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>
65 #define NR_STRIPES 256
66 #define STRIPE_SIZE PAGE_SIZE
67 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
68 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
69 #define IO_THRESHOLD 1
70 #define BYPASS_THRESHOLD 1
71 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
72 #define HASH_MASK (NR_HASH - 1)
74 static inline struct hlist_head
*stripe_hash(struct r5conf
*conf
, sector_t sect
)
76 int hash
= (sect
>> STRIPE_SHIFT
) & HASH_MASK
;
77 return &conf
->stripe_hashtbl
[hash
];
80 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
81 * order without overlap. There may be several bio's per stripe+device, and
82 * a bio could span several devices.
83 * When walking this list for a particular stripe+device, we must never proceed
84 * beyond a bio that extends past this device, as the next bio might no longer
86 * This function is used to determine the 'next' bio in the list, given the sector
87 * of the current stripe+device
89 static inline struct bio
*r5_next_bio(struct bio
*bio
, sector_t sector
)
91 int sectors
= bio
->bi_size
>> 9;
92 if (bio
->bi_sector
+ sectors
< sector
+ STRIPE_SECTORS
)
99 * We maintain a biased count of active stripes in the bottom 16 bits of
100 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
102 static inline int raid5_bi_processed_stripes(struct bio
*bio
)
104 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
105 return (atomic_read(segments
) >> 16) & 0xffff;
108 static inline int raid5_dec_bi_active_stripes(struct bio
*bio
)
110 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
111 return atomic_sub_return(1, segments
) & 0xffff;
114 static inline void raid5_inc_bi_active_stripes(struct bio
*bio
)
116 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
117 atomic_inc(segments
);
120 static inline void raid5_set_bi_processed_stripes(struct bio
*bio
,
123 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
127 old
= atomic_read(segments
);
128 new = (old
& 0xffff) | (cnt
<< 16);
129 } while (atomic_cmpxchg(segments
, old
, new) != old
);
132 static inline void raid5_set_bi_stripes(struct bio
*bio
, unsigned int cnt
)
134 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
135 atomic_set(segments
, cnt
);
138 /* Find first data disk in a raid6 stripe */
139 static inline int raid6_d0(struct stripe_head
*sh
)
142 /* ddf always start from first device */
144 /* md starts just after Q block */
145 if (sh
->qd_idx
== sh
->disks
- 1)
148 return sh
->qd_idx
+ 1;
150 static inline int raid6_next_disk(int disk
, int raid_disks
)
153 return (disk
< raid_disks
) ? disk
: 0;
156 /* When walking through the disks in a raid5, starting at raid6_d0,
157 * We need to map each disk to a 'slot', where the data disks are slot
158 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
159 * is raid_disks-1. This help does that mapping.
161 static int raid6_idx_to_slot(int idx
, struct stripe_head
*sh
,
162 int *count
, int syndrome_disks
)
168 if (idx
== sh
->pd_idx
)
169 return syndrome_disks
;
170 if (idx
== sh
->qd_idx
)
171 return syndrome_disks
+ 1;
177 static void return_io(struct bio
*return_bi
)
179 struct bio
*bi
= return_bi
;
182 return_bi
= bi
->bi_next
;
190 static void print_raid5_conf (struct r5conf
*conf
);
192 static int stripe_operations_active(struct stripe_head
*sh
)
194 return sh
->check_state
|| sh
->reconstruct_state
||
195 test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
) ||
196 test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
199 static void do_release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
)
201 BUG_ON(!list_empty(&sh
->lru
));
202 BUG_ON(atomic_read(&conf
->active_stripes
)==0);
203 if (test_bit(STRIPE_HANDLE
, &sh
->state
)) {
204 if (test_bit(STRIPE_DELAYED
, &sh
->state
) &&
205 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
206 list_add_tail(&sh
->lru
, &conf
->delayed_list
);
207 else if (test_bit(STRIPE_BIT_DELAY
, &sh
->state
) &&
208 sh
->bm_seq
- conf
->seq_write
> 0)
209 list_add_tail(&sh
->lru
, &conf
->bitmap_list
);
211 clear_bit(STRIPE_DELAYED
, &sh
->state
);
212 clear_bit(STRIPE_BIT_DELAY
, &sh
->state
);
213 list_add_tail(&sh
->lru
, &conf
->handle_list
);
215 md_wakeup_thread(conf
->mddev
->thread
);
217 BUG_ON(stripe_operations_active(sh
));
218 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
219 if (atomic_dec_return(&conf
->preread_active_stripes
)
221 md_wakeup_thread(conf
->mddev
->thread
);
222 atomic_dec(&conf
->active_stripes
);
223 if (!test_bit(STRIPE_EXPANDING
, &sh
->state
)) {
224 list_add_tail(&sh
->lru
, &conf
->inactive_list
);
225 wake_up(&conf
->wait_for_stripe
);
226 if (conf
->retry_read_aligned
)
227 md_wakeup_thread(conf
->mddev
->thread
);
232 static void __release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
)
234 if (atomic_dec_and_test(&sh
->count
))
235 do_release_stripe(conf
, sh
);
238 static void release_stripe(struct stripe_head
*sh
)
240 struct r5conf
*conf
= sh
->raid_conf
;
243 local_irq_save(flags
);
244 if (atomic_dec_and_lock(&sh
->count
, &conf
->device_lock
)) {
245 do_release_stripe(conf
, sh
);
246 spin_unlock(&conf
->device_lock
);
248 local_irq_restore(flags
);
251 static inline void remove_hash(struct stripe_head
*sh
)
253 pr_debug("remove_hash(), stripe %llu\n",
254 (unsigned long long)sh
->sector
);
256 hlist_del_init(&sh
->hash
);
259 static inline void insert_hash(struct r5conf
*conf
, struct stripe_head
*sh
)
261 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
263 pr_debug("insert_hash(), stripe %llu\n",
264 (unsigned long long)sh
->sector
);
266 hlist_add_head(&sh
->hash
, hp
);
270 /* find an idle stripe, make sure it is unhashed, and return it. */
271 static struct stripe_head
*get_free_stripe(struct r5conf
*conf
)
273 struct stripe_head
*sh
= NULL
;
274 struct list_head
*first
;
276 if (list_empty(&conf
->inactive_list
))
278 first
= conf
->inactive_list
.next
;
279 sh
= list_entry(first
, struct stripe_head
, lru
);
280 list_del_init(first
);
282 atomic_inc(&conf
->active_stripes
);
287 static void shrink_buffers(struct stripe_head
*sh
)
291 int num
= sh
->raid_conf
->pool_size
;
293 for (i
= 0; i
< num
; i
++) {
297 sh
->dev
[i
].page
= NULL
;
302 static int grow_buffers(struct stripe_head
*sh
)
305 int num
= sh
->raid_conf
->pool_size
;
307 for (i
= 0; i
< num
; i
++) {
310 if (!(page
= alloc_page(GFP_KERNEL
))) {
313 sh
->dev
[i
].page
= page
;
318 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
);
319 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
320 struct stripe_head
*sh
);
322 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int previous
)
324 struct r5conf
*conf
= sh
->raid_conf
;
327 BUG_ON(atomic_read(&sh
->count
) != 0);
328 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
329 BUG_ON(stripe_operations_active(sh
));
331 pr_debug("init_stripe called, stripe %llu\n",
332 (unsigned long long)sh
->sector
);
336 sh
->generation
= conf
->generation
- previous
;
337 sh
->disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
339 stripe_set_idx(sector
, conf
, previous
, sh
);
343 for (i
= sh
->disks
; i
--; ) {
344 struct r5dev
*dev
= &sh
->dev
[i
];
346 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
347 test_bit(R5_LOCKED
, &dev
->flags
)) {
348 printk(KERN_ERR
"sector=%llx i=%d %p %p %p %p %d\n",
349 (unsigned long long)sh
->sector
, i
, dev
->toread
,
350 dev
->read
, dev
->towrite
, dev
->written
,
351 test_bit(R5_LOCKED
, &dev
->flags
));
355 raid5_build_block(sh
, i
, previous
);
357 insert_hash(conf
, sh
);
360 static struct stripe_head
*__find_stripe(struct r5conf
*conf
, sector_t sector
,
363 struct stripe_head
*sh
;
364 struct hlist_node
*hn
;
366 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
367 hlist_for_each_entry(sh
, hn
, stripe_hash(conf
, sector
), hash
)
368 if (sh
->sector
== sector
&& sh
->generation
== generation
)
370 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
375 * Need to check if array has failed when deciding whether to:
377 * - remove non-faulty devices
380 * This determination is simple when no reshape is happening.
381 * However if there is a reshape, we need to carefully check
382 * both the before and after sections.
383 * This is because some failed devices may only affect one
384 * of the two sections, and some non-in_sync devices may
385 * be insync in the section most affected by failed devices.
387 static int calc_degraded(struct r5conf
*conf
)
389 int degraded
, degraded2
;
394 for (i
= 0; i
< conf
->previous_raid_disks
; i
++) {
395 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
396 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
397 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
398 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
400 else if (test_bit(In_sync
, &rdev
->flags
))
403 /* not in-sync or faulty.
404 * If the reshape increases the number of devices,
405 * this is being recovered by the reshape, so
406 * this 'previous' section is not in_sync.
407 * If the number of devices is being reduced however,
408 * the device can only be part of the array if
409 * we are reverting a reshape, so this section will
412 if (conf
->raid_disks
>= conf
->previous_raid_disks
)
416 if (conf
->raid_disks
== conf
->previous_raid_disks
)
420 for (i
= 0; i
< conf
->raid_disks
; i
++) {
421 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
422 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
423 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
424 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
426 else if (test_bit(In_sync
, &rdev
->flags
))
429 /* not in-sync or faulty.
430 * If reshape increases the number of devices, this
431 * section has already been recovered, else it
432 * almost certainly hasn't.
434 if (conf
->raid_disks
<= conf
->previous_raid_disks
)
438 if (degraded2
> degraded
)
443 static int has_failed(struct r5conf
*conf
)
447 if (conf
->mddev
->reshape_position
== MaxSector
)
448 return conf
->mddev
->degraded
> conf
->max_degraded
;
450 degraded
= calc_degraded(conf
);
451 if (degraded
> conf
->max_degraded
)
456 static struct stripe_head
*
457 get_active_stripe(struct r5conf
*conf
, sector_t sector
,
458 int previous
, int noblock
, int noquiesce
)
460 struct stripe_head
*sh
;
462 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
464 spin_lock_irq(&conf
->device_lock
);
467 wait_event_lock_irq(conf
->wait_for_stripe
,
468 conf
->quiesce
== 0 || noquiesce
,
469 conf
->device_lock
, /* nothing */);
470 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
472 if (!conf
->inactive_blocked
)
473 sh
= get_free_stripe(conf
);
474 if (noblock
&& sh
== NULL
)
477 conf
->inactive_blocked
= 1;
478 wait_event_lock_irq(conf
->wait_for_stripe
,
479 !list_empty(&conf
->inactive_list
) &&
480 (atomic_read(&conf
->active_stripes
)
481 < (conf
->max_nr_stripes
*3/4)
482 || !conf
->inactive_blocked
),
485 conf
->inactive_blocked
= 0;
487 init_stripe(sh
, sector
, previous
);
489 if (atomic_read(&sh
->count
)) {
490 BUG_ON(!list_empty(&sh
->lru
)
491 && !test_bit(STRIPE_EXPANDING
, &sh
->state
)
492 && !test_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
));
494 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
495 atomic_inc(&conf
->active_stripes
);
496 if (list_empty(&sh
->lru
) &&
497 !test_bit(STRIPE_EXPANDING
, &sh
->state
))
499 list_del_init(&sh
->lru
);
502 } while (sh
== NULL
);
505 atomic_inc(&sh
->count
);
507 spin_unlock_irq(&conf
->device_lock
);
511 /* Determine if 'data_offset' or 'new_data_offset' should be used
512 * in this stripe_head.
514 static int use_new_offset(struct r5conf
*conf
, struct stripe_head
*sh
)
516 sector_t progress
= conf
->reshape_progress
;
517 /* Need a memory barrier to make sure we see the value
518 * of conf->generation, or ->data_offset that was set before
519 * reshape_progress was updated.
522 if (progress
== MaxSector
)
524 if (sh
->generation
== conf
->generation
- 1)
526 /* We are in a reshape, and this is a new-generation stripe,
527 * so use new_data_offset.
533 raid5_end_read_request(struct bio
*bi
, int error
);
535 raid5_end_write_request(struct bio
*bi
, int error
);
537 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
539 struct r5conf
*conf
= sh
->raid_conf
;
540 int i
, disks
= sh
->disks
;
544 for (i
= disks
; i
--; ) {
546 int replace_only
= 0;
547 struct bio
*bi
, *rbi
;
548 struct md_rdev
*rdev
, *rrdev
= NULL
;
549 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
550 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
554 if (test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
556 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
558 else if (test_and_clear_bit(R5_WantReplace
,
559 &sh
->dev
[i
].flags
)) {
564 if (test_and_clear_bit(R5_SyncIO
, &sh
->dev
[i
].flags
))
567 bi
= &sh
->dev
[i
].req
;
568 rbi
= &sh
->dev
[i
].rreq
; /* For writing to replacement */
573 bi
->bi_end_io
= raid5_end_write_request
;
574 rbi
->bi_end_io
= raid5_end_write_request
;
576 bi
->bi_end_io
= raid5_end_read_request
;
579 rrdev
= rcu_dereference(conf
->disks
[i
].replacement
);
580 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
581 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
590 /* We raced and saw duplicates */
593 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
) && rrdev
)
598 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
601 atomic_inc(&rdev
->nr_pending
);
602 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
605 atomic_inc(&rrdev
->nr_pending
);
608 /* We have already checked bad blocks for reads. Now
609 * need to check for writes. We never accept write errors
610 * on the replacement, so we don't to check rrdev.
612 while ((rw
& WRITE
) && rdev
&&
613 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
616 int bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
617 &first_bad
, &bad_sectors
);
622 set_bit(BlockedBadBlocks
, &rdev
->flags
);
623 if (!conf
->mddev
->external
&&
624 conf
->mddev
->flags
) {
625 /* It is very unlikely, but we might
626 * still need to write out the
627 * bad block log - better give it
629 md_check_recovery(conf
->mddev
);
632 * Because md_wait_for_blocked_rdev
633 * will dec nr_pending, we must
634 * increment it first.
636 atomic_inc(&rdev
->nr_pending
);
637 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
639 /* Acknowledged bad block - skip the write */
640 rdev_dec_pending(rdev
, conf
->mddev
);
646 if (s
->syncing
|| s
->expanding
|| s
->expanded
648 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
650 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
652 bi
->bi_bdev
= rdev
->bdev
;
653 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
654 __func__
, (unsigned long long)sh
->sector
,
656 atomic_inc(&sh
->count
);
657 if (use_new_offset(conf
, sh
))
658 bi
->bi_sector
= (sh
->sector
659 + rdev
->new_data_offset
);
661 bi
->bi_sector
= (sh
->sector
662 + rdev
->data_offset
);
663 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
664 bi
->bi_rw
|= REQ_FLUSH
;
666 bi
->bi_flags
= 1 << BIO_UPTODATE
;
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
;
673 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
674 generic_make_request(bi
);
677 if (s
->syncing
|| s
->expanding
|| s
->expanded
679 md_sync_acct(rrdev
->bdev
, STRIPE_SECTORS
);
681 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
683 rbi
->bi_bdev
= rrdev
->bdev
;
684 pr_debug("%s: for %llu schedule op %ld on "
685 "replacement disc %d\n",
686 __func__
, (unsigned long long)sh
->sector
,
688 atomic_inc(&sh
->count
);
689 if (use_new_offset(conf
, sh
))
690 rbi
->bi_sector
= (sh
->sector
691 + rrdev
->new_data_offset
);
693 rbi
->bi_sector
= (sh
->sector
694 + rrdev
->data_offset
);
695 rbi
->bi_flags
= 1 << BIO_UPTODATE
;
697 rbi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
698 rbi
->bi_io_vec
[0].bv_offset
= 0;
699 rbi
->bi_size
= STRIPE_SIZE
;
701 generic_make_request(rbi
);
703 if (!rdev
&& !rrdev
) {
705 set_bit(STRIPE_DEGRADED
, &sh
->state
);
706 pr_debug("skip op %ld on disc %d for sector %llu\n",
707 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
708 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
709 set_bit(STRIPE_HANDLE
, &sh
->state
);
714 static struct dma_async_tx_descriptor
*
715 async_copy_data(int frombio
, struct bio
*bio
, struct page
*page
,
716 sector_t sector
, struct dma_async_tx_descriptor
*tx
)
719 struct page
*bio_page
;
722 struct async_submit_ctl submit
;
723 enum async_tx_flags flags
= 0;
725 if (bio
->bi_sector
>= sector
)
726 page_offset
= (signed)(bio
->bi_sector
- sector
) * 512;
728 page_offset
= (signed)(sector
- bio
->bi_sector
) * -512;
731 flags
|= ASYNC_TX_FENCE
;
732 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
734 bio_for_each_segment(bvl
, bio
, i
) {
735 int len
= bvl
->bv_len
;
739 if (page_offset
< 0) {
740 b_offset
= -page_offset
;
741 page_offset
+= b_offset
;
745 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
746 clen
= STRIPE_SIZE
- page_offset
;
751 b_offset
+= bvl
->bv_offset
;
752 bio_page
= bvl
->bv_page
;
754 tx
= async_memcpy(page
, bio_page
, page_offset
,
755 b_offset
, clen
, &submit
);
757 tx
= async_memcpy(bio_page
, page
, b_offset
,
758 page_offset
, clen
, &submit
);
760 /* chain the operations */
761 submit
.depend_tx
= tx
;
763 if (clen
< len
) /* hit end of page */
771 static void ops_complete_biofill(void *stripe_head_ref
)
773 struct stripe_head
*sh
= stripe_head_ref
;
774 struct bio
*return_bi
= NULL
;
777 pr_debug("%s: stripe %llu\n", __func__
,
778 (unsigned long long)sh
->sector
);
780 /* clear completed biofills */
781 for (i
= sh
->disks
; i
--; ) {
782 struct r5dev
*dev
= &sh
->dev
[i
];
784 /* acknowledge completion of a biofill operation */
785 /* and check if we need to reply to a read request,
786 * new R5_Wantfill requests are held off until
787 * !STRIPE_BIOFILL_RUN
789 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
790 struct bio
*rbi
, *rbi2
;
795 while (rbi
&& rbi
->bi_sector
<
796 dev
->sector
+ STRIPE_SECTORS
) {
797 rbi2
= r5_next_bio(rbi
, dev
->sector
);
798 if (!raid5_dec_bi_active_stripes(rbi
)) {
799 rbi
->bi_next
= return_bi
;
806 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
808 return_io(return_bi
);
810 set_bit(STRIPE_HANDLE
, &sh
->state
);
814 static void ops_run_biofill(struct stripe_head
*sh
)
816 struct dma_async_tx_descriptor
*tx
= NULL
;
817 struct async_submit_ctl submit
;
820 pr_debug("%s: stripe %llu\n", __func__
,
821 (unsigned long long)sh
->sector
);
823 for (i
= sh
->disks
; i
--; ) {
824 struct r5dev
*dev
= &sh
->dev
[i
];
825 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
827 spin_lock_irq(&sh
->stripe_lock
);
828 dev
->read
= rbi
= dev
->toread
;
830 spin_unlock_irq(&sh
->stripe_lock
);
831 while (rbi
&& rbi
->bi_sector
<
832 dev
->sector
+ STRIPE_SECTORS
) {
833 tx
= async_copy_data(0, rbi
, dev
->page
,
835 rbi
= r5_next_bio(rbi
, dev
->sector
);
840 atomic_inc(&sh
->count
);
841 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
842 async_trigger_callback(&submit
);
845 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
852 tgt
= &sh
->dev
[target
];
853 set_bit(R5_UPTODATE
, &tgt
->flags
);
854 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
855 clear_bit(R5_Wantcompute
, &tgt
->flags
);
858 static void ops_complete_compute(void *stripe_head_ref
)
860 struct stripe_head
*sh
= stripe_head_ref
;
862 pr_debug("%s: stripe %llu\n", __func__
,
863 (unsigned long long)sh
->sector
);
865 /* mark the computed target(s) as uptodate */
866 mark_target_uptodate(sh
, sh
->ops
.target
);
867 mark_target_uptodate(sh
, sh
->ops
.target2
);
869 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
870 if (sh
->check_state
== check_state_compute_run
)
871 sh
->check_state
= check_state_compute_result
;
872 set_bit(STRIPE_HANDLE
, &sh
->state
);
876 /* return a pointer to the address conversion region of the scribble buffer */
877 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
878 struct raid5_percpu
*percpu
)
880 return percpu
->scribble
+ sizeof(struct page
*) * (sh
->disks
+ 2);
883 static struct dma_async_tx_descriptor
*
884 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
886 int disks
= sh
->disks
;
887 struct page
**xor_srcs
= percpu
->scribble
;
888 int target
= sh
->ops
.target
;
889 struct r5dev
*tgt
= &sh
->dev
[target
];
890 struct page
*xor_dest
= tgt
->page
;
892 struct dma_async_tx_descriptor
*tx
;
893 struct async_submit_ctl submit
;
896 pr_debug("%s: stripe %llu block: %d\n",
897 __func__
, (unsigned long long)sh
->sector
, target
);
898 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
900 for (i
= disks
; i
--; )
902 xor_srcs
[count
++] = sh
->dev
[i
].page
;
904 atomic_inc(&sh
->count
);
906 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
907 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
));
908 if (unlikely(count
== 1))
909 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
911 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
916 /* set_syndrome_sources - populate source buffers for gen_syndrome
917 * @srcs - (struct page *) array of size sh->disks
918 * @sh - stripe_head to parse
920 * Populates srcs in proper layout order for the stripe and returns the
921 * 'count' of sources to be used in a call to async_gen_syndrome. The P
922 * destination buffer is recorded in srcs[count] and the Q destination
923 * is recorded in srcs[count+1]].
925 static int set_syndrome_sources(struct page
**srcs
, struct stripe_head
*sh
)
927 int disks
= sh
->disks
;
928 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
929 int d0_idx
= raid6_d0(sh
);
933 for (i
= 0; i
< disks
; i
++)
939 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
941 srcs
[slot
] = sh
->dev
[i
].page
;
942 i
= raid6_next_disk(i
, disks
);
943 } while (i
!= d0_idx
);
945 return syndrome_disks
;
948 static struct dma_async_tx_descriptor
*
949 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
951 int disks
= sh
->disks
;
952 struct page
**blocks
= percpu
->scribble
;
954 int qd_idx
= sh
->qd_idx
;
955 struct dma_async_tx_descriptor
*tx
;
956 struct async_submit_ctl submit
;
962 if (sh
->ops
.target
< 0)
963 target
= sh
->ops
.target2
;
964 else if (sh
->ops
.target2
< 0)
965 target
= sh
->ops
.target
;
967 /* we should only have one valid target */
970 pr_debug("%s: stripe %llu block: %d\n",
971 __func__
, (unsigned long long)sh
->sector
, target
);
973 tgt
= &sh
->dev
[target
];
974 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
977 atomic_inc(&sh
->count
);
979 if (target
== qd_idx
) {
980 count
= set_syndrome_sources(blocks
, sh
);
981 blocks
[count
] = NULL
; /* regenerating p is not necessary */
982 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
983 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
984 ops_complete_compute
, sh
,
985 to_addr_conv(sh
, percpu
));
986 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
988 /* Compute any data- or p-drive using XOR */
990 for (i
= disks
; i
-- ; ) {
991 if (i
== target
|| i
== qd_idx
)
993 blocks
[count
++] = sh
->dev
[i
].page
;
996 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
997 NULL
, ops_complete_compute
, sh
,
998 to_addr_conv(sh
, percpu
));
999 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
1005 static struct dma_async_tx_descriptor
*
1006 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1008 int i
, count
, disks
= sh
->disks
;
1009 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1010 int d0_idx
= raid6_d0(sh
);
1011 int faila
= -1, failb
= -1;
1012 int target
= sh
->ops
.target
;
1013 int target2
= sh
->ops
.target2
;
1014 struct r5dev
*tgt
= &sh
->dev
[target
];
1015 struct r5dev
*tgt2
= &sh
->dev
[target2
];
1016 struct dma_async_tx_descriptor
*tx
;
1017 struct page
**blocks
= percpu
->scribble
;
1018 struct async_submit_ctl submit
;
1020 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1021 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
1022 BUG_ON(target
< 0 || target2
< 0);
1023 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1024 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
1026 /* we need to open-code set_syndrome_sources to handle the
1027 * slot number conversion for 'faila' and 'failb'
1029 for (i
= 0; i
< disks
; i
++)
1034 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1036 blocks
[slot
] = sh
->dev
[i
].page
;
1042 i
= raid6_next_disk(i
, disks
);
1043 } while (i
!= d0_idx
);
1045 BUG_ON(faila
== failb
);
1048 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1049 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1051 atomic_inc(&sh
->count
);
1053 if (failb
== syndrome_disks
+1) {
1054 /* Q disk is one of the missing disks */
1055 if (faila
== syndrome_disks
) {
1056 /* Missing P+Q, just recompute */
1057 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1058 ops_complete_compute
, sh
,
1059 to_addr_conv(sh
, percpu
));
1060 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
1061 STRIPE_SIZE
, &submit
);
1065 int qd_idx
= sh
->qd_idx
;
1067 /* Missing D+Q: recompute D from P, then recompute Q */
1068 if (target
== qd_idx
)
1069 data_target
= target2
;
1071 data_target
= target
;
1074 for (i
= disks
; i
-- ; ) {
1075 if (i
== data_target
|| i
== qd_idx
)
1077 blocks
[count
++] = sh
->dev
[i
].page
;
1079 dest
= sh
->dev
[data_target
].page
;
1080 init_async_submit(&submit
,
1081 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1083 to_addr_conv(sh
, percpu
));
1084 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
1087 count
= set_syndrome_sources(blocks
, sh
);
1088 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1089 ops_complete_compute
, sh
,
1090 to_addr_conv(sh
, percpu
));
1091 return async_gen_syndrome(blocks
, 0, count
+2,
1092 STRIPE_SIZE
, &submit
);
1095 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1096 ops_complete_compute
, sh
,
1097 to_addr_conv(sh
, percpu
));
1098 if (failb
== syndrome_disks
) {
1099 /* We're missing D+P. */
1100 return async_raid6_datap_recov(syndrome_disks
+2,
1104 /* We're missing D+D. */
1105 return async_raid6_2data_recov(syndrome_disks
+2,
1106 STRIPE_SIZE
, faila
, failb
,
1113 static void ops_complete_prexor(void *stripe_head_ref
)
1115 struct stripe_head
*sh
= stripe_head_ref
;
1117 pr_debug("%s: stripe %llu\n", __func__
,
1118 (unsigned long long)sh
->sector
);
1121 static struct dma_async_tx_descriptor
*
1122 ops_run_prexor(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1123 struct dma_async_tx_descriptor
*tx
)
1125 int disks
= sh
->disks
;
1126 struct page
**xor_srcs
= percpu
->scribble
;
1127 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1128 struct async_submit_ctl submit
;
1130 /* existing parity data subtracted */
1131 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1133 pr_debug("%s: stripe %llu\n", __func__
,
1134 (unsigned long long)sh
->sector
);
1136 for (i
= disks
; i
--; ) {
1137 struct r5dev
*dev
= &sh
->dev
[i
];
1138 /* Only process blocks that are known to be uptodate */
1139 if (test_bit(R5_Wantdrain
, &dev
->flags
))
1140 xor_srcs
[count
++] = dev
->page
;
1143 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1144 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
));
1145 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1150 static struct dma_async_tx_descriptor
*
1151 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1153 int disks
= sh
->disks
;
1156 pr_debug("%s: stripe %llu\n", __func__
,
1157 (unsigned long long)sh
->sector
);
1159 for (i
= disks
; i
--; ) {
1160 struct r5dev
*dev
= &sh
->dev
[i
];
1163 if (test_and_clear_bit(R5_Wantdrain
, &dev
->flags
)) {
1166 spin_lock_irq(&sh
->stripe_lock
);
1167 chosen
= dev
->towrite
;
1168 dev
->towrite
= NULL
;
1169 BUG_ON(dev
->written
);
1170 wbi
= dev
->written
= chosen
;
1171 spin_unlock_irq(&sh
->stripe_lock
);
1173 while (wbi
&& wbi
->bi_sector
<
1174 dev
->sector
+ STRIPE_SECTORS
) {
1175 if (wbi
->bi_rw
& REQ_FUA
)
1176 set_bit(R5_WantFUA
, &dev
->flags
);
1177 if (wbi
->bi_rw
& REQ_SYNC
)
1178 set_bit(R5_SyncIO
, &dev
->flags
);
1179 if (wbi
->bi_rw
& REQ_DISCARD
)
1180 set_bit(R5_Discard
, &dev
->flags
);
1182 tx
= async_copy_data(1, wbi
, dev
->page
,
1184 wbi
= r5_next_bio(wbi
, dev
->sector
);
1192 static void ops_complete_reconstruct(void *stripe_head_ref
)
1194 struct stripe_head
*sh
= stripe_head_ref
;
1195 int disks
= sh
->disks
;
1196 int pd_idx
= sh
->pd_idx
;
1197 int qd_idx
= sh
->qd_idx
;
1199 bool fua
= false, sync
= false, discard
= false;
1201 pr_debug("%s: stripe %llu\n", __func__
,
1202 (unsigned long long)sh
->sector
);
1204 for (i
= disks
; i
--; ) {
1205 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1206 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
1207 discard
|= test_bit(R5_Discard
, &sh
->dev
[i
].flags
);
1210 for (i
= disks
; i
--; ) {
1211 struct r5dev
*dev
= &sh
->dev
[i
];
1213 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1215 set_bit(R5_UPTODATE
, &dev
->flags
);
1217 set_bit(R5_WantFUA
, &dev
->flags
);
1219 set_bit(R5_SyncIO
, &dev
->flags
);
1223 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1224 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1225 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1226 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1228 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1229 sh
->reconstruct_state
= reconstruct_state_result
;
1232 set_bit(STRIPE_HANDLE
, &sh
->state
);
1237 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1238 struct dma_async_tx_descriptor
*tx
)
1240 int disks
= sh
->disks
;
1241 struct page
**xor_srcs
= percpu
->scribble
;
1242 struct async_submit_ctl submit
;
1243 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1244 struct page
*xor_dest
;
1246 unsigned long flags
;
1248 pr_debug("%s: stripe %llu\n", __func__
,
1249 (unsigned long long)sh
->sector
);
1251 for (i
= 0; i
< sh
->disks
; i
++) {
1254 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1257 if (i
>= sh
->disks
) {
1258 atomic_inc(&sh
->count
);
1259 set_bit(R5_Discard
, &sh
->dev
[pd_idx
].flags
);
1260 ops_complete_reconstruct(sh
);
1263 /* check if prexor is active which means only process blocks
1264 * that are part of a read-modify-write (written)
1266 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1268 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1269 for (i
= disks
; i
--; ) {
1270 struct r5dev
*dev
= &sh
->dev
[i
];
1272 xor_srcs
[count
++] = dev
->page
;
1275 xor_dest
= sh
->dev
[pd_idx
].page
;
1276 for (i
= disks
; i
--; ) {
1277 struct r5dev
*dev
= &sh
->dev
[i
];
1279 xor_srcs
[count
++] = dev
->page
;
1283 /* 1/ if we prexor'd then the dest is reused as a source
1284 * 2/ if we did not prexor then we are redoing the parity
1285 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1286 * for the synchronous xor case
1288 flags
= ASYNC_TX_ACK
|
1289 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1291 atomic_inc(&sh
->count
);
1293 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, sh
,
1294 to_addr_conv(sh
, percpu
));
1295 if (unlikely(count
== 1))
1296 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1298 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1302 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1303 struct dma_async_tx_descriptor
*tx
)
1305 struct async_submit_ctl submit
;
1306 struct page
**blocks
= percpu
->scribble
;
1309 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1311 for (i
= 0; i
< sh
->disks
; i
++) {
1312 if (sh
->pd_idx
== i
|| sh
->qd_idx
== i
)
1314 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1317 if (i
>= sh
->disks
) {
1318 atomic_inc(&sh
->count
);
1319 set_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
1320 set_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
1321 ops_complete_reconstruct(sh
);
1325 count
= set_syndrome_sources(blocks
, sh
);
1327 atomic_inc(&sh
->count
);
1329 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_reconstruct
,
1330 sh
, to_addr_conv(sh
, percpu
));
1331 async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1334 static void ops_complete_check(void *stripe_head_ref
)
1336 struct stripe_head
*sh
= stripe_head_ref
;
1338 pr_debug("%s: stripe %llu\n", __func__
,
1339 (unsigned long long)sh
->sector
);
1341 sh
->check_state
= check_state_check_result
;
1342 set_bit(STRIPE_HANDLE
, &sh
->state
);
1346 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1348 int disks
= sh
->disks
;
1349 int pd_idx
= sh
->pd_idx
;
1350 int qd_idx
= sh
->qd_idx
;
1351 struct page
*xor_dest
;
1352 struct page
**xor_srcs
= percpu
->scribble
;
1353 struct dma_async_tx_descriptor
*tx
;
1354 struct async_submit_ctl submit
;
1358 pr_debug("%s: stripe %llu\n", __func__
,
1359 (unsigned long long)sh
->sector
);
1362 xor_dest
= sh
->dev
[pd_idx
].page
;
1363 xor_srcs
[count
++] = xor_dest
;
1364 for (i
= disks
; i
--; ) {
1365 if (i
== pd_idx
|| i
== qd_idx
)
1367 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1370 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1371 to_addr_conv(sh
, percpu
));
1372 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1373 &sh
->ops
.zero_sum_result
, &submit
);
1375 atomic_inc(&sh
->count
);
1376 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1377 tx
= async_trigger_callback(&submit
);
1380 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1382 struct page
**srcs
= percpu
->scribble
;
1383 struct async_submit_ctl submit
;
1386 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1387 (unsigned long long)sh
->sector
, checkp
);
1389 count
= set_syndrome_sources(srcs
, sh
);
1393 atomic_inc(&sh
->count
);
1394 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1395 sh
, to_addr_conv(sh
, percpu
));
1396 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1397 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1400 static void __raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1402 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1403 struct dma_async_tx_descriptor
*tx
= NULL
;
1404 struct r5conf
*conf
= sh
->raid_conf
;
1405 int level
= conf
->level
;
1406 struct raid5_percpu
*percpu
;
1410 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1411 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1412 ops_run_biofill(sh
);
1416 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1418 tx
= ops_run_compute5(sh
, percpu
);
1420 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
1421 tx
= ops_run_compute6_1(sh
, percpu
);
1423 tx
= ops_run_compute6_2(sh
, percpu
);
1425 /* terminate the chain if reconstruct is not set to be run */
1426 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
1430 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
))
1431 tx
= ops_run_prexor(sh
, percpu
, tx
);
1433 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
1434 tx
= ops_run_biodrain(sh
, tx
);
1438 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
1440 ops_run_reconstruct5(sh
, percpu
, tx
);
1442 ops_run_reconstruct6(sh
, percpu
, tx
);
1445 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
1446 if (sh
->check_state
== check_state_run
)
1447 ops_run_check_p(sh
, percpu
);
1448 else if (sh
->check_state
== check_state_run_q
)
1449 ops_run_check_pq(sh
, percpu
, 0);
1450 else if (sh
->check_state
== check_state_run_pq
)
1451 ops_run_check_pq(sh
, percpu
, 1);
1457 for (i
= disks
; i
--; ) {
1458 struct r5dev
*dev
= &sh
->dev
[i
];
1459 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
1460 wake_up(&sh
->raid_conf
->wait_for_overlap
);
1465 #ifdef CONFIG_MULTICORE_RAID456
1466 static void async_run_ops(void *param
, async_cookie_t cookie
)
1468 struct stripe_head
*sh
= param
;
1469 unsigned long ops_request
= sh
->ops
.request
;
1471 clear_bit_unlock(STRIPE_OPS_REQ_PENDING
, &sh
->state
);
1472 wake_up(&sh
->ops
.wait_for_ops
);
1474 __raid_run_ops(sh
, ops_request
);
1478 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1480 /* since handle_stripe can be called outside of raid5d context
1481 * we need to ensure sh->ops.request is de-staged before another
1484 wait_event(sh
->ops
.wait_for_ops
,
1485 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING
, &sh
->state
));
1486 sh
->ops
.request
= ops_request
;
1488 atomic_inc(&sh
->count
);
1489 async_schedule(async_run_ops
, sh
);
1492 #define raid_run_ops __raid_run_ops
1495 static int grow_one_stripe(struct r5conf
*conf
)
1497 struct stripe_head
*sh
;
1498 sh
= kmem_cache_zalloc(conf
->slab_cache
, GFP_KERNEL
);
1502 sh
->raid_conf
= conf
;
1503 #ifdef CONFIG_MULTICORE_RAID456
1504 init_waitqueue_head(&sh
->ops
.wait_for_ops
);
1507 spin_lock_init(&sh
->stripe_lock
);
1509 if (grow_buffers(sh
)) {
1511 kmem_cache_free(conf
->slab_cache
, sh
);
1514 /* we just created an active stripe so... */
1515 atomic_set(&sh
->count
, 1);
1516 atomic_inc(&conf
->active_stripes
);
1517 INIT_LIST_HEAD(&sh
->lru
);
1522 static int grow_stripes(struct r5conf
*conf
, int num
)
1524 struct kmem_cache
*sc
;
1525 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
1527 if (conf
->mddev
->gendisk
)
1528 sprintf(conf
->cache_name
[0],
1529 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
1531 sprintf(conf
->cache_name
[0],
1532 "raid%d-%p", conf
->level
, conf
->mddev
);
1533 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
1535 conf
->active_name
= 0;
1536 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
1537 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
1541 conf
->slab_cache
= sc
;
1542 conf
->pool_size
= devs
;
1544 if (!grow_one_stripe(conf
))
1550 * scribble_len - return the required size of the scribble region
1551 * @num - total number of disks in the array
1553 * The size must be enough to contain:
1554 * 1/ a struct page pointer for each device in the array +2
1555 * 2/ room to convert each entry in (1) to its corresponding dma
1556 * (dma_map_page()) or page (page_address()) address.
1558 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1559 * calculate over all devices (not just the data blocks), using zeros in place
1560 * of the P and Q blocks.
1562 static size_t scribble_len(int num
)
1566 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
1571 static int resize_stripes(struct r5conf
*conf
, int newsize
)
1573 /* Make all the stripes able to hold 'newsize' devices.
1574 * New slots in each stripe get 'page' set to a new page.
1576 * This happens in stages:
1577 * 1/ create a new kmem_cache and allocate the required number of
1579 * 2/ gather all the old stripe_heads and transfer the pages across
1580 * to the new stripe_heads. This will have the side effect of
1581 * freezing the array as once all stripe_heads have been collected,
1582 * no IO will be possible. Old stripe heads are freed once their
1583 * pages have been transferred over, and the old kmem_cache is
1584 * freed when all stripes are done.
1585 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1586 * we simple return a failre status - no need to clean anything up.
1587 * 4/ allocate new pages for the new slots in the new stripe_heads.
1588 * If this fails, we don't bother trying the shrink the
1589 * stripe_heads down again, we just leave them as they are.
1590 * As each stripe_head is processed the new one is released into
1593 * Once step2 is started, we cannot afford to wait for a write,
1594 * so we use GFP_NOIO allocations.
1596 struct stripe_head
*osh
, *nsh
;
1597 LIST_HEAD(newstripes
);
1598 struct disk_info
*ndisks
;
1601 struct kmem_cache
*sc
;
1604 if (newsize
<= conf
->pool_size
)
1605 return 0; /* never bother to shrink */
1607 err
= md_allow_write(conf
->mddev
);
1612 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
1613 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
1618 for (i
= conf
->max_nr_stripes
; i
; i
--) {
1619 nsh
= kmem_cache_zalloc(sc
, GFP_KERNEL
);
1623 nsh
->raid_conf
= conf
;
1624 #ifdef CONFIG_MULTICORE_RAID456
1625 init_waitqueue_head(&nsh
->ops
.wait_for_ops
);
1627 spin_lock_init(&nsh
->stripe_lock
);
1629 list_add(&nsh
->lru
, &newstripes
);
1632 /* didn't get enough, give up */
1633 while (!list_empty(&newstripes
)) {
1634 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1635 list_del(&nsh
->lru
);
1636 kmem_cache_free(sc
, nsh
);
1638 kmem_cache_destroy(sc
);
1641 /* Step 2 - Must use GFP_NOIO now.
1642 * OK, we have enough stripes, start collecting inactive
1643 * stripes and copying them over
1645 list_for_each_entry(nsh
, &newstripes
, lru
) {
1646 spin_lock_irq(&conf
->device_lock
);
1647 wait_event_lock_irq(conf
->wait_for_stripe
,
1648 !list_empty(&conf
->inactive_list
),
1651 osh
= get_free_stripe(conf
);
1652 spin_unlock_irq(&conf
->device_lock
);
1653 atomic_set(&nsh
->count
, 1);
1654 for(i
=0; i
<conf
->pool_size
; i
++)
1655 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
1656 for( ; i
<newsize
; i
++)
1657 nsh
->dev
[i
].page
= NULL
;
1658 kmem_cache_free(conf
->slab_cache
, osh
);
1660 kmem_cache_destroy(conf
->slab_cache
);
1663 * At this point, we are holding all the stripes so the array
1664 * is completely stalled, so now is a good time to resize
1665 * conf->disks and the scribble region
1667 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
1669 for (i
=0; i
<conf
->raid_disks
; i
++)
1670 ndisks
[i
] = conf
->disks
[i
];
1672 conf
->disks
= ndisks
;
1677 conf
->scribble_len
= scribble_len(newsize
);
1678 for_each_present_cpu(cpu
) {
1679 struct raid5_percpu
*percpu
;
1682 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1683 scribble
= kmalloc(conf
->scribble_len
, GFP_NOIO
);
1686 kfree(percpu
->scribble
);
1687 percpu
->scribble
= scribble
;
1695 /* Step 4, return new stripes to service */
1696 while(!list_empty(&newstripes
)) {
1697 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1698 list_del_init(&nsh
->lru
);
1700 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
1701 if (nsh
->dev
[i
].page
== NULL
) {
1702 struct page
*p
= alloc_page(GFP_NOIO
);
1703 nsh
->dev
[i
].page
= p
;
1707 release_stripe(nsh
);
1709 /* critical section pass, GFP_NOIO no longer needed */
1711 conf
->slab_cache
= sc
;
1712 conf
->active_name
= 1-conf
->active_name
;
1713 conf
->pool_size
= newsize
;
1717 static int drop_one_stripe(struct r5conf
*conf
)
1719 struct stripe_head
*sh
;
1721 spin_lock_irq(&conf
->device_lock
);
1722 sh
= get_free_stripe(conf
);
1723 spin_unlock_irq(&conf
->device_lock
);
1726 BUG_ON(atomic_read(&sh
->count
));
1728 kmem_cache_free(conf
->slab_cache
, sh
);
1729 atomic_dec(&conf
->active_stripes
);
1733 static void shrink_stripes(struct r5conf
*conf
)
1735 while (drop_one_stripe(conf
))
1738 if (conf
->slab_cache
)
1739 kmem_cache_destroy(conf
->slab_cache
);
1740 conf
->slab_cache
= NULL
;
1743 static void raid5_end_read_request(struct bio
* bi
, int error
)
1745 struct stripe_head
*sh
= bi
->bi_private
;
1746 struct r5conf
*conf
= sh
->raid_conf
;
1747 int disks
= sh
->disks
, i
;
1748 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1749 char b
[BDEVNAME_SIZE
];
1750 struct md_rdev
*rdev
= NULL
;
1753 for (i
=0 ; i
<disks
; i
++)
1754 if (bi
== &sh
->dev
[i
].req
)
1757 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1758 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1764 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
1765 /* If replacement finished while this request was outstanding,
1766 * 'replacement' might be NULL already.
1767 * In that case it moved down to 'rdev'.
1768 * rdev is not removed until all requests are finished.
1770 rdev
= conf
->disks
[i
].replacement
;
1772 rdev
= conf
->disks
[i
].rdev
;
1774 if (use_new_offset(conf
, sh
))
1775 s
= sh
->sector
+ rdev
->new_data_offset
;
1777 s
= sh
->sector
+ rdev
->data_offset
;
1779 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1780 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
1781 /* Note that this cannot happen on a
1782 * replacement device. We just fail those on
1787 "md/raid:%s: read error corrected"
1788 " (%lu sectors at %llu on %s)\n",
1789 mdname(conf
->mddev
), STRIPE_SECTORS
,
1790 (unsigned long long)s
,
1791 bdevname(rdev
->bdev
, b
));
1792 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
1793 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1794 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1795 } else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
1796 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
1798 if (atomic_read(&rdev
->read_errors
))
1799 atomic_set(&rdev
->read_errors
, 0);
1801 const char *bdn
= bdevname(rdev
->bdev
, b
);
1805 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1806 atomic_inc(&rdev
->read_errors
);
1807 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
1810 "md/raid:%s: read error on replacement device "
1811 "(sector %llu on %s).\n",
1812 mdname(conf
->mddev
),
1813 (unsigned long long)s
,
1815 else if (conf
->mddev
->degraded
>= conf
->max_degraded
) {
1819 "md/raid:%s: read error not correctable "
1820 "(sector %llu on %s).\n",
1821 mdname(conf
->mddev
),
1822 (unsigned long long)s
,
1824 } else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
)) {
1829 "md/raid:%s: read error NOT corrected!! "
1830 "(sector %llu on %s).\n",
1831 mdname(conf
->mddev
),
1832 (unsigned long long)s
,
1834 } else if (atomic_read(&rdev
->read_errors
)
1835 > conf
->max_nr_stripes
)
1837 "md/raid:%s: Too many read errors, failing device %s.\n",
1838 mdname(conf
->mddev
), bdn
);
1842 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
)) {
1843 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1844 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
1846 set_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
1848 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1849 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1851 && test_bit(In_sync
, &rdev
->flags
)
1852 && rdev_set_badblocks(
1853 rdev
, sh
->sector
, STRIPE_SECTORS
, 0)))
1854 md_error(conf
->mddev
, rdev
);
1857 rdev_dec_pending(rdev
, conf
->mddev
);
1858 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1859 set_bit(STRIPE_HANDLE
, &sh
->state
);
1863 static void raid5_end_write_request(struct bio
*bi
, int error
)
1865 struct stripe_head
*sh
= bi
->bi_private
;
1866 struct r5conf
*conf
= sh
->raid_conf
;
1867 int disks
= sh
->disks
, i
;
1868 struct md_rdev
*uninitialized_var(rdev
);
1869 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1872 int replacement
= 0;
1874 for (i
= 0 ; i
< disks
; i
++) {
1875 if (bi
== &sh
->dev
[i
].req
) {
1876 rdev
= conf
->disks
[i
].rdev
;
1879 if (bi
== &sh
->dev
[i
].rreq
) {
1880 rdev
= conf
->disks
[i
].replacement
;
1884 /* rdev was removed and 'replacement'
1885 * replaced it. rdev is not removed
1886 * until all requests are finished.
1888 rdev
= conf
->disks
[i
].rdev
;
1892 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1893 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1902 md_error(conf
->mddev
, rdev
);
1903 else if (is_badblock(rdev
, sh
->sector
,
1905 &first_bad
, &bad_sectors
))
1906 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
1909 set_bit(WriteErrorSeen
, &rdev
->flags
);
1910 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
1911 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1912 set_bit(MD_RECOVERY_NEEDED
,
1913 &rdev
->mddev
->recovery
);
1914 } else if (is_badblock(rdev
, sh
->sector
,
1916 &first_bad
, &bad_sectors
))
1917 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
1919 rdev_dec_pending(rdev
, conf
->mddev
);
1921 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
1922 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1923 set_bit(STRIPE_HANDLE
, &sh
->state
);
1927 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
);
1929 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
1931 struct r5dev
*dev
= &sh
->dev
[i
];
1933 bio_init(&dev
->req
);
1934 dev
->req
.bi_io_vec
= &dev
->vec
;
1936 dev
->req
.bi_max_vecs
++;
1937 dev
->req
.bi_private
= sh
;
1938 dev
->vec
.bv_page
= dev
->page
;
1940 bio_init(&dev
->rreq
);
1941 dev
->rreq
.bi_io_vec
= &dev
->rvec
;
1942 dev
->rreq
.bi_vcnt
++;
1943 dev
->rreq
.bi_max_vecs
++;
1944 dev
->rreq
.bi_private
= sh
;
1945 dev
->rvec
.bv_page
= dev
->page
;
1948 dev
->sector
= compute_blocknr(sh
, i
, previous
);
1951 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1953 char b
[BDEVNAME_SIZE
];
1954 struct r5conf
*conf
= mddev
->private;
1955 unsigned long flags
;
1956 pr_debug("raid456: error called\n");
1958 spin_lock_irqsave(&conf
->device_lock
, flags
);
1959 clear_bit(In_sync
, &rdev
->flags
);
1960 mddev
->degraded
= calc_degraded(conf
);
1961 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1962 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1964 set_bit(Blocked
, &rdev
->flags
);
1965 set_bit(Faulty
, &rdev
->flags
);
1966 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1968 "md/raid:%s: Disk failure on %s, disabling device.\n"
1969 "md/raid:%s: Operation continuing on %d devices.\n",
1971 bdevname(rdev
->bdev
, b
),
1973 conf
->raid_disks
- mddev
->degraded
);
1977 * Input: a 'big' sector number,
1978 * Output: index of the data and parity disk, and the sector # in them.
1980 static sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
1981 int previous
, int *dd_idx
,
1982 struct stripe_head
*sh
)
1984 sector_t stripe
, stripe2
;
1985 sector_t chunk_number
;
1986 unsigned int chunk_offset
;
1989 sector_t new_sector
;
1990 int algorithm
= previous
? conf
->prev_algo
1992 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
1993 : conf
->chunk_sectors
;
1994 int raid_disks
= previous
? conf
->previous_raid_disks
1996 int data_disks
= raid_disks
- conf
->max_degraded
;
1998 /* First compute the information on this sector */
2001 * Compute the chunk number and the sector offset inside the chunk
2003 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
2004 chunk_number
= r_sector
;
2007 * Compute the stripe number
2009 stripe
= chunk_number
;
2010 *dd_idx
= sector_div(stripe
, data_disks
);
2013 * Select the parity disk based on the user selected algorithm.
2015 pd_idx
= qd_idx
= -1;
2016 switch(conf
->level
) {
2018 pd_idx
= data_disks
;
2021 switch (algorithm
) {
2022 case ALGORITHM_LEFT_ASYMMETRIC
:
2023 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2024 if (*dd_idx
>= pd_idx
)
2027 case ALGORITHM_RIGHT_ASYMMETRIC
:
2028 pd_idx
= sector_div(stripe2
, raid_disks
);
2029 if (*dd_idx
>= pd_idx
)
2032 case ALGORITHM_LEFT_SYMMETRIC
:
2033 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2034 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2036 case ALGORITHM_RIGHT_SYMMETRIC
:
2037 pd_idx
= sector_div(stripe2
, raid_disks
);
2038 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2040 case ALGORITHM_PARITY_0
:
2044 case ALGORITHM_PARITY_N
:
2045 pd_idx
= data_disks
;
2053 switch (algorithm
) {
2054 case ALGORITHM_LEFT_ASYMMETRIC
:
2055 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2056 qd_idx
= pd_idx
+ 1;
2057 if (pd_idx
== raid_disks
-1) {
2058 (*dd_idx
)++; /* Q D D D P */
2060 } else if (*dd_idx
>= pd_idx
)
2061 (*dd_idx
) += 2; /* D D P Q D */
2063 case ALGORITHM_RIGHT_ASYMMETRIC
:
2064 pd_idx
= sector_div(stripe2
, raid_disks
);
2065 qd_idx
= pd_idx
+ 1;
2066 if (pd_idx
== raid_disks
-1) {
2067 (*dd_idx
)++; /* Q D D D P */
2069 } else if (*dd_idx
>= pd_idx
)
2070 (*dd_idx
) += 2; /* D D P Q D */
2072 case ALGORITHM_LEFT_SYMMETRIC
:
2073 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2074 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2075 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2077 case ALGORITHM_RIGHT_SYMMETRIC
:
2078 pd_idx
= sector_div(stripe2
, raid_disks
);
2079 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2080 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2083 case ALGORITHM_PARITY_0
:
2088 case ALGORITHM_PARITY_N
:
2089 pd_idx
= data_disks
;
2090 qd_idx
= data_disks
+ 1;
2093 case ALGORITHM_ROTATING_ZERO_RESTART
:
2094 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2095 * of blocks for computing Q is different.
2097 pd_idx
= sector_div(stripe2
, raid_disks
);
2098 qd_idx
= pd_idx
+ 1;
2099 if (pd_idx
== raid_disks
-1) {
2100 (*dd_idx
)++; /* Q D D D P */
2102 } else if (*dd_idx
>= pd_idx
)
2103 (*dd_idx
) += 2; /* D D P Q D */
2107 case ALGORITHM_ROTATING_N_RESTART
:
2108 /* Same a left_asymmetric, by first stripe is
2109 * D D D P Q rather than
2113 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2114 qd_idx
= pd_idx
+ 1;
2115 if (pd_idx
== raid_disks
-1) {
2116 (*dd_idx
)++; /* Q D D D P */
2118 } else if (*dd_idx
>= pd_idx
)
2119 (*dd_idx
) += 2; /* D D P Q D */
2123 case ALGORITHM_ROTATING_N_CONTINUE
:
2124 /* Same as left_symmetric but Q is before P */
2125 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2126 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
2127 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2131 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2132 /* RAID5 left_asymmetric, with Q on last device */
2133 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2134 if (*dd_idx
>= pd_idx
)
2136 qd_idx
= raid_disks
- 1;
2139 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2140 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2141 if (*dd_idx
>= pd_idx
)
2143 qd_idx
= raid_disks
- 1;
2146 case ALGORITHM_LEFT_SYMMETRIC_6
:
2147 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2148 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2149 qd_idx
= raid_disks
- 1;
2152 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2153 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2154 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2155 qd_idx
= raid_disks
- 1;
2158 case ALGORITHM_PARITY_0_6
:
2161 qd_idx
= raid_disks
- 1;
2171 sh
->pd_idx
= pd_idx
;
2172 sh
->qd_idx
= qd_idx
;
2173 sh
->ddf_layout
= ddf_layout
;
2176 * Finally, compute the new sector number
2178 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
2183 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
2185 struct r5conf
*conf
= sh
->raid_conf
;
2186 int raid_disks
= sh
->disks
;
2187 int data_disks
= raid_disks
- conf
->max_degraded
;
2188 sector_t new_sector
= sh
->sector
, check
;
2189 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2190 : conf
->chunk_sectors
;
2191 int algorithm
= previous
? conf
->prev_algo
2195 sector_t chunk_number
;
2196 int dummy1
, dd_idx
= i
;
2198 struct stripe_head sh2
;
2201 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
2202 stripe
= new_sector
;
2204 if (i
== sh
->pd_idx
)
2206 switch(conf
->level
) {
2209 switch (algorithm
) {
2210 case ALGORITHM_LEFT_ASYMMETRIC
:
2211 case ALGORITHM_RIGHT_ASYMMETRIC
:
2215 case ALGORITHM_LEFT_SYMMETRIC
:
2216 case ALGORITHM_RIGHT_SYMMETRIC
:
2219 i
-= (sh
->pd_idx
+ 1);
2221 case ALGORITHM_PARITY_0
:
2224 case ALGORITHM_PARITY_N
:
2231 if (i
== sh
->qd_idx
)
2232 return 0; /* It is the Q disk */
2233 switch (algorithm
) {
2234 case ALGORITHM_LEFT_ASYMMETRIC
:
2235 case ALGORITHM_RIGHT_ASYMMETRIC
:
2236 case ALGORITHM_ROTATING_ZERO_RESTART
:
2237 case ALGORITHM_ROTATING_N_RESTART
:
2238 if (sh
->pd_idx
== raid_disks
-1)
2239 i
--; /* Q D D D P */
2240 else if (i
> sh
->pd_idx
)
2241 i
-= 2; /* D D P Q D */
2243 case ALGORITHM_LEFT_SYMMETRIC
:
2244 case ALGORITHM_RIGHT_SYMMETRIC
:
2245 if (sh
->pd_idx
== raid_disks
-1)
2246 i
--; /* Q D D D P */
2251 i
-= (sh
->pd_idx
+ 2);
2254 case ALGORITHM_PARITY_0
:
2257 case ALGORITHM_PARITY_N
:
2259 case ALGORITHM_ROTATING_N_CONTINUE
:
2260 /* Like left_symmetric, but P is before Q */
2261 if (sh
->pd_idx
== 0)
2262 i
--; /* P D D D Q */
2267 i
-= (sh
->pd_idx
+ 1);
2270 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2271 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2275 case ALGORITHM_LEFT_SYMMETRIC_6
:
2276 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2278 i
+= data_disks
+ 1;
2279 i
-= (sh
->pd_idx
+ 1);
2281 case ALGORITHM_PARITY_0_6
:
2290 chunk_number
= stripe
* data_disks
+ i
;
2291 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2293 check
= raid5_compute_sector(conf
, r_sector
,
2294 previous
, &dummy1
, &sh2
);
2295 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2296 || sh2
.qd_idx
!= sh
->qd_idx
) {
2297 printk(KERN_ERR
"md/raid:%s: compute_blocknr: map not correct\n",
2298 mdname(conf
->mddev
));
2306 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2307 int rcw
, int expand
)
2309 int i
, pd_idx
= sh
->pd_idx
, disks
= sh
->disks
;
2310 struct r5conf
*conf
= sh
->raid_conf
;
2311 int level
= conf
->level
;
2314 /* if we are not expanding this is a proper write request, and
2315 * there will be bios with new data to be drained into the
2319 sh
->reconstruct_state
= reconstruct_state_drain_run
;
2320 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2322 sh
->reconstruct_state
= reconstruct_state_run
;
2324 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2326 for (i
= disks
; i
--; ) {
2327 struct r5dev
*dev
= &sh
->dev
[i
];
2330 set_bit(R5_LOCKED
, &dev
->flags
);
2331 set_bit(R5_Wantdrain
, &dev
->flags
);
2333 clear_bit(R5_UPTODATE
, &dev
->flags
);
2337 if (s
->locked
+ conf
->max_degraded
== disks
)
2338 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2339 atomic_inc(&conf
->pending_full_writes
);
2342 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
2343 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
2345 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
2346 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
2347 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2348 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2350 for (i
= disks
; i
--; ) {
2351 struct r5dev
*dev
= &sh
->dev
[i
];
2356 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2357 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2358 set_bit(R5_Wantdrain
, &dev
->flags
);
2359 set_bit(R5_LOCKED
, &dev
->flags
);
2360 clear_bit(R5_UPTODATE
, &dev
->flags
);
2366 /* keep the parity disk(s) locked while asynchronous operations
2369 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
2370 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2374 int qd_idx
= sh
->qd_idx
;
2375 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
2377 set_bit(R5_LOCKED
, &dev
->flags
);
2378 clear_bit(R5_UPTODATE
, &dev
->flags
);
2382 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2383 __func__
, (unsigned long long)sh
->sector
,
2384 s
->locked
, s
->ops_request
);
2388 * Each stripe/dev can have one or more bion attached.
2389 * toread/towrite point to the first in a chain.
2390 * The bi_next chain must be in order.
2392 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
, int forwrite
)
2395 struct r5conf
*conf
= sh
->raid_conf
;
2398 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2399 (unsigned long long)bi
->bi_sector
,
2400 (unsigned long long)sh
->sector
);
2403 * If several bio share a stripe. The bio bi_phys_segments acts as a
2404 * reference count to avoid race. The reference count should already be
2405 * increased before this function is called (for example, in
2406 * make_request()), so other bio sharing this stripe will not free the
2407 * stripe. If a stripe is owned by one stripe, the stripe lock will
2410 spin_lock_irq(&sh
->stripe_lock
);
2412 bip
= &sh
->dev
[dd_idx
].towrite
;
2416 bip
= &sh
->dev
[dd_idx
].toread
;
2417 while (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
) {
2418 if ((*bip
)->bi_sector
+ ((*bip
)->bi_size
>> 9) > bi
->bi_sector
)
2420 bip
= & (*bip
)->bi_next
;
2422 if (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
+ ((bi
->bi_size
)>>9))
2425 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
2429 raid5_inc_bi_active_stripes(bi
);
2432 /* check if page is covered */
2433 sector_t sector
= sh
->dev
[dd_idx
].sector
;
2434 for (bi
=sh
->dev
[dd_idx
].towrite
;
2435 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
2436 bi
&& bi
->bi_sector
<= sector
;
2437 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
2438 if (bi
->bi_sector
+ (bi
->bi_size
>>9) >= sector
)
2439 sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
2441 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
2442 set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
);
2445 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2446 (unsigned long long)(*bip
)->bi_sector
,
2447 (unsigned long long)sh
->sector
, dd_idx
);
2448 spin_unlock_irq(&sh
->stripe_lock
);
2450 if (conf
->mddev
->bitmap
&& firstwrite
) {
2451 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
2453 sh
->bm_seq
= conf
->seq_flush
+1;
2454 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
2459 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
2460 spin_unlock_irq(&sh
->stripe_lock
);
2464 static void end_reshape(struct r5conf
*conf
);
2466 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
2467 struct stripe_head
*sh
)
2469 int sectors_per_chunk
=
2470 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
2472 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
2473 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
2475 raid5_compute_sector(conf
,
2476 stripe
* (disks
- conf
->max_degraded
)
2477 *sectors_per_chunk
+ chunk_offset
,
2483 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
2484 struct stripe_head_state
*s
, int disks
,
2485 struct bio
**return_bi
)
2488 for (i
= disks
; i
--; ) {
2492 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2493 struct md_rdev
*rdev
;
2495 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2496 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
2497 atomic_inc(&rdev
->nr_pending
);
2502 if (!rdev_set_badblocks(
2506 md_error(conf
->mddev
, rdev
);
2507 rdev_dec_pending(rdev
, conf
->mddev
);
2510 spin_lock_irq(&sh
->stripe_lock
);
2511 /* fail all writes first */
2512 bi
= sh
->dev
[i
].towrite
;
2513 sh
->dev
[i
].towrite
= NULL
;
2514 spin_unlock_irq(&sh
->stripe_lock
);
2518 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2519 wake_up(&conf
->wait_for_overlap
);
2521 while (bi
&& bi
->bi_sector
<
2522 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2523 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2524 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2525 if (!raid5_dec_bi_active_stripes(bi
)) {
2526 md_write_end(conf
->mddev
);
2527 bi
->bi_next
= *return_bi
;
2533 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2534 STRIPE_SECTORS
, 0, 0);
2536 /* and fail all 'written' */
2537 bi
= sh
->dev
[i
].written
;
2538 sh
->dev
[i
].written
= NULL
;
2539 if (bi
) bitmap_end
= 1;
2540 while (bi
&& bi
->bi_sector
<
2541 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2542 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2543 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2544 if (!raid5_dec_bi_active_stripes(bi
)) {
2545 md_write_end(conf
->mddev
);
2546 bi
->bi_next
= *return_bi
;
2552 /* fail any reads if this device is non-operational and
2553 * the data has not reached the cache yet.
2555 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
2556 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
2557 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
2558 spin_lock_irq(&sh
->stripe_lock
);
2559 bi
= sh
->dev
[i
].toread
;
2560 sh
->dev
[i
].toread
= NULL
;
2561 spin_unlock_irq(&sh
->stripe_lock
);
2562 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2563 wake_up(&conf
->wait_for_overlap
);
2564 while (bi
&& bi
->bi_sector
<
2565 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2566 struct bio
*nextbi
=
2567 r5_next_bio(bi
, sh
->dev
[i
].sector
);
2568 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2569 if (!raid5_dec_bi_active_stripes(bi
)) {
2570 bi
->bi_next
= *return_bi
;
2577 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2578 STRIPE_SECTORS
, 0, 0);
2579 /* If we were in the middle of a write the parity block might
2580 * still be locked - so just clear all R5_LOCKED flags
2582 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2585 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2586 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2587 md_wakeup_thread(conf
->mddev
->thread
);
2591 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
2592 struct stripe_head_state
*s
)
2597 clear_bit(STRIPE_SYNCING
, &sh
->state
);
2600 /* There is nothing more to do for sync/check/repair.
2601 * Don't even need to abort as that is handled elsewhere
2602 * if needed, and not always wanted e.g. if there is a known
2604 * For recover/replace we need to record a bad block on all
2605 * non-sync devices, or abort the recovery
2607 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
2608 /* During recovery devices cannot be removed, so
2609 * locking and refcounting of rdevs is not needed
2611 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2612 struct md_rdev
*rdev
= conf
->disks
[i
].rdev
;
2614 && !test_bit(Faulty
, &rdev
->flags
)
2615 && !test_bit(In_sync
, &rdev
->flags
)
2616 && !rdev_set_badblocks(rdev
, sh
->sector
,
2619 rdev
= conf
->disks
[i
].replacement
;
2621 && !test_bit(Faulty
, &rdev
->flags
)
2622 && !test_bit(In_sync
, &rdev
->flags
)
2623 && !rdev_set_badblocks(rdev
, sh
->sector
,
2628 conf
->recovery_disabled
=
2629 conf
->mddev
->recovery_disabled
;
2631 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, !abort
);
2634 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
2636 struct md_rdev
*rdev
;
2638 /* Doing recovery so rcu locking not required */
2639 rdev
= sh
->raid_conf
->disks
[disk_idx
].replacement
;
2641 && !test_bit(Faulty
, &rdev
->flags
)
2642 && !test_bit(In_sync
, &rdev
->flags
)
2643 && (rdev
->recovery_offset
<= sh
->sector
2644 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
2650 /* fetch_block - checks the given member device to see if its data needs
2651 * to be read or computed to satisfy a request.
2653 * Returns 1 when no more member devices need to be checked, otherwise returns
2654 * 0 to tell the loop in handle_stripe_fill to continue
2656 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2657 int disk_idx
, int disks
)
2659 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
2660 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
2661 &sh
->dev
[s
->failed_num
[1]] };
2663 /* is the data in this block needed, and can we get it? */
2664 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2665 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2667 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2668 s
->syncing
|| s
->expanding
||
2669 (s
->replacing
&& want_replace(sh
, disk_idx
)) ||
2670 (s
->failed
>= 1 && fdev
[0]->toread
) ||
2671 (s
->failed
>= 2 && fdev
[1]->toread
) ||
2672 (sh
->raid_conf
->level
<= 5 && s
->failed
&& fdev
[0]->towrite
&&
2673 !test_bit(R5_OVERWRITE
, &fdev
[0]->flags
)) ||
2674 (sh
->raid_conf
->level
== 6 && s
->failed
&& s
->to_write
))) {
2675 /* we would like to get this block, possibly by computing it,
2676 * otherwise read it if the backing disk is insync
2678 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
2679 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
2680 if ((s
->uptodate
== disks
- 1) &&
2681 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
2682 disk_idx
== s
->failed_num
[1]))) {
2683 /* have disk failed, and we're requested to fetch it;
2686 pr_debug("Computing stripe %llu block %d\n",
2687 (unsigned long long)sh
->sector
, disk_idx
);
2688 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2689 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2690 set_bit(R5_Wantcompute
, &dev
->flags
);
2691 sh
->ops
.target
= disk_idx
;
2692 sh
->ops
.target2
= -1; /* no 2nd target */
2694 /* Careful: from this point on 'uptodate' is in the eye
2695 * of raid_run_ops which services 'compute' operations
2696 * before writes. R5_Wantcompute flags a block that will
2697 * be R5_UPTODATE by the time it is needed for a
2698 * subsequent operation.
2702 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
2703 /* Computing 2-failure is *very* expensive; only
2704 * do it if failed >= 2
2707 for (other
= disks
; other
--; ) {
2708 if (other
== disk_idx
)
2710 if (!test_bit(R5_UPTODATE
,
2711 &sh
->dev
[other
].flags
))
2715 pr_debug("Computing stripe %llu blocks %d,%d\n",
2716 (unsigned long long)sh
->sector
,
2718 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2719 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2720 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
2721 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
2722 sh
->ops
.target
= disk_idx
;
2723 sh
->ops
.target2
= other
;
2727 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2728 set_bit(R5_LOCKED
, &dev
->flags
);
2729 set_bit(R5_Wantread
, &dev
->flags
);
2731 pr_debug("Reading block %d (sync=%d)\n",
2732 disk_idx
, s
->syncing
);
2740 * handle_stripe_fill - read or compute data to satisfy pending requests.
2742 static void handle_stripe_fill(struct stripe_head
*sh
,
2743 struct stripe_head_state
*s
,
2748 /* look for blocks to read/compute, skip this if a compute
2749 * is already in flight, or if the stripe contents are in the
2750 * midst of changing due to a write
2752 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2753 !sh
->reconstruct_state
)
2754 for (i
= disks
; i
--; )
2755 if (fetch_block(sh
, s
, i
, disks
))
2757 set_bit(STRIPE_HANDLE
, &sh
->state
);
2761 /* handle_stripe_clean_event
2762 * any written block on an uptodate or failed drive can be returned.
2763 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2764 * never LOCKED, so we don't need to test 'failed' directly.
2766 static void handle_stripe_clean_event(struct r5conf
*conf
,
2767 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
2772 for (i
= disks
; i
--; )
2773 if (sh
->dev
[i
].written
) {
2775 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2776 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2777 test_bit(R5_Discard
, &dev
->flags
))) {
2778 /* We can return any write requests */
2779 struct bio
*wbi
, *wbi2
;
2780 pr_debug("Return write for disc %d\n", i
);
2781 if (test_and_clear_bit(R5_Discard
, &dev
->flags
))
2782 clear_bit(R5_UPTODATE
, &dev
->flags
);
2784 dev
->written
= NULL
;
2785 while (wbi
&& wbi
->bi_sector
<
2786 dev
->sector
+ STRIPE_SECTORS
) {
2787 wbi2
= r5_next_bio(wbi
, dev
->sector
);
2788 if (!raid5_dec_bi_active_stripes(wbi
)) {
2789 md_write_end(conf
->mddev
);
2790 wbi
->bi_next
= *return_bi
;
2795 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2797 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
2800 } else if (test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
2801 clear_bit(R5_Discard
, &sh
->dev
[i
].flags
);
2803 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2804 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2805 md_wakeup_thread(conf
->mddev
->thread
);
2808 static void handle_stripe_dirtying(struct r5conf
*conf
,
2809 struct stripe_head
*sh
,
2810 struct stripe_head_state
*s
,
2813 int rmw
= 0, rcw
= 0, i
;
2814 sector_t recovery_cp
= conf
->mddev
->recovery_cp
;
2816 /* RAID6 requires 'rcw' in current implementation.
2817 * Otherwise, check whether resync is now happening or should start.
2818 * If yes, then the array is dirty (after unclean shutdown or
2819 * initial creation), so parity in some stripes might be inconsistent.
2820 * In this case, we need to always do reconstruct-write, to ensure
2821 * that in case of drive failure or read-error correction, we
2822 * generate correct data from the parity.
2824 if (conf
->max_degraded
== 2 ||
2825 (recovery_cp
< MaxSector
&& sh
->sector
>= recovery_cp
)) {
2826 /* Calculate the real rcw later - for now make it
2827 * look like rcw is cheaper
2830 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
2831 conf
->max_degraded
, (unsigned long long)recovery_cp
,
2832 (unsigned long long)sh
->sector
);
2833 } else for (i
= disks
; i
--; ) {
2834 /* would I have to read this buffer for read_modify_write */
2835 struct r5dev
*dev
= &sh
->dev
[i
];
2836 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2837 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2838 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2839 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2840 if (test_bit(R5_Insync
, &dev
->flags
))
2843 rmw
+= 2*disks
; /* cannot read it */
2845 /* Would I have to read this buffer for reconstruct_write */
2846 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) && i
!= sh
->pd_idx
&&
2847 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2848 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2849 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2850 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2855 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2856 (unsigned long long)sh
->sector
, rmw
, rcw
);
2857 set_bit(STRIPE_HANDLE
, &sh
->state
);
2858 if (rmw
< rcw
&& rmw
> 0)
2859 /* prefer read-modify-write, but need to get some data */
2860 for (i
= disks
; i
--; ) {
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 test_bit(R5_Insync
, &dev
->flags
)) {
2868 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2869 pr_debug("Read_old block "
2870 "%d for r-m-w\n", i
);
2871 set_bit(R5_LOCKED
, &dev
->flags
);
2872 set_bit(R5_Wantread
, &dev
->flags
);
2875 set_bit(STRIPE_DELAYED
, &sh
->state
);
2876 set_bit(STRIPE_HANDLE
, &sh
->state
);
2880 if (rcw
<= rmw
&& rcw
> 0) {
2881 /* want reconstruct write, but need to get some data */
2883 for (i
= disks
; i
--; ) {
2884 struct r5dev
*dev
= &sh
->dev
[i
];
2885 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2886 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
2887 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2888 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2889 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2891 if (!test_bit(R5_Insync
, &dev
->flags
))
2892 continue; /* it's a failed drive */
2894 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2895 pr_debug("Read_old block "
2896 "%d for Reconstruct\n", i
);
2897 set_bit(R5_LOCKED
, &dev
->flags
);
2898 set_bit(R5_Wantread
, &dev
->flags
);
2901 set_bit(STRIPE_DELAYED
, &sh
->state
);
2902 set_bit(STRIPE_HANDLE
, &sh
->state
);
2907 /* now if nothing is locked, and if we have enough data,
2908 * we can start a write request
2910 /* since handle_stripe can be called at any time we need to handle the
2911 * case where a compute block operation has been submitted and then a
2912 * subsequent call wants to start a write request. raid_run_ops only
2913 * handles the case where compute block and reconstruct are requested
2914 * simultaneously. If this is not the case then new writes need to be
2915 * held off until the compute completes.
2917 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
2918 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
2919 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
2920 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
2923 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
2924 struct stripe_head_state
*s
, int disks
)
2926 struct r5dev
*dev
= NULL
;
2928 set_bit(STRIPE_HANDLE
, &sh
->state
);
2930 switch (sh
->check_state
) {
2931 case check_state_idle
:
2932 /* start a new check operation if there are no failures */
2933 if (s
->failed
== 0) {
2934 BUG_ON(s
->uptodate
!= disks
);
2935 sh
->check_state
= check_state_run
;
2936 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2937 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2941 dev
= &sh
->dev
[s
->failed_num
[0]];
2943 case check_state_compute_result
:
2944 sh
->check_state
= check_state_idle
;
2946 dev
= &sh
->dev
[sh
->pd_idx
];
2948 /* check that a write has not made the stripe insync */
2949 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2952 /* either failed parity check, or recovery is happening */
2953 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
2954 BUG_ON(s
->uptodate
!= disks
);
2956 set_bit(R5_LOCKED
, &dev
->flags
);
2958 set_bit(R5_Wantwrite
, &dev
->flags
);
2960 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2961 set_bit(STRIPE_INSYNC
, &sh
->state
);
2963 case check_state_run
:
2964 break; /* we will be called again upon completion */
2965 case check_state_check_result
:
2966 sh
->check_state
= check_state_idle
;
2968 /* if a failure occurred during the check operation, leave
2969 * STRIPE_INSYNC not set and let the stripe be handled again
2974 /* handle a successful check operation, if parity is correct
2975 * we are done. Otherwise update the mismatch count and repair
2976 * parity if !MD_RECOVERY_CHECK
2978 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
2979 /* parity is correct (on disc,
2980 * not in buffer any more)
2982 set_bit(STRIPE_INSYNC
, &sh
->state
);
2984 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
2985 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2986 /* don't try to repair!! */
2987 set_bit(STRIPE_INSYNC
, &sh
->state
);
2989 sh
->check_state
= check_state_compute_run
;
2990 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2991 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2992 set_bit(R5_Wantcompute
,
2993 &sh
->dev
[sh
->pd_idx
].flags
);
2994 sh
->ops
.target
= sh
->pd_idx
;
2995 sh
->ops
.target2
= -1;
3000 case check_state_compute_run
:
3003 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3004 __func__
, sh
->check_state
,
3005 (unsigned long long) sh
->sector
);
3011 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
3012 struct stripe_head_state
*s
,
3015 int pd_idx
= sh
->pd_idx
;
3016 int qd_idx
= sh
->qd_idx
;
3019 set_bit(STRIPE_HANDLE
, &sh
->state
);
3021 BUG_ON(s
->failed
> 2);
3023 /* Want to check and possibly repair P and Q.
3024 * However there could be one 'failed' device, in which
3025 * case we can only check one of them, possibly using the
3026 * other to generate missing data
3029 switch (sh
->check_state
) {
3030 case check_state_idle
:
3031 /* start a new check operation if there are < 2 failures */
3032 if (s
->failed
== s
->q_failed
) {
3033 /* The only possible failed device holds Q, so it
3034 * makes sense to check P (If anything else were failed,
3035 * we would have used P to recreate it).
3037 sh
->check_state
= check_state_run
;
3039 if (!s
->q_failed
&& s
->failed
< 2) {
3040 /* Q is not failed, and we didn't use it to generate
3041 * anything, so it makes sense to check it
3043 if (sh
->check_state
== check_state_run
)
3044 sh
->check_state
= check_state_run_pq
;
3046 sh
->check_state
= check_state_run_q
;
3049 /* discard potentially stale zero_sum_result */
3050 sh
->ops
.zero_sum_result
= 0;
3052 if (sh
->check_state
== check_state_run
) {
3053 /* async_xor_zero_sum destroys the contents of P */
3054 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
3057 if (sh
->check_state
>= check_state_run
&&
3058 sh
->check_state
<= check_state_run_pq
) {
3059 /* async_syndrome_zero_sum preserves P and Q, so
3060 * no need to mark them !uptodate here
3062 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3066 /* we have 2-disk failure */
3067 BUG_ON(s
->failed
!= 2);
3069 case check_state_compute_result
:
3070 sh
->check_state
= check_state_idle
;
3072 /* check that a write has not made the stripe insync */
3073 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3076 /* now write out any block on a failed drive,
3077 * or P or Q if they were recomputed
3079 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
3080 if (s
->failed
== 2) {
3081 dev
= &sh
->dev
[s
->failed_num
[1]];
3083 set_bit(R5_LOCKED
, &dev
->flags
);
3084 set_bit(R5_Wantwrite
, &dev
->flags
);
3086 if (s
->failed
>= 1) {
3087 dev
= &sh
->dev
[s
->failed_num
[0]];
3089 set_bit(R5_LOCKED
, &dev
->flags
);
3090 set_bit(R5_Wantwrite
, &dev
->flags
);
3092 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3093 dev
= &sh
->dev
[pd_idx
];
3095 set_bit(R5_LOCKED
, &dev
->flags
);
3096 set_bit(R5_Wantwrite
, &dev
->flags
);
3098 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3099 dev
= &sh
->dev
[qd_idx
];
3101 set_bit(R5_LOCKED
, &dev
->flags
);
3102 set_bit(R5_Wantwrite
, &dev
->flags
);
3104 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3106 set_bit(STRIPE_INSYNC
, &sh
->state
);
3108 case check_state_run
:
3109 case check_state_run_q
:
3110 case check_state_run_pq
:
3111 break; /* we will be called again upon completion */
3112 case check_state_check_result
:
3113 sh
->check_state
= check_state_idle
;
3115 /* handle a successful check operation, if parity is correct
3116 * we are done. Otherwise update the mismatch count and repair
3117 * parity if !MD_RECOVERY_CHECK
3119 if (sh
->ops
.zero_sum_result
== 0) {
3120 /* both parities are correct */
3122 set_bit(STRIPE_INSYNC
, &sh
->state
);
3124 /* in contrast to the raid5 case we can validate
3125 * parity, but still have a failure to write
3128 sh
->check_state
= check_state_compute_result
;
3129 /* Returning at this point means that we may go
3130 * off and bring p and/or q uptodate again so
3131 * we make sure to check zero_sum_result again
3132 * to verify if p or q need writeback
3136 atomic64_add(STRIPE_SECTORS
, &conf
->mddev
->resync_mismatches
);
3137 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3138 /* don't try to repair!! */
3139 set_bit(STRIPE_INSYNC
, &sh
->state
);
3141 int *target
= &sh
->ops
.target
;
3143 sh
->ops
.target
= -1;
3144 sh
->ops
.target2
= -1;
3145 sh
->check_state
= check_state_compute_run
;
3146 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3147 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3148 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3149 set_bit(R5_Wantcompute
,
3150 &sh
->dev
[pd_idx
].flags
);
3152 target
= &sh
->ops
.target2
;
3155 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3156 set_bit(R5_Wantcompute
,
3157 &sh
->dev
[qd_idx
].flags
);
3164 case check_state_compute_run
:
3167 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3168 __func__
, sh
->check_state
,
3169 (unsigned long long) sh
->sector
);
3174 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
3178 /* We have read all the blocks in this stripe and now we need to
3179 * copy some of them into a target stripe for expand.
3181 struct dma_async_tx_descriptor
*tx
= NULL
;
3182 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3183 for (i
= 0; i
< sh
->disks
; i
++)
3184 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
3186 struct stripe_head
*sh2
;
3187 struct async_submit_ctl submit
;
3189 sector_t bn
= compute_blocknr(sh
, i
, 1);
3190 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
3192 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
3194 /* so far only the early blocks of this stripe
3195 * have been requested. When later blocks
3196 * get requested, we will try again
3199 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
3200 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
3201 /* must have already done this block */
3202 release_stripe(sh2
);
3206 /* place all the copies on one channel */
3207 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
3208 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
3209 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
3212 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
3213 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
3214 for (j
= 0; j
< conf
->raid_disks
; j
++)
3215 if (j
!= sh2
->pd_idx
&&
3217 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
3219 if (j
== conf
->raid_disks
) {
3220 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
3221 set_bit(STRIPE_HANDLE
, &sh2
->state
);
3223 release_stripe(sh2
);
3226 /* done submitting copies, wait for them to complete */
3229 dma_wait_for_async_tx(tx
);
3234 * handle_stripe - do things to a stripe.
3236 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3237 * state of various bits to see what needs to be done.
3239 * return some read requests which now have data
3240 * return some write requests which are safely on storage
3241 * schedule a read on some buffers
3242 * schedule a write of some buffers
3243 * return confirmation of parity correctness
3247 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
3249 struct r5conf
*conf
= sh
->raid_conf
;
3250 int disks
= sh
->disks
;
3253 int do_recovery
= 0;
3255 memset(s
, 0, sizeof(*s
));
3257 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3258 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3259 s
->failed_num
[0] = -1;
3260 s
->failed_num
[1] = -1;
3262 /* Now to look around and see what can be done */
3264 for (i
=disks
; i
--; ) {
3265 struct md_rdev
*rdev
;
3272 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3274 dev
->toread
, dev
->towrite
, dev
->written
);
3275 /* maybe we can reply to a read
3277 * new wantfill requests are only permitted while
3278 * ops_complete_biofill is guaranteed to be inactive
3280 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
3281 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
3282 set_bit(R5_Wantfill
, &dev
->flags
);
3284 /* now count some things */
3285 if (test_bit(R5_LOCKED
, &dev
->flags
))
3287 if (test_bit(R5_UPTODATE
, &dev
->flags
))
3289 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
3291 BUG_ON(s
->compute
> 2);
3294 if (test_bit(R5_Wantfill
, &dev
->flags
))
3296 else if (dev
->toread
)
3300 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3305 /* Prefer to use the replacement for reads, but only
3306 * if it is recovered enough and has no bad blocks.
3308 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
3309 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
3310 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
3311 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3312 &first_bad
, &bad_sectors
))
3313 set_bit(R5_ReadRepl
, &dev
->flags
);
3316 set_bit(R5_NeedReplace
, &dev
->flags
);
3317 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3318 clear_bit(R5_ReadRepl
, &dev
->flags
);
3320 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
3323 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3324 &first_bad
, &bad_sectors
);
3325 if (s
->blocked_rdev
== NULL
3326 && (test_bit(Blocked
, &rdev
->flags
)
3329 set_bit(BlockedBadBlocks
,
3331 s
->blocked_rdev
= rdev
;
3332 atomic_inc(&rdev
->nr_pending
);
3335 clear_bit(R5_Insync
, &dev
->flags
);
3339 /* also not in-sync */
3340 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
3341 test_bit(R5_UPTODATE
, &dev
->flags
)) {
3342 /* treat as in-sync, but with a read error
3343 * which we can now try to correct
3345 set_bit(R5_Insync
, &dev
->flags
);
3346 set_bit(R5_ReadError
, &dev
->flags
);
3348 } else if (test_bit(In_sync
, &rdev
->flags
))
3349 set_bit(R5_Insync
, &dev
->flags
);
3350 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
3351 /* in sync if before recovery_offset */
3352 set_bit(R5_Insync
, &dev
->flags
);
3353 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
3354 test_bit(R5_Expanded
, &dev
->flags
))
3355 /* If we've reshaped into here, we assume it is Insync.
3356 * We will shortly update recovery_offset to make
3359 set_bit(R5_Insync
, &dev
->flags
);
3361 if (rdev
&& test_bit(R5_WriteError
, &dev
->flags
)) {
3362 /* This flag does not apply to '.replacement'
3363 * only to .rdev, so make sure to check that*/
3364 struct md_rdev
*rdev2
= rcu_dereference(
3365 conf
->disks
[i
].rdev
);
3367 clear_bit(R5_Insync
, &dev
->flags
);
3368 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3369 s
->handle_bad_blocks
= 1;
3370 atomic_inc(&rdev2
->nr_pending
);
3372 clear_bit(R5_WriteError
, &dev
->flags
);
3374 if (rdev
&& test_bit(R5_MadeGood
, &dev
->flags
)) {
3375 /* This flag does not apply to '.replacement'
3376 * only to .rdev, so make sure to check that*/
3377 struct md_rdev
*rdev2
= rcu_dereference(
3378 conf
->disks
[i
].rdev
);
3379 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3380 s
->handle_bad_blocks
= 1;
3381 atomic_inc(&rdev2
->nr_pending
);
3383 clear_bit(R5_MadeGood
, &dev
->flags
);
3385 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
3386 struct md_rdev
*rdev2
= rcu_dereference(
3387 conf
->disks
[i
].replacement
);
3388 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3389 s
->handle_bad_blocks
= 1;
3390 atomic_inc(&rdev2
->nr_pending
);
3392 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
3394 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3395 /* The ReadError flag will just be confusing now */
3396 clear_bit(R5_ReadError
, &dev
->flags
);
3397 clear_bit(R5_ReWrite
, &dev
->flags
);
3399 if (test_bit(R5_ReadError
, &dev
->flags
))
3400 clear_bit(R5_Insync
, &dev
->flags
);
3401 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3403 s
->failed_num
[s
->failed
] = i
;
3405 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
3409 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
3410 /* If there is a failed device being replaced,
3411 * we must be recovering.
3412 * else if we are after recovery_cp, we must be syncing
3413 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3414 * else we can only be replacing
3415 * sync and recovery both need to read all devices, and so
3416 * use the same flag.
3419 sh
->sector
>= conf
->mddev
->recovery_cp
||
3420 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
3428 static void handle_stripe(struct stripe_head
*sh
)
3430 struct stripe_head_state s
;
3431 struct r5conf
*conf
= sh
->raid_conf
;
3434 int disks
= sh
->disks
;
3435 struct r5dev
*pdev
, *qdev
;
3437 clear_bit(STRIPE_HANDLE
, &sh
->state
);
3438 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
3439 /* already being handled, ensure it gets handled
3440 * again when current action finishes */
3441 set_bit(STRIPE_HANDLE
, &sh
->state
);
3445 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
3446 set_bit(STRIPE_SYNCING
, &sh
->state
);
3447 clear_bit(STRIPE_INSYNC
, &sh
->state
);
3449 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3451 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3452 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3453 (unsigned long long)sh
->sector
, sh
->state
,
3454 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
3455 sh
->check_state
, sh
->reconstruct_state
);
3457 analyse_stripe(sh
, &s
);
3459 if (s
.handle_bad_blocks
) {
3460 set_bit(STRIPE_HANDLE
, &sh
->state
);
3464 if (unlikely(s
.blocked_rdev
)) {
3465 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
3466 s
.replacing
|| s
.to_write
|| s
.written
) {
3467 set_bit(STRIPE_HANDLE
, &sh
->state
);
3470 /* There is nothing for the blocked_rdev to block */
3471 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
3472 s
.blocked_rdev
= NULL
;
3475 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
3476 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
3477 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
3480 pr_debug("locked=%d uptodate=%d to_read=%d"
3481 " to_write=%d failed=%d failed_num=%d,%d\n",
3482 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
3483 s
.failed_num
[0], s
.failed_num
[1]);
3484 /* check if the array has lost more than max_degraded devices and,
3485 * if so, some requests might need to be failed.
3487 if (s
.failed
> conf
->max_degraded
) {
3488 sh
->check_state
= 0;
3489 sh
->reconstruct_state
= 0;
3490 if (s
.to_read
+s
.to_write
+s
.written
)
3491 handle_failed_stripe(conf
, sh
, &s
, disks
, &s
.return_bi
);
3492 if (s
.syncing
+ s
.replacing
)
3493 handle_failed_sync(conf
, sh
, &s
);
3496 /* Now we check to see if any write operations have recently
3500 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
3502 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
3503 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
3504 sh
->reconstruct_state
= reconstruct_state_idle
;
3506 /* All the 'written' buffers and the parity block are ready to
3507 * be written back to disk
3509 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
) &&
3510 !test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
));
3511 BUG_ON(sh
->qd_idx
>= 0 &&
3512 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
) &&
3513 !test_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
));
3514 for (i
= disks
; i
--; ) {
3515 struct r5dev
*dev
= &sh
->dev
[i
];
3516 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
3517 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3519 pr_debug("Writing block %d\n", i
);
3520 set_bit(R5_Wantwrite
, &dev
->flags
);
3523 if (!test_bit(R5_Insync
, &dev
->flags
) ||
3524 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3526 set_bit(STRIPE_INSYNC
, &sh
->state
);
3529 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3530 s
.dec_preread_active
= 1;
3534 * might be able to return some write requests if the parity blocks
3535 * are safe, or on a failed drive
3537 pdev
= &sh
->dev
[sh
->pd_idx
];
3538 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
3539 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
3540 qdev
= &sh
->dev
[sh
->qd_idx
];
3541 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
3542 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
3546 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
3547 && !test_bit(R5_LOCKED
, &pdev
->flags
)
3548 && (test_bit(R5_UPTODATE
, &pdev
->flags
) ||
3549 test_bit(R5_Discard
, &pdev
->flags
))))) &&
3550 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
3551 && !test_bit(R5_LOCKED
, &qdev
->flags
)
3552 && (test_bit(R5_UPTODATE
, &qdev
->flags
) ||
3553 test_bit(R5_Discard
, &qdev
->flags
))))))
3554 handle_stripe_clean_event(conf
, sh
, disks
, &s
.return_bi
);
3556 /* Now we might consider reading some blocks, either to check/generate
3557 * parity, or to satisfy requests
3558 * or to load a block that is being partially written.
3560 if (s
.to_read
|| s
.non_overwrite
3561 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
3562 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
3565 handle_stripe_fill(sh
, &s
, disks
);
3567 /* Now to consider new write requests and what else, if anything
3568 * should be read. We do not handle new writes when:
3569 * 1/ A 'write' operation (copy+xor) is already in flight.
3570 * 2/ A 'check' operation is in flight, as it may clobber the parity
3573 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
3574 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
3576 /* maybe we need to check and possibly fix the parity for this stripe
3577 * Any reads will already have been scheduled, so we just see if enough
3578 * data is available. The parity check is held off while parity
3579 * dependent operations are in flight.
3581 if (sh
->check_state
||
3582 (s
.syncing
&& s
.locked
== 0 &&
3583 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
3584 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
3585 if (conf
->level
== 6)
3586 handle_parity_checks6(conf
, sh
, &s
, disks
);
3588 handle_parity_checks5(conf
, sh
, &s
, disks
);
3591 if (s
.replacing
&& s
.locked
== 0
3592 && !test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3593 /* Write out to replacement devices where possible */
3594 for (i
= 0; i
< conf
->raid_disks
; i
++)
3595 if (test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
) &&
3596 test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
3597 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
3598 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3601 set_bit(STRIPE_INSYNC
, &sh
->state
);
3603 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
3604 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3605 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3606 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3609 /* If the failed drives are just a ReadError, then we might need
3610 * to progress the repair/check process
3612 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
3613 for (i
= 0; i
< s
.failed
; i
++) {
3614 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
3615 if (test_bit(R5_ReadError
, &dev
->flags
)
3616 && !test_bit(R5_LOCKED
, &dev
->flags
)
3617 && test_bit(R5_UPTODATE
, &dev
->flags
)
3619 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3620 set_bit(R5_Wantwrite
, &dev
->flags
);
3621 set_bit(R5_ReWrite
, &dev
->flags
);
3622 set_bit(R5_LOCKED
, &dev
->flags
);
3625 /* let's read it back */
3626 set_bit(R5_Wantread
, &dev
->flags
);
3627 set_bit(R5_LOCKED
, &dev
->flags
);
3634 /* Finish reconstruct operations initiated by the expansion process */
3635 if (sh
->reconstruct_state
== reconstruct_state_result
) {
3636 struct stripe_head
*sh_src
3637 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
3638 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
3639 /* sh cannot be written until sh_src has been read.
3640 * so arrange for sh to be delayed a little
3642 set_bit(STRIPE_DELAYED
, &sh
->state
);
3643 set_bit(STRIPE_HANDLE
, &sh
->state
);
3644 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
3646 atomic_inc(&conf
->preread_active_stripes
);
3647 release_stripe(sh_src
);
3651 release_stripe(sh_src
);
3653 sh
->reconstruct_state
= reconstruct_state_idle
;
3654 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3655 for (i
= conf
->raid_disks
; i
--; ) {
3656 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3657 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3662 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
3663 !sh
->reconstruct_state
) {
3664 /* Need to write out all blocks after computing parity */
3665 sh
->disks
= conf
->raid_disks
;
3666 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
3667 schedule_reconstruction(sh
, &s
, 1, 1);
3668 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
3669 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3670 atomic_dec(&conf
->reshape_stripes
);
3671 wake_up(&conf
->wait_for_overlap
);
3672 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3675 if (s
.expanding
&& s
.locked
== 0 &&
3676 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
3677 handle_stripe_expansion(conf
, sh
);
3680 /* wait for this device to become unblocked */
3681 if (unlikely(s
.blocked_rdev
)) {
3682 if (conf
->mddev
->external
)
3683 md_wait_for_blocked_rdev(s
.blocked_rdev
,
3686 /* Internal metadata will immediately
3687 * be written by raid5d, so we don't
3688 * need to wait here.
3690 rdev_dec_pending(s
.blocked_rdev
,
3694 if (s
.handle_bad_blocks
)
3695 for (i
= disks
; i
--; ) {
3696 struct md_rdev
*rdev
;
3697 struct r5dev
*dev
= &sh
->dev
[i
];
3698 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
3699 /* We own a safe reference to the rdev */
3700 rdev
= conf
->disks
[i
].rdev
;
3701 if (!rdev_set_badblocks(rdev
, sh
->sector
,
3703 md_error(conf
->mddev
, rdev
);
3704 rdev_dec_pending(rdev
, conf
->mddev
);
3706 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
3707 rdev
= conf
->disks
[i
].rdev
;
3708 rdev_clear_badblocks(rdev
, sh
->sector
,
3710 rdev_dec_pending(rdev
, conf
->mddev
);
3712 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
3713 rdev
= conf
->disks
[i
].replacement
;
3715 /* rdev have been moved down */
3716 rdev
= conf
->disks
[i
].rdev
;
3717 rdev_clear_badblocks(rdev
, sh
->sector
,
3719 rdev_dec_pending(rdev
, conf
->mddev
);
3724 raid_run_ops(sh
, s
.ops_request
);
3728 if (s
.dec_preread_active
) {
3729 /* We delay this until after ops_run_io so that if make_request
3730 * is waiting on a flush, it won't continue until the writes
3731 * have actually been submitted.
3733 atomic_dec(&conf
->preread_active_stripes
);
3734 if (atomic_read(&conf
->preread_active_stripes
) <
3736 md_wakeup_thread(conf
->mddev
->thread
);
3739 return_io(s
.return_bi
);
3741 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
3744 static void raid5_activate_delayed(struct r5conf
*conf
)
3746 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
3747 while (!list_empty(&conf
->delayed_list
)) {
3748 struct list_head
*l
= conf
->delayed_list
.next
;
3749 struct stripe_head
*sh
;
3750 sh
= list_entry(l
, struct stripe_head
, lru
);
3752 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3753 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3754 atomic_inc(&conf
->preread_active_stripes
);
3755 list_add_tail(&sh
->lru
, &conf
->hold_list
);
3760 static void activate_bit_delay(struct r5conf
*conf
)
3762 /* device_lock is held */
3763 struct list_head head
;
3764 list_add(&head
, &conf
->bitmap_list
);
3765 list_del_init(&conf
->bitmap_list
);
3766 while (!list_empty(&head
)) {
3767 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
3768 list_del_init(&sh
->lru
);
3769 atomic_inc(&sh
->count
);
3770 __release_stripe(conf
, sh
);
3774 int md_raid5_congested(struct mddev
*mddev
, int bits
)
3776 struct r5conf
*conf
= mddev
->private;
3778 /* No difference between reads and writes. Just check
3779 * how busy the stripe_cache is
3782 if (conf
->inactive_blocked
)
3786 if (list_empty_careful(&conf
->inactive_list
))
3791 EXPORT_SYMBOL_GPL(md_raid5_congested
);
3793 static int raid5_congested(void *data
, int bits
)
3795 struct mddev
*mddev
= data
;
3797 return mddev_congested(mddev
, bits
) ||
3798 md_raid5_congested(mddev
, bits
);
3801 /* We want read requests to align with chunks where possible,
3802 * but write requests don't need to.
3804 static int raid5_mergeable_bvec(struct request_queue
*q
,
3805 struct bvec_merge_data
*bvm
,
3806 struct bio_vec
*biovec
)
3808 struct mddev
*mddev
= q
->queuedata
;
3809 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
3811 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3812 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
3814 if ((bvm
->bi_rw
& 1) == WRITE
)
3815 return biovec
->bv_len
; /* always allow writes to be mergeable */
3817 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3818 chunk_sectors
= mddev
->new_chunk_sectors
;
3819 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
3820 if (max
< 0) max
= 0;
3821 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
3822 return biovec
->bv_len
;
3828 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
3830 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
3831 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3832 unsigned int bio_sectors
= bio
->bi_size
>> 9;
3834 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3835 chunk_sectors
= mddev
->new_chunk_sectors
;
3836 return chunk_sectors
>=
3837 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
3841 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3842 * later sampled by raid5d.
3844 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
3846 unsigned long flags
;
3848 spin_lock_irqsave(&conf
->device_lock
, flags
);
3850 bi
->bi_next
= conf
->retry_read_aligned_list
;
3851 conf
->retry_read_aligned_list
= bi
;
3853 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3854 md_wakeup_thread(conf
->mddev
->thread
);
3858 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
)
3862 bi
= conf
->retry_read_aligned
;
3864 conf
->retry_read_aligned
= NULL
;
3867 bi
= conf
->retry_read_aligned_list
;
3869 conf
->retry_read_aligned_list
= bi
->bi_next
;
3872 * this sets the active strip count to 1 and the processed
3873 * strip count to zero (upper 8 bits)
3875 raid5_set_bi_stripes(bi
, 1); /* biased count of active stripes */
3883 * The "raid5_align_endio" should check if the read succeeded and if it
3884 * did, call bio_endio on the original bio (having bio_put the new bio
3886 * If the read failed..
3888 static void raid5_align_endio(struct bio
*bi
, int error
)
3890 struct bio
* raid_bi
= bi
->bi_private
;
3891 struct mddev
*mddev
;
3892 struct r5conf
*conf
;
3893 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3894 struct md_rdev
*rdev
;
3898 rdev
= (void*)raid_bi
->bi_next
;
3899 raid_bi
->bi_next
= NULL
;
3900 mddev
= rdev
->mddev
;
3901 conf
= mddev
->private;
3903 rdev_dec_pending(rdev
, conf
->mddev
);
3905 if (!error
&& uptodate
) {
3906 bio_endio(raid_bi
, 0);
3907 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
3908 wake_up(&conf
->wait_for_stripe
);
3913 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3915 add_bio_to_retry(raid_bi
, conf
);
3918 static int bio_fits_rdev(struct bio
*bi
)
3920 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
3922 if ((bi
->bi_size
>>9) > queue_max_sectors(q
))
3924 blk_recount_segments(q
, bi
);
3925 if (bi
->bi_phys_segments
> queue_max_segments(q
))
3928 if (q
->merge_bvec_fn
)
3929 /* it's too hard to apply the merge_bvec_fn at this stage,
3938 static int chunk_aligned_read(struct mddev
*mddev
, struct bio
* raid_bio
)
3940 struct r5conf
*conf
= mddev
->private;
3942 struct bio
* align_bi
;
3943 struct md_rdev
*rdev
;
3944 sector_t end_sector
;
3946 if (!in_chunk_boundary(mddev
, raid_bio
)) {
3947 pr_debug("chunk_aligned_read : non aligned\n");
3951 * use bio_clone_mddev to make a copy of the bio
3953 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
3957 * set bi_end_io to a new function, and set bi_private to the
3960 align_bi
->bi_end_io
= raid5_align_endio
;
3961 align_bi
->bi_private
= raid_bio
;
3965 align_bi
->bi_sector
= raid5_compute_sector(conf
, raid_bio
->bi_sector
,
3969 end_sector
= align_bi
->bi_sector
+ (align_bi
->bi_size
>> 9);
3971 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
3972 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
3973 rdev
->recovery_offset
< end_sector
) {
3974 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
3976 (test_bit(Faulty
, &rdev
->flags
) ||
3977 !(test_bit(In_sync
, &rdev
->flags
) ||
3978 rdev
->recovery_offset
>= end_sector
)))
3985 atomic_inc(&rdev
->nr_pending
);
3987 raid_bio
->bi_next
= (void*)rdev
;
3988 align_bi
->bi_bdev
= rdev
->bdev
;
3989 align_bi
->bi_flags
&= ~(1 << BIO_SEG_VALID
);
3991 if (!bio_fits_rdev(align_bi
) ||
3992 is_badblock(rdev
, align_bi
->bi_sector
, align_bi
->bi_size
>>9,
3993 &first_bad
, &bad_sectors
)) {
3994 /* too big in some way, or has a known bad block */
3996 rdev_dec_pending(rdev
, mddev
);
4000 /* No reshape active, so we can trust rdev->data_offset */
4001 align_bi
->bi_sector
+= rdev
->data_offset
;
4003 spin_lock_irq(&conf
->device_lock
);
4004 wait_event_lock_irq(conf
->wait_for_stripe
,
4006 conf
->device_lock
, /* nothing */);
4007 atomic_inc(&conf
->active_aligned_reads
);
4008 spin_unlock_irq(&conf
->device_lock
);
4010 generic_make_request(align_bi
);
4019 /* __get_priority_stripe - get the next stripe to process
4021 * Full stripe writes are allowed to pass preread active stripes up until
4022 * the bypass_threshold is exceeded. In general the bypass_count
4023 * increments when the handle_list is handled before the hold_list; however, it
4024 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4025 * stripe with in flight i/o. The bypass_count will be reset when the
4026 * head of the hold_list has changed, i.e. the head was promoted to the
4029 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
)
4031 struct stripe_head
*sh
;
4033 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4035 list_empty(&conf
->handle_list
) ? "empty" : "busy",
4036 list_empty(&conf
->hold_list
) ? "empty" : "busy",
4037 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
4039 if (!list_empty(&conf
->handle_list
)) {
4040 sh
= list_entry(conf
->handle_list
.next
, typeof(*sh
), lru
);
4042 if (list_empty(&conf
->hold_list
))
4043 conf
->bypass_count
= 0;
4044 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
4045 if (conf
->hold_list
.next
== conf
->last_hold
)
4046 conf
->bypass_count
++;
4048 conf
->last_hold
= conf
->hold_list
.next
;
4049 conf
->bypass_count
-= conf
->bypass_threshold
;
4050 if (conf
->bypass_count
< 0)
4051 conf
->bypass_count
= 0;
4054 } else if (!list_empty(&conf
->hold_list
) &&
4055 ((conf
->bypass_threshold
&&
4056 conf
->bypass_count
> conf
->bypass_threshold
) ||
4057 atomic_read(&conf
->pending_full_writes
) == 0)) {
4058 sh
= list_entry(conf
->hold_list
.next
,
4060 conf
->bypass_count
-= conf
->bypass_threshold
;
4061 if (conf
->bypass_count
< 0)
4062 conf
->bypass_count
= 0;
4066 list_del_init(&sh
->lru
);
4067 atomic_inc(&sh
->count
);
4068 BUG_ON(atomic_read(&sh
->count
) != 1);
4072 struct raid5_plug_cb
{
4073 struct blk_plug_cb cb
;
4074 struct list_head list
;
4077 static void raid5_unplug(struct blk_plug_cb
*blk_cb
, bool from_schedule
)
4079 struct raid5_plug_cb
*cb
= container_of(
4080 blk_cb
, struct raid5_plug_cb
, cb
);
4081 struct stripe_head
*sh
;
4082 struct mddev
*mddev
= cb
->cb
.data
;
4083 struct r5conf
*conf
= mddev
->private;
4085 if (cb
->list
.next
&& !list_empty(&cb
->list
)) {
4086 spin_lock_irq(&conf
->device_lock
);
4087 while (!list_empty(&cb
->list
)) {
4088 sh
= list_first_entry(&cb
->list
, struct stripe_head
, lru
);
4089 list_del_init(&sh
->lru
);
4091 * avoid race release_stripe_plug() sees
4092 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4093 * is still in our list
4095 smp_mb__before_clear_bit();
4096 clear_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
);
4097 __release_stripe(conf
, sh
);
4099 spin_unlock_irq(&conf
->device_lock
);
4104 static void release_stripe_plug(struct mddev
*mddev
,
4105 struct stripe_head
*sh
)
4107 struct blk_plug_cb
*blk_cb
= blk_check_plugged(
4108 raid5_unplug
, mddev
,
4109 sizeof(struct raid5_plug_cb
));
4110 struct raid5_plug_cb
*cb
;
4117 cb
= container_of(blk_cb
, struct raid5_plug_cb
, cb
);
4119 if (cb
->list
.next
== NULL
)
4120 INIT_LIST_HEAD(&cb
->list
);
4122 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
))
4123 list_add_tail(&sh
->lru
, &cb
->list
);
4128 static void make_discard_request(struct mddev
*mddev
, struct bio
*bi
)
4130 struct r5conf
*conf
= mddev
->private;
4131 sector_t logical_sector
, last_sector
;
4132 struct stripe_head
*sh
;
4136 if (mddev
->reshape_position
!= MaxSector
)
4137 /* Skip discard while reshape is happening */
4140 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4141 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
4144 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
4146 stripe_sectors
= conf
->chunk_sectors
*
4147 (conf
->raid_disks
- conf
->max_degraded
);
4148 logical_sector
= DIV_ROUND_UP_SECTOR_T(logical_sector
,
4150 sector_div(last_sector
, stripe_sectors
);
4152 logical_sector
*= conf
->chunk_sectors
;
4153 last_sector
*= conf
->chunk_sectors
;
4155 for (; logical_sector
< last_sector
;
4156 logical_sector
+= STRIPE_SECTORS
) {
4160 sh
= get_active_stripe(conf
, logical_sector
, 0, 0, 0);
4161 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
4162 TASK_UNINTERRUPTIBLE
);
4163 spin_lock_irq(&sh
->stripe_lock
);
4164 for (d
= 0; d
< conf
->raid_disks
; d
++) {
4165 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
4167 if (sh
->dev
[d
].towrite
|| sh
->dev
[d
].toread
) {
4168 set_bit(R5_Overlap
, &sh
->dev
[d
].flags
);
4169 spin_unlock_irq(&sh
->stripe_lock
);
4175 finish_wait(&conf
->wait_for_overlap
, &w
);
4176 for (d
= 0; d
< conf
->raid_disks
; d
++) {
4177 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
4179 sh
->dev
[d
].towrite
= bi
;
4180 set_bit(R5_OVERWRITE
, &sh
->dev
[d
].flags
);
4181 raid5_inc_bi_active_stripes(bi
);
4183 spin_unlock_irq(&sh
->stripe_lock
);
4184 if (conf
->mddev
->bitmap
) {
4186 d
< conf
->raid_disks
- conf
->max_degraded
;
4188 bitmap_startwrite(mddev
->bitmap
,
4192 sh
->bm_seq
= conf
->seq_flush
+ 1;
4193 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
4196 set_bit(STRIPE_HANDLE
, &sh
->state
);
4197 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4198 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4199 atomic_inc(&conf
->preread_active_stripes
);
4200 release_stripe_plug(mddev
, sh
);
4203 remaining
= raid5_dec_bi_active_stripes(bi
);
4204 if (remaining
== 0) {
4205 md_write_end(mddev
);
4210 static void make_request(struct mddev
*mddev
, struct bio
* bi
)
4212 struct r5conf
*conf
= mddev
->private;
4214 sector_t new_sector
;
4215 sector_t logical_sector
, last_sector
;
4216 struct stripe_head
*sh
;
4217 const int rw
= bio_data_dir(bi
);
4220 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
4221 md_flush_request(mddev
, bi
);
4225 md_write_start(mddev
, bi
);
4228 mddev
->reshape_position
== MaxSector
&&
4229 chunk_aligned_read(mddev
,bi
))
4232 if (unlikely(bi
->bi_rw
& REQ_DISCARD
)) {
4233 make_discard_request(mddev
, bi
);
4237 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4238 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
4240 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
4242 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
4248 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
4249 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
4250 /* spinlock is needed as reshape_progress may be
4251 * 64bit on a 32bit platform, and so it might be
4252 * possible to see a half-updated value
4253 * Of course reshape_progress could change after
4254 * the lock is dropped, so once we get a reference
4255 * to the stripe that we think it is, we will have
4258 spin_lock_irq(&conf
->device_lock
);
4259 if (mddev
->reshape_backwards
4260 ? logical_sector
< conf
->reshape_progress
4261 : logical_sector
>= conf
->reshape_progress
) {
4264 if (mddev
->reshape_backwards
4265 ? logical_sector
< conf
->reshape_safe
4266 : logical_sector
>= conf
->reshape_safe
) {
4267 spin_unlock_irq(&conf
->device_lock
);
4272 spin_unlock_irq(&conf
->device_lock
);
4275 new_sector
= raid5_compute_sector(conf
, logical_sector
,
4278 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4279 (unsigned long long)new_sector
,
4280 (unsigned long long)logical_sector
);
4282 sh
= get_active_stripe(conf
, new_sector
, previous
,
4283 (bi
->bi_rw
&RWA_MASK
), 0);
4285 if (unlikely(previous
)) {
4286 /* expansion might have moved on while waiting for a
4287 * stripe, so we must do the range check again.
4288 * Expansion could still move past after this
4289 * test, but as we are holding a reference to
4290 * 'sh', we know that if that happens,
4291 * STRIPE_EXPANDING will get set and the expansion
4292 * won't proceed until we finish with the stripe.
4295 spin_lock_irq(&conf
->device_lock
);
4296 if (mddev
->reshape_backwards
4297 ? logical_sector
>= conf
->reshape_progress
4298 : logical_sector
< conf
->reshape_progress
)
4299 /* mismatch, need to try again */
4301 spin_unlock_irq(&conf
->device_lock
);
4310 logical_sector
>= mddev
->suspend_lo
&&
4311 logical_sector
< mddev
->suspend_hi
) {
4313 /* As the suspend_* range is controlled by
4314 * userspace, we want an interruptible
4317 flush_signals(current
);
4318 prepare_to_wait(&conf
->wait_for_overlap
,
4319 &w
, TASK_INTERRUPTIBLE
);
4320 if (logical_sector
>= mddev
->suspend_lo
&&
4321 logical_sector
< mddev
->suspend_hi
)
4326 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
4327 !add_stripe_bio(sh
, bi
, dd_idx
, rw
)) {
4328 /* Stripe is busy expanding or
4329 * add failed due to overlap. Flush everything
4332 md_wakeup_thread(mddev
->thread
);
4337 finish_wait(&conf
->wait_for_overlap
, &w
);
4338 set_bit(STRIPE_HANDLE
, &sh
->state
);
4339 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4340 if ((bi
->bi_rw
& REQ_SYNC
) &&
4341 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4342 atomic_inc(&conf
->preread_active_stripes
);
4343 release_stripe_plug(mddev
, sh
);
4345 /* cannot get stripe for read-ahead, just give-up */
4346 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
4347 finish_wait(&conf
->wait_for_overlap
, &w
);
4352 remaining
= raid5_dec_bi_active_stripes(bi
);
4353 if (remaining
== 0) {
4356 md_write_end(mddev
);
4362 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
4364 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
4366 /* reshaping is quite different to recovery/resync so it is
4367 * handled quite separately ... here.
4369 * On each call to sync_request, we gather one chunk worth of
4370 * destination stripes and flag them as expanding.
4371 * Then we find all the source stripes and request reads.
4372 * As the reads complete, handle_stripe will copy the data
4373 * into the destination stripe and release that stripe.
4375 struct r5conf
*conf
= mddev
->private;
4376 struct stripe_head
*sh
;
4377 sector_t first_sector
, last_sector
;
4378 int raid_disks
= conf
->previous_raid_disks
;
4379 int data_disks
= raid_disks
- conf
->max_degraded
;
4380 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
4383 sector_t writepos
, readpos
, safepos
;
4384 sector_t stripe_addr
;
4385 int reshape_sectors
;
4386 struct list_head stripes
;
4388 if (sector_nr
== 0) {
4389 /* If restarting in the middle, skip the initial sectors */
4390 if (mddev
->reshape_backwards
&&
4391 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
4392 sector_nr
= raid5_size(mddev
, 0, 0)
4393 - conf
->reshape_progress
;
4394 } else if (!mddev
->reshape_backwards
&&
4395 conf
->reshape_progress
> 0)
4396 sector_nr
= conf
->reshape_progress
;
4397 sector_div(sector_nr
, new_data_disks
);
4399 mddev
->curr_resync_completed
= sector_nr
;
4400 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4406 /* We need to process a full chunk at a time.
4407 * If old and new chunk sizes differ, we need to process the
4410 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
)
4411 reshape_sectors
= mddev
->new_chunk_sectors
;
4413 reshape_sectors
= mddev
->chunk_sectors
;
4415 /* We update the metadata at least every 10 seconds, or when
4416 * the data about to be copied would over-write the source of
4417 * the data at the front of the range. i.e. one new_stripe
4418 * along from reshape_progress new_maps to after where
4419 * reshape_safe old_maps to
4421 writepos
= conf
->reshape_progress
;
4422 sector_div(writepos
, new_data_disks
);
4423 readpos
= conf
->reshape_progress
;
4424 sector_div(readpos
, data_disks
);
4425 safepos
= conf
->reshape_safe
;
4426 sector_div(safepos
, data_disks
);
4427 if (mddev
->reshape_backwards
) {
4428 writepos
-= min_t(sector_t
, reshape_sectors
, writepos
);
4429 readpos
+= reshape_sectors
;
4430 safepos
+= reshape_sectors
;
4432 writepos
+= reshape_sectors
;
4433 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
4434 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
4437 /* Having calculated the 'writepos' possibly use it
4438 * to set 'stripe_addr' which is where we will write to.
4440 if (mddev
->reshape_backwards
) {
4441 BUG_ON(conf
->reshape_progress
== 0);
4442 stripe_addr
= writepos
;
4443 BUG_ON((mddev
->dev_sectors
&
4444 ~((sector_t
)reshape_sectors
- 1))
4445 - reshape_sectors
- stripe_addr
4448 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
4449 stripe_addr
= sector_nr
;
4452 /* 'writepos' is the most advanced device address we might write.
4453 * 'readpos' is the least advanced device address we might read.
4454 * 'safepos' is the least address recorded in the metadata as having
4456 * If there is a min_offset_diff, these are adjusted either by
4457 * increasing the safepos/readpos if diff is negative, or
4458 * increasing writepos if diff is positive.
4459 * If 'readpos' is then behind 'writepos', there is no way that we can
4460 * ensure safety in the face of a crash - that must be done by userspace
4461 * making a backup of the data. So in that case there is no particular
4462 * rush to update metadata.
4463 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4464 * update the metadata to advance 'safepos' to match 'readpos' so that
4465 * we can be safe in the event of a crash.
4466 * So we insist on updating metadata if safepos is behind writepos and
4467 * readpos is beyond writepos.
4468 * In any case, update the metadata every 10 seconds.
4469 * Maybe that number should be configurable, but I'm not sure it is
4470 * worth it.... maybe it could be a multiple of safemode_delay???
4472 if (conf
->min_offset_diff
< 0) {
4473 safepos
+= -conf
->min_offset_diff
;
4474 readpos
+= -conf
->min_offset_diff
;
4476 writepos
+= conf
->min_offset_diff
;
4478 if ((mddev
->reshape_backwards
4479 ? (safepos
> writepos
&& readpos
< writepos
)
4480 : (safepos
< writepos
&& readpos
> writepos
)) ||
4481 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4482 /* Cannot proceed until we've updated the superblock... */
4483 wait_event(conf
->wait_for_overlap
,
4484 atomic_read(&conf
->reshape_stripes
)==0);
4485 mddev
->reshape_position
= conf
->reshape_progress
;
4486 mddev
->curr_resync_completed
= sector_nr
;
4487 conf
->reshape_checkpoint
= jiffies
;
4488 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4489 md_wakeup_thread(mddev
->thread
);
4490 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4491 kthread_should_stop());
4492 spin_lock_irq(&conf
->device_lock
);
4493 conf
->reshape_safe
= mddev
->reshape_position
;
4494 spin_unlock_irq(&conf
->device_lock
);
4495 wake_up(&conf
->wait_for_overlap
);
4496 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4499 INIT_LIST_HEAD(&stripes
);
4500 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
4502 int skipped_disk
= 0;
4503 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
4504 set_bit(STRIPE_EXPANDING
, &sh
->state
);
4505 atomic_inc(&conf
->reshape_stripes
);
4506 /* If any of this stripe is beyond the end of the old
4507 * array, then we need to zero those blocks
4509 for (j
=sh
->disks
; j
--;) {
4511 if (j
== sh
->pd_idx
)
4513 if (conf
->level
== 6 &&
4516 s
= compute_blocknr(sh
, j
, 0);
4517 if (s
< raid5_size(mddev
, 0, 0)) {
4521 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
4522 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
4523 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
4525 if (!skipped_disk
) {
4526 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4527 set_bit(STRIPE_HANDLE
, &sh
->state
);
4529 list_add(&sh
->lru
, &stripes
);
4531 spin_lock_irq(&conf
->device_lock
);
4532 if (mddev
->reshape_backwards
)
4533 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
4535 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
4536 spin_unlock_irq(&conf
->device_lock
);
4537 /* Ok, those stripe are ready. We can start scheduling
4538 * reads on the source stripes.
4539 * The source stripes are determined by mapping the first and last
4540 * block on the destination stripes.
4543 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
4546 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
4547 * new_data_disks
- 1),
4549 if (last_sector
>= mddev
->dev_sectors
)
4550 last_sector
= mddev
->dev_sectors
- 1;
4551 while (first_sector
<= last_sector
) {
4552 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
4553 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
4554 set_bit(STRIPE_HANDLE
, &sh
->state
);
4556 first_sector
+= STRIPE_SECTORS
;
4558 /* Now that the sources are clearly marked, we can release
4559 * the destination stripes
4561 while (!list_empty(&stripes
)) {
4562 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
4563 list_del_init(&sh
->lru
);
4566 /* If this takes us to the resync_max point where we have to pause,
4567 * then we need to write out the superblock.
4569 sector_nr
+= reshape_sectors
;
4570 if ((sector_nr
- mddev
->curr_resync_completed
) * 2
4571 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
4572 /* Cannot proceed until we've updated the superblock... */
4573 wait_event(conf
->wait_for_overlap
,
4574 atomic_read(&conf
->reshape_stripes
) == 0);
4575 mddev
->reshape_position
= conf
->reshape_progress
;
4576 mddev
->curr_resync_completed
= sector_nr
;
4577 conf
->reshape_checkpoint
= jiffies
;
4578 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4579 md_wakeup_thread(mddev
->thread
);
4580 wait_event(mddev
->sb_wait
,
4581 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
4582 || kthread_should_stop());
4583 spin_lock_irq(&conf
->device_lock
);
4584 conf
->reshape_safe
= mddev
->reshape_position
;
4585 spin_unlock_irq(&conf
->device_lock
);
4586 wake_up(&conf
->wait_for_overlap
);
4587 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4589 return reshape_sectors
;
4592 /* FIXME go_faster isn't used */
4593 static inline sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
4595 struct r5conf
*conf
= mddev
->private;
4596 struct stripe_head
*sh
;
4597 sector_t max_sector
= mddev
->dev_sectors
;
4598 sector_t sync_blocks
;
4599 int still_degraded
= 0;
4602 if (sector_nr
>= max_sector
) {
4603 /* just being told to finish up .. nothing much to do */
4605 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
4610 if (mddev
->curr_resync
< max_sector
) /* aborted */
4611 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
4613 else /* completed sync */
4615 bitmap_close_sync(mddev
->bitmap
);
4620 /* Allow raid5_quiesce to complete */
4621 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
4623 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
4624 return reshape_request(mddev
, sector_nr
, skipped
);
4626 /* No need to check resync_max as we never do more than one
4627 * stripe, and as resync_max will always be on a chunk boundary,
4628 * if the check in md_do_sync didn't fire, there is no chance
4629 * of overstepping resync_max here
4632 /* if there is too many failed drives and we are trying
4633 * to resync, then assert that we are finished, because there is
4634 * nothing we can do.
4636 if (mddev
->degraded
>= conf
->max_degraded
&&
4637 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
4638 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
4642 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
4643 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
4644 !conf
->fullsync
&& sync_blocks
>= STRIPE_SECTORS
) {
4645 /* we can skip this block, and probably more */
4646 sync_blocks
/= STRIPE_SECTORS
;
4648 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
4651 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
4653 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
4655 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
4656 /* make sure we don't swamp the stripe cache if someone else
4657 * is trying to get access
4659 schedule_timeout_uninterruptible(1);
4661 /* Need to check if array will still be degraded after recovery/resync
4662 * We don't need to check the 'failed' flag as when that gets set,
4665 for (i
= 0; i
< conf
->raid_disks
; i
++)
4666 if (conf
->disks
[i
].rdev
== NULL
)
4669 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
4671 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
4676 return STRIPE_SECTORS
;
4679 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
)
4681 /* We may not be able to submit a whole bio at once as there
4682 * may not be enough stripe_heads available.
4683 * We cannot pre-allocate enough stripe_heads as we may need
4684 * more than exist in the cache (if we allow ever large chunks).
4685 * So we do one stripe head at a time and record in
4686 * ->bi_hw_segments how many have been done.
4688 * We *know* that this entire raid_bio is in one chunk, so
4689 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4691 struct stripe_head
*sh
;
4693 sector_t sector
, logical_sector
, last_sector
;
4698 logical_sector
= raid_bio
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4699 sector
= raid5_compute_sector(conf
, logical_sector
,
4701 last_sector
= raid_bio
->bi_sector
+ (raid_bio
->bi_size
>>9);
4703 for (; logical_sector
< last_sector
;
4704 logical_sector
+= STRIPE_SECTORS
,
4705 sector
+= STRIPE_SECTORS
,
4708 if (scnt
< raid5_bi_processed_stripes(raid_bio
))
4709 /* already done this stripe */
4712 sh
= get_active_stripe(conf
, sector
, 0, 1, 0);
4715 /* failed to get a stripe - must wait */
4716 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
4717 conf
->retry_read_aligned
= raid_bio
;
4721 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0)) {
4723 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
4724 conf
->retry_read_aligned
= raid_bio
;
4728 set_bit(R5_ReadNoMerge
, &sh
->dev
[dd_idx
].flags
);
4733 remaining
= raid5_dec_bi_active_stripes(raid_bio
);
4735 bio_endio(raid_bio
, 0);
4736 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4737 wake_up(&conf
->wait_for_stripe
);
4741 #define MAX_STRIPE_BATCH 8
4742 static int handle_active_stripes(struct r5conf
*conf
)
4744 struct stripe_head
*batch
[MAX_STRIPE_BATCH
], *sh
;
4745 int i
, batch_size
= 0;
4747 while (batch_size
< MAX_STRIPE_BATCH
&&
4748 (sh
= __get_priority_stripe(conf
)) != NULL
)
4749 batch
[batch_size
++] = sh
;
4751 if (batch_size
== 0)
4753 spin_unlock_irq(&conf
->device_lock
);
4755 for (i
= 0; i
< batch_size
; i
++)
4756 handle_stripe(batch
[i
]);
4760 spin_lock_irq(&conf
->device_lock
);
4761 for (i
= 0; i
< batch_size
; i
++)
4762 __release_stripe(conf
, batch
[i
]);
4767 * This is our raid5 kernel thread.
4769 * We scan the hash table for stripes which can be handled now.
4770 * During the scan, completed stripes are saved for us by the interrupt
4771 * handler, so that they will not have to wait for our next wakeup.
4773 static void raid5d(struct md_thread
*thread
)
4775 struct mddev
*mddev
= thread
->mddev
;
4776 struct r5conf
*conf
= mddev
->private;
4778 struct blk_plug plug
;
4780 pr_debug("+++ raid5d active\n");
4782 md_check_recovery(mddev
);
4784 blk_start_plug(&plug
);
4786 spin_lock_irq(&conf
->device_lock
);
4792 !list_empty(&conf
->bitmap_list
)) {
4793 /* Now is a good time to flush some bitmap updates */
4795 spin_unlock_irq(&conf
->device_lock
);
4796 bitmap_unplug(mddev
->bitmap
);
4797 spin_lock_irq(&conf
->device_lock
);
4798 conf
->seq_write
= conf
->seq_flush
;
4799 activate_bit_delay(conf
);
4801 raid5_activate_delayed(conf
);
4803 while ((bio
= remove_bio_from_retry(conf
))) {
4805 spin_unlock_irq(&conf
->device_lock
);
4806 ok
= retry_aligned_read(conf
, bio
);
4807 spin_lock_irq(&conf
->device_lock
);
4813 batch_size
= handle_active_stripes(conf
);
4816 handled
+= batch_size
;
4818 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
)) {
4819 spin_unlock_irq(&conf
->device_lock
);
4820 md_check_recovery(mddev
);
4821 spin_lock_irq(&conf
->device_lock
);
4824 pr_debug("%d stripes handled\n", handled
);
4826 spin_unlock_irq(&conf
->device_lock
);
4828 async_tx_issue_pending_all();
4829 blk_finish_plug(&plug
);
4831 pr_debug("--- raid5d inactive\n");
4835 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
4837 struct r5conf
*conf
= mddev
->private;
4839 return sprintf(page
, "%d\n", conf
->max_nr_stripes
);
4845 raid5_set_cache_size(struct mddev
*mddev
, int size
)
4847 struct r5conf
*conf
= mddev
->private;
4850 if (size
<= 16 || size
> 32768)
4852 while (size
< conf
->max_nr_stripes
) {
4853 if (drop_one_stripe(conf
))
4854 conf
->max_nr_stripes
--;
4858 err
= md_allow_write(mddev
);
4861 while (size
> conf
->max_nr_stripes
) {
4862 if (grow_one_stripe(conf
))
4863 conf
->max_nr_stripes
++;
4868 EXPORT_SYMBOL(raid5_set_cache_size
);
4871 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
4873 struct r5conf
*conf
= mddev
->private;
4877 if (len
>= PAGE_SIZE
)
4882 if (strict_strtoul(page
, 10, &new))
4884 err
= raid5_set_cache_size(mddev
, new);
4890 static struct md_sysfs_entry
4891 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
4892 raid5_show_stripe_cache_size
,
4893 raid5_store_stripe_cache_size
);
4896 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
4898 struct r5conf
*conf
= mddev
->private;
4900 return sprintf(page
, "%d\n", conf
->bypass_threshold
);
4906 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
4908 struct r5conf
*conf
= mddev
->private;
4910 if (len
>= PAGE_SIZE
)
4915 if (strict_strtoul(page
, 10, &new))
4917 if (new > conf
->max_nr_stripes
)
4919 conf
->bypass_threshold
= new;
4923 static struct md_sysfs_entry
4924 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
4926 raid5_show_preread_threshold
,
4927 raid5_store_preread_threshold
);
4930 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
4932 struct r5conf
*conf
= mddev
->private;
4934 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
4939 static struct md_sysfs_entry
4940 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
4942 static struct attribute
*raid5_attrs
[] = {
4943 &raid5_stripecache_size
.attr
,
4944 &raid5_stripecache_active
.attr
,
4945 &raid5_preread_bypass_threshold
.attr
,
4948 static struct attribute_group raid5_attrs_group
= {
4950 .attrs
= raid5_attrs
,
4954 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
4956 struct r5conf
*conf
= mddev
->private;
4959 sectors
= mddev
->dev_sectors
;
4961 /* size is defined by the smallest of previous and new size */
4962 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
4964 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
4965 sectors
&= ~((sector_t
)mddev
->new_chunk_sectors
- 1);
4966 return sectors
* (raid_disks
- conf
->max_degraded
);
4969 static void raid5_free_percpu(struct r5conf
*conf
)
4971 struct raid5_percpu
*percpu
;
4978 for_each_possible_cpu(cpu
) {
4979 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4980 safe_put_page(percpu
->spare_page
);
4981 kfree(percpu
->scribble
);
4983 #ifdef CONFIG_HOTPLUG_CPU
4984 unregister_cpu_notifier(&conf
->cpu_notify
);
4988 free_percpu(conf
->percpu
);
4991 static void free_conf(struct r5conf
*conf
)
4993 shrink_stripes(conf
);
4994 raid5_free_percpu(conf
);
4996 kfree(conf
->stripe_hashtbl
);
5000 #ifdef CONFIG_HOTPLUG_CPU
5001 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
5004 struct r5conf
*conf
= container_of(nfb
, struct r5conf
, cpu_notify
);
5005 long cpu
= (long)hcpu
;
5006 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
5009 case CPU_UP_PREPARE
:
5010 case CPU_UP_PREPARE_FROZEN
:
5011 if (conf
->level
== 6 && !percpu
->spare_page
)
5012 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
5013 if (!percpu
->scribble
)
5014 percpu
->scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
5016 if (!percpu
->scribble
||
5017 (conf
->level
== 6 && !percpu
->spare_page
)) {
5018 safe_put_page(percpu
->spare_page
);
5019 kfree(percpu
->scribble
);
5020 pr_err("%s: failed memory allocation for cpu%ld\n",
5022 return notifier_from_errno(-ENOMEM
);
5026 case CPU_DEAD_FROZEN
:
5027 safe_put_page(percpu
->spare_page
);
5028 kfree(percpu
->scribble
);
5029 percpu
->spare_page
= NULL
;
5030 percpu
->scribble
= NULL
;
5039 static int raid5_alloc_percpu(struct r5conf
*conf
)
5042 struct page
*spare_page
;
5043 struct raid5_percpu __percpu
*allcpus
;
5047 allcpus
= alloc_percpu(struct raid5_percpu
);
5050 conf
->percpu
= allcpus
;
5054 for_each_present_cpu(cpu
) {
5055 if (conf
->level
== 6) {
5056 spare_page
= alloc_page(GFP_KERNEL
);
5061 per_cpu_ptr(conf
->percpu
, cpu
)->spare_page
= spare_page
;
5063 scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
5068 per_cpu_ptr(conf
->percpu
, cpu
)->scribble
= scribble
;
5070 #ifdef CONFIG_HOTPLUG_CPU
5071 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
5072 conf
->cpu_notify
.priority
= 0;
5074 err
= register_cpu_notifier(&conf
->cpu_notify
);
5081 static struct r5conf
*setup_conf(struct mddev
*mddev
)
5083 struct r5conf
*conf
;
5084 int raid_disk
, memory
, max_disks
;
5085 struct md_rdev
*rdev
;
5086 struct disk_info
*disk
;
5089 if (mddev
->new_level
!= 5
5090 && mddev
->new_level
!= 4
5091 && mddev
->new_level
!= 6) {
5092 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5093 mdname(mddev
), mddev
->new_level
);
5094 return ERR_PTR(-EIO
);
5096 if ((mddev
->new_level
== 5
5097 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
5098 (mddev
->new_level
== 6
5099 && !algorithm_valid_raid6(mddev
->new_layout
))) {
5100 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
5101 mdname(mddev
), mddev
->new_layout
);
5102 return ERR_PTR(-EIO
);
5104 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
5105 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5106 mdname(mddev
), mddev
->raid_disks
);
5107 return ERR_PTR(-EINVAL
);
5110 if (!mddev
->new_chunk_sectors
||
5111 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
5112 !is_power_of_2(mddev
->new_chunk_sectors
)) {
5113 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
5114 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
5115 return ERR_PTR(-EINVAL
);
5118 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
5121 spin_lock_init(&conf
->device_lock
);
5122 init_waitqueue_head(&conf
->wait_for_stripe
);
5123 init_waitqueue_head(&conf
->wait_for_overlap
);
5124 INIT_LIST_HEAD(&conf
->handle_list
);
5125 INIT_LIST_HEAD(&conf
->hold_list
);
5126 INIT_LIST_HEAD(&conf
->delayed_list
);
5127 INIT_LIST_HEAD(&conf
->bitmap_list
);
5128 INIT_LIST_HEAD(&conf
->inactive_list
);
5129 atomic_set(&conf
->active_stripes
, 0);
5130 atomic_set(&conf
->preread_active_stripes
, 0);
5131 atomic_set(&conf
->active_aligned_reads
, 0);
5132 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
5133 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
5135 conf
->raid_disks
= mddev
->raid_disks
;
5136 if (mddev
->reshape_position
== MaxSector
)
5137 conf
->previous_raid_disks
= mddev
->raid_disks
;
5139 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
5140 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
5141 conf
->scribble_len
= scribble_len(max_disks
);
5143 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
5148 conf
->mddev
= mddev
;
5150 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
5153 conf
->level
= mddev
->new_level
;
5154 if (raid5_alloc_percpu(conf
) != 0)
5157 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
5159 rdev_for_each(rdev
, mddev
) {
5160 raid_disk
= rdev
->raid_disk
;
5161 if (raid_disk
>= max_disks
5164 disk
= conf
->disks
+ raid_disk
;
5166 if (test_bit(Replacement
, &rdev
->flags
)) {
5167 if (disk
->replacement
)
5169 disk
->replacement
= rdev
;
5176 if (test_bit(In_sync
, &rdev
->flags
)) {
5177 char b
[BDEVNAME_SIZE
];
5178 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
5180 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
5181 } else if (rdev
->saved_raid_disk
!= raid_disk
)
5182 /* Cannot rely on bitmap to complete recovery */
5186 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
5187 conf
->level
= mddev
->new_level
;
5188 if (conf
->level
== 6)
5189 conf
->max_degraded
= 2;
5191 conf
->max_degraded
= 1;
5192 conf
->algorithm
= mddev
->new_layout
;
5193 conf
->max_nr_stripes
= NR_STRIPES
;
5194 conf
->reshape_progress
= mddev
->reshape_position
;
5195 if (conf
->reshape_progress
!= MaxSector
) {
5196 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
5197 conf
->prev_algo
= mddev
->layout
;
5200 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
5201 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
5202 if (grow_stripes(conf
, conf
->max_nr_stripes
)) {
5204 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5205 mdname(mddev
), memory
);
5208 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
5209 mdname(mddev
), memory
);
5211 sprintf(pers_name
, "raid%d", mddev
->new_level
);
5212 conf
->thread
= md_register_thread(raid5d
, mddev
, pers_name
);
5213 if (!conf
->thread
) {
5215 "md/raid:%s: couldn't allocate thread.\n",
5225 return ERR_PTR(-EIO
);
5227 return ERR_PTR(-ENOMEM
);
5231 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
5234 case ALGORITHM_PARITY_0
:
5235 if (raid_disk
< max_degraded
)
5238 case ALGORITHM_PARITY_N
:
5239 if (raid_disk
>= raid_disks
- max_degraded
)
5242 case ALGORITHM_PARITY_0_6
:
5243 if (raid_disk
== 0 ||
5244 raid_disk
== raid_disks
- 1)
5247 case ALGORITHM_LEFT_ASYMMETRIC_6
:
5248 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
5249 case ALGORITHM_LEFT_SYMMETRIC_6
:
5250 case ALGORITHM_RIGHT_SYMMETRIC_6
:
5251 if (raid_disk
== raid_disks
- 1)
5257 static int run(struct mddev
*mddev
)
5259 struct r5conf
*conf
;
5260 int working_disks
= 0;
5261 int dirty_parity_disks
= 0;
5262 struct md_rdev
*rdev
;
5263 sector_t reshape_offset
= 0;
5265 long long min_offset_diff
= 0;
5268 if (mddev
->recovery_cp
!= MaxSector
)
5269 printk(KERN_NOTICE
"md/raid:%s: not clean"
5270 " -- starting background reconstruction\n",
5273 rdev_for_each(rdev
, mddev
) {
5275 if (rdev
->raid_disk
< 0)
5277 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
5279 min_offset_diff
= diff
;
5281 } else if (mddev
->reshape_backwards
&&
5282 diff
< min_offset_diff
)
5283 min_offset_diff
= diff
;
5284 else if (!mddev
->reshape_backwards
&&
5285 diff
> min_offset_diff
)
5286 min_offset_diff
= diff
;
5289 if (mddev
->reshape_position
!= MaxSector
) {
5290 /* Check that we can continue the reshape.
5291 * Difficulties arise if the stripe we would write to
5292 * next is at or after the stripe we would read from next.
5293 * For a reshape that changes the number of devices, this
5294 * is only possible for a very short time, and mdadm makes
5295 * sure that time appears to have past before assembling
5296 * the array. So we fail if that time hasn't passed.
5297 * For a reshape that keeps the number of devices the same
5298 * mdadm must be monitoring the reshape can keeping the
5299 * critical areas read-only and backed up. It will start
5300 * the array in read-only mode, so we check for that.
5302 sector_t here_new
, here_old
;
5304 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
5306 if (mddev
->new_level
!= mddev
->level
) {
5307 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
5308 "required - aborting.\n",
5312 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
5313 /* reshape_position must be on a new-stripe boundary, and one
5314 * further up in new geometry must map after here in old
5317 here_new
= mddev
->reshape_position
;
5318 if (sector_div(here_new
, mddev
->new_chunk_sectors
*
5319 (mddev
->raid_disks
- max_degraded
))) {
5320 printk(KERN_ERR
"md/raid:%s: reshape_position not "
5321 "on a stripe boundary\n", mdname(mddev
));
5324 reshape_offset
= here_new
* mddev
->new_chunk_sectors
;
5325 /* here_new is the stripe we will write to */
5326 here_old
= mddev
->reshape_position
;
5327 sector_div(here_old
, mddev
->chunk_sectors
*
5328 (old_disks
-max_degraded
));
5329 /* here_old is the first stripe that we might need to read
5331 if (mddev
->delta_disks
== 0) {
5332 if ((here_new
* mddev
->new_chunk_sectors
!=
5333 here_old
* mddev
->chunk_sectors
)) {
5334 printk(KERN_ERR
"md/raid:%s: reshape position is"
5335 " confused - aborting\n", mdname(mddev
));
5338 /* We cannot be sure it is safe to start an in-place
5339 * reshape. It is only safe if user-space is monitoring
5340 * and taking constant backups.
5341 * mdadm always starts a situation like this in
5342 * readonly mode so it can take control before
5343 * allowing any writes. So just check for that.
5345 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
5346 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
5347 /* not really in-place - so OK */;
5348 else if (mddev
->ro
== 0) {
5349 printk(KERN_ERR
"md/raid:%s: in-place reshape "
5350 "must be started in read-only mode "
5355 } else if (mddev
->reshape_backwards
5356 ? (here_new
* mddev
->new_chunk_sectors
+ min_offset_diff
<=
5357 here_old
* mddev
->chunk_sectors
)
5358 : (here_new
* mddev
->new_chunk_sectors
>=
5359 here_old
* mddev
->chunk_sectors
+ (-min_offset_diff
))) {
5360 /* Reading from the same stripe as writing to - bad */
5361 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
5362 "auto-recovery - aborting.\n",
5366 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
5368 /* OK, we should be able to continue; */
5370 BUG_ON(mddev
->level
!= mddev
->new_level
);
5371 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
5372 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
5373 BUG_ON(mddev
->delta_disks
!= 0);
5376 if (mddev
->private == NULL
)
5377 conf
= setup_conf(mddev
);
5379 conf
= mddev
->private;
5382 return PTR_ERR(conf
);
5384 conf
->min_offset_diff
= min_offset_diff
;
5385 mddev
->thread
= conf
->thread
;
5386 conf
->thread
= NULL
;
5387 mddev
->private = conf
;
5389 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
5391 rdev
= conf
->disks
[i
].rdev
;
5392 if (!rdev
&& conf
->disks
[i
].replacement
) {
5393 /* The replacement is all we have yet */
5394 rdev
= conf
->disks
[i
].replacement
;
5395 conf
->disks
[i
].replacement
= NULL
;
5396 clear_bit(Replacement
, &rdev
->flags
);
5397 conf
->disks
[i
].rdev
= rdev
;
5401 if (conf
->disks
[i
].replacement
&&
5402 conf
->reshape_progress
!= MaxSector
) {
5403 /* replacements and reshape simply do not mix. */
5404 printk(KERN_ERR
"md: cannot handle concurrent "
5405 "replacement and reshape.\n");
5408 if (test_bit(In_sync
, &rdev
->flags
)) {
5412 /* This disc is not fully in-sync. However if it
5413 * just stored parity (beyond the recovery_offset),
5414 * when we don't need to be concerned about the
5415 * array being dirty.
5416 * When reshape goes 'backwards', we never have
5417 * partially completed devices, so we only need
5418 * to worry about reshape going forwards.
5420 /* Hack because v0.91 doesn't store recovery_offset properly. */
5421 if (mddev
->major_version
== 0 &&
5422 mddev
->minor_version
> 90)
5423 rdev
->recovery_offset
= reshape_offset
;
5425 if (rdev
->recovery_offset
< reshape_offset
) {
5426 /* We need to check old and new layout */
5427 if (!only_parity(rdev
->raid_disk
,
5430 conf
->max_degraded
))
5433 if (!only_parity(rdev
->raid_disk
,
5435 conf
->previous_raid_disks
,
5436 conf
->max_degraded
))
5438 dirty_parity_disks
++;
5442 * 0 for a fully functional array, 1 or 2 for a degraded array.
5444 mddev
->degraded
= calc_degraded(conf
);
5446 if (has_failed(conf
)) {
5447 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
5448 " (%d/%d failed)\n",
5449 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
5453 /* device size must be a multiple of chunk size */
5454 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
5455 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
5457 if (mddev
->degraded
> dirty_parity_disks
&&
5458 mddev
->recovery_cp
!= MaxSector
) {
5459 if (mddev
->ok_start_degraded
)
5461 "md/raid:%s: starting dirty degraded array"
5462 " - data corruption possible.\n",
5466 "md/raid:%s: cannot start dirty degraded array.\n",
5472 if (mddev
->degraded
== 0)
5473 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
5474 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
5475 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
5478 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
5479 " out of %d devices, algorithm %d\n",
5480 mdname(mddev
), conf
->level
,
5481 mddev
->raid_disks
- mddev
->degraded
,
5482 mddev
->raid_disks
, mddev
->new_layout
);
5484 print_raid5_conf(conf
);
5486 if (conf
->reshape_progress
!= MaxSector
) {
5487 conf
->reshape_safe
= conf
->reshape_progress
;
5488 atomic_set(&conf
->reshape_stripes
, 0);
5489 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5490 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5491 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5492 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5493 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5498 /* Ok, everything is just fine now */
5499 if (mddev
->to_remove
== &raid5_attrs_group
)
5500 mddev
->to_remove
= NULL
;
5501 else if (mddev
->kobj
.sd
&&
5502 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
5504 "raid5: failed to create sysfs attributes for %s\n",
5506 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5510 bool discard_supported
= true;
5511 /* read-ahead size must cover two whole stripes, which
5512 * is 2 * (datadisks) * chunksize where 'n' is the
5513 * number of raid devices
5515 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
5516 int stripe
= data_disks
*
5517 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
5518 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
5519 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
5521 blk_queue_merge_bvec(mddev
->queue
, raid5_mergeable_bvec
);
5523 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
5524 mddev
->queue
->backing_dev_info
.congested_fn
= raid5_congested
;
5526 chunk_size
= mddev
->chunk_sectors
<< 9;
5527 blk_queue_io_min(mddev
->queue
, chunk_size
);
5528 blk_queue_io_opt(mddev
->queue
, chunk_size
*
5529 (conf
->raid_disks
- conf
->max_degraded
));
5531 * We can only discard a whole stripe. It doesn't make sense to
5532 * discard data disk but write parity disk
5534 stripe
= stripe
* PAGE_SIZE
;
5535 /* Round up to power of 2, as discard handling
5536 * currently assumes that */
5537 while ((stripe
-1) & stripe
)
5538 stripe
= (stripe
| (stripe
-1)) + 1;
5539 mddev
->queue
->limits
.discard_alignment
= stripe
;
5540 mddev
->queue
->limits
.discard_granularity
= stripe
;
5542 * unaligned part of discard request will be ignored, so can't
5543 * guarantee discard_zerors_data
5545 mddev
->queue
->limits
.discard_zeroes_data
= 0;
5547 rdev_for_each(rdev
, mddev
) {
5548 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
5549 rdev
->data_offset
<< 9);
5550 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
5551 rdev
->new_data_offset
<< 9);
5553 * discard_zeroes_data is required, otherwise data
5554 * could be lost. Consider a scenario: discard a stripe
5555 * (the stripe could be inconsistent if
5556 * discard_zeroes_data is 0); write one disk of the
5557 * stripe (the stripe could be inconsistent again
5558 * depending on which disks are used to calculate
5559 * parity); the disk is broken; The stripe data of this
5562 if (!blk_queue_discard(bdev_get_queue(rdev
->bdev
)) ||
5563 !bdev_get_queue(rdev
->bdev
)->
5564 limits
.discard_zeroes_data
)
5565 discard_supported
= false;
5568 if (discard_supported
&&
5569 mddev
->queue
->limits
.max_discard_sectors
>= stripe
&&
5570 mddev
->queue
->limits
.discard_granularity
>= stripe
)
5571 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
5574 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
5580 md_unregister_thread(&mddev
->thread
);
5581 print_raid5_conf(conf
);
5583 mddev
->private = NULL
;
5584 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
5588 static int stop(struct mddev
*mddev
)
5590 struct r5conf
*conf
= mddev
->private;
5592 md_unregister_thread(&mddev
->thread
);
5594 mddev
->queue
->backing_dev_info
.congested_fn
= NULL
;
5596 mddev
->private = NULL
;
5597 mddev
->to_remove
= &raid5_attrs_group
;
5601 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
5603 struct r5conf
*conf
= mddev
->private;
5606 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
5607 mddev
->chunk_sectors
/ 2, mddev
->layout
);
5608 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
5609 for (i
= 0; i
< conf
->raid_disks
; i
++)
5610 seq_printf (seq
, "%s",
5611 conf
->disks
[i
].rdev
&&
5612 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
5613 seq_printf (seq
, "]");
5616 static void print_raid5_conf (struct r5conf
*conf
)
5619 struct disk_info
*tmp
;
5621 printk(KERN_DEBUG
"RAID conf printout:\n");
5623 printk("(conf==NULL)\n");
5626 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
5628 conf
->raid_disks
- conf
->mddev
->degraded
);
5630 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5631 char b
[BDEVNAME_SIZE
];
5632 tmp
= conf
->disks
+ i
;
5634 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
5635 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
5636 bdevname(tmp
->rdev
->bdev
, b
));
5640 static int raid5_spare_active(struct mddev
*mddev
)
5643 struct r5conf
*conf
= mddev
->private;
5644 struct disk_info
*tmp
;
5646 unsigned long flags
;
5648 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5649 tmp
= conf
->disks
+ i
;
5650 if (tmp
->replacement
5651 && tmp
->replacement
->recovery_offset
== MaxSector
5652 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
5653 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
5654 /* Replacement has just become active. */
5656 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
5659 /* Replaced device not technically faulty,
5660 * but we need to be sure it gets removed
5661 * and never re-added.
5663 set_bit(Faulty
, &tmp
->rdev
->flags
);
5664 sysfs_notify_dirent_safe(
5665 tmp
->rdev
->sysfs_state
);
5667 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
5668 } else if (tmp
->rdev
5669 && tmp
->rdev
->recovery_offset
== MaxSector
5670 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
5671 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
5673 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
5676 spin_lock_irqsave(&conf
->device_lock
, flags
);
5677 mddev
->degraded
= calc_degraded(conf
);
5678 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5679 print_raid5_conf(conf
);
5683 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
5685 struct r5conf
*conf
= mddev
->private;
5687 int number
= rdev
->raid_disk
;
5688 struct md_rdev
**rdevp
;
5689 struct disk_info
*p
= conf
->disks
+ number
;
5691 print_raid5_conf(conf
);
5692 if (rdev
== p
->rdev
)
5694 else if (rdev
== p
->replacement
)
5695 rdevp
= &p
->replacement
;
5699 if (number
>= conf
->raid_disks
&&
5700 conf
->reshape_progress
== MaxSector
)
5701 clear_bit(In_sync
, &rdev
->flags
);
5703 if (test_bit(In_sync
, &rdev
->flags
) ||
5704 atomic_read(&rdev
->nr_pending
)) {
5708 /* Only remove non-faulty devices if recovery
5711 if (!test_bit(Faulty
, &rdev
->flags
) &&
5712 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
5713 !has_failed(conf
) &&
5714 (!p
->replacement
|| p
->replacement
== rdev
) &&
5715 number
< conf
->raid_disks
) {
5721 if (atomic_read(&rdev
->nr_pending
)) {
5722 /* lost the race, try later */
5725 } else if (p
->replacement
) {
5726 /* We must have just cleared 'rdev' */
5727 p
->rdev
= p
->replacement
;
5728 clear_bit(Replacement
, &p
->replacement
->flags
);
5729 smp_mb(); /* Make sure other CPUs may see both as identical
5730 * but will never see neither - if they are careful
5732 p
->replacement
= NULL
;
5733 clear_bit(WantReplacement
, &rdev
->flags
);
5735 /* We might have just removed the Replacement as faulty-
5736 * clear the bit just in case
5738 clear_bit(WantReplacement
, &rdev
->flags
);
5741 print_raid5_conf(conf
);
5745 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
5747 struct r5conf
*conf
= mddev
->private;
5750 struct disk_info
*p
;
5752 int last
= conf
->raid_disks
- 1;
5754 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
5757 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
5758 /* no point adding a device */
5761 if (rdev
->raid_disk
>= 0)
5762 first
= last
= rdev
->raid_disk
;
5765 * find the disk ... but prefer rdev->saved_raid_disk
5768 if (rdev
->saved_raid_disk
>= 0 &&
5769 rdev
->saved_raid_disk
>= first
&&
5770 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
5771 first
= rdev
->saved_raid_disk
;
5773 for (disk
= first
; disk
<= last
; disk
++) {
5774 p
= conf
->disks
+ disk
;
5775 if (p
->rdev
== NULL
) {
5776 clear_bit(In_sync
, &rdev
->flags
);
5777 rdev
->raid_disk
= disk
;
5779 if (rdev
->saved_raid_disk
!= disk
)
5781 rcu_assign_pointer(p
->rdev
, rdev
);
5785 for (disk
= first
; disk
<= last
; disk
++) {
5786 p
= conf
->disks
+ disk
;
5787 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
5788 p
->replacement
== NULL
) {
5789 clear_bit(In_sync
, &rdev
->flags
);
5790 set_bit(Replacement
, &rdev
->flags
);
5791 rdev
->raid_disk
= disk
;
5794 rcu_assign_pointer(p
->replacement
, rdev
);
5799 print_raid5_conf(conf
);
5803 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
5805 /* no resync is happening, and there is enough space
5806 * on all devices, so we can resize.
5807 * We need to make sure resync covers any new space.
5808 * If the array is shrinking we should possibly wait until
5809 * any io in the removed space completes, but it hardly seems
5813 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
5814 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
5815 if (mddev
->external_size
&&
5816 mddev
->array_sectors
> newsize
)
5818 if (mddev
->bitmap
) {
5819 int ret
= bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
5823 md_set_array_sectors(mddev
, newsize
);
5824 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5825 revalidate_disk(mddev
->gendisk
);
5826 if (sectors
> mddev
->dev_sectors
&&
5827 mddev
->recovery_cp
> mddev
->dev_sectors
) {
5828 mddev
->recovery_cp
= mddev
->dev_sectors
;
5829 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
5831 mddev
->dev_sectors
= sectors
;
5832 mddev
->resync_max_sectors
= sectors
;
5836 static int check_stripe_cache(struct mddev
*mddev
)
5838 /* Can only proceed if there are plenty of stripe_heads.
5839 * We need a minimum of one full stripe,, and for sensible progress
5840 * it is best to have about 4 times that.
5841 * If we require 4 times, then the default 256 4K stripe_heads will
5842 * allow for chunk sizes up to 256K, which is probably OK.
5843 * If the chunk size is greater, user-space should request more
5844 * stripe_heads first.
5846 struct r5conf
*conf
= mddev
->private;
5847 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5848 > conf
->max_nr_stripes
||
5849 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5850 > conf
->max_nr_stripes
) {
5851 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5853 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
5860 static int check_reshape(struct mddev
*mddev
)
5862 struct r5conf
*conf
= mddev
->private;
5864 if (mddev
->delta_disks
== 0 &&
5865 mddev
->new_layout
== mddev
->layout
&&
5866 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
5867 return 0; /* nothing to do */
5868 if (has_failed(conf
))
5870 if (mddev
->delta_disks
< 0) {
5871 /* We might be able to shrink, but the devices must
5872 * be made bigger first.
5873 * For raid6, 4 is the minimum size.
5874 * Otherwise 2 is the minimum
5877 if (mddev
->level
== 6)
5879 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
5883 if (!check_stripe_cache(mddev
))
5886 return resize_stripes(conf
, (conf
->previous_raid_disks
5887 + mddev
->delta_disks
));
5890 static int raid5_start_reshape(struct mddev
*mddev
)
5892 struct r5conf
*conf
= mddev
->private;
5893 struct md_rdev
*rdev
;
5895 unsigned long flags
;
5897 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
5900 if (!check_stripe_cache(mddev
))
5903 if (has_failed(conf
))
5906 rdev_for_each(rdev
, mddev
) {
5907 if (!test_bit(In_sync
, &rdev
->flags
)
5908 && !test_bit(Faulty
, &rdev
->flags
))
5912 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
5913 /* Not enough devices even to make a degraded array
5918 /* Refuse to reduce size of the array. Any reductions in
5919 * array size must be through explicit setting of array_size
5922 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
5923 < mddev
->array_sectors
) {
5924 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
5925 "before number of disks\n", mdname(mddev
));
5929 atomic_set(&conf
->reshape_stripes
, 0);
5930 spin_lock_irq(&conf
->device_lock
);
5931 conf
->previous_raid_disks
= conf
->raid_disks
;
5932 conf
->raid_disks
+= mddev
->delta_disks
;
5933 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
5934 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
5935 conf
->prev_algo
= conf
->algorithm
;
5936 conf
->algorithm
= mddev
->new_layout
;
5938 /* Code that selects data_offset needs to see the generation update
5939 * if reshape_progress has been set - so a memory barrier needed.
5942 if (mddev
->reshape_backwards
)
5943 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
5945 conf
->reshape_progress
= 0;
5946 conf
->reshape_safe
= conf
->reshape_progress
;
5947 spin_unlock_irq(&conf
->device_lock
);
5949 /* Add some new drives, as many as will fit.
5950 * We know there are enough to make the newly sized array work.
5951 * Don't add devices if we are reducing the number of
5952 * devices in the array. This is because it is not possible
5953 * to correctly record the "partially reconstructed" state of
5954 * such devices during the reshape and confusion could result.
5956 if (mddev
->delta_disks
>= 0) {
5957 rdev_for_each(rdev
, mddev
)
5958 if (rdev
->raid_disk
< 0 &&
5959 !test_bit(Faulty
, &rdev
->flags
)) {
5960 if (raid5_add_disk(mddev
, rdev
) == 0) {
5962 >= conf
->previous_raid_disks
)
5963 set_bit(In_sync
, &rdev
->flags
);
5965 rdev
->recovery_offset
= 0;
5967 if (sysfs_link_rdev(mddev
, rdev
))
5968 /* Failure here is OK */;
5970 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
5971 && !test_bit(Faulty
, &rdev
->flags
)) {
5972 /* This is a spare that was manually added */
5973 set_bit(In_sync
, &rdev
->flags
);
5976 /* When a reshape changes the number of devices,
5977 * ->degraded is measured against the larger of the
5978 * pre and post number of devices.
5980 spin_lock_irqsave(&conf
->device_lock
, flags
);
5981 mddev
->degraded
= calc_degraded(conf
);
5982 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5984 mddev
->raid_disks
= conf
->raid_disks
;
5985 mddev
->reshape_position
= conf
->reshape_progress
;
5986 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5988 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5989 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5990 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5991 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5992 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5994 if (!mddev
->sync_thread
) {
5995 mddev
->recovery
= 0;
5996 spin_lock_irq(&conf
->device_lock
);
5997 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
5998 rdev_for_each(rdev
, mddev
)
5999 rdev
->new_data_offset
= rdev
->data_offset
;
6001 conf
->reshape_progress
= MaxSector
;
6002 mddev
->reshape_position
= MaxSector
;
6003 spin_unlock_irq(&conf
->device_lock
);
6006 conf
->reshape_checkpoint
= jiffies
;
6007 md_wakeup_thread(mddev
->sync_thread
);
6008 md_new_event(mddev
);
6012 /* This is called from the reshape thread and should make any
6013 * changes needed in 'conf'
6015 static void end_reshape(struct r5conf
*conf
)
6018 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
6019 struct md_rdev
*rdev
;
6021 spin_lock_irq(&conf
->device_lock
);
6022 conf
->previous_raid_disks
= conf
->raid_disks
;
6023 rdev_for_each(rdev
, conf
->mddev
)
6024 rdev
->data_offset
= rdev
->new_data_offset
;
6026 conf
->reshape_progress
= MaxSector
;
6027 spin_unlock_irq(&conf
->device_lock
);
6028 wake_up(&conf
->wait_for_overlap
);
6030 /* read-ahead size must cover two whole stripes, which is
6031 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6033 if (conf
->mddev
->queue
) {
6034 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
6035 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
6037 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
6038 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
6043 /* This is called from the raid5d thread with mddev_lock held.
6044 * It makes config changes to the device.
6046 static void raid5_finish_reshape(struct mddev
*mddev
)
6048 struct r5conf
*conf
= mddev
->private;
6050 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
6052 if (mddev
->delta_disks
> 0) {
6053 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
6054 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
6055 revalidate_disk(mddev
->gendisk
);
6058 spin_lock_irq(&conf
->device_lock
);
6059 mddev
->degraded
= calc_degraded(conf
);
6060 spin_unlock_irq(&conf
->device_lock
);
6061 for (d
= conf
->raid_disks
;
6062 d
< conf
->raid_disks
- mddev
->delta_disks
;
6064 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
6066 clear_bit(In_sync
, &rdev
->flags
);
6067 rdev
= conf
->disks
[d
].replacement
;
6069 clear_bit(In_sync
, &rdev
->flags
);
6072 mddev
->layout
= conf
->algorithm
;
6073 mddev
->chunk_sectors
= conf
->chunk_sectors
;
6074 mddev
->reshape_position
= MaxSector
;
6075 mddev
->delta_disks
= 0;
6076 mddev
->reshape_backwards
= 0;
6080 static void raid5_quiesce(struct mddev
*mddev
, int state
)
6082 struct r5conf
*conf
= mddev
->private;
6085 case 2: /* resume for a suspend */
6086 wake_up(&conf
->wait_for_overlap
);
6089 case 1: /* stop all writes */
6090 spin_lock_irq(&conf
->device_lock
);
6091 /* '2' tells resync/reshape to pause so that all
6092 * active stripes can drain
6095 wait_event_lock_irq(conf
->wait_for_stripe
,
6096 atomic_read(&conf
->active_stripes
) == 0 &&
6097 atomic_read(&conf
->active_aligned_reads
) == 0,
6098 conf
->device_lock
, /* nothing */);
6100 spin_unlock_irq(&conf
->device_lock
);
6101 /* allow reshape to continue */
6102 wake_up(&conf
->wait_for_overlap
);
6105 case 0: /* re-enable writes */
6106 spin_lock_irq(&conf
->device_lock
);
6108 wake_up(&conf
->wait_for_stripe
);
6109 wake_up(&conf
->wait_for_overlap
);
6110 spin_unlock_irq(&conf
->device_lock
);
6116 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
6118 struct r0conf
*raid0_conf
= mddev
->private;
6121 /* for raid0 takeover only one zone is supported */
6122 if (raid0_conf
->nr_strip_zones
> 1) {
6123 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6125 return ERR_PTR(-EINVAL
);
6128 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
6129 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
6130 mddev
->dev_sectors
= sectors
;
6131 mddev
->new_level
= level
;
6132 mddev
->new_layout
= ALGORITHM_PARITY_N
;
6133 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
6134 mddev
->raid_disks
+= 1;
6135 mddev
->delta_disks
= 1;
6136 /* make sure it will be not marked as dirty */
6137 mddev
->recovery_cp
= MaxSector
;
6139 return setup_conf(mddev
);
6143 static void *raid5_takeover_raid1(struct mddev
*mddev
)
6147 if (mddev
->raid_disks
!= 2 ||
6148 mddev
->degraded
> 1)
6149 return ERR_PTR(-EINVAL
);
6151 /* Should check if there are write-behind devices? */
6153 chunksect
= 64*2; /* 64K by default */
6155 /* The array must be an exact multiple of chunksize */
6156 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
6159 if ((chunksect
<<9) < STRIPE_SIZE
)
6160 /* array size does not allow a suitable chunk size */
6161 return ERR_PTR(-EINVAL
);
6163 mddev
->new_level
= 5;
6164 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
6165 mddev
->new_chunk_sectors
= chunksect
;
6167 return setup_conf(mddev
);
6170 static void *raid5_takeover_raid6(struct mddev
*mddev
)
6174 switch (mddev
->layout
) {
6175 case ALGORITHM_LEFT_ASYMMETRIC_6
:
6176 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
6178 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
6179 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
6181 case ALGORITHM_LEFT_SYMMETRIC_6
:
6182 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
6184 case ALGORITHM_RIGHT_SYMMETRIC_6
:
6185 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
6187 case ALGORITHM_PARITY_0_6
:
6188 new_layout
= ALGORITHM_PARITY_0
;
6190 case ALGORITHM_PARITY_N
:
6191 new_layout
= ALGORITHM_PARITY_N
;
6194 return ERR_PTR(-EINVAL
);
6196 mddev
->new_level
= 5;
6197 mddev
->new_layout
= new_layout
;
6198 mddev
->delta_disks
= -1;
6199 mddev
->raid_disks
-= 1;
6200 return setup_conf(mddev
);
6204 static int raid5_check_reshape(struct mddev
*mddev
)
6206 /* For a 2-drive array, the layout and chunk size can be changed
6207 * immediately as not restriping is needed.
6208 * For larger arrays we record the new value - after validation
6209 * to be used by a reshape pass.
6211 struct r5conf
*conf
= mddev
->private;
6212 int new_chunk
= mddev
->new_chunk_sectors
;
6214 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
6216 if (new_chunk
> 0) {
6217 if (!is_power_of_2(new_chunk
))
6219 if (new_chunk
< (PAGE_SIZE
>>9))
6221 if (mddev
->array_sectors
& (new_chunk
-1))
6222 /* not factor of array size */
6226 /* They look valid */
6228 if (mddev
->raid_disks
== 2) {
6229 /* can make the change immediately */
6230 if (mddev
->new_layout
>= 0) {
6231 conf
->algorithm
= mddev
->new_layout
;
6232 mddev
->layout
= mddev
->new_layout
;
6234 if (new_chunk
> 0) {
6235 conf
->chunk_sectors
= new_chunk
;
6236 mddev
->chunk_sectors
= new_chunk
;
6238 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
6239 md_wakeup_thread(mddev
->thread
);
6241 return check_reshape(mddev
);
6244 static int raid6_check_reshape(struct mddev
*mddev
)
6246 int new_chunk
= mddev
->new_chunk_sectors
;
6248 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
6250 if (new_chunk
> 0) {
6251 if (!is_power_of_2(new_chunk
))
6253 if (new_chunk
< (PAGE_SIZE
>> 9))
6255 if (mddev
->array_sectors
& (new_chunk
-1))
6256 /* not factor of array size */
6260 /* They look valid */
6261 return check_reshape(mddev
);
6264 static void *raid5_takeover(struct mddev
*mddev
)
6266 /* raid5 can take over:
6267 * raid0 - if there is only one strip zone - make it a raid4 layout
6268 * raid1 - if there are two drives. We need to know the chunk size
6269 * raid4 - trivial - just use a raid4 layout.
6270 * raid6 - Providing it is a *_6 layout
6272 if (mddev
->level
== 0)
6273 return raid45_takeover_raid0(mddev
, 5);
6274 if (mddev
->level
== 1)
6275 return raid5_takeover_raid1(mddev
);
6276 if (mddev
->level
== 4) {
6277 mddev
->new_layout
= ALGORITHM_PARITY_N
;
6278 mddev
->new_level
= 5;
6279 return setup_conf(mddev
);
6281 if (mddev
->level
== 6)
6282 return raid5_takeover_raid6(mddev
);
6284 return ERR_PTR(-EINVAL
);
6287 static void *raid4_takeover(struct mddev
*mddev
)
6289 /* raid4 can take over:
6290 * raid0 - if there is only one strip zone
6291 * raid5 - if layout is right
6293 if (mddev
->level
== 0)
6294 return raid45_takeover_raid0(mddev
, 4);
6295 if (mddev
->level
== 5 &&
6296 mddev
->layout
== ALGORITHM_PARITY_N
) {
6297 mddev
->new_layout
= 0;
6298 mddev
->new_level
= 4;
6299 return setup_conf(mddev
);
6301 return ERR_PTR(-EINVAL
);
6304 static struct md_personality raid5_personality
;
6306 static void *raid6_takeover(struct mddev
*mddev
)
6308 /* Currently can only take over a raid5. We map the
6309 * personality to an equivalent raid6 personality
6310 * with the Q block at the end.
6314 if (mddev
->pers
!= &raid5_personality
)
6315 return ERR_PTR(-EINVAL
);
6316 if (mddev
->degraded
> 1)
6317 return ERR_PTR(-EINVAL
);
6318 if (mddev
->raid_disks
> 253)
6319 return ERR_PTR(-EINVAL
);
6320 if (mddev
->raid_disks
< 3)
6321 return ERR_PTR(-EINVAL
);
6323 switch (mddev
->layout
) {
6324 case ALGORITHM_LEFT_ASYMMETRIC
:
6325 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
6327 case ALGORITHM_RIGHT_ASYMMETRIC
:
6328 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
6330 case ALGORITHM_LEFT_SYMMETRIC
:
6331 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
6333 case ALGORITHM_RIGHT_SYMMETRIC
:
6334 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
6336 case ALGORITHM_PARITY_0
:
6337 new_layout
= ALGORITHM_PARITY_0_6
;
6339 case ALGORITHM_PARITY_N
:
6340 new_layout
= ALGORITHM_PARITY_N
;
6343 return ERR_PTR(-EINVAL
);
6345 mddev
->new_level
= 6;
6346 mddev
->new_layout
= new_layout
;
6347 mddev
->delta_disks
= 1;
6348 mddev
->raid_disks
+= 1;
6349 return setup_conf(mddev
);
6353 static struct md_personality raid6_personality
=
6357 .owner
= THIS_MODULE
,
6358 .make_request
= make_request
,
6362 .error_handler
= error
,
6363 .hot_add_disk
= raid5_add_disk
,
6364 .hot_remove_disk
= raid5_remove_disk
,
6365 .spare_active
= raid5_spare_active
,
6366 .sync_request
= sync_request
,
6367 .resize
= raid5_resize
,
6369 .check_reshape
= raid6_check_reshape
,
6370 .start_reshape
= raid5_start_reshape
,
6371 .finish_reshape
= raid5_finish_reshape
,
6372 .quiesce
= raid5_quiesce
,
6373 .takeover
= raid6_takeover
,
6375 static struct md_personality raid5_personality
=
6379 .owner
= THIS_MODULE
,
6380 .make_request
= make_request
,
6384 .error_handler
= error
,
6385 .hot_add_disk
= raid5_add_disk
,
6386 .hot_remove_disk
= raid5_remove_disk
,
6387 .spare_active
= raid5_spare_active
,
6388 .sync_request
= sync_request
,
6389 .resize
= raid5_resize
,
6391 .check_reshape
= raid5_check_reshape
,
6392 .start_reshape
= raid5_start_reshape
,
6393 .finish_reshape
= raid5_finish_reshape
,
6394 .quiesce
= raid5_quiesce
,
6395 .takeover
= raid5_takeover
,
6398 static struct md_personality raid4_personality
=
6402 .owner
= THIS_MODULE
,
6403 .make_request
= make_request
,
6407 .error_handler
= error
,
6408 .hot_add_disk
= raid5_add_disk
,
6409 .hot_remove_disk
= raid5_remove_disk
,
6410 .spare_active
= raid5_spare_active
,
6411 .sync_request
= sync_request
,
6412 .resize
= raid5_resize
,
6414 .check_reshape
= raid5_check_reshape
,
6415 .start_reshape
= raid5_start_reshape
,
6416 .finish_reshape
= raid5_finish_reshape
,
6417 .quiesce
= raid5_quiesce
,
6418 .takeover
= raid4_takeover
,
6421 static int __init
raid5_init(void)
6423 register_md_personality(&raid6_personality
);
6424 register_md_personality(&raid5_personality
);
6425 register_md_personality(&raid4_personality
);
6429 static void raid5_exit(void)
6431 unregister_md_personality(&raid6_personality
);
6432 unregister_md_personality(&raid5_personality
);
6433 unregister_md_personality(&raid4_personality
);
6436 module_init(raid5_init
);
6437 module_exit(raid5_exit
);
6438 MODULE_LICENSE("GPL");
6439 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6440 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6441 MODULE_ALIAS("md-raid5");
6442 MODULE_ALIAS("md-raid4");
6443 MODULE_ALIAS("md-level-5");
6444 MODULE_ALIAS("md-level-4");
6445 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6446 MODULE_ALIAS("md-raid6");
6447 MODULE_ALIAS("md-level-6");
6449 /* This used to be two separate modules, they were: */
6450 MODULE_ALIAS("raid5");
6451 MODULE_ALIAS("raid6");