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
5100 case CPU_UP_PREPARE
:
5101 case CPU_UP_PREPARE_FROZEN
:
5102 if (alloc_scratch_buffer(conf
, percpu
)) {
5103 pr_err("%s: failed memory allocation for cpu%ld\n",
5105 return notifier_from_errno(-ENOMEM
);
5109 case CPU_DEAD_FROZEN
:
5110 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
5119 static int raid5_alloc_percpu(struct r5conf
*conf
)
5124 conf
->percpu
= alloc_percpu(struct raid5_percpu
);
5128 #ifdef CONFIG_HOTPLUG_CPU
5129 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
5130 conf
->cpu_notify
.priority
= 0;
5131 err
= register_cpu_notifier(&conf
->cpu_notify
);
5137 for_each_present_cpu(cpu
) {
5138 err
= alloc_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
5140 pr_err("%s: failed memory allocation for cpu%ld\n",
5150 static struct r5conf
*setup_conf(struct mddev
*mddev
)
5152 struct r5conf
*conf
;
5153 int raid_disk
, memory
, max_disks
;
5154 struct md_rdev
*rdev
;
5155 struct disk_info
*disk
;
5158 if (mddev
->new_level
!= 5
5159 && mddev
->new_level
!= 4
5160 && mddev
->new_level
!= 6) {
5161 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5162 mdname(mddev
), mddev
->new_level
);
5163 return ERR_PTR(-EIO
);
5165 if ((mddev
->new_level
== 5
5166 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
5167 (mddev
->new_level
== 6
5168 && !algorithm_valid_raid6(mddev
->new_layout
))) {
5169 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
5170 mdname(mddev
), mddev
->new_layout
);
5171 return ERR_PTR(-EIO
);
5173 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
5174 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5175 mdname(mddev
), mddev
->raid_disks
);
5176 return ERR_PTR(-EINVAL
);
5179 if (!mddev
->new_chunk_sectors
||
5180 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
5181 !is_power_of_2(mddev
->new_chunk_sectors
)) {
5182 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
5183 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
5184 return ERR_PTR(-EINVAL
);
5187 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
5190 spin_lock_init(&conf
->device_lock
);
5191 init_waitqueue_head(&conf
->wait_for_stripe
);
5192 init_waitqueue_head(&conf
->wait_for_overlap
);
5193 INIT_LIST_HEAD(&conf
->handle_list
);
5194 INIT_LIST_HEAD(&conf
->hold_list
);
5195 INIT_LIST_HEAD(&conf
->delayed_list
);
5196 INIT_LIST_HEAD(&conf
->bitmap_list
);
5197 INIT_LIST_HEAD(&conf
->inactive_list
);
5198 atomic_set(&conf
->active_stripes
, 0);
5199 atomic_set(&conf
->preread_active_stripes
, 0);
5200 atomic_set(&conf
->active_aligned_reads
, 0);
5201 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
5202 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
5204 conf
->raid_disks
= mddev
->raid_disks
;
5205 if (mddev
->reshape_position
== MaxSector
)
5206 conf
->previous_raid_disks
= mddev
->raid_disks
;
5208 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
5209 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
5210 conf
->scribble_len
= scribble_len(max_disks
);
5212 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
5217 conf
->mddev
= mddev
;
5219 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
5222 conf
->level
= mddev
->new_level
;
5223 if (raid5_alloc_percpu(conf
) != 0)
5226 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
5228 rdev_for_each(rdev
, mddev
) {
5229 raid_disk
= rdev
->raid_disk
;
5230 if (raid_disk
>= max_disks
5233 disk
= conf
->disks
+ raid_disk
;
5235 if (test_bit(Replacement
, &rdev
->flags
)) {
5236 if (disk
->replacement
)
5238 disk
->replacement
= rdev
;
5245 if (test_bit(In_sync
, &rdev
->flags
)) {
5246 char b
[BDEVNAME_SIZE
];
5247 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
5249 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
5250 } else if (rdev
->saved_raid_disk
!= raid_disk
)
5251 /* Cannot rely on bitmap to complete recovery */
5255 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
5256 conf
->level
= mddev
->new_level
;
5257 if (conf
->level
== 6)
5258 conf
->max_degraded
= 2;
5260 conf
->max_degraded
= 1;
5261 conf
->algorithm
= mddev
->new_layout
;
5262 conf
->max_nr_stripes
= NR_STRIPES
;
5263 conf
->reshape_progress
= mddev
->reshape_position
;
5264 if (conf
->reshape_progress
!= MaxSector
) {
5265 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
5266 conf
->prev_algo
= mddev
->layout
;
5269 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
5270 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
5271 if (grow_stripes(conf
, conf
->max_nr_stripes
)) {
5273 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5274 mdname(mddev
), memory
);
5277 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
5278 mdname(mddev
), memory
);
5280 sprintf(pers_name
, "raid%d", mddev
->new_level
);
5281 conf
->thread
= md_register_thread(raid5d
, mddev
, pers_name
);
5282 if (!conf
->thread
) {
5284 "md/raid:%s: couldn't allocate thread.\n",
5294 return ERR_PTR(-EIO
);
5296 return ERR_PTR(-ENOMEM
);
5300 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
5303 case ALGORITHM_PARITY_0
:
5304 if (raid_disk
< max_degraded
)
5307 case ALGORITHM_PARITY_N
:
5308 if (raid_disk
>= raid_disks
- max_degraded
)
5311 case ALGORITHM_PARITY_0_6
:
5312 if (raid_disk
== 0 ||
5313 raid_disk
== raid_disks
- 1)
5316 case ALGORITHM_LEFT_ASYMMETRIC_6
:
5317 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
5318 case ALGORITHM_LEFT_SYMMETRIC_6
:
5319 case ALGORITHM_RIGHT_SYMMETRIC_6
:
5320 if (raid_disk
== raid_disks
- 1)
5326 static int run(struct mddev
*mddev
)
5328 struct r5conf
*conf
;
5329 int working_disks
= 0;
5330 int dirty_parity_disks
= 0;
5331 struct md_rdev
*rdev
;
5332 sector_t reshape_offset
= 0;
5334 long long min_offset_diff
= 0;
5337 if (mddev
->recovery_cp
!= MaxSector
)
5338 printk(KERN_NOTICE
"md/raid:%s: not clean"
5339 " -- starting background reconstruction\n",
5342 rdev_for_each(rdev
, mddev
) {
5344 if (rdev
->raid_disk
< 0)
5346 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
5348 min_offset_diff
= diff
;
5350 } else if (mddev
->reshape_backwards
&&
5351 diff
< min_offset_diff
)
5352 min_offset_diff
= diff
;
5353 else if (!mddev
->reshape_backwards
&&
5354 diff
> min_offset_diff
)
5355 min_offset_diff
= diff
;
5358 if (mddev
->reshape_position
!= MaxSector
) {
5359 /* Check that we can continue the reshape.
5360 * Difficulties arise if the stripe we would write to
5361 * next is at or after the stripe we would read from next.
5362 * For a reshape that changes the number of devices, this
5363 * is only possible for a very short time, and mdadm makes
5364 * sure that time appears to have past before assembling
5365 * the array. So we fail if that time hasn't passed.
5366 * For a reshape that keeps the number of devices the same
5367 * mdadm must be monitoring the reshape can keeping the
5368 * critical areas read-only and backed up. It will start
5369 * the array in read-only mode, so we check for that.
5371 sector_t here_new
, here_old
;
5373 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
5375 if (mddev
->new_level
!= mddev
->level
) {
5376 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
5377 "required - aborting.\n",
5381 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
5382 /* reshape_position must be on a new-stripe boundary, and one
5383 * further up in new geometry must map after here in old
5386 here_new
= mddev
->reshape_position
;
5387 if (sector_div(here_new
, mddev
->new_chunk_sectors
*
5388 (mddev
->raid_disks
- max_degraded
))) {
5389 printk(KERN_ERR
"md/raid:%s: reshape_position not "
5390 "on a stripe boundary\n", mdname(mddev
));
5393 reshape_offset
= here_new
* mddev
->new_chunk_sectors
;
5394 /* here_new is the stripe we will write to */
5395 here_old
= mddev
->reshape_position
;
5396 sector_div(here_old
, mddev
->chunk_sectors
*
5397 (old_disks
-max_degraded
));
5398 /* here_old is the first stripe that we might need to read
5400 if (mddev
->delta_disks
== 0) {
5401 if ((here_new
* mddev
->new_chunk_sectors
!=
5402 here_old
* mddev
->chunk_sectors
)) {
5403 printk(KERN_ERR
"md/raid:%s: reshape position is"
5404 " confused - aborting\n", mdname(mddev
));
5407 /* We cannot be sure it is safe to start an in-place
5408 * reshape. It is only safe if user-space is monitoring
5409 * and taking constant backups.
5410 * mdadm always starts a situation like this in
5411 * readonly mode so it can take control before
5412 * allowing any writes. So just check for that.
5414 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
5415 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
5416 /* not really in-place - so OK */;
5417 else if (mddev
->ro
== 0) {
5418 printk(KERN_ERR
"md/raid:%s: in-place reshape "
5419 "must be started in read-only mode "
5424 } else if (mddev
->reshape_backwards
5425 ? (here_new
* mddev
->new_chunk_sectors
+ min_offset_diff
<=
5426 here_old
* mddev
->chunk_sectors
)
5427 : (here_new
* mddev
->new_chunk_sectors
>=
5428 here_old
* mddev
->chunk_sectors
+ (-min_offset_diff
))) {
5429 /* Reading from the same stripe as writing to - bad */
5430 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
5431 "auto-recovery - aborting.\n",
5435 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
5437 /* OK, we should be able to continue; */
5439 BUG_ON(mddev
->level
!= mddev
->new_level
);
5440 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
5441 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
5442 BUG_ON(mddev
->delta_disks
!= 0);
5445 if (mddev
->private == NULL
)
5446 conf
= setup_conf(mddev
);
5448 conf
= mddev
->private;
5451 return PTR_ERR(conf
);
5453 conf
->min_offset_diff
= min_offset_diff
;
5454 mddev
->thread
= conf
->thread
;
5455 conf
->thread
= NULL
;
5456 mddev
->private = conf
;
5458 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
5460 rdev
= conf
->disks
[i
].rdev
;
5461 if (!rdev
&& conf
->disks
[i
].replacement
) {
5462 /* The replacement is all we have yet */
5463 rdev
= conf
->disks
[i
].replacement
;
5464 conf
->disks
[i
].replacement
= NULL
;
5465 clear_bit(Replacement
, &rdev
->flags
);
5466 conf
->disks
[i
].rdev
= rdev
;
5470 if (conf
->disks
[i
].replacement
&&
5471 conf
->reshape_progress
!= MaxSector
) {
5472 /* replacements and reshape simply do not mix. */
5473 printk(KERN_ERR
"md: cannot handle concurrent "
5474 "replacement and reshape.\n");
5477 if (test_bit(In_sync
, &rdev
->flags
)) {
5481 /* This disc is not fully in-sync. However if it
5482 * just stored parity (beyond the recovery_offset),
5483 * when we don't need to be concerned about the
5484 * array being dirty.
5485 * When reshape goes 'backwards', we never have
5486 * partially completed devices, so we only need
5487 * to worry about reshape going forwards.
5489 /* Hack because v0.91 doesn't store recovery_offset properly. */
5490 if (mddev
->major_version
== 0 &&
5491 mddev
->minor_version
> 90)
5492 rdev
->recovery_offset
= reshape_offset
;
5494 if (rdev
->recovery_offset
< reshape_offset
) {
5495 /* We need to check old and new layout */
5496 if (!only_parity(rdev
->raid_disk
,
5499 conf
->max_degraded
))
5502 if (!only_parity(rdev
->raid_disk
,
5504 conf
->previous_raid_disks
,
5505 conf
->max_degraded
))
5507 dirty_parity_disks
++;
5511 * 0 for a fully functional array, 1 or 2 for a degraded array.
5513 mddev
->degraded
= calc_degraded(conf
);
5515 if (has_failed(conf
)) {
5516 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
5517 " (%d/%d failed)\n",
5518 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
5522 /* device size must be a multiple of chunk size */
5523 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
5524 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
5526 if (mddev
->degraded
> dirty_parity_disks
&&
5527 mddev
->recovery_cp
!= MaxSector
) {
5528 if (mddev
->ok_start_degraded
)
5530 "md/raid:%s: starting dirty degraded array"
5531 " - data corruption possible.\n",
5535 "md/raid:%s: cannot start dirty degraded array.\n",
5541 if (mddev
->degraded
== 0)
5542 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
5543 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
5544 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
5547 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
5548 " out of %d devices, algorithm %d\n",
5549 mdname(mddev
), conf
->level
,
5550 mddev
->raid_disks
- mddev
->degraded
,
5551 mddev
->raid_disks
, mddev
->new_layout
);
5553 print_raid5_conf(conf
);
5555 if (conf
->reshape_progress
!= MaxSector
) {
5556 conf
->reshape_safe
= conf
->reshape_progress
;
5557 atomic_set(&conf
->reshape_stripes
, 0);
5558 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5559 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5560 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5561 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5562 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5567 /* Ok, everything is just fine now */
5568 if (mddev
->to_remove
== &raid5_attrs_group
)
5569 mddev
->to_remove
= NULL
;
5570 else if (mddev
->kobj
.sd
&&
5571 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
5573 "raid5: failed to create sysfs attributes for %s\n",
5575 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5579 bool discard_supported
= true;
5580 /* read-ahead size must cover two whole stripes, which
5581 * is 2 * (datadisks) * chunksize where 'n' is the
5582 * number of raid devices
5584 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
5585 int stripe
= data_disks
*
5586 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
5587 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
5588 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
5590 blk_queue_merge_bvec(mddev
->queue
, raid5_mergeable_bvec
);
5592 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
5593 mddev
->queue
->backing_dev_info
.congested_fn
= raid5_congested
;
5595 chunk_size
= mddev
->chunk_sectors
<< 9;
5596 blk_queue_io_min(mddev
->queue
, chunk_size
);
5597 blk_queue_io_opt(mddev
->queue
, chunk_size
*
5598 (conf
->raid_disks
- conf
->max_degraded
));
5600 * We can only discard a whole stripe. It doesn't make sense to
5601 * discard data disk but write parity disk
5603 stripe
= stripe
* PAGE_SIZE
;
5604 /* Round up to power of 2, as discard handling
5605 * currently assumes that */
5606 while ((stripe
-1) & stripe
)
5607 stripe
= (stripe
| (stripe
-1)) + 1;
5608 mddev
->queue
->limits
.discard_alignment
= stripe
;
5609 mddev
->queue
->limits
.discard_granularity
= stripe
;
5611 * unaligned part of discard request will be ignored, so can't
5612 * guarantee discard_zerors_data
5614 mddev
->queue
->limits
.discard_zeroes_data
= 0;
5616 blk_queue_max_write_same_sectors(mddev
->queue
, 0);
5618 rdev_for_each(rdev
, mddev
) {
5619 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
5620 rdev
->data_offset
<< 9);
5621 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
5622 rdev
->new_data_offset
<< 9);
5624 * discard_zeroes_data is required, otherwise data
5625 * could be lost. Consider a scenario: discard a stripe
5626 * (the stripe could be inconsistent if
5627 * discard_zeroes_data is 0); write one disk of the
5628 * stripe (the stripe could be inconsistent again
5629 * depending on which disks are used to calculate
5630 * parity); the disk is broken; The stripe data of this
5633 if (!blk_queue_discard(bdev_get_queue(rdev
->bdev
)) ||
5634 !bdev_get_queue(rdev
->bdev
)->
5635 limits
.discard_zeroes_data
)
5636 discard_supported
= false;
5639 if (discard_supported
&&
5640 mddev
->queue
->limits
.max_discard_sectors
>= stripe
&&
5641 mddev
->queue
->limits
.discard_granularity
>= stripe
)
5642 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
5645 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
5651 md_unregister_thread(&mddev
->thread
);
5652 print_raid5_conf(conf
);
5654 mddev
->private = NULL
;
5655 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
5659 static int stop(struct mddev
*mddev
)
5661 struct r5conf
*conf
= mddev
->private;
5663 md_unregister_thread(&mddev
->thread
);
5665 mddev
->queue
->backing_dev_info
.congested_fn
= NULL
;
5667 mddev
->private = NULL
;
5668 mddev
->to_remove
= &raid5_attrs_group
;
5672 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
5674 struct r5conf
*conf
= mddev
->private;
5677 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
5678 mddev
->chunk_sectors
/ 2, mddev
->layout
);
5679 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
5680 for (i
= 0; i
< conf
->raid_disks
; i
++)
5681 seq_printf (seq
, "%s",
5682 conf
->disks
[i
].rdev
&&
5683 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
5684 seq_printf (seq
, "]");
5687 static void print_raid5_conf (struct r5conf
*conf
)
5690 struct disk_info
*tmp
;
5692 printk(KERN_DEBUG
"RAID conf printout:\n");
5694 printk("(conf==NULL)\n");
5697 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
5699 conf
->raid_disks
- conf
->mddev
->degraded
);
5701 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5702 char b
[BDEVNAME_SIZE
];
5703 tmp
= conf
->disks
+ i
;
5705 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
5706 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
5707 bdevname(tmp
->rdev
->bdev
, b
));
5711 static int raid5_spare_active(struct mddev
*mddev
)
5714 struct r5conf
*conf
= mddev
->private;
5715 struct disk_info
*tmp
;
5717 unsigned long flags
;
5719 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5720 tmp
= conf
->disks
+ i
;
5721 if (tmp
->replacement
5722 && tmp
->replacement
->recovery_offset
== MaxSector
5723 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
5724 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
5725 /* Replacement has just become active. */
5727 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
5730 /* Replaced device not technically faulty,
5731 * but we need to be sure it gets removed
5732 * and never re-added.
5734 set_bit(Faulty
, &tmp
->rdev
->flags
);
5735 sysfs_notify_dirent_safe(
5736 tmp
->rdev
->sysfs_state
);
5738 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
5739 } else if (tmp
->rdev
5740 && tmp
->rdev
->recovery_offset
== MaxSector
5741 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
5742 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
5744 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
5747 spin_lock_irqsave(&conf
->device_lock
, flags
);
5748 mddev
->degraded
= calc_degraded(conf
);
5749 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5750 print_raid5_conf(conf
);
5754 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
5756 struct r5conf
*conf
= mddev
->private;
5758 int number
= rdev
->raid_disk
;
5759 struct md_rdev
**rdevp
;
5760 struct disk_info
*p
= conf
->disks
+ number
;
5762 print_raid5_conf(conf
);
5763 if (rdev
== p
->rdev
)
5765 else if (rdev
== p
->replacement
)
5766 rdevp
= &p
->replacement
;
5770 if (number
>= conf
->raid_disks
&&
5771 conf
->reshape_progress
== MaxSector
)
5772 clear_bit(In_sync
, &rdev
->flags
);
5774 if (test_bit(In_sync
, &rdev
->flags
) ||
5775 atomic_read(&rdev
->nr_pending
)) {
5779 /* Only remove non-faulty devices if recovery
5782 if (!test_bit(Faulty
, &rdev
->flags
) &&
5783 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
5784 !has_failed(conf
) &&
5785 (!p
->replacement
|| p
->replacement
== rdev
) &&
5786 number
< conf
->raid_disks
) {
5792 if (atomic_read(&rdev
->nr_pending
)) {
5793 /* lost the race, try later */
5796 } else if (p
->replacement
) {
5797 /* We must have just cleared 'rdev' */
5798 p
->rdev
= p
->replacement
;
5799 clear_bit(Replacement
, &p
->replacement
->flags
);
5800 smp_mb(); /* Make sure other CPUs may see both as identical
5801 * but will never see neither - if they are careful
5803 p
->replacement
= NULL
;
5804 clear_bit(WantReplacement
, &rdev
->flags
);
5806 /* We might have just removed the Replacement as faulty-
5807 * clear the bit just in case
5809 clear_bit(WantReplacement
, &rdev
->flags
);
5812 print_raid5_conf(conf
);
5816 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
5818 struct r5conf
*conf
= mddev
->private;
5821 struct disk_info
*p
;
5823 int last
= conf
->raid_disks
- 1;
5825 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
5828 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
5829 /* no point adding a device */
5832 if (rdev
->raid_disk
>= 0)
5833 first
= last
= rdev
->raid_disk
;
5836 * find the disk ... but prefer rdev->saved_raid_disk
5839 if (rdev
->saved_raid_disk
>= 0 &&
5840 rdev
->saved_raid_disk
>= first
&&
5841 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
5842 first
= rdev
->saved_raid_disk
;
5844 for (disk
= first
; disk
<= last
; disk
++) {
5845 p
= conf
->disks
+ disk
;
5846 if (p
->rdev
== NULL
) {
5847 clear_bit(In_sync
, &rdev
->flags
);
5848 rdev
->raid_disk
= disk
;
5850 if (rdev
->saved_raid_disk
!= disk
)
5852 rcu_assign_pointer(p
->rdev
, rdev
);
5856 for (disk
= first
; disk
<= last
; disk
++) {
5857 p
= conf
->disks
+ disk
;
5858 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
5859 p
->replacement
== NULL
) {
5860 clear_bit(In_sync
, &rdev
->flags
);
5861 set_bit(Replacement
, &rdev
->flags
);
5862 rdev
->raid_disk
= disk
;
5865 rcu_assign_pointer(p
->replacement
, rdev
);
5870 print_raid5_conf(conf
);
5874 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
5876 /* no resync is happening, and there is enough space
5877 * on all devices, so we can resize.
5878 * We need to make sure resync covers any new space.
5879 * If the array is shrinking we should possibly wait until
5880 * any io in the removed space completes, but it hardly seems
5884 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
5885 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
5886 if (mddev
->external_size
&&
5887 mddev
->array_sectors
> newsize
)
5889 if (mddev
->bitmap
) {
5890 int ret
= bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
5894 md_set_array_sectors(mddev
, newsize
);
5895 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5896 revalidate_disk(mddev
->gendisk
);
5897 if (sectors
> mddev
->dev_sectors
&&
5898 mddev
->recovery_cp
> mddev
->dev_sectors
) {
5899 mddev
->recovery_cp
= mddev
->dev_sectors
;
5900 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
5902 mddev
->dev_sectors
= sectors
;
5903 mddev
->resync_max_sectors
= sectors
;
5907 static int check_stripe_cache(struct mddev
*mddev
)
5909 /* Can only proceed if there are plenty of stripe_heads.
5910 * We need a minimum of one full stripe,, and for sensible progress
5911 * it is best to have about 4 times that.
5912 * If we require 4 times, then the default 256 4K stripe_heads will
5913 * allow for chunk sizes up to 256K, which is probably OK.
5914 * If the chunk size is greater, user-space should request more
5915 * stripe_heads first.
5917 struct r5conf
*conf
= mddev
->private;
5918 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5919 > conf
->max_nr_stripes
||
5920 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5921 > conf
->max_nr_stripes
) {
5922 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5924 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
5931 static int check_reshape(struct mddev
*mddev
)
5933 struct r5conf
*conf
= mddev
->private;
5935 if (mddev
->delta_disks
== 0 &&
5936 mddev
->new_layout
== mddev
->layout
&&
5937 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
5938 return 0; /* nothing to do */
5939 if (has_failed(conf
))
5941 if (mddev
->delta_disks
< 0) {
5942 /* We might be able to shrink, but the devices must
5943 * be made bigger first.
5944 * For raid6, 4 is the minimum size.
5945 * Otherwise 2 is the minimum
5948 if (mddev
->level
== 6)
5950 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
5954 if (!check_stripe_cache(mddev
))
5957 return resize_stripes(conf
, (conf
->previous_raid_disks
5958 + mddev
->delta_disks
));
5961 static int raid5_start_reshape(struct mddev
*mddev
)
5963 struct r5conf
*conf
= mddev
->private;
5964 struct md_rdev
*rdev
;
5966 unsigned long flags
;
5968 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
5971 if (!check_stripe_cache(mddev
))
5974 if (has_failed(conf
))
5977 rdev_for_each(rdev
, mddev
) {
5978 if (!test_bit(In_sync
, &rdev
->flags
)
5979 && !test_bit(Faulty
, &rdev
->flags
))
5983 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
5984 /* Not enough devices even to make a degraded array
5989 /* Refuse to reduce size of the array. Any reductions in
5990 * array size must be through explicit setting of array_size
5993 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
5994 < mddev
->array_sectors
) {
5995 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
5996 "before number of disks\n", mdname(mddev
));
6000 atomic_set(&conf
->reshape_stripes
, 0);
6001 spin_lock_irq(&conf
->device_lock
);
6002 conf
->previous_raid_disks
= conf
->raid_disks
;
6003 conf
->raid_disks
+= mddev
->delta_disks
;
6004 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
6005 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
6006 conf
->prev_algo
= conf
->algorithm
;
6007 conf
->algorithm
= mddev
->new_layout
;
6009 /* Code that selects data_offset needs to see the generation update
6010 * if reshape_progress has been set - so a memory barrier needed.
6013 if (mddev
->reshape_backwards
)
6014 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
6016 conf
->reshape_progress
= 0;
6017 conf
->reshape_safe
= conf
->reshape_progress
;
6018 spin_unlock_irq(&conf
->device_lock
);
6020 /* Add some new drives, as many as will fit.
6021 * We know there are enough to make the newly sized array work.
6022 * Don't add devices if we are reducing the number of
6023 * devices in the array. This is because it is not possible
6024 * to correctly record the "partially reconstructed" state of
6025 * such devices during the reshape and confusion could result.
6027 if (mddev
->delta_disks
>= 0) {
6028 rdev_for_each(rdev
, mddev
)
6029 if (rdev
->raid_disk
< 0 &&
6030 !test_bit(Faulty
, &rdev
->flags
)) {
6031 if (raid5_add_disk(mddev
, rdev
) == 0) {
6033 >= conf
->previous_raid_disks
)
6034 set_bit(In_sync
, &rdev
->flags
);
6036 rdev
->recovery_offset
= 0;
6038 if (sysfs_link_rdev(mddev
, rdev
))
6039 /* Failure here is OK */;
6041 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
6042 && !test_bit(Faulty
, &rdev
->flags
)) {
6043 /* This is a spare that was manually added */
6044 set_bit(In_sync
, &rdev
->flags
);
6047 /* When a reshape changes the number of devices,
6048 * ->degraded is measured against the larger of the
6049 * pre and post number of devices.
6051 spin_lock_irqsave(&conf
->device_lock
, flags
);
6052 mddev
->degraded
= calc_degraded(conf
);
6053 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
6055 mddev
->raid_disks
= conf
->raid_disks
;
6056 mddev
->reshape_position
= conf
->reshape_progress
;
6057 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
6059 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
6060 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
6061 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
6062 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
6063 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
6065 if (!mddev
->sync_thread
) {
6066 mddev
->recovery
= 0;
6067 spin_lock_irq(&conf
->device_lock
);
6068 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
6069 rdev_for_each(rdev
, mddev
)
6070 rdev
->new_data_offset
= rdev
->data_offset
;
6072 conf
->reshape_progress
= MaxSector
;
6073 mddev
->reshape_position
= MaxSector
;
6074 spin_unlock_irq(&conf
->device_lock
);
6077 conf
->reshape_checkpoint
= jiffies
;
6078 md_wakeup_thread(mddev
->sync_thread
);
6079 md_new_event(mddev
);
6083 /* This is called from the reshape thread and should make any
6084 * changes needed in 'conf'
6086 static void end_reshape(struct r5conf
*conf
)
6089 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
6090 struct md_rdev
*rdev
;
6092 spin_lock_irq(&conf
->device_lock
);
6093 conf
->previous_raid_disks
= conf
->raid_disks
;
6094 rdev_for_each(rdev
, conf
->mddev
)
6095 rdev
->data_offset
= rdev
->new_data_offset
;
6097 conf
->reshape_progress
= MaxSector
;
6098 spin_unlock_irq(&conf
->device_lock
);
6099 wake_up(&conf
->wait_for_overlap
);
6101 /* read-ahead size must cover two whole stripes, which is
6102 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6104 if (conf
->mddev
->queue
) {
6105 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
6106 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
6108 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
6109 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
6114 /* This is called from the raid5d thread with mddev_lock held.
6115 * It makes config changes to the device.
6117 static void raid5_finish_reshape(struct mddev
*mddev
)
6119 struct r5conf
*conf
= mddev
->private;
6121 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
6123 if (mddev
->delta_disks
> 0) {
6124 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
6125 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
6126 revalidate_disk(mddev
->gendisk
);
6129 spin_lock_irq(&conf
->device_lock
);
6130 mddev
->degraded
= calc_degraded(conf
);
6131 spin_unlock_irq(&conf
->device_lock
);
6132 for (d
= conf
->raid_disks
;
6133 d
< conf
->raid_disks
- mddev
->delta_disks
;
6135 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
6137 clear_bit(In_sync
, &rdev
->flags
);
6138 rdev
= conf
->disks
[d
].replacement
;
6140 clear_bit(In_sync
, &rdev
->flags
);
6143 mddev
->layout
= conf
->algorithm
;
6144 mddev
->chunk_sectors
= conf
->chunk_sectors
;
6145 mddev
->reshape_position
= MaxSector
;
6146 mddev
->delta_disks
= 0;
6147 mddev
->reshape_backwards
= 0;
6151 static void raid5_quiesce(struct mddev
*mddev
, int state
)
6153 struct r5conf
*conf
= mddev
->private;
6156 case 2: /* resume for a suspend */
6157 wake_up(&conf
->wait_for_overlap
);
6160 case 1: /* stop all writes */
6161 spin_lock_irq(&conf
->device_lock
);
6162 /* '2' tells resync/reshape to pause so that all
6163 * active stripes can drain
6166 wait_event_lock_irq(conf
->wait_for_stripe
,
6167 atomic_read(&conf
->active_stripes
) == 0 &&
6168 atomic_read(&conf
->active_aligned_reads
) == 0,
6171 spin_unlock_irq(&conf
->device_lock
);
6172 /* allow reshape to continue */
6173 wake_up(&conf
->wait_for_overlap
);
6176 case 0: /* re-enable writes */
6177 spin_lock_irq(&conf
->device_lock
);
6179 wake_up(&conf
->wait_for_stripe
);
6180 wake_up(&conf
->wait_for_overlap
);
6181 spin_unlock_irq(&conf
->device_lock
);
6187 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
6189 struct r0conf
*raid0_conf
= mddev
->private;
6192 /* for raid0 takeover only one zone is supported */
6193 if (raid0_conf
->nr_strip_zones
> 1) {
6194 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6196 return ERR_PTR(-EINVAL
);
6199 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
6200 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
6201 mddev
->dev_sectors
= sectors
;
6202 mddev
->new_level
= level
;
6203 mddev
->new_layout
= ALGORITHM_PARITY_N
;
6204 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
6205 mddev
->raid_disks
+= 1;
6206 mddev
->delta_disks
= 1;
6207 /* make sure it will be not marked as dirty */
6208 mddev
->recovery_cp
= MaxSector
;
6210 return setup_conf(mddev
);
6214 static void *raid5_takeover_raid1(struct mddev
*mddev
)
6218 if (mddev
->raid_disks
!= 2 ||
6219 mddev
->degraded
> 1)
6220 return ERR_PTR(-EINVAL
);
6222 /* Should check if there are write-behind devices? */
6224 chunksect
= 64*2; /* 64K by default */
6226 /* The array must be an exact multiple of chunksize */
6227 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
6230 if ((chunksect
<<9) < STRIPE_SIZE
)
6231 /* array size does not allow a suitable chunk size */
6232 return ERR_PTR(-EINVAL
);
6234 mddev
->new_level
= 5;
6235 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
6236 mddev
->new_chunk_sectors
= chunksect
;
6238 return setup_conf(mddev
);
6241 static void *raid5_takeover_raid6(struct mddev
*mddev
)
6245 switch (mddev
->layout
) {
6246 case ALGORITHM_LEFT_ASYMMETRIC_6
:
6247 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
6249 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
6250 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
6252 case ALGORITHM_LEFT_SYMMETRIC_6
:
6253 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
6255 case ALGORITHM_RIGHT_SYMMETRIC_6
:
6256 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
6258 case ALGORITHM_PARITY_0_6
:
6259 new_layout
= ALGORITHM_PARITY_0
;
6261 case ALGORITHM_PARITY_N
:
6262 new_layout
= ALGORITHM_PARITY_N
;
6265 return ERR_PTR(-EINVAL
);
6267 mddev
->new_level
= 5;
6268 mddev
->new_layout
= new_layout
;
6269 mddev
->delta_disks
= -1;
6270 mddev
->raid_disks
-= 1;
6271 return setup_conf(mddev
);
6275 static int raid5_check_reshape(struct mddev
*mddev
)
6277 /* For a 2-drive array, the layout and chunk size can be changed
6278 * immediately as not restriping is needed.
6279 * For larger arrays we record the new value - after validation
6280 * to be used by a reshape pass.
6282 struct r5conf
*conf
= mddev
->private;
6283 int new_chunk
= mddev
->new_chunk_sectors
;
6285 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
6287 if (new_chunk
> 0) {
6288 if (!is_power_of_2(new_chunk
))
6290 if (new_chunk
< (PAGE_SIZE
>>9))
6292 if (mddev
->array_sectors
& (new_chunk
-1))
6293 /* not factor of array size */
6297 /* They look valid */
6299 if (mddev
->raid_disks
== 2) {
6300 /* can make the change immediately */
6301 if (mddev
->new_layout
>= 0) {
6302 conf
->algorithm
= mddev
->new_layout
;
6303 mddev
->layout
= mddev
->new_layout
;
6305 if (new_chunk
> 0) {
6306 conf
->chunk_sectors
= new_chunk
;
6307 mddev
->chunk_sectors
= new_chunk
;
6309 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
6310 md_wakeup_thread(mddev
->thread
);
6312 return check_reshape(mddev
);
6315 static int raid6_check_reshape(struct mddev
*mddev
)
6317 int new_chunk
= mddev
->new_chunk_sectors
;
6319 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
6321 if (new_chunk
> 0) {
6322 if (!is_power_of_2(new_chunk
))
6324 if (new_chunk
< (PAGE_SIZE
>> 9))
6326 if (mddev
->array_sectors
& (new_chunk
-1))
6327 /* not factor of array size */
6331 /* They look valid */
6332 return check_reshape(mddev
);
6335 static void *raid5_takeover(struct mddev
*mddev
)
6337 /* raid5 can take over:
6338 * raid0 - if there is only one strip zone - make it a raid4 layout
6339 * raid1 - if there are two drives. We need to know the chunk size
6340 * raid4 - trivial - just use a raid4 layout.
6341 * raid6 - Providing it is a *_6 layout
6343 if (mddev
->level
== 0)
6344 return raid45_takeover_raid0(mddev
, 5);
6345 if (mddev
->level
== 1)
6346 return raid5_takeover_raid1(mddev
);
6347 if (mddev
->level
== 4) {
6348 mddev
->new_layout
= ALGORITHM_PARITY_N
;
6349 mddev
->new_level
= 5;
6350 return setup_conf(mddev
);
6352 if (mddev
->level
== 6)
6353 return raid5_takeover_raid6(mddev
);
6355 return ERR_PTR(-EINVAL
);
6358 static void *raid4_takeover(struct mddev
*mddev
)
6360 /* raid4 can take over:
6361 * raid0 - if there is only one strip zone
6362 * raid5 - if layout is right
6364 if (mddev
->level
== 0)
6365 return raid45_takeover_raid0(mddev
, 4);
6366 if (mddev
->level
== 5 &&
6367 mddev
->layout
== ALGORITHM_PARITY_N
) {
6368 mddev
->new_layout
= 0;
6369 mddev
->new_level
= 4;
6370 return setup_conf(mddev
);
6372 return ERR_PTR(-EINVAL
);
6375 static struct md_personality raid5_personality
;
6377 static void *raid6_takeover(struct mddev
*mddev
)
6379 /* Currently can only take over a raid5. We map the
6380 * personality to an equivalent raid6 personality
6381 * with the Q block at the end.
6385 if (mddev
->pers
!= &raid5_personality
)
6386 return ERR_PTR(-EINVAL
);
6387 if (mddev
->degraded
> 1)
6388 return ERR_PTR(-EINVAL
);
6389 if (mddev
->raid_disks
> 253)
6390 return ERR_PTR(-EINVAL
);
6391 if (mddev
->raid_disks
< 3)
6392 return ERR_PTR(-EINVAL
);
6394 switch (mddev
->layout
) {
6395 case ALGORITHM_LEFT_ASYMMETRIC
:
6396 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
6398 case ALGORITHM_RIGHT_ASYMMETRIC
:
6399 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
6401 case ALGORITHM_LEFT_SYMMETRIC
:
6402 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
6404 case ALGORITHM_RIGHT_SYMMETRIC
:
6405 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
6407 case ALGORITHM_PARITY_0
:
6408 new_layout
= ALGORITHM_PARITY_0_6
;
6410 case ALGORITHM_PARITY_N
:
6411 new_layout
= ALGORITHM_PARITY_N
;
6414 return ERR_PTR(-EINVAL
);
6416 mddev
->new_level
= 6;
6417 mddev
->new_layout
= new_layout
;
6418 mddev
->delta_disks
= 1;
6419 mddev
->raid_disks
+= 1;
6420 return setup_conf(mddev
);
6424 static struct md_personality raid6_personality
=
6428 .owner
= THIS_MODULE
,
6429 .make_request
= make_request
,
6433 .error_handler
= error
,
6434 .hot_add_disk
= raid5_add_disk
,
6435 .hot_remove_disk
= raid5_remove_disk
,
6436 .spare_active
= raid5_spare_active
,
6437 .sync_request
= sync_request
,
6438 .resize
= raid5_resize
,
6440 .check_reshape
= raid6_check_reshape
,
6441 .start_reshape
= raid5_start_reshape
,
6442 .finish_reshape
= raid5_finish_reshape
,
6443 .quiesce
= raid5_quiesce
,
6444 .takeover
= raid6_takeover
,
6446 static struct md_personality raid5_personality
=
6450 .owner
= THIS_MODULE
,
6451 .make_request
= make_request
,
6455 .error_handler
= error
,
6456 .hot_add_disk
= raid5_add_disk
,
6457 .hot_remove_disk
= raid5_remove_disk
,
6458 .spare_active
= raid5_spare_active
,
6459 .sync_request
= sync_request
,
6460 .resize
= raid5_resize
,
6462 .check_reshape
= raid5_check_reshape
,
6463 .start_reshape
= raid5_start_reshape
,
6464 .finish_reshape
= raid5_finish_reshape
,
6465 .quiesce
= raid5_quiesce
,
6466 .takeover
= raid5_takeover
,
6469 static struct md_personality raid4_personality
=
6473 .owner
= THIS_MODULE
,
6474 .make_request
= make_request
,
6478 .error_handler
= error
,
6479 .hot_add_disk
= raid5_add_disk
,
6480 .hot_remove_disk
= raid5_remove_disk
,
6481 .spare_active
= raid5_spare_active
,
6482 .sync_request
= sync_request
,
6483 .resize
= raid5_resize
,
6485 .check_reshape
= raid5_check_reshape
,
6486 .start_reshape
= raid5_start_reshape
,
6487 .finish_reshape
= raid5_finish_reshape
,
6488 .quiesce
= raid5_quiesce
,
6489 .takeover
= raid4_takeover
,
6492 static int __init
raid5_init(void)
6494 register_md_personality(&raid6_personality
);
6495 register_md_personality(&raid5_personality
);
6496 register_md_personality(&raid4_personality
);
6500 static void raid5_exit(void)
6502 unregister_md_personality(&raid6_personality
);
6503 unregister_md_personality(&raid5_personality
);
6504 unregister_md_personality(&raid4_personality
);
6507 module_init(raid5_init
);
6508 module_exit(raid5_exit
);
6509 MODULE_LICENSE("GPL");
6510 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6511 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6512 MODULE_ALIAS("md-raid5");
6513 MODULE_ALIAS("md-raid4");
6514 MODULE_ALIAS("md-level-5");
6515 MODULE_ALIAS("md-level-4");
6516 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6517 MODULE_ALIAS("md-raid6");
6518 MODULE_ALIAS("md-level-6");
6520 /* This used to be two separate modules, they were: */
6521 MODULE_ALIAS("raid5");
6522 MODULE_ALIAS("raid6");