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
))
398 else if (test_bit(In_sync
, &rdev
->flags
))
401 /* not in-sync or faulty.
402 * If the reshape increases the number of devices,
403 * this is being recovered by the reshape, so
404 * this 'previous' section is not in_sync.
405 * If the number of devices is being reduced however,
406 * the device can only be part of the array if
407 * we are reverting a reshape, so this section will
410 if (conf
->raid_disks
>= conf
->previous_raid_disks
)
414 if (conf
->raid_disks
== conf
->previous_raid_disks
)
418 for (i
= 0; i
< conf
->raid_disks
; i
++) {
419 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
420 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
422 else if (test_bit(In_sync
, &rdev
->flags
))
425 /* not in-sync or faulty.
426 * If reshape increases the number of devices, this
427 * section has already been recovered, else it
428 * almost certainly hasn't.
430 if (conf
->raid_disks
<= conf
->previous_raid_disks
)
434 if (degraded2
> degraded
)
439 static int has_failed(struct r5conf
*conf
)
443 if (conf
->mddev
->reshape_position
== MaxSector
)
444 return conf
->mddev
->degraded
> conf
->max_degraded
;
446 degraded
= calc_degraded(conf
);
447 if (degraded
> conf
->max_degraded
)
452 static struct stripe_head
*
453 get_active_stripe(struct r5conf
*conf
, sector_t sector
,
454 int previous
, int noblock
, int noquiesce
)
456 struct stripe_head
*sh
;
458 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
460 spin_lock_irq(&conf
->device_lock
);
463 wait_event_lock_irq(conf
->wait_for_stripe
,
464 conf
->quiesce
== 0 || noquiesce
,
465 conf
->device_lock
, /* nothing */);
466 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
468 if (!conf
->inactive_blocked
)
469 sh
= get_free_stripe(conf
);
470 if (noblock
&& sh
== NULL
)
473 conf
->inactive_blocked
= 1;
474 wait_event_lock_irq(conf
->wait_for_stripe
,
475 !list_empty(&conf
->inactive_list
) &&
476 (atomic_read(&conf
->active_stripes
)
477 < (conf
->max_nr_stripes
*3/4)
478 || !conf
->inactive_blocked
),
481 conf
->inactive_blocked
= 0;
483 init_stripe(sh
, sector
, previous
);
485 if (atomic_read(&sh
->count
)) {
486 BUG_ON(!list_empty(&sh
->lru
)
487 && !test_bit(STRIPE_EXPANDING
, &sh
->state
));
489 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
490 atomic_inc(&conf
->active_stripes
);
491 if (list_empty(&sh
->lru
) &&
492 !test_bit(STRIPE_EXPANDING
, &sh
->state
))
494 list_del_init(&sh
->lru
);
497 } while (sh
== NULL
);
500 atomic_inc(&sh
->count
);
502 spin_unlock_irq(&conf
->device_lock
);
506 /* Determine if 'data_offset' or 'new_data_offset' should be used
507 * in this stripe_head.
509 static int use_new_offset(struct r5conf
*conf
, struct stripe_head
*sh
)
511 sector_t progress
= conf
->reshape_progress
;
512 /* Need a memory barrier to make sure we see the value
513 * of conf->generation, or ->data_offset that was set before
514 * reshape_progress was updated.
517 if (progress
== MaxSector
)
519 if (sh
->generation
== conf
->generation
- 1)
521 /* We are in a reshape, and this is a new-generation stripe,
522 * so use new_data_offset.
528 raid5_end_read_request(struct bio
*bi
, int error
);
530 raid5_end_write_request(struct bio
*bi
, int error
);
532 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
534 struct r5conf
*conf
= sh
->raid_conf
;
535 int i
, disks
= sh
->disks
;
539 for (i
= disks
; i
--; ) {
541 int replace_only
= 0;
542 struct bio
*bi
, *rbi
;
543 struct md_rdev
*rdev
, *rrdev
= NULL
;
544 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
545 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
549 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
551 else if (test_and_clear_bit(R5_WantReplace
,
552 &sh
->dev
[i
].flags
)) {
557 if (test_and_clear_bit(R5_SyncIO
, &sh
->dev
[i
].flags
))
560 bi
= &sh
->dev
[i
].req
;
561 rbi
= &sh
->dev
[i
].rreq
; /* For writing to replacement */
566 bi
->bi_end_io
= raid5_end_write_request
;
567 rbi
->bi_end_io
= raid5_end_write_request
;
569 bi
->bi_end_io
= raid5_end_read_request
;
572 rrdev
= rcu_dereference(conf
->disks
[i
].replacement
);
573 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
574 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
583 /* We raced and saw duplicates */
586 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
) && rrdev
)
591 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
594 atomic_inc(&rdev
->nr_pending
);
595 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
598 atomic_inc(&rrdev
->nr_pending
);
601 /* We have already checked bad blocks for reads. Now
602 * need to check for writes. We never accept write errors
603 * on the replacement, so we don't to check rrdev.
605 while ((rw
& WRITE
) && rdev
&&
606 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
609 int bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
610 &first_bad
, &bad_sectors
);
615 set_bit(BlockedBadBlocks
, &rdev
->flags
);
616 if (!conf
->mddev
->external
&&
617 conf
->mddev
->flags
) {
618 /* It is very unlikely, but we might
619 * still need to write out the
620 * bad block log - better give it
622 md_check_recovery(conf
->mddev
);
625 * Because md_wait_for_blocked_rdev
626 * will dec nr_pending, we must
627 * increment it first.
629 atomic_inc(&rdev
->nr_pending
);
630 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
632 /* Acknowledged bad block - skip the write */
633 rdev_dec_pending(rdev
, conf
->mddev
);
639 if (s
->syncing
|| s
->expanding
|| s
->expanded
641 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
643 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
645 bi
->bi_bdev
= rdev
->bdev
;
646 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
647 __func__
, (unsigned long long)sh
->sector
,
649 atomic_inc(&sh
->count
);
650 if (use_new_offset(conf
, sh
))
651 bi
->bi_sector
= (sh
->sector
652 + rdev
->new_data_offset
);
654 bi
->bi_sector
= (sh
->sector
655 + rdev
->data_offset
);
656 bi
->bi_flags
= 1 << BIO_UPTODATE
;
658 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
659 bi
->bi_io_vec
[0].bv_offset
= 0;
660 bi
->bi_size
= STRIPE_SIZE
;
663 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
664 generic_make_request(bi
);
667 if (s
->syncing
|| s
->expanding
|| s
->expanded
669 md_sync_acct(rrdev
->bdev
, STRIPE_SECTORS
);
671 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
673 rbi
->bi_bdev
= rrdev
->bdev
;
674 pr_debug("%s: for %llu schedule op %ld on "
675 "replacement disc %d\n",
676 __func__
, (unsigned long long)sh
->sector
,
678 atomic_inc(&sh
->count
);
679 if (use_new_offset(conf
, sh
))
680 rbi
->bi_sector
= (sh
->sector
681 + rrdev
->new_data_offset
);
683 rbi
->bi_sector
= (sh
->sector
684 + rrdev
->data_offset
);
685 rbi
->bi_flags
= 1 << BIO_UPTODATE
;
687 rbi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
688 rbi
->bi_io_vec
[0].bv_offset
= 0;
689 rbi
->bi_size
= STRIPE_SIZE
;
691 generic_make_request(rbi
);
693 if (!rdev
&& !rrdev
) {
695 set_bit(STRIPE_DEGRADED
, &sh
->state
);
696 pr_debug("skip op %ld on disc %d for sector %llu\n",
697 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
698 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
699 set_bit(STRIPE_HANDLE
, &sh
->state
);
704 static struct dma_async_tx_descriptor
*
705 async_copy_data(int frombio
, struct bio
*bio
, struct page
*page
,
706 sector_t sector
, struct dma_async_tx_descriptor
*tx
)
709 struct page
*bio_page
;
712 struct async_submit_ctl submit
;
713 enum async_tx_flags flags
= 0;
715 if (bio
->bi_sector
>= sector
)
716 page_offset
= (signed)(bio
->bi_sector
- sector
) * 512;
718 page_offset
= (signed)(sector
- bio
->bi_sector
) * -512;
721 flags
|= ASYNC_TX_FENCE
;
722 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
724 bio_for_each_segment(bvl
, bio
, i
) {
725 int len
= bvl
->bv_len
;
729 if (page_offset
< 0) {
730 b_offset
= -page_offset
;
731 page_offset
+= b_offset
;
735 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
736 clen
= STRIPE_SIZE
- page_offset
;
741 b_offset
+= bvl
->bv_offset
;
742 bio_page
= bvl
->bv_page
;
744 tx
= async_memcpy(page
, bio_page
, page_offset
,
745 b_offset
, clen
, &submit
);
747 tx
= async_memcpy(bio_page
, page
, b_offset
,
748 page_offset
, clen
, &submit
);
750 /* chain the operations */
751 submit
.depend_tx
= tx
;
753 if (clen
< len
) /* hit end of page */
761 static void ops_complete_biofill(void *stripe_head_ref
)
763 struct stripe_head
*sh
= stripe_head_ref
;
764 struct bio
*return_bi
= NULL
;
767 pr_debug("%s: stripe %llu\n", __func__
,
768 (unsigned long long)sh
->sector
);
770 /* clear completed biofills */
771 for (i
= sh
->disks
; i
--; ) {
772 struct r5dev
*dev
= &sh
->dev
[i
];
774 /* acknowledge completion of a biofill operation */
775 /* and check if we need to reply to a read request,
776 * new R5_Wantfill requests are held off until
777 * !STRIPE_BIOFILL_RUN
779 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
780 struct bio
*rbi
, *rbi2
;
785 while (rbi
&& rbi
->bi_sector
<
786 dev
->sector
+ STRIPE_SECTORS
) {
787 rbi2
= r5_next_bio(rbi
, dev
->sector
);
788 if (!raid5_dec_bi_active_stripes(rbi
)) {
789 rbi
->bi_next
= return_bi
;
796 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
798 return_io(return_bi
);
800 set_bit(STRIPE_HANDLE
, &sh
->state
);
804 static void ops_run_biofill(struct stripe_head
*sh
)
806 struct dma_async_tx_descriptor
*tx
= NULL
;
807 struct async_submit_ctl submit
;
810 pr_debug("%s: stripe %llu\n", __func__
,
811 (unsigned long long)sh
->sector
);
813 for (i
= sh
->disks
; i
--; ) {
814 struct r5dev
*dev
= &sh
->dev
[i
];
815 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
817 spin_lock_irq(&sh
->stripe_lock
);
818 dev
->read
= rbi
= dev
->toread
;
820 spin_unlock_irq(&sh
->stripe_lock
);
821 while (rbi
&& rbi
->bi_sector
<
822 dev
->sector
+ STRIPE_SECTORS
) {
823 tx
= async_copy_data(0, rbi
, dev
->page
,
825 rbi
= r5_next_bio(rbi
, dev
->sector
);
830 atomic_inc(&sh
->count
);
831 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
832 async_trigger_callback(&submit
);
835 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
842 tgt
= &sh
->dev
[target
];
843 set_bit(R5_UPTODATE
, &tgt
->flags
);
844 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
845 clear_bit(R5_Wantcompute
, &tgt
->flags
);
848 static void ops_complete_compute(void *stripe_head_ref
)
850 struct stripe_head
*sh
= stripe_head_ref
;
852 pr_debug("%s: stripe %llu\n", __func__
,
853 (unsigned long long)sh
->sector
);
855 /* mark the computed target(s) as uptodate */
856 mark_target_uptodate(sh
, sh
->ops
.target
);
857 mark_target_uptodate(sh
, sh
->ops
.target2
);
859 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
860 if (sh
->check_state
== check_state_compute_run
)
861 sh
->check_state
= check_state_compute_result
;
862 set_bit(STRIPE_HANDLE
, &sh
->state
);
866 /* return a pointer to the address conversion region of the scribble buffer */
867 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
868 struct raid5_percpu
*percpu
)
870 return percpu
->scribble
+ sizeof(struct page
*) * (sh
->disks
+ 2);
873 static struct dma_async_tx_descriptor
*
874 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
876 int disks
= sh
->disks
;
877 struct page
**xor_srcs
= percpu
->scribble
;
878 int target
= sh
->ops
.target
;
879 struct r5dev
*tgt
= &sh
->dev
[target
];
880 struct page
*xor_dest
= tgt
->page
;
882 struct dma_async_tx_descriptor
*tx
;
883 struct async_submit_ctl submit
;
886 pr_debug("%s: stripe %llu block: %d\n",
887 __func__
, (unsigned long long)sh
->sector
, target
);
888 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
890 for (i
= disks
; i
--; )
892 xor_srcs
[count
++] = sh
->dev
[i
].page
;
894 atomic_inc(&sh
->count
);
896 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
897 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
));
898 if (unlikely(count
== 1))
899 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
901 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
906 /* set_syndrome_sources - populate source buffers for gen_syndrome
907 * @srcs - (struct page *) array of size sh->disks
908 * @sh - stripe_head to parse
910 * Populates srcs in proper layout order for the stripe and returns the
911 * 'count' of sources to be used in a call to async_gen_syndrome. The P
912 * destination buffer is recorded in srcs[count] and the Q destination
913 * is recorded in srcs[count+1]].
915 static int set_syndrome_sources(struct page
**srcs
, struct stripe_head
*sh
)
917 int disks
= sh
->disks
;
918 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
919 int d0_idx
= raid6_d0(sh
);
923 for (i
= 0; i
< disks
; i
++)
929 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
931 srcs
[slot
] = sh
->dev
[i
].page
;
932 i
= raid6_next_disk(i
, disks
);
933 } while (i
!= d0_idx
);
935 return syndrome_disks
;
938 static struct dma_async_tx_descriptor
*
939 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
941 int disks
= sh
->disks
;
942 struct page
**blocks
= percpu
->scribble
;
944 int qd_idx
= sh
->qd_idx
;
945 struct dma_async_tx_descriptor
*tx
;
946 struct async_submit_ctl submit
;
952 if (sh
->ops
.target
< 0)
953 target
= sh
->ops
.target2
;
954 else if (sh
->ops
.target2
< 0)
955 target
= sh
->ops
.target
;
957 /* we should only have one valid target */
960 pr_debug("%s: stripe %llu block: %d\n",
961 __func__
, (unsigned long long)sh
->sector
, target
);
963 tgt
= &sh
->dev
[target
];
964 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
967 atomic_inc(&sh
->count
);
969 if (target
== qd_idx
) {
970 count
= set_syndrome_sources(blocks
, sh
);
971 blocks
[count
] = NULL
; /* regenerating p is not necessary */
972 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
973 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
974 ops_complete_compute
, sh
,
975 to_addr_conv(sh
, percpu
));
976 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
978 /* Compute any data- or p-drive using XOR */
980 for (i
= disks
; i
-- ; ) {
981 if (i
== target
|| i
== qd_idx
)
983 blocks
[count
++] = sh
->dev
[i
].page
;
986 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
987 NULL
, ops_complete_compute
, sh
,
988 to_addr_conv(sh
, percpu
));
989 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
995 static struct dma_async_tx_descriptor
*
996 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
998 int i
, count
, disks
= sh
->disks
;
999 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1000 int d0_idx
= raid6_d0(sh
);
1001 int faila
= -1, failb
= -1;
1002 int target
= sh
->ops
.target
;
1003 int target2
= sh
->ops
.target2
;
1004 struct r5dev
*tgt
= &sh
->dev
[target
];
1005 struct r5dev
*tgt2
= &sh
->dev
[target2
];
1006 struct dma_async_tx_descriptor
*tx
;
1007 struct page
**blocks
= percpu
->scribble
;
1008 struct async_submit_ctl submit
;
1010 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1011 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
1012 BUG_ON(target
< 0 || target2
< 0);
1013 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1014 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
1016 /* we need to open-code set_syndrome_sources to handle the
1017 * slot number conversion for 'faila' and 'failb'
1019 for (i
= 0; i
< disks
; i
++)
1024 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1026 blocks
[slot
] = sh
->dev
[i
].page
;
1032 i
= raid6_next_disk(i
, disks
);
1033 } while (i
!= d0_idx
);
1035 BUG_ON(faila
== failb
);
1038 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1039 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1041 atomic_inc(&sh
->count
);
1043 if (failb
== syndrome_disks
+1) {
1044 /* Q disk is one of the missing disks */
1045 if (faila
== syndrome_disks
) {
1046 /* Missing P+Q, just recompute */
1047 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1048 ops_complete_compute
, sh
,
1049 to_addr_conv(sh
, percpu
));
1050 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
1051 STRIPE_SIZE
, &submit
);
1055 int qd_idx
= sh
->qd_idx
;
1057 /* Missing D+Q: recompute D from P, then recompute Q */
1058 if (target
== qd_idx
)
1059 data_target
= target2
;
1061 data_target
= target
;
1064 for (i
= disks
; i
-- ; ) {
1065 if (i
== data_target
|| i
== qd_idx
)
1067 blocks
[count
++] = sh
->dev
[i
].page
;
1069 dest
= sh
->dev
[data_target
].page
;
1070 init_async_submit(&submit
,
1071 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1073 to_addr_conv(sh
, percpu
));
1074 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
1077 count
= set_syndrome_sources(blocks
, sh
);
1078 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1079 ops_complete_compute
, sh
,
1080 to_addr_conv(sh
, percpu
));
1081 return async_gen_syndrome(blocks
, 0, count
+2,
1082 STRIPE_SIZE
, &submit
);
1085 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1086 ops_complete_compute
, sh
,
1087 to_addr_conv(sh
, percpu
));
1088 if (failb
== syndrome_disks
) {
1089 /* We're missing D+P. */
1090 return async_raid6_datap_recov(syndrome_disks
+2,
1094 /* We're missing D+D. */
1095 return async_raid6_2data_recov(syndrome_disks
+2,
1096 STRIPE_SIZE
, faila
, failb
,
1103 static void ops_complete_prexor(void *stripe_head_ref
)
1105 struct stripe_head
*sh
= stripe_head_ref
;
1107 pr_debug("%s: stripe %llu\n", __func__
,
1108 (unsigned long long)sh
->sector
);
1111 static struct dma_async_tx_descriptor
*
1112 ops_run_prexor(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1113 struct dma_async_tx_descriptor
*tx
)
1115 int disks
= sh
->disks
;
1116 struct page
**xor_srcs
= percpu
->scribble
;
1117 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1118 struct async_submit_ctl submit
;
1120 /* existing parity data subtracted */
1121 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1123 pr_debug("%s: stripe %llu\n", __func__
,
1124 (unsigned long long)sh
->sector
);
1126 for (i
= disks
; i
--; ) {
1127 struct r5dev
*dev
= &sh
->dev
[i
];
1128 /* Only process blocks that are known to be uptodate */
1129 if (test_bit(R5_Wantdrain
, &dev
->flags
))
1130 xor_srcs
[count
++] = dev
->page
;
1133 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1134 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
));
1135 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1140 static struct dma_async_tx_descriptor
*
1141 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1143 int disks
= sh
->disks
;
1146 pr_debug("%s: stripe %llu\n", __func__
,
1147 (unsigned long long)sh
->sector
);
1149 for (i
= disks
; i
--; ) {
1150 struct r5dev
*dev
= &sh
->dev
[i
];
1153 if (test_and_clear_bit(R5_Wantdrain
, &dev
->flags
)) {
1156 spin_lock_irq(&sh
->stripe_lock
);
1157 chosen
= dev
->towrite
;
1158 dev
->towrite
= NULL
;
1159 BUG_ON(dev
->written
);
1160 wbi
= dev
->written
= chosen
;
1161 spin_unlock_irq(&sh
->stripe_lock
);
1163 while (wbi
&& wbi
->bi_sector
<
1164 dev
->sector
+ STRIPE_SECTORS
) {
1165 if (wbi
->bi_rw
& REQ_FUA
)
1166 set_bit(R5_WantFUA
, &dev
->flags
);
1167 if (wbi
->bi_rw
& REQ_SYNC
)
1168 set_bit(R5_SyncIO
, &dev
->flags
);
1169 tx
= async_copy_data(1, wbi
, dev
->page
,
1171 wbi
= r5_next_bio(wbi
, dev
->sector
);
1179 static void ops_complete_reconstruct(void *stripe_head_ref
)
1181 struct stripe_head
*sh
= stripe_head_ref
;
1182 int disks
= sh
->disks
;
1183 int pd_idx
= sh
->pd_idx
;
1184 int qd_idx
= sh
->qd_idx
;
1186 bool fua
= false, sync
= false;
1188 pr_debug("%s: stripe %llu\n", __func__
,
1189 (unsigned long long)sh
->sector
);
1191 for (i
= disks
; i
--; ) {
1192 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1193 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
1196 for (i
= disks
; i
--; ) {
1197 struct r5dev
*dev
= &sh
->dev
[i
];
1199 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1200 set_bit(R5_UPTODATE
, &dev
->flags
);
1202 set_bit(R5_WantFUA
, &dev
->flags
);
1204 set_bit(R5_SyncIO
, &dev
->flags
);
1208 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1209 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1210 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1211 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1213 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1214 sh
->reconstruct_state
= reconstruct_state_result
;
1217 set_bit(STRIPE_HANDLE
, &sh
->state
);
1222 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1223 struct dma_async_tx_descriptor
*tx
)
1225 int disks
= sh
->disks
;
1226 struct page
**xor_srcs
= percpu
->scribble
;
1227 struct async_submit_ctl submit
;
1228 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1229 struct page
*xor_dest
;
1231 unsigned long flags
;
1233 pr_debug("%s: stripe %llu\n", __func__
,
1234 (unsigned long long)sh
->sector
);
1236 /* check if prexor is active which means only process blocks
1237 * that are part of a read-modify-write (written)
1239 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1241 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1242 for (i
= disks
; i
--; ) {
1243 struct r5dev
*dev
= &sh
->dev
[i
];
1245 xor_srcs
[count
++] = dev
->page
;
1248 xor_dest
= sh
->dev
[pd_idx
].page
;
1249 for (i
= disks
; i
--; ) {
1250 struct r5dev
*dev
= &sh
->dev
[i
];
1252 xor_srcs
[count
++] = dev
->page
;
1256 /* 1/ if we prexor'd then the dest is reused as a source
1257 * 2/ if we did not prexor then we are redoing the parity
1258 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1259 * for the synchronous xor case
1261 flags
= ASYNC_TX_ACK
|
1262 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1264 atomic_inc(&sh
->count
);
1266 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, sh
,
1267 to_addr_conv(sh
, percpu
));
1268 if (unlikely(count
== 1))
1269 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1271 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1275 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1276 struct dma_async_tx_descriptor
*tx
)
1278 struct async_submit_ctl submit
;
1279 struct page
**blocks
= percpu
->scribble
;
1282 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1284 count
= set_syndrome_sources(blocks
, sh
);
1286 atomic_inc(&sh
->count
);
1288 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_reconstruct
,
1289 sh
, to_addr_conv(sh
, percpu
));
1290 async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1293 static void ops_complete_check(void *stripe_head_ref
)
1295 struct stripe_head
*sh
= stripe_head_ref
;
1297 pr_debug("%s: stripe %llu\n", __func__
,
1298 (unsigned long long)sh
->sector
);
1300 sh
->check_state
= check_state_check_result
;
1301 set_bit(STRIPE_HANDLE
, &sh
->state
);
1305 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1307 int disks
= sh
->disks
;
1308 int pd_idx
= sh
->pd_idx
;
1309 int qd_idx
= sh
->qd_idx
;
1310 struct page
*xor_dest
;
1311 struct page
**xor_srcs
= percpu
->scribble
;
1312 struct dma_async_tx_descriptor
*tx
;
1313 struct async_submit_ctl submit
;
1317 pr_debug("%s: stripe %llu\n", __func__
,
1318 (unsigned long long)sh
->sector
);
1321 xor_dest
= sh
->dev
[pd_idx
].page
;
1322 xor_srcs
[count
++] = xor_dest
;
1323 for (i
= disks
; i
--; ) {
1324 if (i
== pd_idx
|| i
== qd_idx
)
1326 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1329 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1330 to_addr_conv(sh
, percpu
));
1331 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1332 &sh
->ops
.zero_sum_result
, &submit
);
1334 atomic_inc(&sh
->count
);
1335 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1336 tx
= async_trigger_callback(&submit
);
1339 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1341 struct page
**srcs
= percpu
->scribble
;
1342 struct async_submit_ctl submit
;
1345 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1346 (unsigned long long)sh
->sector
, checkp
);
1348 count
= set_syndrome_sources(srcs
, sh
);
1352 atomic_inc(&sh
->count
);
1353 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1354 sh
, to_addr_conv(sh
, percpu
));
1355 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1356 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1359 static void __raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1361 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1362 struct dma_async_tx_descriptor
*tx
= NULL
;
1363 struct r5conf
*conf
= sh
->raid_conf
;
1364 int level
= conf
->level
;
1365 struct raid5_percpu
*percpu
;
1369 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1370 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1371 ops_run_biofill(sh
);
1375 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1377 tx
= ops_run_compute5(sh
, percpu
);
1379 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
1380 tx
= ops_run_compute6_1(sh
, percpu
);
1382 tx
= ops_run_compute6_2(sh
, percpu
);
1384 /* terminate the chain if reconstruct is not set to be run */
1385 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
1389 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
))
1390 tx
= ops_run_prexor(sh
, percpu
, tx
);
1392 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
1393 tx
= ops_run_biodrain(sh
, tx
);
1397 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
1399 ops_run_reconstruct5(sh
, percpu
, tx
);
1401 ops_run_reconstruct6(sh
, percpu
, tx
);
1404 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
1405 if (sh
->check_state
== check_state_run
)
1406 ops_run_check_p(sh
, percpu
);
1407 else if (sh
->check_state
== check_state_run_q
)
1408 ops_run_check_pq(sh
, percpu
, 0);
1409 else if (sh
->check_state
== check_state_run_pq
)
1410 ops_run_check_pq(sh
, percpu
, 1);
1416 for (i
= disks
; i
--; ) {
1417 struct r5dev
*dev
= &sh
->dev
[i
];
1418 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
1419 wake_up(&sh
->raid_conf
->wait_for_overlap
);
1424 #ifdef CONFIG_MULTICORE_RAID456
1425 static void async_run_ops(void *param
, async_cookie_t cookie
)
1427 struct stripe_head
*sh
= param
;
1428 unsigned long ops_request
= sh
->ops
.request
;
1430 clear_bit_unlock(STRIPE_OPS_REQ_PENDING
, &sh
->state
);
1431 wake_up(&sh
->ops
.wait_for_ops
);
1433 __raid_run_ops(sh
, ops_request
);
1437 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1439 /* since handle_stripe can be called outside of raid5d context
1440 * we need to ensure sh->ops.request is de-staged before another
1443 wait_event(sh
->ops
.wait_for_ops
,
1444 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING
, &sh
->state
));
1445 sh
->ops
.request
= ops_request
;
1447 atomic_inc(&sh
->count
);
1448 async_schedule(async_run_ops
, sh
);
1451 #define raid_run_ops __raid_run_ops
1454 static int grow_one_stripe(struct r5conf
*conf
)
1456 struct stripe_head
*sh
;
1457 sh
= kmem_cache_zalloc(conf
->slab_cache
, GFP_KERNEL
);
1461 sh
->raid_conf
= conf
;
1462 #ifdef CONFIG_MULTICORE_RAID456
1463 init_waitqueue_head(&sh
->ops
.wait_for_ops
);
1466 spin_lock_init(&sh
->stripe_lock
);
1468 if (grow_buffers(sh
)) {
1470 kmem_cache_free(conf
->slab_cache
, sh
);
1473 /* we just created an active stripe so... */
1474 atomic_set(&sh
->count
, 1);
1475 atomic_inc(&conf
->active_stripes
);
1476 INIT_LIST_HEAD(&sh
->lru
);
1481 static int grow_stripes(struct r5conf
*conf
, int num
)
1483 struct kmem_cache
*sc
;
1484 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
1486 if (conf
->mddev
->gendisk
)
1487 sprintf(conf
->cache_name
[0],
1488 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
1490 sprintf(conf
->cache_name
[0],
1491 "raid%d-%p", conf
->level
, conf
->mddev
);
1492 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
1494 conf
->active_name
= 0;
1495 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
1496 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
1500 conf
->slab_cache
= sc
;
1501 conf
->pool_size
= devs
;
1503 if (!grow_one_stripe(conf
))
1509 * scribble_len - return the required size of the scribble region
1510 * @num - total number of disks in the array
1512 * The size must be enough to contain:
1513 * 1/ a struct page pointer for each device in the array +2
1514 * 2/ room to convert each entry in (1) to its corresponding dma
1515 * (dma_map_page()) or page (page_address()) address.
1517 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1518 * calculate over all devices (not just the data blocks), using zeros in place
1519 * of the P and Q blocks.
1521 static size_t scribble_len(int num
)
1525 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
1530 static int resize_stripes(struct r5conf
*conf
, int newsize
)
1532 /* Make all the stripes able to hold 'newsize' devices.
1533 * New slots in each stripe get 'page' set to a new page.
1535 * This happens in stages:
1536 * 1/ create a new kmem_cache and allocate the required number of
1538 * 2/ gather all the old stripe_heads and tranfer the pages across
1539 * to the new stripe_heads. This will have the side effect of
1540 * freezing the array as once all stripe_heads have been collected,
1541 * no IO will be possible. Old stripe heads are freed once their
1542 * pages have been transferred over, and the old kmem_cache is
1543 * freed when all stripes are done.
1544 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1545 * we simple return a failre status - no need to clean anything up.
1546 * 4/ allocate new pages for the new slots in the new stripe_heads.
1547 * If this fails, we don't bother trying the shrink the
1548 * stripe_heads down again, we just leave them as they are.
1549 * As each stripe_head is processed the new one is released into
1552 * Once step2 is started, we cannot afford to wait for a write,
1553 * so we use GFP_NOIO allocations.
1555 struct stripe_head
*osh
, *nsh
;
1556 LIST_HEAD(newstripes
);
1557 struct disk_info
*ndisks
;
1560 struct kmem_cache
*sc
;
1563 if (newsize
<= conf
->pool_size
)
1564 return 0; /* never bother to shrink */
1566 err
= md_allow_write(conf
->mddev
);
1571 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
1572 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
1577 for (i
= conf
->max_nr_stripes
; i
; i
--) {
1578 nsh
= kmem_cache_zalloc(sc
, GFP_KERNEL
);
1582 nsh
->raid_conf
= conf
;
1583 #ifdef CONFIG_MULTICORE_RAID456
1584 init_waitqueue_head(&nsh
->ops
.wait_for_ops
);
1587 list_add(&nsh
->lru
, &newstripes
);
1590 /* didn't get enough, give up */
1591 while (!list_empty(&newstripes
)) {
1592 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1593 list_del(&nsh
->lru
);
1594 kmem_cache_free(sc
, nsh
);
1596 kmem_cache_destroy(sc
);
1599 /* Step 2 - Must use GFP_NOIO now.
1600 * OK, we have enough stripes, start collecting inactive
1601 * stripes and copying them over
1603 list_for_each_entry(nsh
, &newstripes
, lru
) {
1604 spin_lock_irq(&conf
->device_lock
);
1605 wait_event_lock_irq(conf
->wait_for_stripe
,
1606 !list_empty(&conf
->inactive_list
),
1609 osh
= get_free_stripe(conf
);
1610 spin_unlock_irq(&conf
->device_lock
);
1611 atomic_set(&nsh
->count
, 1);
1612 for(i
=0; i
<conf
->pool_size
; i
++)
1613 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
1614 for( ; i
<newsize
; i
++)
1615 nsh
->dev
[i
].page
= NULL
;
1616 kmem_cache_free(conf
->slab_cache
, osh
);
1618 kmem_cache_destroy(conf
->slab_cache
);
1621 * At this point, we are holding all the stripes so the array
1622 * is completely stalled, so now is a good time to resize
1623 * conf->disks and the scribble region
1625 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
1627 for (i
=0; i
<conf
->raid_disks
; i
++)
1628 ndisks
[i
] = conf
->disks
[i
];
1630 conf
->disks
= ndisks
;
1635 conf
->scribble_len
= scribble_len(newsize
);
1636 for_each_present_cpu(cpu
) {
1637 struct raid5_percpu
*percpu
;
1640 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1641 scribble
= kmalloc(conf
->scribble_len
, GFP_NOIO
);
1644 kfree(percpu
->scribble
);
1645 percpu
->scribble
= scribble
;
1653 /* Step 4, return new stripes to service */
1654 while(!list_empty(&newstripes
)) {
1655 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1656 list_del_init(&nsh
->lru
);
1658 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
1659 if (nsh
->dev
[i
].page
== NULL
) {
1660 struct page
*p
= alloc_page(GFP_NOIO
);
1661 nsh
->dev
[i
].page
= p
;
1665 release_stripe(nsh
);
1667 /* critical section pass, GFP_NOIO no longer needed */
1669 conf
->slab_cache
= sc
;
1670 conf
->active_name
= 1-conf
->active_name
;
1671 conf
->pool_size
= newsize
;
1675 static int drop_one_stripe(struct r5conf
*conf
)
1677 struct stripe_head
*sh
;
1679 spin_lock_irq(&conf
->device_lock
);
1680 sh
= get_free_stripe(conf
);
1681 spin_unlock_irq(&conf
->device_lock
);
1684 BUG_ON(atomic_read(&sh
->count
));
1686 kmem_cache_free(conf
->slab_cache
, sh
);
1687 atomic_dec(&conf
->active_stripes
);
1691 static void shrink_stripes(struct r5conf
*conf
)
1693 while (drop_one_stripe(conf
))
1696 if (conf
->slab_cache
)
1697 kmem_cache_destroy(conf
->slab_cache
);
1698 conf
->slab_cache
= NULL
;
1701 static void raid5_end_read_request(struct bio
* bi
, int error
)
1703 struct stripe_head
*sh
= bi
->bi_private
;
1704 struct r5conf
*conf
= sh
->raid_conf
;
1705 int disks
= sh
->disks
, i
;
1706 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1707 char b
[BDEVNAME_SIZE
];
1708 struct md_rdev
*rdev
= NULL
;
1711 for (i
=0 ; i
<disks
; i
++)
1712 if (bi
== &sh
->dev
[i
].req
)
1715 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1716 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1722 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
1723 /* If replacement finished while this request was outstanding,
1724 * 'replacement' might be NULL already.
1725 * In that case it moved down to 'rdev'.
1726 * rdev is not removed until all requests are finished.
1728 rdev
= conf
->disks
[i
].replacement
;
1730 rdev
= conf
->disks
[i
].rdev
;
1732 if (use_new_offset(conf
, sh
))
1733 s
= sh
->sector
+ rdev
->new_data_offset
;
1735 s
= sh
->sector
+ rdev
->data_offset
;
1737 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1738 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
1739 /* Note that this cannot happen on a
1740 * replacement device. We just fail those on
1745 "md/raid:%s: read error corrected"
1746 " (%lu sectors at %llu on %s)\n",
1747 mdname(conf
->mddev
), STRIPE_SECTORS
,
1748 (unsigned long long)s
,
1749 bdevname(rdev
->bdev
, b
));
1750 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
1751 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1752 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1754 if (atomic_read(&rdev
->read_errors
))
1755 atomic_set(&rdev
->read_errors
, 0);
1757 const char *bdn
= bdevname(rdev
->bdev
, b
);
1761 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1762 atomic_inc(&rdev
->read_errors
);
1763 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
1766 "md/raid:%s: read error on replacement device "
1767 "(sector %llu on %s).\n",
1768 mdname(conf
->mddev
),
1769 (unsigned long long)s
,
1771 else if (conf
->mddev
->degraded
>= conf
->max_degraded
) {
1775 "md/raid:%s: read error not correctable "
1776 "(sector %llu on %s).\n",
1777 mdname(conf
->mddev
),
1778 (unsigned long long)s
,
1780 } else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
)) {
1785 "md/raid:%s: read error NOT corrected!! "
1786 "(sector %llu on %s).\n",
1787 mdname(conf
->mddev
),
1788 (unsigned long long)s
,
1790 } else if (atomic_read(&rdev
->read_errors
)
1791 > conf
->max_nr_stripes
)
1793 "md/raid:%s: Too many read errors, failing device %s.\n",
1794 mdname(conf
->mddev
), bdn
);
1798 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1800 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1801 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1803 && test_bit(In_sync
, &rdev
->flags
)
1804 && rdev_set_badblocks(
1805 rdev
, sh
->sector
, STRIPE_SECTORS
, 0)))
1806 md_error(conf
->mddev
, rdev
);
1809 rdev_dec_pending(rdev
, conf
->mddev
);
1810 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1811 set_bit(STRIPE_HANDLE
, &sh
->state
);
1815 static void raid5_end_write_request(struct bio
*bi
, int error
)
1817 struct stripe_head
*sh
= bi
->bi_private
;
1818 struct r5conf
*conf
= sh
->raid_conf
;
1819 int disks
= sh
->disks
, i
;
1820 struct md_rdev
*uninitialized_var(rdev
);
1821 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1824 int replacement
= 0;
1826 for (i
= 0 ; i
< disks
; i
++) {
1827 if (bi
== &sh
->dev
[i
].req
) {
1828 rdev
= conf
->disks
[i
].rdev
;
1831 if (bi
== &sh
->dev
[i
].rreq
) {
1832 rdev
= conf
->disks
[i
].replacement
;
1836 /* rdev was removed and 'replacement'
1837 * replaced it. rdev is not removed
1838 * until all requests are finished.
1840 rdev
= conf
->disks
[i
].rdev
;
1844 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1845 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1854 md_error(conf
->mddev
, rdev
);
1855 else if (is_badblock(rdev
, sh
->sector
,
1857 &first_bad
, &bad_sectors
))
1858 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
1861 set_bit(WriteErrorSeen
, &rdev
->flags
);
1862 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
1863 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1864 set_bit(MD_RECOVERY_NEEDED
,
1865 &rdev
->mddev
->recovery
);
1866 } else if (is_badblock(rdev
, sh
->sector
,
1868 &first_bad
, &bad_sectors
))
1869 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
1871 rdev_dec_pending(rdev
, conf
->mddev
);
1873 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
1874 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1875 set_bit(STRIPE_HANDLE
, &sh
->state
);
1879 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
);
1881 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
1883 struct r5dev
*dev
= &sh
->dev
[i
];
1885 bio_init(&dev
->req
);
1886 dev
->req
.bi_io_vec
= &dev
->vec
;
1888 dev
->req
.bi_max_vecs
++;
1889 dev
->req
.bi_private
= sh
;
1890 dev
->vec
.bv_page
= dev
->page
;
1892 bio_init(&dev
->rreq
);
1893 dev
->rreq
.bi_io_vec
= &dev
->rvec
;
1894 dev
->rreq
.bi_vcnt
++;
1895 dev
->rreq
.bi_max_vecs
++;
1896 dev
->rreq
.bi_private
= sh
;
1897 dev
->rvec
.bv_page
= dev
->page
;
1900 dev
->sector
= compute_blocknr(sh
, i
, previous
);
1903 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1905 char b
[BDEVNAME_SIZE
];
1906 struct r5conf
*conf
= mddev
->private;
1907 unsigned long flags
;
1908 pr_debug("raid456: error called\n");
1910 spin_lock_irqsave(&conf
->device_lock
, flags
);
1911 clear_bit(In_sync
, &rdev
->flags
);
1912 mddev
->degraded
= calc_degraded(conf
);
1913 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1914 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1916 set_bit(Blocked
, &rdev
->flags
);
1917 set_bit(Faulty
, &rdev
->flags
);
1918 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1920 "md/raid:%s: Disk failure on %s, disabling device.\n"
1921 "md/raid:%s: Operation continuing on %d devices.\n",
1923 bdevname(rdev
->bdev
, b
),
1925 conf
->raid_disks
- mddev
->degraded
);
1929 * Input: a 'big' sector number,
1930 * Output: index of the data and parity disk, and the sector # in them.
1932 static sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
1933 int previous
, int *dd_idx
,
1934 struct stripe_head
*sh
)
1936 sector_t stripe
, stripe2
;
1937 sector_t chunk_number
;
1938 unsigned int chunk_offset
;
1941 sector_t new_sector
;
1942 int algorithm
= previous
? conf
->prev_algo
1944 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
1945 : conf
->chunk_sectors
;
1946 int raid_disks
= previous
? conf
->previous_raid_disks
1948 int data_disks
= raid_disks
- conf
->max_degraded
;
1950 /* First compute the information on this sector */
1953 * Compute the chunk number and the sector offset inside the chunk
1955 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
1956 chunk_number
= r_sector
;
1959 * Compute the stripe number
1961 stripe
= chunk_number
;
1962 *dd_idx
= sector_div(stripe
, data_disks
);
1965 * Select the parity disk based on the user selected algorithm.
1967 pd_idx
= qd_idx
= -1;
1968 switch(conf
->level
) {
1970 pd_idx
= data_disks
;
1973 switch (algorithm
) {
1974 case ALGORITHM_LEFT_ASYMMETRIC
:
1975 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
1976 if (*dd_idx
>= pd_idx
)
1979 case ALGORITHM_RIGHT_ASYMMETRIC
:
1980 pd_idx
= sector_div(stripe2
, raid_disks
);
1981 if (*dd_idx
>= pd_idx
)
1984 case ALGORITHM_LEFT_SYMMETRIC
:
1985 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
1986 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1988 case ALGORITHM_RIGHT_SYMMETRIC
:
1989 pd_idx
= sector_div(stripe2
, raid_disks
);
1990 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1992 case ALGORITHM_PARITY_0
:
1996 case ALGORITHM_PARITY_N
:
1997 pd_idx
= data_disks
;
2005 switch (algorithm
) {
2006 case ALGORITHM_LEFT_ASYMMETRIC
:
2007 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2008 qd_idx
= pd_idx
+ 1;
2009 if (pd_idx
== raid_disks
-1) {
2010 (*dd_idx
)++; /* Q D D D P */
2012 } else if (*dd_idx
>= pd_idx
)
2013 (*dd_idx
) += 2; /* D D P Q D */
2015 case ALGORITHM_RIGHT_ASYMMETRIC
:
2016 pd_idx
= sector_div(stripe2
, raid_disks
);
2017 qd_idx
= pd_idx
+ 1;
2018 if (pd_idx
== raid_disks
-1) {
2019 (*dd_idx
)++; /* Q D D D P */
2021 } else if (*dd_idx
>= pd_idx
)
2022 (*dd_idx
) += 2; /* D D P Q D */
2024 case ALGORITHM_LEFT_SYMMETRIC
:
2025 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2026 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2027 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2029 case ALGORITHM_RIGHT_SYMMETRIC
:
2030 pd_idx
= sector_div(stripe2
, raid_disks
);
2031 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2032 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2035 case ALGORITHM_PARITY_0
:
2040 case ALGORITHM_PARITY_N
:
2041 pd_idx
= data_disks
;
2042 qd_idx
= data_disks
+ 1;
2045 case ALGORITHM_ROTATING_ZERO_RESTART
:
2046 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2047 * of blocks for computing Q is different.
2049 pd_idx
= sector_div(stripe2
, raid_disks
);
2050 qd_idx
= pd_idx
+ 1;
2051 if (pd_idx
== raid_disks
-1) {
2052 (*dd_idx
)++; /* Q D D D P */
2054 } else if (*dd_idx
>= pd_idx
)
2055 (*dd_idx
) += 2; /* D D P Q D */
2059 case ALGORITHM_ROTATING_N_RESTART
:
2060 /* Same a left_asymmetric, by first stripe is
2061 * D D D P Q rather than
2065 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2066 qd_idx
= pd_idx
+ 1;
2067 if (pd_idx
== raid_disks
-1) {
2068 (*dd_idx
)++; /* Q D D D P */
2070 } else if (*dd_idx
>= pd_idx
)
2071 (*dd_idx
) += 2; /* D D P Q D */
2075 case ALGORITHM_ROTATING_N_CONTINUE
:
2076 /* Same as left_symmetric but Q is before P */
2077 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2078 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
2079 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2083 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2084 /* RAID5 left_asymmetric, with Q on last device */
2085 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2086 if (*dd_idx
>= pd_idx
)
2088 qd_idx
= raid_disks
- 1;
2091 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2092 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2093 if (*dd_idx
>= pd_idx
)
2095 qd_idx
= raid_disks
- 1;
2098 case ALGORITHM_LEFT_SYMMETRIC_6
:
2099 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2100 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2101 qd_idx
= raid_disks
- 1;
2104 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2105 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2106 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2107 qd_idx
= raid_disks
- 1;
2110 case ALGORITHM_PARITY_0_6
:
2113 qd_idx
= raid_disks
- 1;
2123 sh
->pd_idx
= pd_idx
;
2124 sh
->qd_idx
= qd_idx
;
2125 sh
->ddf_layout
= ddf_layout
;
2128 * Finally, compute the new sector number
2130 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
2135 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
2137 struct r5conf
*conf
= sh
->raid_conf
;
2138 int raid_disks
= sh
->disks
;
2139 int data_disks
= raid_disks
- conf
->max_degraded
;
2140 sector_t new_sector
= sh
->sector
, check
;
2141 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2142 : conf
->chunk_sectors
;
2143 int algorithm
= previous
? conf
->prev_algo
2147 sector_t chunk_number
;
2148 int dummy1
, dd_idx
= i
;
2150 struct stripe_head sh2
;
2153 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
2154 stripe
= new_sector
;
2156 if (i
== sh
->pd_idx
)
2158 switch(conf
->level
) {
2161 switch (algorithm
) {
2162 case ALGORITHM_LEFT_ASYMMETRIC
:
2163 case ALGORITHM_RIGHT_ASYMMETRIC
:
2167 case ALGORITHM_LEFT_SYMMETRIC
:
2168 case ALGORITHM_RIGHT_SYMMETRIC
:
2171 i
-= (sh
->pd_idx
+ 1);
2173 case ALGORITHM_PARITY_0
:
2176 case ALGORITHM_PARITY_N
:
2183 if (i
== sh
->qd_idx
)
2184 return 0; /* It is the Q disk */
2185 switch (algorithm
) {
2186 case ALGORITHM_LEFT_ASYMMETRIC
:
2187 case ALGORITHM_RIGHT_ASYMMETRIC
:
2188 case ALGORITHM_ROTATING_ZERO_RESTART
:
2189 case ALGORITHM_ROTATING_N_RESTART
:
2190 if (sh
->pd_idx
== raid_disks
-1)
2191 i
--; /* Q D D D P */
2192 else if (i
> sh
->pd_idx
)
2193 i
-= 2; /* D D P Q D */
2195 case ALGORITHM_LEFT_SYMMETRIC
:
2196 case ALGORITHM_RIGHT_SYMMETRIC
:
2197 if (sh
->pd_idx
== raid_disks
-1)
2198 i
--; /* Q D D D P */
2203 i
-= (sh
->pd_idx
+ 2);
2206 case ALGORITHM_PARITY_0
:
2209 case ALGORITHM_PARITY_N
:
2211 case ALGORITHM_ROTATING_N_CONTINUE
:
2212 /* Like left_symmetric, but P is before Q */
2213 if (sh
->pd_idx
== 0)
2214 i
--; /* P D D D Q */
2219 i
-= (sh
->pd_idx
+ 1);
2222 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2223 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2227 case ALGORITHM_LEFT_SYMMETRIC_6
:
2228 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2230 i
+= data_disks
+ 1;
2231 i
-= (sh
->pd_idx
+ 1);
2233 case ALGORITHM_PARITY_0_6
:
2242 chunk_number
= stripe
* data_disks
+ i
;
2243 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2245 check
= raid5_compute_sector(conf
, r_sector
,
2246 previous
, &dummy1
, &sh2
);
2247 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2248 || sh2
.qd_idx
!= sh
->qd_idx
) {
2249 printk(KERN_ERR
"md/raid:%s: compute_blocknr: map not correct\n",
2250 mdname(conf
->mddev
));
2258 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2259 int rcw
, int expand
)
2261 int i
, pd_idx
= sh
->pd_idx
, disks
= sh
->disks
;
2262 struct r5conf
*conf
= sh
->raid_conf
;
2263 int level
= conf
->level
;
2266 /* if we are not expanding this is a proper write request, and
2267 * there will be bios with new data to be drained into the
2271 sh
->reconstruct_state
= reconstruct_state_drain_run
;
2272 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2274 sh
->reconstruct_state
= reconstruct_state_run
;
2276 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2278 for (i
= disks
; i
--; ) {
2279 struct r5dev
*dev
= &sh
->dev
[i
];
2282 set_bit(R5_LOCKED
, &dev
->flags
);
2283 set_bit(R5_Wantdrain
, &dev
->flags
);
2285 clear_bit(R5_UPTODATE
, &dev
->flags
);
2289 if (s
->locked
+ conf
->max_degraded
== disks
)
2290 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2291 atomic_inc(&conf
->pending_full_writes
);
2294 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
2295 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
2297 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
2298 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
2299 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2300 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2302 for (i
= disks
; i
--; ) {
2303 struct r5dev
*dev
= &sh
->dev
[i
];
2308 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2309 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2310 set_bit(R5_Wantdrain
, &dev
->flags
);
2311 set_bit(R5_LOCKED
, &dev
->flags
);
2312 clear_bit(R5_UPTODATE
, &dev
->flags
);
2318 /* keep the parity disk(s) locked while asynchronous operations
2321 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
2322 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2326 int qd_idx
= sh
->qd_idx
;
2327 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
2329 set_bit(R5_LOCKED
, &dev
->flags
);
2330 clear_bit(R5_UPTODATE
, &dev
->flags
);
2334 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2335 __func__
, (unsigned long long)sh
->sector
,
2336 s
->locked
, s
->ops_request
);
2340 * Each stripe/dev can have one or more bion attached.
2341 * toread/towrite point to the first in a chain.
2342 * The bi_next chain must be in order.
2344 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
, int forwrite
)
2347 struct r5conf
*conf
= sh
->raid_conf
;
2350 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2351 (unsigned long long)bi
->bi_sector
,
2352 (unsigned long long)sh
->sector
);
2355 * If several bio share a stripe. The bio bi_phys_segments acts as a
2356 * reference count to avoid race. The reference count should already be
2357 * increased before this function is called (for example, in
2358 * make_request()), so other bio sharing this stripe will not free the
2359 * stripe. If a stripe is owned by one stripe, the stripe lock will
2362 spin_lock_irq(&sh
->stripe_lock
);
2364 bip
= &sh
->dev
[dd_idx
].towrite
;
2368 bip
= &sh
->dev
[dd_idx
].toread
;
2369 while (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
) {
2370 if ((*bip
)->bi_sector
+ ((*bip
)->bi_size
>> 9) > bi
->bi_sector
)
2372 bip
= & (*bip
)->bi_next
;
2374 if (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
+ ((bi
->bi_size
)>>9))
2377 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
2381 raid5_inc_bi_active_stripes(bi
);
2384 /* check if page is covered */
2385 sector_t sector
= sh
->dev
[dd_idx
].sector
;
2386 for (bi
=sh
->dev
[dd_idx
].towrite
;
2387 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
2388 bi
&& bi
->bi_sector
<= sector
;
2389 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
2390 if (bi
->bi_sector
+ (bi
->bi_size
>>9) >= sector
)
2391 sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
2393 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
2394 set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
);
2396 spin_unlock_irq(&sh
->stripe_lock
);
2398 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2399 (unsigned long long)(*bip
)->bi_sector
,
2400 (unsigned long long)sh
->sector
, dd_idx
);
2402 if (conf
->mddev
->bitmap
&& firstwrite
) {
2403 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
2405 sh
->bm_seq
= conf
->seq_flush
+1;
2406 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
2411 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
2412 spin_unlock_irq(&sh
->stripe_lock
);
2416 static void end_reshape(struct r5conf
*conf
);
2418 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
2419 struct stripe_head
*sh
)
2421 int sectors_per_chunk
=
2422 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
2424 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
2425 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
2427 raid5_compute_sector(conf
,
2428 stripe
* (disks
- conf
->max_degraded
)
2429 *sectors_per_chunk
+ chunk_offset
,
2435 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
2436 struct stripe_head_state
*s
, int disks
,
2437 struct bio
**return_bi
)
2440 for (i
= disks
; i
--; ) {
2444 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2445 struct md_rdev
*rdev
;
2447 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2448 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
2449 atomic_inc(&rdev
->nr_pending
);
2454 if (!rdev_set_badblocks(
2458 md_error(conf
->mddev
, rdev
);
2459 rdev_dec_pending(rdev
, conf
->mddev
);
2462 spin_lock_irq(&sh
->stripe_lock
);
2463 /* fail all writes first */
2464 bi
= sh
->dev
[i
].towrite
;
2465 sh
->dev
[i
].towrite
= NULL
;
2466 spin_unlock_irq(&sh
->stripe_lock
);
2472 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2473 wake_up(&conf
->wait_for_overlap
);
2475 while (bi
&& bi
->bi_sector
<
2476 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2477 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2478 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2479 if (!raid5_dec_bi_active_stripes(bi
)) {
2480 md_write_end(conf
->mddev
);
2481 bi
->bi_next
= *return_bi
;
2487 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2488 STRIPE_SECTORS
, 0, 0);
2490 /* and fail all 'written' */
2491 bi
= sh
->dev
[i
].written
;
2492 sh
->dev
[i
].written
= NULL
;
2493 if (bi
) bitmap_end
= 1;
2494 while (bi
&& bi
->bi_sector
<
2495 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2496 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2497 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2498 if (!raid5_dec_bi_active_stripes(bi
)) {
2499 md_write_end(conf
->mddev
);
2500 bi
->bi_next
= *return_bi
;
2506 /* fail any reads if this device is non-operational and
2507 * the data has not reached the cache yet.
2509 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
2510 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
2511 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
2512 bi
= sh
->dev
[i
].toread
;
2513 sh
->dev
[i
].toread
= NULL
;
2514 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2515 wake_up(&conf
->wait_for_overlap
);
2516 if (bi
) s
->to_read
--;
2517 while (bi
&& bi
->bi_sector
<
2518 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2519 struct bio
*nextbi
=
2520 r5_next_bio(bi
, sh
->dev
[i
].sector
);
2521 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2522 if (!raid5_dec_bi_active_stripes(bi
)) {
2523 bi
->bi_next
= *return_bi
;
2530 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2531 STRIPE_SECTORS
, 0, 0);
2532 /* If we were in the middle of a write the parity block might
2533 * still be locked - so just clear all R5_LOCKED flags
2535 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2538 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2539 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2540 md_wakeup_thread(conf
->mddev
->thread
);
2544 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
2545 struct stripe_head_state
*s
)
2550 clear_bit(STRIPE_SYNCING
, &sh
->state
);
2553 /* There is nothing more to do for sync/check/repair.
2554 * Don't even need to abort as that is handled elsewhere
2555 * if needed, and not always wanted e.g. if there is a known
2557 * For recover/replace we need to record a bad block on all
2558 * non-sync devices, or abort the recovery
2560 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
2561 /* During recovery devices cannot be removed, so
2562 * locking and refcounting of rdevs is not needed
2564 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2565 struct md_rdev
*rdev
= conf
->disks
[i
].rdev
;
2567 && !test_bit(Faulty
, &rdev
->flags
)
2568 && !test_bit(In_sync
, &rdev
->flags
)
2569 && !rdev_set_badblocks(rdev
, sh
->sector
,
2572 rdev
= conf
->disks
[i
].replacement
;
2574 && !test_bit(Faulty
, &rdev
->flags
)
2575 && !test_bit(In_sync
, &rdev
->flags
)
2576 && !rdev_set_badblocks(rdev
, sh
->sector
,
2581 conf
->recovery_disabled
=
2582 conf
->mddev
->recovery_disabled
;
2584 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, !abort
);
2587 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
2589 struct md_rdev
*rdev
;
2591 /* Doing recovery so rcu locking not required */
2592 rdev
= sh
->raid_conf
->disks
[disk_idx
].replacement
;
2594 && !test_bit(Faulty
, &rdev
->flags
)
2595 && !test_bit(In_sync
, &rdev
->flags
)
2596 && (rdev
->recovery_offset
<= sh
->sector
2597 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
2603 /* fetch_block - checks the given member device to see if its data needs
2604 * to be read or computed to satisfy a request.
2606 * Returns 1 when no more member devices need to be checked, otherwise returns
2607 * 0 to tell the loop in handle_stripe_fill to continue
2609 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2610 int disk_idx
, int disks
)
2612 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
2613 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
2614 &sh
->dev
[s
->failed_num
[1]] };
2616 /* is the data in this block needed, and can we get it? */
2617 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2618 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2620 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2621 s
->syncing
|| s
->expanding
||
2622 (s
->replacing
&& want_replace(sh
, disk_idx
)) ||
2623 (s
->failed
>= 1 && fdev
[0]->toread
) ||
2624 (s
->failed
>= 2 && fdev
[1]->toread
) ||
2625 (sh
->raid_conf
->level
<= 5 && s
->failed
&& fdev
[0]->towrite
&&
2626 !test_bit(R5_OVERWRITE
, &fdev
[0]->flags
)) ||
2627 (sh
->raid_conf
->level
== 6 && s
->failed
&& s
->to_write
))) {
2628 /* we would like to get this block, possibly by computing it,
2629 * otherwise read it if the backing disk is insync
2631 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
2632 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
2633 if ((s
->uptodate
== disks
- 1) &&
2634 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
2635 disk_idx
== s
->failed_num
[1]))) {
2636 /* have disk failed, and we're requested to fetch it;
2639 pr_debug("Computing stripe %llu block %d\n",
2640 (unsigned long long)sh
->sector
, disk_idx
);
2641 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2642 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2643 set_bit(R5_Wantcompute
, &dev
->flags
);
2644 sh
->ops
.target
= disk_idx
;
2645 sh
->ops
.target2
= -1; /* no 2nd target */
2647 /* Careful: from this point on 'uptodate' is in the eye
2648 * of raid_run_ops which services 'compute' operations
2649 * before writes. R5_Wantcompute flags a block that will
2650 * be R5_UPTODATE by the time it is needed for a
2651 * subsequent operation.
2655 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
2656 /* Computing 2-failure is *very* expensive; only
2657 * do it if failed >= 2
2660 for (other
= disks
; other
--; ) {
2661 if (other
== disk_idx
)
2663 if (!test_bit(R5_UPTODATE
,
2664 &sh
->dev
[other
].flags
))
2668 pr_debug("Computing stripe %llu blocks %d,%d\n",
2669 (unsigned long long)sh
->sector
,
2671 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2672 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2673 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
2674 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
2675 sh
->ops
.target
= disk_idx
;
2676 sh
->ops
.target2
= other
;
2680 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2681 set_bit(R5_LOCKED
, &dev
->flags
);
2682 set_bit(R5_Wantread
, &dev
->flags
);
2684 pr_debug("Reading block %d (sync=%d)\n",
2685 disk_idx
, s
->syncing
);
2693 * handle_stripe_fill - read or compute data to satisfy pending requests.
2695 static void handle_stripe_fill(struct stripe_head
*sh
,
2696 struct stripe_head_state
*s
,
2701 /* look for blocks to read/compute, skip this if a compute
2702 * is already in flight, or if the stripe contents are in the
2703 * midst of changing due to a write
2705 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2706 !sh
->reconstruct_state
)
2707 for (i
= disks
; i
--; )
2708 if (fetch_block(sh
, s
, i
, disks
))
2710 set_bit(STRIPE_HANDLE
, &sh
->state
);
2714 /* handle_stripe_clean_event
2715 * any written block on an uptodate or failed drive can be returned.
2716 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2717 * never LOCKED, so we don't need to test 'failed' directly.
2719 static void handle_stripe_clean_event(struct r5conf
*conf
,
2720 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
2725 for (i
= disks
; i
--; )
2726 if (sh
->dev
[i
].written
) {
2728 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2729 test_bit(R5_UPTODATE
, &dev
->flags
)) {
2730 /* We can return any write requests */
2731 struct bio
*wbi
, *wbi2
;
2732 pr_debug("Return write for disc %d\n", i
);
2734 dev
->written
= NULL
;
2735 while (wbi
&& wbi
->bi_sector
<
2736 dev
->sector
+ STRIPE_SECTORS
) {
2737 wbi2
= r5_next_bio(wbi
, dev
->sector
);
2738 if (!raid5_dec_bi_active_stripes(wbi
)) {
2739 md_write_end(conf
->mddev
);
2740 wbi
->bi_next
= *return_bi
;
2745 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2747 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
2752 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2753 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2754 md_wakeup_thread(conf
->mddev
->thread
);
2757 static void handle_stripe_dirtying(struct r5conf
*conf
,
2758 struct stripe_head
*sh
,
2759 struct stripe_head_state
*s
,
2762 int rmw
= 0, rcw
= 0, i
;
2763 if (conf
->max_degraded
== 2) {
2764 /* RAID6 requires 'rcw' in current implementation
2765 * Calculate the real rcw later - for now fake it
2766 * look like rcw is cheaper
2769 } else for (i
= disks
; i
--; ) {
2770 /* would I have to read this buffer for read_modify_write */
2771 struct r5dev
*dev
= &sh
->dev
[i
];
2772 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2773 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2774 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2775 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2776 if (test_bit(R5_Insync
, &dev
->flags
))
2779 rmw
+= 2*disks
; /* cannot read it */
2781 /* Would I have to read this buffer for reconstruct_write */
2782 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) && i
!= sh
->pd_idx
&&
2783 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2784 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2785 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2786 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2791 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2792 (unsigned long long)sh
->sector
, rmw
, rcw
);
2793 set_bit(STRIPE_HANDLE
, &sh
->state
);
2794 if (rmw
< rcw
&& rmw
> 0)
2795 /* prefer read-modify-write, but need to get some data */
2796 for (i
= disks
; i
--; ) {
2797 struct r5dev
*dev
= &sh
->dev
[i
];
2798 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2799 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2800 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2801 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2802 test_bit(R5_Insync
, &dev
->flags
)) {
2804 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2805 pr_debug("Read_old block "
2806 "%d for r-m-w\n", i
);
2807 set_bit(R5_LOCKED
, &dev
->flags
);
2808 set_bit(R5_Wantread
, &dev
->flags
);
2811 set_bit(STRIPE_DELAYED
, &sh
->state
);
2812 set_bit(STRIPE_HANDLE
, &sh
->state
);
2816 if (rcw
<= rmw
&& rcw
> 0) {
2817 /* want reconstruct write, but need to get some data */
2819 for (i
= disks
; i
--; ) {
2820 struct r5dev
*dev
= &sh
->dev
[i
];
2821 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2822 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
2823 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2824 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2825 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2827 if (!test_bit(R5_Insync
, &dev
->flags
))
2828 continue; /* it's a failed drive */
2830 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2831 pr_debug("Read_old block "
2832 "%d for Reconstruct\n", i
);
2833 set_bit(R5_LOCKED
, &dev
->flags
);
2834 set_bit(R5_Wantread
, &dev
->flags
);
2837 set_bit(STRIPE_DELAYED
, &sh
->state
);
2838 set_bit(STRIPE_HANDLE
, &sh
->state
);
2843 /* now if nothing is locked, and if we have enough data,
2844 * we can start a write request
2846 /* since handle_stripe can be called at any time we need to handle the
2847 * case where a compute block operation has been submitted and then a
2848 * subsequent call wants to start a write request. raid_run_ops only
2849 * handles the case where compute block and reconstruct are requested
2850 * simultaneously. If this is not the case then new writes need to be
2851 * held off until the compute completes.
2853 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
2854 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
2855 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
2856 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
2859 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
2860 struct stripe_head_state
*s
, int disks
)
2862 struct r5dev
*dev
= NULL
;
2864 set_bit(STRIPE_HANDLE
, &sh
->state
);
2866 switch (sh
->check_state
) {
2867 case check_state_idle
:
2868 /* start a new check operation if there are no failures */
2869 if (s
->failed
== 0) {
2870 BUG_ON(s
->uptodate
!= disks
);
2871 sh
->check_state
= check_state_run
;
2872 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2873 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2877 dev
= &sh
->dev
[s
->failed_num
[0]];
2879 case check_state_compute_result
:
2880 sh
->check_state
= check_state_idle
;
2882 dev
= &sh
->dev
[sh
->pd_idx
];
2884 /* check that a write has not made the stripe insync */
2885 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2888 /* either failed parity check, or recovery is happening */
2889 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
2890 BUG_ON(s
->uptodate
!= disks
);
2892 set_bit(R5_LOCKED
, &dev
->flags
);
2894 set_bit(R5_Wantwrite
, &dev
->flags
);
2896 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2897 set_bit(STRIPE_INSYNC
, &sh
->state
);
2899 case check_state_run
:
2900 break; /* we will be called again upon completion */
2901 case check_state_check_result
:
2902 sh
->check_state
= check_state_idle
;
2904 /* if a failure occurred during the check operation, leave
2905 * STRIPE_INSYNC not set and let the stripe be handled again
2910 /* handle a successful check operation, if parity is correct
2911 * we are done. Otherwise update the mismatch count and repair
2912 * parity if !MD_RECOVERY_CHECK
2914 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
2915 /* parity is correct (on disc,
2916 * not in buffer any more)
2918 set_bit(STRIPE_INSYNC
, &sh
->state
);
2920 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2921 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2922 /* don't try to repair!! */
2923 set_bit(STRIPE_INSYNC
, &sh
->state
);
2925 sh
->check_state
= check_state_compute_run
;
2926 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2927 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2928 set_bit(R5_Wantcompute
,
2929 &sh
->dev
[sh
->pd_idx
].flags
);
2930 sh
->ops
.target
= sh
->pd_idx
;
2931 sh
->ops
.target2
= -1;
2936 case check_state_compute_run
:
2939 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
2940 __func__
, sh
->check_state
,
2941 (unsigned long long) sh
->sector
);
2947 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
2948 struct stripe_head_state
*s
,
2951 int pd_idx
= sh
->pd_idx
;
2952 int qd_idx
= sh
->qd_idx
;
2955 set_bit(STRIPE_HANDLE
, &sh
->state
);
2957 BUG_ON(s
->failed
> 2);
2959 /* Want to check and possibly repair P and Q.
2960 * However there could be one 'failed' device, in which
2961 * case we can only check one of them, possibly using the
2962 * other to generate missing data
2965 switch (sh
->check_state
) {
2966 case check_state_idle
:
2967 /* start a new check operation if there are < 2 failures */
2968 if (s
->failed
== s
->q_failed
) {
2969 /* The only possible failed device holds Q, so it
2970 * makes sense to check P (If anything else were failed,
2971 * we would have used P to recreate it).
2973 sh
->check_state
= check_state_run
;
2975 if (!s
->q_failed
&& s
->failed
< 2) {
2976 /* Q is not failed, and we didn't use it to generate
2977 * anything, so it makes sense to check it
2979 if (sh
->check_state
== check_state_run
)
2980 sh
->check_state
= check_state_run_pq
;
2982 sh
->check_state
= check_state_run_q
;
2985 /* discard potentially stale zero_sum_result */
2986 sh
->ops
.zero_sum_result
= 0;
2988 if (sh
->check_state
== check_state_run
) {
2989 /* async_xor_zero_sum destroys the contents of P */
2990 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2993 if (sh
->check_state
>= check_state_run
&&
2994 sh
->check_state
<= check_state_run_pq
) {
2995 /* async_syndrome_zero_sum preserves P and Q, so
2996 * no need to mark them !uptodate here
2998 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3002 /* we have 2-disk failure */
3003 BUG_ON(s
->failed
!= 2);
3005 case check_state_compute_result
:
3006 sh
->check_state
= check_state_idle
;
3008 /* check that a write has not made the stripe insync */
3009 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3012 /* now write out any block on a failed drive,
3013 * or P or Q if they were recomputed
3015 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
3016 if (s
->failed
== 2) {
3017 dev
= &sh
->dev
[s
->failed_num
[1]];
3019 set_bit(R5_LOCKED
, &dev
->flags
);
3020 set_bit(R5_Wantwrite
, &dev
->flags
);
3022 if (s
->failed
>= 1) {
3023 dev
= &sh
->dev
[s
->failed_num
[0]];
3025 set_bit(R5_LOCKED
, &dev
->flags
);
3026 set_bit(R5_Wantwrite
, &dev
->flags
);
3028 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3029 dev
= &sh
->dev
[pd_idx
];
3031 set_bit(R5_LOCKED
, &dev
->flags
);
3032 set_bit(R5_Wantwrite
, &dev
->flags
);
3034 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3035 dev
= &sh
->dev
[qd_idx
];
3037 set_bit(R5_LOCKED
, &dev
->flags
);
3038 set_bit(R5_Wantwrite
, &dev
->flags
);
3040 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3042 set_bit(STRIPE_INSYNC
, &sh
->state
);
3044 case check_state_run
:
3045 case check_state_run_q
:
3046 case check_state_run_pq
:
3047 break; /* we will be called again upon completion */
3048 case check_state_check_result
:
3049 sh
->check_state
= check_state_idle
;
3051 /* handle a successful check operation, if parity is correct
3052 * we are done. Otherwise update the mismatch count and repair
3053 * parity if !MD_RECOVERY_CHECK
3055 if (sh
->ops
.zero_sum_result
== 0) {
3056 /* both parities are correct */
3058 set_bit(STRIPE_INSYNC
, &sh
->state
);
3060 /* in contrast to the raid5 case we can validate
3061 * parity, but still have a failure to write
3064 sh
->check_state
= check_state_compute_result
;
3065 /* Returning at this point means that we may go
3066 * off and bring p and/or q uptodate again so
3067 * we make sure to check zero_sum_result again
3068 * to verify if p or q need writeback
3072 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
3073 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3074 /* don't try to repair!! */
3075 set_bit(STRIPE_INSYNC
, &sh
->state
);
3077 int *target
= &sh
->ops
.target
;
3079 sh
->ops
.target
= -1;
3080 sh
->ops
.target2
= -1;
3081 sh
->check_state
= check_state_compute_run
;
3082 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3083 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3084 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3085 set_bit(R5_Wantcompute
,
3086 &sh
->dev
[pd_idx
].flags
);
3088 target
= &sh
->ops
.target2
;
3091 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3092 set_bit(R5_Wantcompute
,
3093 &sh
->dev
[qd_idx
].flags
);
3100 case check_state_compute_run
:
3103 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3104 __func__
, sh
->check_state
,
3105 (unsigned long long) sh
->sector
);
3110 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
3114 /* We have read all the blocks in this stripe and now we need to
3115 * copy some of them into a target stripe for expand.
3117 struct dma_async_tx_descriptor
*tx
= NULL
;
3118 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3119 for (i
= 0; i
< sh
->disks
; i
++)
3120 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
3122 struct stripe_head
*sh2
;
3123 struct async_submit_ctl submit
;
3125 sector_t bn
= compute_blocknr(sh
, i
, 1);
3126 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
3128 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
3130 /* so far only the early blocks of this stripe
3131 * have been requested. When later blocks
3132 * get requested, we will try again
3135 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
3136 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
3137 /* must have already done this block */
3138 release_stripe(sh2
);
3142 /* place all the copies on one channel */
3143 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
3144 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
3145 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
3148 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
3149 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
3150 for (j
= 0; j
< conf
->raid_disks
; j
++)
3151 if (j
!= sh2
->pd_idx
&&
3153 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
3155 if (j
== conf
->raid_disks
) {
3156 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
3157 set_bit(STRIPE_HANDLE
, &sh2
->state
);
3159 release_stripe(sh2
);
3162 /* done submitting copies, wait for them to complete */
3165 dma_wait_for_async_tx(tx
);
3170 * handle_stripe - do things to a stripe.
3172 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3173 * state of various bits to see what needs to be done.
3175 * return some read requests which now have data
3176 * return some write requests which are safely on storage
3177 * schedule a read on some buffers
3178 * schedule a write of some buffers
3179 * return confirmation of parity correctness
3183 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
3185 struct r5conf
*conf
= sh
->raid_conf
;
3186 int disks
= sh
->disks
;
3189 int do_recovery
= 0;
3191 memset(s
, 0, sizeof(*s
));
3193 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3194 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3195 s
->failed_num
[0] = -1;
3196 s
->failed_num
[1] = -1;
3198 /* Now to look around and see what can be done */
3200 for (i
=disks
; i
--; ) {
3201 struct md_rdev
*rdev
;
3208 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3210 dev
->toread
, dev
->towrite
, dev
->written
);
3211 /* maybe we can reply to a read
3213 * new wantfill requests are only permitted while
3214 * ops_complete_biofill is guaranteed to be inactive
3216 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
3217 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
3218 set_bit(R5_Wantfill
, &dev
->flags
);
3220 /* now count some things */
3221 if (test_bit(R5_LOCKED
, &dev
->flags
))
3223 if (test_bit(R5_UPTODATE
, &dev
->flags
))
3225 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
3227 BUG_ON(s
->compute
> 2);
3230 if (test_bit(R5_Wantfill
, &dev
->flags
))
3232 else if (dev
->toread
)
3236 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3241 /* Prefer to use the replacement for reads, but only
3242 * if it is recovered enough and has no bad blocks.
3244 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
3245 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
3246 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
3247 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3248 &first_bad
, &bad_sectors
))
3249 set_bit(R5_ReadRepl
, &dev
->flags
);
3252 set_bit(R5_NeedReplace
, &dev
->flags
);
3253 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3254 clear_bit(R5_ReadRepl
, &dev
->flags
);
3256 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
3259 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3260 &first_bad
, &bad_sectors
);
3261 if (s
->blocked_rdev
== NULL
3262 && (test_bit(Blocked
, &rdev
->flags
)
3265 set_bit(BlockedBadBlocks
,
3267 s
->blocked_rdev
= rdev
;
3268 atomic_inc(&rdev
->nr_pending
);
3271 clear_bit(R5_Insync
, &dev
->flags
);
3275 /* also not in-sync */
3276 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
3277 test_bit(R5_UPTODATE
, &dev
->flags
)) {
3278 /* treat as in-sync, but with a read error
3279 * which we can now try to correct
3281 set_bit(R5_Insync
, &dev
->flags
);
3282 set_bit(R5_ReadError
, &dev
->flags
);
3284 } else if (test_bit(In_sync
, &rdev
->flags
))
3285 set_bit(R5_Insync
, &dev
->flags
);
3286 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
3287 /* in sync if before recovery_offset */
3288 set_bit(R5_Insync
, &dev
->flags
);
3289 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
3290 test_bit(R5_Expanded
, &dev
->flags
))
3291 /* If we've reshaped into here, we assume it is Insync.
3292 * We will shortly update recovery_offset to make
3295 set_bit(R5_Insync
, &dev
->flags
);
3297 if (rdev
&& test_bit(R5_WriteError
, &dev
->flags
)) {
3298 /* This flag does not apply to '.replacement'
3299 * only to .rdev, so make sure to check that*/
3300 struct md_rdev
*rdev2
= rcu_dereference(
3301 conf
->disks
[i
].rdev
);
3303 clear_bit(R5_Insync
, &dev
->flags
);
3304 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3305 s
->handle_bad_blocks
= 1;
3306 atomic_inc(&rdev2
->nr_pending
);
3308 clear_bit(R5_WriteError
, &dev
->flags
);
3310 if (rdev
&& test_bit(R5_MadeGood
, &dev
->flags
)) {
3311 /* This flag does not apply to '.replacement'
3312 * only to .rdev, so make sure to check that*/
3313 struct md_rdev
*rdev2
= rcu_dereference(
3314 conf
->disks
[i
].rdev
);
3315 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3316 s
->handle_bad_blocks
= 1;
3317 atomic_inc(&rdev2
->nr_pending
);
3319 clear_bit(R5_MadeGood
, &dev
->flags
);
3321 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
3322 struct md_rdev
*rdev2
= rcu_dereference(
3323 conf
->disks
[i
].replacement
);
3324 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3325 s
->handle_bad_blocks
= 1;
3326 atomic_inc(&rdev2
->nr_pending
);
3328 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
3330 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3331 /* The ReadError flag will just be confusing now */
3332 clear_bit(R5_ReadError
, &dev
->flags
);
3333 clear_bit(R5_ReWrite
, &dev
->flags
);
3335 if (test_bit(R5_ReadError
, &dev
->flags
))
3336 clear_bit(R5_Insync
, &dev
->flags
);
3337 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3339 s
->failed_num
[s
->failed
] = i
;
3341 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
3345 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
3346 /* If there is a failed device being replaced,
3347 * we must be recovering.
3348 * else if we are after recovery_cp, we must be syncing
3349 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3350 * else we can only be replacing
3351 * sync and recovery both need to read all devices, and so
3352 * use the same flag.
3355 sh
->sector
>= conf
->mddev
->recovery_cp
||
3356 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
3364 static void handle_stripe(struct stripe_head
*sh
)
3366 struct stripe_head_state s
;
3367 struct r5conf
*conf
= sh
->raid_conf
;
3370 int disks
= sh
->disks
;
3371 struct r5dev
*pdev
, *qdev
;
3373 clear_bit(STRIPE_HANDLE
, &sh
->state
);
3374 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
3375 /* already being handled, ensure it gets handled
3376 * again when current action finishes */
3377 set_bit(STRIPE_HANDLE
, &sh
->state
);
3381 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
3382 set_bit(STRIPE_SYNCING
, &sh
->state
);
3383 clear_bit(STRIPE_INSYNC
, &sh
->state
);
3385 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3387 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3388 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3389 (unsigned long long)sh
->sector
, sh
->state
,
3390 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
3391 sh
->check_state
, sh
->reconstruct_state
);
3393 analyse_stripe(sh
, &s
);
3395 if (s
.handle_bad_blocks
) {
3396 set_bit(STRIPE_HANDLE
, &sh
->state
);
3400 if (unlikely(s
.blocked_rdev
)) {
3401 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
3402 s
.replacing
|| s
.to_write
|| s
.written
) {
3403 set_bit(STRIPE_HANDLE
, &sh
->state
);
3406 /* There is nothing for the blocked_rdev to block */
3407 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
3408 s
.blocked_rdev
= NULL
;
3411 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
3412 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
3413 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
3416 pr_debug("locked=%d uptodate=%d to_read=%d"
3417 " to_write=%d failed=%d failed_num=%d,%d\n",
3418 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
3419 s
.failed_num
[0], s
.failed_num
[1]);
3420 /* check if the array has lost more than max_degraded devices and,
3421 * if so, some requests might need to be failed.
3423 if (s
.failed
> conf
->max_degraded
) {
3424 sh
->check_state
= 0;
3425 sh
->reconstruct_state
= 0;
3426 if (s
.to_read
+s
.to_write
+s
.written
)
3427 handle_failed_stripe(conf
, sh
, &s
, disks
, &s
.return_bi
);
3428 if (s
.syncing
+ s
.replacing
)
3429 handle_failed_sync(conf
, sh
, &s
);
3433 * might be able to return some write requests if the parity blocks
3434 * are safe, or on a failed drive
3436 pdev
= &sh
->dev
[sh
->pd_idx
];
3437 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
3438 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
3439 qdev
= &sh
->dev
[sh
->qd_idx
];
3440 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
3441 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
3445 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
3446 && !test_bit(R5_LOCKED
, &pdev
->flags
)
3447 && test_bit(R5_UPTODATE
, &pdev
->flags
)))) &&
3448 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
3449 && !test_bit(R5_LOCKED
, &qdev
->flags
)
3450 && test_bit(R5_UPTODATE
, &qdev
->flags
)))))
3451 handle_stripe_clean_event(conf
, sh
, disks
, &s
.return_bi
);
3453 /* Now we might consider reading some blocks, either to check/generate
3454 * parity, or to satisfy requests
3455 * or to load a block that is being partially written.
3457 if (s
.to_read
|| s
.non_overwrite
3458 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
3459 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
3462 handle_stripe_fill(sh
, &s
, disks
);
3464 /* Now we check to see if any write operations have recently
3468 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
3470 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
3471 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
3472 sh
->reconstruct_state
= reconstruct_state_idle
;
3474 /* All the 'written' buffers and the parity block are ready to
3475 * be written back to disk
3477 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
));
3478 BUG_ON(sh
->qd_idx
>= 0 &&
3479 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
));
3480 for (i
= disks
; i
--; ) {
3481 struct r5dev
*dev
= &sh
->dev
[i
];
3482 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
3483 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3485 pr_debug("Writing block %d\n", i
);
3486 set_bit(R5_Wantwrite
, &dev
->flags
);
3489 if (!test_bit(R5_Insync
, &dev
->flags
) ||
3490 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3492 set_bit(STRIPE_INSYNC
, &sh
->state
);
3495 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3496 s
.dec_preread_active
= 1;
3499 /* Now to consider new write requests and what else, if anything
3500 * should be read. We do not handle new writes when:
3501 * 1/ A 'write' operation (copy+xor) is already in flight.
3502 * 2/ A 'check' operation is in flight, as it may clobber the parity
3505 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
3506 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
3508 /* maybe we need to check and possibly fix the parity for this stripe
3509 * Any reads will already have been scheduled, so we just see if enough
3510 * data is available. The parity check is held off while parity
3511 * dependent operations are in flight.
3513 if (sh
->check_state
||
3514 (s
.syncing
&& s
.locked
== 0 &&
3515 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
3516 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
3517 if (conf
->level
== 6)
3518 handle_parity_checks6(conf
, sh
, &s
, disks
);
3520 handle_parity_checks5(conf
, sh
, &s
, disks
);
3523 if (s
.replacing
&& s
.locked
== 0
3524 && !test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3525 /* Write out to replacement devices where possible */
3526 for (i
= 0; i
< conf
->raid_disks
; i
++)
3527 if (test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
) &&
3528 test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
3529 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
3530 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3533 set_bit(STRIPE_INSYNC
, &sh
->state
);
3535 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
3536 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3537 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3538 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3541 /* If the failed drives are just a ReadError, then we might need
3542 * to progress the repair/check process
3544 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
3545 for (i
= 0; i
< s
.failed
; i
++) {
3546 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
3547 if (test_bit(R5_ReadError
, &dev
->flags
)
3548 && !test_bit(R5_LOCKED
, &dev
->flags
)
3549 && test_bit(R5_UPTODATE
, &dev
->flags
)
3551 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3552 set_bit(R5_Wantwrite
, &dev
->flags
);
3553 set_bit(R5_ReWrite
, &dev
->flags
);
3554 set_bit(R5_LOCKED
, &dev
->flags
);
3557 /* let's read it back */
3558 set_bit(R5_Wantread
, &dev
->flags
);
3559 set_bit(R5_LOCKED
, &dev
->flags
);
3566 /* Finish reconstruct operations initiated by the expansion process */
3567 if (sh
->reconstruct_state
== reconstruct_state_result
) {
3568 struct stripe_head
*sh_src
3569 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
3570 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
3571 /* sh cannot be written until sh_src has been read.
3572 * so arrange for sh to be delayed a little
3574 set_bit(STRIPE_DELAYED
, &sh
->state
);
3575 set_bit(STRIPE_HANDLE
, &sh
->state
);
3576 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
3578 atomic_inc(&conf
->preread_active_stripes
);
3579 release_stripe(sh_src
);
3583 release_stripe(sh_src
);
3585 sh
->reconstruct_state
= reconstruct_state_idle
;
3586 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3587 for (i
= conf
->raid_disks
; i
--; ) {
3588 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3589 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3594 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
3595 !sh
->reconstruct_state
) {
3596 /* Need to write out all blocks after computing parity */
3597 sh
->disks
= conf
->raid_disks
;
3598 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
3599 schedule_reconstruction(sh
, &s
, 1, 1);
3600 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
3601 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3602 atomic_dec(&conf
->reshape_stripes
);
3603 wake_up(&conf
->wait_for_overlap
);
3604 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3607 if (s
.expanding
&& s
.locked
== 0 &&
3608 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
3609 handle_stripe_expansion(conf
, sh
);
3612 /* wait for this device to become unblocked */
3613 if (unlikely(s
.blocked_rdev
)) {
3614 if (conf
->mddev
->external
)
3615 md_wait_for_blocked_rdev(s
.blocked_rdev
,
3618 /* Internal metadata will immediately
3619 * be written by raid5d, so we don't
3620 * need to wait here.
3622 rdev_dec_pending(s
.blocked_rdev
,
3626 if (s
.handle_bad_blocks
)
3627 for (i
= disks
; i
--; ) {
3628 struct md_rdev
*rdev
;
3629 struct r5dev
*dev
= &sh
->dev
[i
];
3630 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
3631 /* We own a safe reference to the rdev */
3632 rdev
= conf
->disks
[i
].rdev
;
3633 if (!rdev_set_badblocks(rdev
, sh
->sector
,
3635 md_error(conf
->mddev
, rdev
);
3636 rdev_dec_pending(rdev
, conf
->mddev
);
3638 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
3639 rdev
= conf
->disks
[i
].rdev
;
3640 rdev_clear_badblocks(rdev
, sh
->sector
,
3642 rdev_dec_pending(rdev
, conf
->mddev
);
3644 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
3645 rdev
= conf
->disks
[i
].replacement
;
3647 /* rdev have been moved down */
3648 rdev
= conf
->disks
[i
].rdev
;
3649 rdev_clear_badblocks(rdev
, sh
->sector
,
3651 rdev_dec_pending(rdev
, conf
->mddev
);
3656 raid_run_ops(sh
, s
.ops_request
);
3660 if (s
.dec_preread_active
) {
3661 /* We delay this until after ops_run_io so that if make_request
3662 * is waiting on a flush, it won't continue until the writes
3663 * have actually been submitted.
3665 atomic_dec(&conf
->preread_active_stripes
);
3666 if (atomic_read(&conf
->preread_active_stripes
) <
3668 md_wakeup_thread(conf
->mddev
->thread
);
3671 return_io(s
.return_bi
);
3673 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
3676 static void raid5_activate_delayed(struct r5conf
*conf
)
3678 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
3679 while (!list_empty(&conf
->delayed_list
)) {
3680 struct list_head
*l
= conf
->delayed_list
.next
;
3681 struct stripe_head
*sh
;
3682 sh
= list_entry(l
, struct stripe_head
, lru
);
3684 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3685 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3686 atomic_inc(&conf
->preread_active_stripes
);
3687 list_add_tail(&sh
->lru
, &conf
->hold_list
);
3692 static void activate_bit_delay(struct r5conf
*conf
)
3694 /* device_lock is held */
3695 struct list_head head
;
3696 list_add(&head
, &conf
->bitmap_list
);
3697 list_del_init(&conf
->bitmap_list
);
3698 while (!list_empty(&head
)) {
3699 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
3700 list_del_init(&sh
->lru
);
3701 atomic_inc(&sh
->count
);
3702 __release_stripe(conf
, sh
);
3706 int md_raid5_congested(struct mddev
*mddev
, int bits
)
3708 struct r5conf
*conf
= mddev
->private;
3710 /* No difference between reads and writes. Just check
3711 * how busy the stripe_cache is
3714 if (conf
->inactive_blocked
)
3718 if (list_empty_careful(&conf
->inactive_list
))
3723 EXPORT_SYMBOL_GPL(md_raid5_congested
);
3725 static int raid5_congested(void *data
, int bits
)
3727 struct mddev
*mddev
= data
;
3729 return mddev_congested(mddev
, bits
) ||
3730 md_raid5_congested(mddev
, bits
);
3733 /* We want read requests to align with chunks where possible,
3734 * but write requests don't need to.
3736 static int raid5_mergeable_bvec(struct request_queue
*q
,
3737 struct bvec_merge_data
*bvm
,
3738 struct bio_vec
*biovec
)
3740 struct mddev
*mddev
= q
->queuedata
;
3741 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
3743 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3744 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
3746 if ((bvm
->bi_rw
& 1) == WRITE
)
3747 return biovec
->bv_len
; /* always allow writes to be mergeable */
3749 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3750 chunk_sectors
= mddev
->new_chunk_sectors
;
3751 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
3752 if (max
< 0) max
= 0;
3753 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
3754 return biovec
->bv_len
;
3760 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
3762 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
3763 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3764 unsigned int bio_sectors
= bio
->bi_size
>> 9;
3766 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3767 chunk_sectors
= mddev
->new_chunk_sectors
;
3768 return chunk_sectors
>=
3769 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
3773 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3774 * later sampled by raid5d.
3776 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
3778 unsigned long flags
;
3780 spin_lock_irqsave(&conf
->device_lock
, flags
);
3782 bi
->bi_next
= conf
->retry_read_aligned_list
;
3783 conf
->retry_read_aligned_list
= bi
;
3785 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3786 md_wakeup_thread(conf
->mddev
->thread
);
3790 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
)
3794 bi
= conf
->retry_read_aligned
;
3796 conf
->retry_read_aligned
= NULL
;
3799 bi
= conf
->retry_read_aligned_list
;
3801 conf
->retry_read_aligned_list
= bi
->bi_next
;
3804 * this sets the active strip count to 1 and the processed
3805 * strip count to zero (upper 8 bits)
3807 raid5_set_bi_stripes(bi
, 1); /* biased count of active stripes */
3815 * The "raid5_align_endio" should check if the read succeeded and if it
3816 * did, call bio_endio on the original bio (having bio_put the new bio
3818 * If the read failed..
3820 static void raid5_align_endio(struct bio
*bi
, int error
)
3822 struct bio
* raid_bi
= bi
->bi_private
;
3823 struct mddev
*mddev
;
3824 struct r5conf
*conf
;
3825 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3826 struct md_rdev
*rdev
;
3830 rdev
= (void*)raid_bi
->bi_next
;
3831 raid_bi
->bi_next
= NULL
;
3832 mddev
= rdev
->mddev
;
3833 conf
= mddev
->private;
3835 rdev_dec_pending(rdev
, conf
->mddev
);
3837 if (!error
&& uptodate
) {
3838 bio_endio(raid_bi
, 0);
3839 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
3840 wake_up(&conf
->wait_for_stripe
);
3845 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3847 add_bio_to_retry(raid_bi
, conf
);
3850 static int bio_fits_rdev(struct bio
*bi
)
3852 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
3854 if ((bi
->bi_size
>>9) > queue_max_sectors(q
))
3856 blk_recount_segments(q
, bi
);
3857 if (bi
->bi_phys_segments
> queue_max_segments(q
))
3860 if (q
->merge_bvec_fn
)
3861 /* it's too hard to apply the merge_bvec_fn at this stage,
3870 static int chunk_aligned_read(struct mddev
*mddev
, struct bio
* raid_bio
)
3872 struct r5conf
*conf
= mddev
->private;
3874 struct bio
* align_bi
;
3875 struct md_rdev
*rdev
;
3876 sector_t end_sector
;
3878 if (!in_chunk_boundary(mddev
, raid_bio
)) {
3879 pr_debug("chunk_aligned_read : non aligned\n");
3883 * use bio_clone_mddev to make a copy of the bio
3885 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
3889 * set bi_end_io to a new function, and set bi_private to the
3892 align_bi
->bi_end_io
= raid5_align_endio
;
3893 align_bi
->bi_private
= raid_bio
;
3897 align_bi
->bi_sector
= raid5_compute_sector(conf
, raid_bio
->bi_sector
,
3901 end_sector
= align_bi
->bi_sector
+ (align_bi
->bi_size
>> 9);
3903 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
3904 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
3905 rdev
->recovery_offset
< end_sector
) {
3906 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
3908 (test_bit(Faulty
, &rdev
->flags
) ||
3909 !(test_bit(In_sync
, &rdev
->flags
) ||
3910 rdev
->recovery_offset
>= end_sector
)))
3917 atomic_inc(&rdev
->nr_pending
);
3919 raid_bio
->bi_next
= (void*)rdev
;
3920 align_bi
->bi_bdev
= rdev
->bdev
;
3921 align_bi
->bi_flags
&= ~(1 << BIO_SEG_VALID
);
3923 if (!bio_fits_rdev(align_bi
) ||
3924 is_badblock(rdev
, align_bi
->bi_sector
, align_bi
->bi_size
>>9,
3925 &first_bad
, &bad_sectors
)) {
3926 /* too big in some way, or has a known bad block */
3928 rdev_dec_pending(rdev
, mddev
);
3932 /* No reshape active, so we can trust rdev->data_offset */
3933 align_bi
->bi_sector
+= rdev
->data_offset
;
3935 spin_lock_irq(&conf
->device_lock
);
3936 wait_event_lock_irq(conf
->wait_for_stripe
,
3938 conf
->device_lock
, /* nothing */);
3939 atomic_inc(&conf
->active_aligned_reads
);
3940 spin_unlock_irq(&conf
->device_lock
);
3942 generic_make_request(align_bi
);
3951 /* __get_priority_stripe - get the next stripe to process
3953 * Full stripe writes are allowed to pass preread active stripes up until
3954 * the bypass_threshold is exceeded. In general the bypass_count
3955 * increments when the handle_list is handled before the hold_list; however, it
3956 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3957 * stripe with in flight i/o. The bypass_count will be reset when the
3958 * head of the hold_list has changed, i.e. the head was promoted to the
3961 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
)
3963 struct stripe_head
*sh
;
3965 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3967 list_empty(&conf
->handle_list
) ? "empty" : "busy",
3968 list_empty(&conf
->hold_list
) ? "empty" : "busy",
3969 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
3971 if (!list_empty(&conf
->handle_list
)) {
3972 sh
= list_entry(conf
->handle_list
.next
, typeof(*sh
), lru
);
3974 if (list_empty(&conf
->hold_list
))
3975 conf
->bypass_count
= 0;
3976 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
3977 if (conf
->hold_list
.next
== conf
->last_hold
)
3978 conf
->bypass_count
++;
3980 conf
->last_hold
= conf
->hold_list
.next
;
3981 conf
->bypass_count
-= conf
->bypass_threshold
;
3982 if (conf
->bypass_count
< 0)
3983 conf
->bypass_count
= 0;
3986 } else if (!list_empty(&conf
->hold_list
) &&
3987 ((conf
->bypass_threshold
&&
3988 conf
->bypass_count
> conf
->bypass_threshold
) ||
3989 atomic_read(&conf
->pending_full_writes
) == 0)) {
3990 sh
= list_entry(conf
->hold_list
.next
,
3992 conf
->bypass_count
-= conf
->bypass_threshold
;
3993 if (conf
->bypass_count
< 0)
3994 conf
->bypass_count
= 0;
3998 list_del_init(&sh
->lru
);
3999 atomic_inc(&sh
->count
);
4000 BUG_ON(atomic_read(&sh
->count
) != 1);
4004 static void make_request(struct mddev
*mddev
, struct bio
* bi
)
4006 struct r5conf
*conf
= mddev
->private;
4008 sector_t new_sector
;
4009 sector_t logical_sector
, last_sector
;
4010 struct stripe_head
*sh
;
4011 const int rw
= bio_data_dir(bi
);
4014 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
4015 md_flush_request(mddev
, bi
);
4019 md_write_start(mddev
, bi
);
4022 mddev
->reshape_position
== MaxSector
&&
4023 chunk_aligned_read(mddev
,bi
))
4026 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4027 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
4029 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
4031 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
4037 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
4038 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
4039 /* spinlock is needed as reshape_progress may be
4040 * 64bit on a 32bit platform, and so it might be
4041 * possible to see a half-updated value
4042 * Of course reshape_progress could change after
4043 * the lock is dropped, so once we get a reference
4044 * to the stripe that we think it is, we will have
4047 spin_lock_irq(&conf
->device_lock
);
4048 if (mddev
->reshape_backwards
4049 ? logical_sector
< conf
->reshape_progress
4050 : logical_sector
>= conf
->reshape_progress
) {
4053 if (mddev
->reshape_backwards
4054 ? logical_sector
< conf
->reshape_safe
4055 : logical_sector
>= conf
->reshape_safe
) {
4056 spin_unlock_irq(&conf
->device_lock
);
4061 spin_unlock_irq(&conf
->device_lock
);
4064 new_sector
= raid5_compute_sector(conf
, logical_sector
,
4067 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4068 (unsigned long long)new_sector
,
4069 (unsigned long long)logical_sector
);
4071 sh
= get_active_stripe(conf
, new_sector
, previous
,
4072 (bi
->bi_rw
&RWA_MASK
), 0);
4074 if (unlikely(previous
)) {
4075 /* expansion might have moved on while waiting for a
4076 * stripe, so we must do the range check again.
4077 * Expansion could still move past after this
4078 * test, but as we are holding a reference to
4079 * 'sh', we know that if that happens,
4080 * STRIPE_EXPANDING will get set and the expansion
4081 * won't proceed until we finish with the stripe.
4084 spin_lock_irq(&conf
->device_lock
);
4085 if (mddev
->reshape_backwards
4086 ? logical_sector
>= conf
->reshape_progress
4087 : logical_sector
< conf
->reshape_progress
)
4088 /* mismatch, need to try again */
4090 spin_unlock_irq(&conf
->device_lock
);
4099 logical_sector
>= mddev
->suspend_lo
&&
4100 logical_sector
< mddev
->suspend_hi
) {
4102 /* As the suspend_* range is controlled by
4103 * userspace, we want an interruptible
4106 flush_signals(current
);
4107 prepare_to_wait(&conf
->wait_for_overlap
,
4108 &w
, TASK_INTERRUPTIBLE
);
4109 if (logical_sector
>= mddev
->suspend_lo
&&
4110 logical_sector
< mddev
->suspend_hi
)
4115 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
4116 !add_stripe_bio(sh
, bi
, dd_idx
, rw
)) {
4117 /* Stripe is busy expanding or
4118 * add failed due to overlap. Flush everything
4121 md_wakeup_thread(mddev
->thread
);
4126 finish_wait(&conf
->wait_for_overlap
, &w
);
4127 set_bit(STRIPE_HANDLE
, &sh
->state
);
4128 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4129 if ((bi
->bi_rw
& REQ_SYNC
) &&
4130 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4131 atomic_inc(&conf
->preread_active_stripes
);
4132 mddev_check_plugged(mddev
);
4135 /* cannot get stripe for read-ahead, just give-up */
4136 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
4137 finish_wait(&conf
->wait_for_overlap
, &w
);
4142 remaining
= raid5_dec_bi_active_stripes(bi
);
4143 if (remaining
== 0) {
4146 md_write_end(mddev
);
4152 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
4154 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
4156 /* reshaping is quite different to recovery/resync so it is
4157 * handled quite separately ... here.
4159 * On each call to sync_request, we gather one chunk worth of
4160 * destination stripes and flag them as expanding.
4161 * Then we find all the source stripes and request reads.
4162 * As the reads complete, handle_stripe will copy the data
4163 * into the destination stripe and release that stripe.
4165 struct r5conf
*conf
= mddev
->private;
4166 struct stripe_head
*sh
;
4167 sector_t first_sector
, last_sector
;
4168 int raid_disks
= conf
->previous_raid_disks
;
4169 int data_disks
= raid_disks
- conf
->max_degraded
;
4170 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
4173 sector_t writepos
, readpos
, safepos
;
4174 sector_t stripe_addr
;
4175 int reshape_sectors
;
4176 struct list_head stripes
;
4178 if (sector_nr
== 0) {
4179 /* If restarting in the middle, skip the initial sectors */
4180 if (mddev
->reshape_backwards
&&
4181 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
4182 sector_nr
= raid5_size(mddev
, 0, 0)
4183 - conf
->reshape_progress
;
4184 } else if (!mddev
->reshape_backwards
&&
4185 conf
->reshape_progress
> 0)
4186 sector_nr
= conf
->reshape_progress
;
4187 sector_div(sector_nr
, new_data_disks
);
4189 mddev
->curr_resync_completed
= sector_nr
;
4190 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4196 /* We need to process a full chunk at a time.
4197 * If old and new chunk sizes differ, we need to process the
4200 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
)
4201 reshape_sectors
= mddev
->new_chunk_sectors
;
4203 reshape_sectors
= mddev
->chunk_sectors
;
4205 /* We update the metadata at least every 10 seconds, or when
4206 * the data about to be copied would over-write the source of
4207 * the data at the front of the range. i.e. one new_stripe
4208 * along from reshape_progress new_maps to after where
4209 * reshape_safe old_maps to
4211 writepos
= conf
->reshape_progress
;
4212 sector_div(writepos
, new_data_disks
);
4213 readpos
= conf
->reshape_progress
;
4214 sector_div(readpos
, data_disks
);
4215 safepos
= conf
->reshape_safe
;
4216 sector_div(safepos
, data_disks
);
4217 if (mddev
->reshape_backwards
) {
4218 writepos
-= min_t(sector_t
, reshape_sectors
, writepos
);
4219 readpos
+= reshape_sectors
;
4220 safepos
+= reshape_sectors
;
4222 writepos
+= reshape_sectors
;
4223 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
4224 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
4227 /* Having calculated the 'writepos' possibly use it
4228 * to set 'stripe_addr' which is where we will write to.
4230 if (mddev
->reshape_backwards
) {
4231 BUG_ON(conf
->reshape_progress
== 0);
4232 stripe_addr
= writepos
;
4233 BUG_ON((mddev
->dev_sectors
&
4234 ~((sector_t
)reshape_sectors
- 1))
4235 - reshape_sectors
- stripe_addr
4238 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
4239 stripe_addr
= sector_nr
;
4242 /* 'writepos' is the most advanced device address we might write.
4243 * 'readpos' is the least advanced device address we might read.
4244 * 'safepos' is the least address recorded in the metadata as having
4246 * If there is a min_offset_diff, these are adjusted either by
4247 * increasing the safepos/readpos if diff is negative, or
4248 * increasing writepos if diff is positive.
4249 * If 'readpos' is then behind 'writepos', there is no way that we can
4250 * ensure safety in the face of a crash - that must be done by userspace
4251 * making a backup of the data. So in that case there is no particular
4252 * rush to update metadata.
4253 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4254 * update the metadata to advance 'safepos' to match 'readpos' so that
4255 * we can be safe in the event of a crash.
4256 * So we insist on updating metadata if safepos is behind writepos and
4257 * readpos is beyond writepos.
4258 * In any case, update the metadata every 10 seconds.
4259 * Maybe that number should be configurable, but I'm not sure it is
4260 * worth it.... maybe it could be a multiple of safemode_delay???
4262 if (conf
->min_offset_diff
< 0) {
4263 safepos
+= -conf
->min_offset_diff
;
4264 readpos
+= -conf
->min_offset_diff
;
4266 writepos
+= conf
->min_offset_diff
;
4268 if ((mddev
->reshape_backwards
4269 ? (safepos
> writepos
&& readpos
< writepos
)
4270 : (safepos
< writepos
&& readpos
> writepos
)) ||
4271 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4272 /* Cannot proceed until we've updated the superblock... */
4273 wait_event(conf
->wait_for_overlap
,
4274 atomic_read(&conf
->reshape_stripes
)==0);
4275 mddev
->reshape_position
= conf
->reshape_progress
;
4276 mddev
->curr_resync_completed
= sector_nr
;
4277 conf
->reshape_checkpoint
= jiffies
;
4278 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4279 md_wakeup_thread(mddev
->thread
);
4280 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4281 kthread_should_stop());
4282 spin_lock_irq(&conf
->device_lock
);
4283 conf
->reshape_safe
= mddev
->reshape_position
;
4284 spin_unlock_irq(&conf
->device_lock
);
4285 wake_up(&conf
->wait_for_overlap
);
4286 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4289 INIT_LIST_HEAD(&stripes
);
4290 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
4292 int skipped_disk
= 0;
4293 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
4294 set_bit(STRIPE_EXPANDING
, &sh
->state
);
4295 atomic_inc(&conf
->reshape_stripes
);
4296 /* If any of this stripe is beyond the end of the old
4297 * array, then we need to zero those blocks
4299 for (j
=sh
->disks
; j
--;) {
4301 if (j
== sh
->pd_idx
)
4303 if (conf
->level
== 6 &&
4306 s
= compute_blocknr(sh
, j
, 0);
4307 if (s
< raid5_size(mddev
, 0, 0)) {
4311 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
4312 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
4313 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
4315 if (!skipped_disk
) {
4316 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4317 set_bit(STRIPE_HANDLE
, &sh
->state
);
4319 list_add(&sh
->lru
, &stripes
);
4321 spin_lock_irq(&conf
->device_lock
);
4322 if (mddev
->reshape_backwards
)
4323 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
4325 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
4326 spin_unlock_irq(&conf
->device_lock
);
4327 /* Ok, those stripe are ready. We can start scheduling
4328 * reads on the source stripes.
4329 * The source stripes are determined by mapping the first and last
4330 * block on the destination stripes.
4333 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
4336 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
4337 * new_data_disks
- 1),
4339 if (last_sector
>= mddev
->dev_sectors
)
4340 last_sector
= mddev
->dev_sectors
- 1;
4341 while (first_sector
<= last_sector
) {
4342 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
4343 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
4344 set_bit(STRIPE_HANDLE
, &sh
->state
);
4346 first_sector
+= STRIPE_SECTORS
;
4348 /* Now that the sources are clearly marked, we can release
4349 * the destination stripes
4351 while (!list_empty(&stripes
)) {
4352 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
4353 list_del_init(&sh
->lru
);
4356 /* If this takes us to the resync_max point where we have to pause,
4357 * then we need to write out the superblock.
4359 sector_nr
+= reshape_sectors
;
4360 if ((sector_nr
- mddev
->curr_resync_completed
) * 2
4361 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
4362 /* Cannot proceed until we've updated the superblock... */
4363 wait_event(conf
->wait_for_overlap
,
4364 atomic_read(&conf
->reshape_stripes
) == 0);
4365 mddev
->reshape_position
= conf
->reshape_progress
;
4366 mddev
->curr_resync_completed
= sector_nr
;
4367 conf
->reshape_checkpoint
= jiffies
;
4368 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4369 md_wakeup_thread(mddev
->thread
);
4370 wait_event(mddev
->sb_wait
,
4371 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
4372 || kthread_should_stop());
4373 spin_lock_irq(&conf
->device_lock
);
4374 conf
->reshape_safe
= mddev
->reshape_position
;
4375 spin_unlock_irq(&conf
->device_lock
);
4376 wake_up(&conf
->wait_for_overlap
);
4377 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4379 return reshape_sectors
;
4382 /* FIXME go_faster isn't used */
4383 static inline sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
4385 struct r5conf
*conf
= mddev
->private;
4386 struct stripe_head
*sh
;
4387 sector_t max_sector
= mddev
->dev_sectors
;
4388 sector_t sync_blocks
;
4389 int still_degraded
= 0;
4392 if (sector_nr
>= max_sector
) {
4393 /* just being told to finish up .. nothing much to do */
4395 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
4400 if (mddev
->curr_resync
< max_sector
) /* aborted */
4401 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
4403 else /* completed sync */
4405 bitmap_close_sync(mddev
->bitmap
);
4410 /* Allow raid5_quiesce to complete */
4411 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
4413 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
4414 return reshape_request(mddev
, sector_nr
, skipped
);
4416 /* No need to check resync_max as we never do more than one
4417 * stripe, and as resync_max will always be on a chunk boundary,
4418 * if the check in md_do_sync didn't fire, there is no chance
4419 * of overstepping resync_max here
4422 /* if there is too many failed drives and we are trying
4423 * to resync, then assert that we are finished, because there is
4424 * nothing we can do.
4426 if (mddev
->degraded
>= conf
->max_degraded
&&
4427 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
4428 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
4432 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
4433 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
4434 !conf
->fullsync
&& sync_blocks
>= STRIPE_SECTORS
) {
4435 /* we can skip this block, and probably more */
4436 sync_blocks
/= STRIPE_SECTORS
;
4438 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
4441 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
4443 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
4445 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
4446 /* make sure we don't swamp the stripe cache if someone else
4447 * is trying to get access
4449 schedule_timeout_uninterruptible(1);
4451 /* Need to check if array will still be degraded after recovery/resync
4452 * We don't need to check the 'failed' flag as when that gets set,
4455 for (i
= 0; i
< conf
->raid_disks
; i
++)
4456 if (conf
->disks
[i
].rdev
== NULL
)
4459 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
4461 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
4466 return STRIPE_SECTORS
;
4469 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
)
4471 /* We may not be able to submit a whole bio at once as there
4472 * may not be enough stripe_heads available.
4473 * We cannot pre-allocate enough stripe_heads as we may need
4474 * more than exist in the cache (if we allow ever large chunks).
4475 * So we do one stripe head at a time and record in
4476 * ->bi_hw_segments how many have been done.
4478 * We *know* that this entire raid_bio is in one chunk, so
4479 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4481 struct stripe_head
*sh
;
4483 sector_t sector
, logical_sector
, last_sector
;
4488 logical_sector
= raid_bio
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4489 sector
= raid5_compute_sector(conf
, logical_sector
,
4491 last_sector
= raid_bio
->bi_sector
+ (raid_bio
->bi_size
>>9);
4493 for (; logical_sector
< last_sector
;
4494 logical_sector
+= STRIPE_SECTORS
,
4495 sector
+= STRIPE_SECTORS
,
4498 if (scnt
< raid5_bi_processed_stripes(raid_bio
))
4499 /* already done this stripe */
4502 sh
= get_active_stripe(conf
, sector
, 0, 1, 0);
4505 /* failed to get a stripe - must wait */
4506 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
4507 conf
->retry_read_aligned
= raid_bio
;
4511 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0)) {
4513 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
4514 conf
->retry_read_aligned
= raid_bio
;
4522 remaining
= raid5_dec_bi_active_stripes(raid_bio
);
4524 bio_endio(raid_bio
, 0);
4525 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4526 wake_up(&conf
->wait_for_stripe
);
4532 * This is our raid5 kernel thread.
4534 * We scan the hash table for stripes which can be handled now.
4535 * During the scan, completed stripes are saved for us by the interrupt
4536 * handler, so that they will not have to wait for our next wakeup.
4538 static void raid5d(struct mddev
*mddev
)
4540 struct stripe_head
*sh
;
4541 struct r5conf
*conf
= mddev
->private;
4543 struct blk_plug plug
;
4545 pr_debug("+++ raid5d active\n");
4547 md_check_recovery(mddev
);
4549 blk_start_plug(&plug
);
4551 spin_lock_irq(&conf
->device_lock
);
4555 if (atomic_read(&mddev
->plug_cnt
) == 0 &&
4556 !list_empty(&conf
->bitmap_list
)) {
4557 /* Now is a good time to flush some bitmap updates */
4559 spin_unlock_irq(&conf
->device_lock
);
4560 bitmap_unplug(mddev
->bitmap
);
4561 spin_lock_irq(&conf
->device_lock
);
4562 conf
->seq_write
= conf
->seq_flush
;
4563 activate_bit_delay(conf
);
4565 if (atomic_read(&mddev
->plug_cnt
) == 0)
4566 raid5_activate_delayed(conf
);
4568 while ((bio
= remove_bio_from_retry(conf
))) {
4570 spin_unlock_irq(&conf
->device_lock
);
4571 ok
= retry_aligned_read(conf
, bio
);
4572 spin_lock_irq(&conf
->device_lock
);
4578 sh
= __get_priority_stripe(conf
);
4582 spin_unlock_irq(&conf
->device_lock
);
4589 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
4590 md_check_recovery(mddev
);
4592 spin_lock_irq(&conf
->device_lock
);
4594 pr_debug("%d stripes handled\n", handled
);
4596 spin_unlock_irq(&conf
->device_lock
);
4598 async_tx_issue_pending_all();
4599 blk_finish_plug(&plug
);
4601 pr_debug("--- raid5d inactive\n");
4605 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
4607 struct r5conf
*conf
= mddev
->private;
4609 return sprintf(page
, "%d\n", conf
->max_nr_stripes
);
4615 raid5_set_cache_size(struct mddev
*mddev
, int size
)
4617 struct r5conf
*conf
= mddev
->private;
4620 if (size
<= 16 || size
> 32768)
4622 while (size
< conf
->max_nr_stripes
) {
4623 if (drop_one_stripe(conf
))
4624 conf
->max_nr_stripes
--;
4628 err
= md_allow_write(mddev
);
4631 while (size
> conf
->max_nr_stripes
) {
4632 if (grow_one_stripe(conf
))
4633 conf
->max_nr_stripes
++;
4638 EXPORT_SYMBOL(raid5_set_cache_size
);
4641 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
4643 struct r5conf
*conf
= mddev
->private;
4647 if (len
>= PAGE_SIZE
)
4652 if (strict_strtoul(page
, 10, &new))
4654 err
= raid5_set_cache_size(mddev
, new);
4660 static struct md_sysfs_entry
4661 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
4662 raid5_show_stripe_cache_size
,
4663 raid5_store_stripe_cache_size
);
4666 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
4668 struct r5conf
*conf
= mddev
->private;
4670 return sprintf(page
, "%d\n", conf
->bypass_threshold
);
4676 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
4678 struct r5conf
*conf
= mddev
->private;
4680 if (len
>= PAGE_SIZE
)
4685 if (strict_strtoul(page
, 10, &new))
4687 if (new > conf
->max_nr_stripes
)
4689 conf
->bypass_threshold
= new;
4693 static struct md_sysfs_entry
4694 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
4696 raid5_show_preread_threshold
,
4697 raid5_store_preread_threshold
);
4700 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
4702 struct r5conf
*conf
= mddev
->private;
4704 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
4709 static struct md_sysfs_entry
4710 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
4712 static struct attribute
*raid5_attrs
[] = {
4713 &raid5_stripecache_size
.attr
,
4714 &raid5_stripecache_active
.attr
,
4715 &raid5_preread_bypass_threshold
.attr
,
4718 static struct attribute_group raid5_attrs_group
= {
4720 .attrs
= raid5_attrs
,
4724 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
4726 struct r5conf
*conf
= mddev
->private;
4729 sectors
= mddev
->dev_sectors
;
4731 /* size is defined by the smallest of previous and new size */
4732 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
4734 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
4735 sectors
&= ~((sector_t
)mddev
->new_chunk_sectors
- 1);
4736 return sectors
* (raid_disks
- conf
->max_degraded
);
4739 static void raid5_free_percpu(struct r5conf
*conf
)
4741 struct raid5_percpu
*percpu
;
4748 for_each_possible_cpu(cpu
) {
4749 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4750 safe_put_page(percpu
->spare_page
);
4751 kfree(percpu
->scribble
);
4753 #ifdef CONFIG_HOTPLUG_CPU
4754 unregister_cpu_notifier(&conf
->cpu_notify
);
4758 free_percpu(conf
->percpu
);
4761 static void free_conf(struct r5conf
*conf
)
4763 shrink_stripes(conf
);
4764 raid5_free_percpu(conf
);
4766 kfree(conf
->stripe_hashtbl
);
4770 #ifdef CONFIG_HOTPLUG_CPU
4771 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
4774 struct r5conf
*conf
= container_of(nfb
, struct r5conf
, cpu_notify
);
4775 long cpu
= (long)hcpu
;
4776 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4779 case CPU_UP_PREPARE
:
4780 case CPU_UP_PREPARE_FROZEN
:
4781 if (conf
->level
== 6 && !percpu
->spare_page
)
4782 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
4783 if (!percpu
->scribble
)
4784 percpu
->scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
4786 if (!percpu
->scribble
||
4787 (conf
->level
== 6 && !percpu
->spare_page
)) {
4788 safe_put_page(percpu
->spare_page
);
4789 kfree(percpu
->scribble
);
4790 pr_err("%s: failed memory allocation for cpu%ld\n",
4792 return notifier_from_errno(-ENOMEM
);
4796 case CPU_DEAD_FROZEN
:
4797 safe_put_page(percpu
->spare_page
);
4798 kfree(percpu
->scribble
);
4799 percpu
->spare_page
= NULL
;
4800 percpu
->scribble
= NULL
;
4809 static int raid5_alloc_percpu(struct r5conf
*conf
)
4812 struct page
*spare_page
;
4813 struct raid5_percpu __percpu
*allcpus
;
4817 allcpus
= alloc_percpu(struct raid5_percpu
);
4820 conf
->percpu
= allcpus
;
4824 for_each_present_cpu(cpu
) {
4825 if (conf
->level
== 6) {
4826 spare_page
= alloc_page(GFP_KERNEL
);
4831 per_cpu_ptr(conf
->percpu
, cpu
)->spare_page
= spare_page
;
4833 scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
4838 per_cpu_ptr(conf
->percpu
, cpu
)->scribble
= scribble
;
4840 #ifdef CONFIG_HOTPLUG_CPU
4841 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
4842 conf
->cpu_notify
.priority
= 0;
4844 err
= register_cpu_notifier(&conf
->cpu_notify
);
4851 static struct r5conf
*setup_conf(struct mddev
*mddev
)
4853 struct r5conf
*conf
;
4854 int raid_disk
, memory
, max_disks
;
4855 struct md_rdev
*rdev
;
4856 struct disk_info
*disk
;
4859 if (mddev
->new_level
!= 5
4860 && mddev
->new_level
!= 4
4861 && mddev
->new_level
!= 6) {
4862 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4863 mdname(mddev
), mddev
->new_level
);
4864 return ERR_PTR(-EIO
);
4866 if ((mddev
->new_level
== 5
4867 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
4868 (mddev
->new_level
== 6
4869 && !algorithm_valid_raid6(mddev
->new_layout
))) {
4870 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
4871 mdname(mddev
), mddev
->new_layout
);
4872 return ERR_PTR(-EIO
);
4874 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
4875 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4876 mdname(mddev
), mddev
->raid_disks
);
4877 return ERR_PTR(-EINVAL
);
4880 if (!mddev
->new_chunk_sectors
||
4881 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
4882 !is_power_of_2(mddev
->new_chunk_sectors
)) {
4883 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
4884 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
4885 return ERR_PTR(-EINVAL
);
4888 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
4891 spin_lock_init(&conf
->device_lock
);
4892 init_waitqueue_head(&conf
->wait_for_stripe
);
4893 init_waitqueue_head(&conf
->wait_for_overlap
);
4894 INIT_LIST_HEAD(&conf
->handle_list
);
4895 INIT_LIST_HEAD(&conf
->hold_list
);
4896 INIT_LIST_HEAD(&conf
->delayed_list
);
4897 INIT_LIST_HEAD(&conf
->bitmap_list
);
4898 INIT_LIST_HEAD(&conf
->inactive_list
);
4899 atomic_set(&conf
->active_stripes
, 0);
4900 atomic_set(&conf
->preread_active_stripes
, 0);
4901 atomic_set(&conf
->active_aligned_reads
, 0);
4902 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
4903 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
4905 conf
->raid_disks
= mddev
->raid_disks
;
4906 if (mddev
->reshape_position
== MaxSector
)
4907 conf
->previous_raid_disks
= mddev
->raid_disks
;
4909 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4910 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
4911 conf
->scribble_len
= scribble_len(max_disks
);
4913 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
4918 conf
->mddev
= mddev
;
4920 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
4923 conf
->level
= mddev
->new_level
;
4924 if (raid5_alloc_percpu(conf
) != 0)
4927 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
4929 rdev_for_each(rdev
, mddev
) {
4930 raid_disk
= rdev
->raid_disk
;
4931 if (raid_disk
>= max_disks
4934 disk
= conf
->disks
+ raid_disk
;
4936 if (test_bit(Replacement
, &rdev
->flags
)) {
4937 if (disk
->replacement
)
4939 disk
->replacement
= rdev
;
4946 if (test_bit(In_sync
, &rdev
->flags
)) {
4947 char b
[BDEVNAME_SIZE
];
4948 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
4950 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
4951 } else if (rdev
->saved_raid_disk
!= raid_disk
)
4952 /* Cannot rely on bitmap to complete recovery */
4956 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
4957 conf
->level
= mddev
->new_level
;
4958 if (conf
->level
== 6)
4959 conf
->max_degraded
= 2;
4961 conf
->max_degraded
= 1;
4962 conf
->algorithm
= mddev
->new_layout
;
4963 conf
->max_nr_stripes
= NR_STRIPES
;
4964 conf
->reshape_progress
= mddev
->reshape_position
;
4965 if (conf
->reshape_progress
!= MaxSector
) {
4966 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
4967 conf
->prev_algo
= mddev
->layout
;
4970 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
4971 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
4972 if (grow_stripes(conf
, conf
->max_nr_stripes
)) {
4974 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4975 mdname(mddev
), memory
);
4978 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
4979 mdname(mddev
), memory
);
4981 sprintf(pers_name
, "raid%d", mddev
->new_level
);
4982 conf
->thread
= md_register_thread(raid5d
, mddev
, pers_name
);
4983 if (!conf
->thread
) {
4985 "md/raid:%s: couldn't allocate thread.\n",
4995 return ERR_PTR(-EIO
);
4997 return ERR_PTR(-ENOMEM
);
5001 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
5004 case ALGORITHM_PARITY_0
:
5005 if (raid_disk
< max_degraded
)
5008 case ALGORITHM_PARITY_N
:
5009 if (raid_disk
>= raid_disks
- max_degraded
)
5012 case ALGORITHM_PARITY_0_6
:
5013 if (raid_disk
== 0 ||
5014 raid_disk
== raid_disks
- 1)
5017 case ALGORITHM_LEFT_ASYMMETRIC_6
:
5018 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
5019 case ALGORITHM_LEFT_SYMMETRIC_6
:
5020 case ALGORITHM_RIGHT_SYMMETRIC_6
:
5021 if (raid_disk
== raid_disks
- 1)
5027 static int run(struct mddev
*mddev
)
5029 struct r5conf
*conf
;
5030 int working_disks
= 0;
5031 int dirty_parity_disks
= 0;
5032 struct md_rdev
*rdev
;
5033 sector_t reshape_offset
= 0;
5035 long long min_offset_diff
= 0;
5038 if (mddev
->recovery_cp
!= MaxSector
)
5039 printk(KERN_NOTICE
"md/raid:%s: not clean"
5040 " -- starting background reconstruction\n",
5043 rdev_for_each(rdev
, mddev
) {
5045 if (rdev
->raid_disk
< 0)
5047 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
5049 min_offset_diff
= diff
;
5051 } else if (mddev
->reshape_backwards
&&
5052 diff
< min_offset_diff
)
5053 min_offset_diff
= diff
;
5054 else if (!mddev
->reshape_backwards
&&
5055 diff
> min_offset_diff
)
5056 min_offset_diff
= diff
;
5059 if (mddev
->reshape_position
!= MaxSector
) {
5060 /* Check that we can continue the reshape.
5061 * Difficulties arise if the stripe we would write to
5062 * next is at or after the stripe we would read from next.
5063 * For a reshape that changes the number of devices, this
5064 * is only possible for a very short time, and mdadm makes
5065 * sure that time appears to have past before assembling
5066 * the array. So we fail if that time hasn't passed.
5067 * For a reshape that keeps the number of devices the same
5068 * mdadm must be monitoring the reshape can keeping the
5069 * critical areas read-only and backed up. It will start
5070 * the array in read-only mode, so we check for that.
5072 sector_t here_new
, here_old
;
5074 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
5076 if (mddev
->new_level
!= mddev
->level
) {
5077 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
5078 "required - aborting.\n",
5082 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
5083 /* reshape_position must be on a new-stripe boundary, and one
5084 * further up in new geometry must map after here in old
5087 here_new
= mddev
->reshape_position
;
5088 if (sector_div(here_new
, mddev
->new_chunk_sectors
*
5089 (mddev
->raid_disks
- max_degraded
))) {
5090 printk(KERN_ERR
"md/raid:%s: reshape_position not "
5091 "on a stripe boundary\n", mdname(mddev
));
5094 reshape_offset
= here_new
* mddev
->new_chunk_sectors
;
5095 /* here_new is the stripe we will write to */
5096 here_old
= mddev
->reshape_position
;
5097 sector_div(here_old
, mddev
->chunk_sectors
*
5098 (old_disks
-max_degraded
));
5099 /* here_old is the first stripe that we might need to read
5101 if (mddev
->delta_disks
== 0) {
5102 if ((here_new
* mddev
->new_chunk_sectors
!=
5103 here_old
* mddev
->chunk_sectors
)) {
5104 printk(KERN_ERR
"md/raid:%s: reshape position is"
5105 " confused - aborting\n", mdname(mddev
));
5108 /* We cannot be sure it is safe to start an in-place
5109 * reshape. It is only safe if user-space is monitoring
5110 * and taking constant backups.
5111 * mdadm always starts a situation like this in
5112 * readonly mode so it can take control before
5113 * allowing any writes. So just check for that.
5115 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
5116 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
5117 /* not really in-place - so OK */;
5118 else if (mddev
->ro
== 0) {
5119 printk(KERN_ERR
"md/raid:%s: in-place reshape "
5120 "must be started in read-only mode "
5125 } else if (mddev
->reshape_backwards
5126 ? (here_new
* mddev
->new_chunk_sectors
+ min_offset_diff
<=
5127 here_old
* mddev
->chunk_sectors
)
5128 : (here_new
* mddev
->new_chunk_sectors
>=
5129 here_old
* mddev
->chunk_sectors
+ (-min_offset_diff
))) {
5130 /* Reading from the same stripe as writing to - bad */
5131 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
5132 "auto-recovery - aborting.\n",
5136 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
5138 /* OK, we should be able to continue; */
5140 BUG_ON(mddev
->level
!= mddev
->new_level
);
5141 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
5142 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
5143 BUG_ON(mddev
->delta_disks
!= 0);
5146 if (mddev
->private == NULL
)
5147 conf
= setup_conf(mddev
);
5149 conf
= mddev
->private;
5152 return PTR_ERR(conf
);
5154 conf
->min_offset_diff
= min_offset_diff
;
5155 mddev
->thread
= conf
->thread
;
5156 conf
->thread
= NULL
;
5157 mddev
->private = conf
;
5159 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
5161 rdev
= conf
->disks
[i
].rdev
;
5162 if (!rdev
&& conf
->disks
[i
].replacement
) {
5163 /* The replacement is all we have yet */
5164 rdev
= conf
->disks
[i
].replacement
;
5165 conf
->disks
[i
].replacement
= NULL
;
5166 clear_bit(Replacement
, &rdev
->flags
);
5167 conf
->disks
[i
].rdev
= rdev
;
5171 if (conf
->disks
[i
].replacement
&&
5172 conf
->reshape_progress
!= MaxSector
) {
5173 /* replacements and reshape simply do not mix. */
5174 printk(KERN_ERR
"md: cannot handle concurrent "
5175 "replacement and reshape.\n");
5178 if (test_bit(In_sync
, &rdev
->flags
)) {
5182 /* This disc is not fully in-sync. However if it
5183 * just stored parity (beyond the recovery_offset),
5184 * when we don't need to be concerned about the
5185 * array being dirty.
5186 * When reshape goes 'backwards', we never have
5187 * partially completed devices, so we only need
5188 * to worry about reshape going forwards.
5190 /* Hack because v0.91 doesn't store recovery_offset properly. */
5191 if (mddev
->major_version
== 0 &&
5192 mddev
->minor_version
> 90)
5193 rdev
->recovery_offset
= reshape_offset
;
5195 if (rdev
->recovery_offset
< reshape_offset
) {
5196 /* We need to check old and new layout */
5197 if (!only_parity(rdev
->raid_disk
,
5200 conf
->max_degraded
))
5203 if (!only_parity(rdev
->raid_disk
,
5205 conf
->previous_raid_disks
,
5206 conf
->max_degraded
))
5208 dirty_parity_disks
++;
5212 * 0 for a fully functional array, 1 or 2 for a degraded array.
5214 mddev
->degraded
= calc_degraded(conf
);
5216 if (has_failed(conf
)) {
5217 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
5218 " (%d/%d failed)\n",
5219 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
5223 /* device size must be a multiple of chunk size */
5224 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
5225 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
5227 if (mddev
->degraded
> dirty_parity_disks
&&
5228 mddev
->recovery_cp
!= MaxSector
) {
5229 if (mddev
->ok_start_degraded
)
5231 "md/raid:%s: starting dirty degraded array"
5232 " - data corruption possible.\n",
5236 "md/raid:%s: cannot start dirty degraded array.\n",
5242 if (mddev
->degraded
== 0)
5243 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
5244 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
5245 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
5248 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
5249 " out of %d devices, algorithm %d\n",
5250 mdname(mddev
), conf
->level
,
5251 mddev
->raid_disks
- mddev
->degraded
,
5252 mddev
->raid_disks
, mddev
->new_layout
);
5254 print_raid5_conf(conf
);
5256 if (conf
->reshape_progress
!= MaxSector
) {
5257 conf
->reshape_safe
= conf
->reshape_progress
;
5258 atomic_set(&conf
->reshape_stripes
, 0);
5259 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5260 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5261 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5262 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5263 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5268 /* Ok, everything is just fine now */
5269 if (mddev
->to_remove
== &raid5_attrs_group
)
5270 mddev
->to_remove
= NULL
;
5271 else if (mddev
->kobj
.sd
&&
5272 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
5274 "raid5: failed to create sysfs attributes for %s\n",
5276 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5280 /* read-ahead size must cover two whole stripes, which
5281 * is 2 * (datadisks) * chunksize where 'n' is the
5282 * number of raid devices
5284 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
5285 int stripe
= data_disks
*
5286 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
5287 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
5288 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
5290 blk_queue_merge_bvec(mddev
->queue
, raid5_mergeable_bvec
);
5292 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
5293 mddev
->queue
->backing_dev_info
.congested_fn
= raid5_congested
;
5295 chunk_size
= mddev
->chunk_sectors
<< 9;
5296 blk_queue_io_min(mddev
->queue
, chunk_size
);
5297 blk_queue_io_opt(mddev
->queue
, chunk_size
*
5298 (conf
->raid_disks
- conf
->max_degraded
));
5300 rdev_for_each(rdev
, mddev
) {
5301 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
5302 rdev
->data_offset
<< 9);
5303 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
5304 rdev
->new_data_offset
<< 9);
5310 md_unregister_thread(&mddev
->thread
);
5311 print_raid5_conf(conf
);
5313 mddev
->private = NULL
;
5314 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
5318 static int stop(struct mddev
*mddev
)
5320 struct r5conf
*conf
= mddev
->private;
5322 md_unregister_thread(&mddev
->thread
);
5324 mddev
->queue
->backing_dev_info
.congested_fn
= NULL
;
5326 mddev
->private = NULL
;
5327 mddev
->to_remove
= &raid5_attrs_group
;
5331 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
5333 struct r5conf
*conf
= mddev
->private;
5336 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
5337 mddev
->chunk_sectors
/ 2, mddev
->layout
);
5338 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
5339 for (i
= 0; i
< conf
->raid_disks
; i
++)
5340 seq_printf (seq
, "%s",
5341 conf
->disks
[i
].rdev
&&
5342 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
5343 seq_printf (seq
, "]");
5346 static void print_raid5_conf (struct r5conf
*conf
)
5349 struct disk_info
*tmp
;
5351 printk(KERN_DEBUG
"RAID conf printout:\n");
5353 printk("(conf==NULL)\n");
5356 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
5358 conf
->raid_disks
- conf
->mddev
->degraded
);
5360 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5361 char b
[BDEVNAME_SIZE
];
5362 tmp
= conf
->disks
+ i
;
5364 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
5365 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
5366 bdevname(tmp
->rdev
->bdev
, b
));
5370 static int raid5_spare_active(struct mddev
*mddev
)
5373 struct r5conf
*conf
= mddev
->private;
5374 struct disk_info
*tmp
;
5376 unsigned long flags
;
5378 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5379 tmp
= conf
->disks
+ i
;
5380 if (tmp
->replacement
5381 && tmp
->replacement
->recovery_offset
== MaxSector
5382 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
5383 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
5384 /* Replacement has just become active. */
5386 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
5389 /* Replaced device not technically faulty,
5390 * but we need to be sure it gets removed
5391 * and never re-added.
5393 set_bit(Faulty
, &tmp
->rdev
->flags
);
5394 sysfs_notify_dirent_safe(
5395 tmp
->rdev
->sysfs_state
);
5397 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
5398 } else if (tmp
->rdev
5399 && tmp
->rdev
->recovery_offset
== MaxSector
5400 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
5401 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
5403 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
5406 spin_lock_irqsave(&conf
->device_lock
, flags
);
5407 mddev
->degraded
= calc_degraded(conf
);
5408 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5409 print_raid5_conf(conf
);
5413 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
5415 struct r5conf
*conf
= mddev
->private;
5417 int number
= rdev
->raid_disk
;
5418 struct md_rdev
**rdevp
;
5419 struct disk_info
*p
= conf
->disks
+ number
;
5421 print_raid5_conf(conf
);
5422 if (rdev
== p
->rdev
)
5424 else if (rdev
== p
->replacement
)
5425 rdevp
= &p
->replacement
;
5429 if (number
>= conf
->raid_disks
&&
5430 conf
->reshape_progress
== MaxSector
)
5431 clear_bit(In_sync
, &rdev
->flags
);
5433 if (test_bit(In_sync
, &rdev
->flags
) ||
5434 atomic_read(&rdev
->nr_pending
)) {
5438 /* Only remove non-faulty devices if recovery
5441 if (!test_bit(Faulty
, &rdev
->flags
) &&
5442 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
5443 !has_failed(conf
) &&
5444 (!p
->replacement
|| p
->replacement
== rdev
) &&
5445 number
< conf
->raid_disks
) {
5451 if (atomic_read(&rdev
->nr_pending
)) {
5452 /* lost the race, try later */
5455 } else if (p
->replacement
) {
5456 /* We must have just cleared 'rdev' */
5457 p
->rdev
= p
->replacement
;
5458 clear_bit(Replacement
, &p
->replacement
->flags
);
5459 smp_mb(); /* Make sure other CPUs may see both as identical
5460 * but will never see neither - if they are careful
5462 p
->replacement
= NULL
;
5463 clear_bit(WantReplacement
, &rdev
->flags
);
5465 /* We might have just removed the Replacement as faulty-
5466 * clear the bit just in case
5468 clear_bit(WantReplacement
, &rdev
->flags
);
5471 print_raid5_conf(conf
);
5475 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
5477 struct r5conf
*conf
= mddev
->private;
5480 struct disk_info
*p
;
5482 int last
= conf
->raid_disks
- 1;
5484 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
5487 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
5488 /* no point adding a device */
5491 if (rdev
->raid_disk
>= 0)
5492 first
= last
= rdev
->raid_disk
;
5495 * find the disk ... but prefer rdev->saved_raid_disk
5498 if (rdev
->saved_raid_disk
>= 0 &&
5499 rdev
->saved_raid_disk
>= first
&&
5500 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
5501 first
= rdev
->saved_raid_disk
;
5503 for (disk
= first
; disk
<= last
; disk
++) {
5504 p
= conf
->disks
+ disk
;
5505 if (p
->rdev
== NULL
) {
5506 clear_bit(In_sync
, &rdev
->flags
);
5507 rdev
->raid_disk
= disk
;
5509 if (rdev
->saved_raid_disk
!= disk
)
5511 rcu_assign_pointer(p
->rdev
, rdev
);
5515 for (disk
= first
; disk
<= last
; disk
++) {
5516 p
= conf
->disks
+ disk
;
5517 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
5518 p
->replacement
== NULL
) {
5519 clear_bit(In_sync
, &rdev
->flags
);
5520 set_bit(Replacement
, &rdev
->flags
);
5521 rdev
->raid_disk
= disk
;
5524 rcu_assign_pointer(p
->replacement
, rdev
);
5529 print_raid5_conf(conf
);
5533 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
5535 /* no resync is happening, and there is enough space
5536 * on all devices, so we can resize.
5537 * We need to make sure resync covers any new space.
5538 * If the array is shrinking we should possibly wait until
5539 * any io in the removed space completes, but it hardly seems
5543 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
5544 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
5545 if (mddev
->external_size
&&
5546 mddev
->array_sectors
> newsize
)
5548 if (mddev
->bitmap
) {
5549 int ret
= bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
5553 md_set_array_sectors(mddev
, newsize
);
5554 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5555 revalidate_disk(mddev
->gendisk
);
5556 if (sectors
> mddev
->dev_sectors
&&
5557 mddev
->recovery_cp
> mddev
->dev_sectors
) {
5558 mddev
->recovery_cp
= mddev
->dev_sectors
;
5559 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
5561 mddev
->dev_sectors
= sectors
;
5562 mddev
->resync_max_sectors
= sectors
;
5566 static int check_stripe_cache(struct mddev
*mddev
)
5568 /* Can only proceed if there are plenty of stripe_heads.
5569 * We need a minimum of one full stripe,, and for sensible progress
5570 * it is best to have about 4 times that.
5571 * If we require 4 times, then the default 256 4K stripe_heads will
5572 * allow for chunk sizes up to 256K, which is probably OK.
5573 * If the chunk size is greater, user-space should request more
5574 * stripe_heads first.
5576 struct r5conf
*conf
= mddev
->private;
5577 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5578 > conf
->max_nr_stripes
||
5579 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5580 > conf
->max_nr_stripes
) {
5581 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5583 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
5590 static int check_reshape(struct mddev
*mddev
)
5592 struct r5conf
*conf
= mddev
->private;
5594 if (mddev
->delta_disks
== 0 &&
5595 mddev
->new_layout
== mddev
->layout
&&
5596 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
5597 return 0; /* nothing to do */
5598 if (has_failed(conf
))
5600 if (mddev
->delta_disks
< 0) {
5601 /* We might be able to shrink, but the devices must
5602 * be made bigger first.
5603 * For raid6, 4 is the minimum size.
5604 * Otherwise 2 is the minimum
5607 if (mddev
->level
== 6)
5609 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
5613 if (!check_stripe_cache(mddev
))
5616 return resize_stripes(conf
, conf
->raid_disks
+ mddev
->delta_disks
);
5619 static int raid5_start_reshape(struct mddev
*mddev
)
5621 struct r5conf
*conf
= mddev
->private;
5622 struct md_rdev
*rdev
;
5624 unsigned long flags
;
5626 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
5629 if (!check_stripe_cache(mddev
))
5632 if (has_failed(conf
))
5635 rdev_for_each(rdev
, mddev
) {
5636 if (!test_bit(In_sync
, &rdev
->flags
)
5637 && !test_bit(Faulty
, &rdev
->flags
))
5641 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
5642 /* Not enough devices even to make a degraded array
5647 /* Refuse to reduce size of the array. Any reductions in
5648 * array size must be through explicit setting of array_size
5651 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
5652 < mddev
->array_sectors
) {
5653 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
5654 "before number of disks\n", mdname(mddev
));
5658 atomic_set(&conf
->reshape_stripes
, 0);
5659 spin_lock_irq(&conf
->device_lock
);
5660 conf
->previous_raid_disks
= conf
->raid_disks
;
5661 conf
->raid_disks
+= mddev
->delta_disks
;
5662 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
5663 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
5664 conf
->prev_algo
= conf
->algorithm
;
5665 conf
->algorithm
= mddev
->new_layout
;
5667 /* Code that selects data_offset needs to see the generation update
5668 * if reshape_progress has been set - so a memory barrier needed.
5671 if (mddev
->reshape_backwards
)
5672 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
5674 conf
->reshape_progress
= 0;
5675 conf
->reshape_safe
= conf
->reshape_progress
;
5676 spin_unlock_irq(&conf
->device_lock
);
5678 /* Add some new drives, as many as will fit.
5679 * We know there are enough to make the newly sized array work.
5680 * Don't add devices if we are reducing the number of
5681 * devices in the array. This is because it is not possible
5682 * to correctly record the "partially reconstructed" state of
5683 * such devices during the reshape and confusion could result.
5685 if (mddev
->delta_disks
>= 0) {
5686 rdev_for_each(rdev
, mddev
)
5687 if (rdev
->raid_disk
< 0 &&
5688 !test_bit(Faulty
, &rdev
->flags
)) {
5689 if (raid5_add_disk(mddev
, rdev
) == 0) {
5691 >= conf
->previous_raid_disks
)
5692 set_bit(In_sync
, &rdev
->flags
);
5694 rdev
->recovery_offset
= 0;
5696 if (sysfs_link_rdev(mddev
, rdev
))
5697 /* Failure here is OK */;
5699 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
5700 && !test_bit(Faulty
, &rdev
->flags
)) {
5701 /* This is a spare that was manually added */
5702 set_bit(In_sync
, &rdev
->flags
);
5705 /* When a reshape changes the number of devices,
5706 * ->degraded is measured against the larger of the
5707 * pre and post number of devices.
5709 spin_lock_irqsave(&conf
->device_lock
, flags
);
5710 mddev
->degraded
= calc_degraded(conf
);
5711 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5713 mddev
->raid_disks
= conf
->raid_disks
;
5714 mddev
->reshape_position
= conf
->reshape_progress
;
5715 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5717 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5718 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5719 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5720 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5721 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5723 if (!mddev
->sync_thread
) {
5724 mddev
->recovery
= 0;
5725 spin_lock_irq(&conf
->device_lock
);
5726 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
5727 rdev_for_each(rdev
, mddev
)
5728 rdev
->new_data_offset
= rdev
->data_offset
;
5730 conf
->reshape_progress
= MaxSector
;
5731 mddev
->reshape_position
= MaxSector
;
5732 spin_unlock_irq(&conf
->device_lock
);
5735 conf
->reshape_checkpoint
= jiffies
;
5736 md_wakeup_thread(mddev
->sync_thread
);
5737 md_new_event(mddev
);
5741 /* This is called from the reshape thread and should make any
5742 * changes needed in 'conf'
5744 static void end_reshape(struct r5conf
*conf
)
5747 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
5748 struct md_rdev
*rdev
;
5750 spin_lock_irq(&conf
->device_lock
);
5751 conf
->previous_raid_disks
= conf
->raid_disks
;
5752 rdev_for_each(rdev
, conf
->mddev
)
5753 rdev
->data_offset
= rdev
->new_data_offset
;
5755 conf
->reshape_progress
= MaxSector
;
5756 spin_unlock_irq(&conf
->device_lock
);
5757 wake_up(&conf
->wait_for_overlap
);
5759 /* read-ahead size must cover two whole stripes, which is
5760 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5762 if (conf
->mddev
->queue
) {
5763 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5764 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
5766 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
5767 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
5772 /* This is called from the raid5d thread with mddev_lock held.
5773 * It makes config changes to the device.
5775 static void raid5_finish_reshape(struct mddev
*mddev
)
5777 struct r5conf
*conf
= mddev
->private;
5779 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
5781 if (mddev
->delta_disks
> 0) {
5782 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5783 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5784 revalidate_disk(mddev
->gendisk
);
5787 spin_lock_irq(&conf
->device_lock
);
5788 mddev
->degraded
= calc_degraded(conf
);
5789 spin_unlock_irq(&conf
->device_lock
);
5790 for (d
= conf
->raid_disks
;
5791 d
< conf
->raid_disks
- mddev
->delta_disks
;
5793 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
5795 clear_bit(In_sync
, &rdev
->flags
);
5796 rdev
= conf
->disks
[d
].replacement
;
5798 clear_bit(In_sync
, &rdev
->flags
);
5801 mddev
->layout
= conf
->algorithm
;
5802 mddev
->chunk_sectors
= conf
->chunk_sectors
;
5803 mddev
->reshape_position
= MaxSector
;
5804 mddev
->delta_disks
= 0;
5805 mddev
->reshape_backwards
= 0;
5809 static void raid5_quiesce(struct mddev
*mddev
, int state
)
5811 struct r5conf
*conf
= mddev
->private;
5814 case 2: /* resume for a suspend */
5815 wake_up(&conf
->wait_for_overlap
);
5818 case 1: /* stop all writes */
5819 spin_lock_irq(&conf
->device_lock
);
5820 /* '2' tells resync/reshape to pause so that all
5821 * active stripes can drain
5824 wait_event_lock_irq(conf
->wait_for_stripe
,
5825 atomic_read(&conf
->active_stripes
) == 0 &&
5826 atomic_read(&conf
->active_aligned_reads
) == 0,
5827 conf
->device_lock
, /* nothing */);
5829 spin_unlock_irq(&conf
->device_lock
);
5830 /* allow reshape to continue */
5831 wake_up(&conf
->wait_for_overlap
);
5834 case 0: /* re-enable writes */
5835 spin_lock_irq(&conf
->device_lock
);
5837 wake_up(&conf
->wait_for_stripe
);
5838 wake_up(&conf
->wait_for_overlap
);
5839 spin_unlock_irq(&conf
->device_lock
);
5845 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
5847 struct r0conf
*raid0_conf
= mddev
->private;
5850 /* for raid0 takeover only one zone is supported */
5851 if (raid0_conf
->nr_strip_zones
> 1) {
5852 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5854 return ERR_PTR(-EINVAL
);
5857 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
5858 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
5859 mddev
->dev_sectors
= sectors
;
5860 mddev
->new_level
= level
;
5861 mddev
->new_layout
= ALGORITHM_PARITY_N
;
5862 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
5863 mddev
->raid_disks
+= 1;
5864 mddev
->delta_disks
= 1;
5865 /* make sure it will be not marked as dirty */
5866 mddev
->recovery_cp
= MaxSector
;
5868 return setup_conf(mddev
);
5872 static void *raid5_takeover_raid1(struct mddev
*mddev
)
5876 if (mddev
->raid_disks
!= 2 ||
5877 mddev
->degraded
> 1)
5878 return ERR_PTR(-EINVAL
);
5880 /* Should check if there are write-behind devices? */
5882 chunksect
= 64*2; /* 64K by default */
5884 /* The array must be an exact multiple of chunksize */
5885 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
5888 if ((chunksect
<<9) < STRIPE_SIZE
)
5889 /* array size does not allow a suitable chunk size */
5890 return ERR_PTR(-EINVAL
);
5892 mddev
->new_level
= 5;
5893 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5894 mddev
->new_chunk_sectors
= chunksect
;
5896 return setup_conf(mddev
);
5899 static void *raid5_takeover_raid6(struct mddev
*mddev
)
5903 switch (mddev
->layout
) {
5904 case ALGORITHM_LEFT_ASYMMETRIC_6
:
5905 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
5907 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
5908 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
5910 case ALGORITHM_LEFT_SYMMETRIC_6
:
5911 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5913 case ALGORITHM_RIGHT_SYMMETRIC_6
:
5914 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
5916 case ALGORITHM_PARITY_0_6
:
5917 new_layout
= ALGORITHM_PARITY_0
;
5919 case ALGORITHM_PARITY_N
:
5920 new_layout
= ALGORITHM_PARITY_N
;
5923 return ERR_PTR(-EINVAL
);
5925 mddev
->new_level
= 5;
5926 mddev
->new_layout
= new_layout
;
5927 mddev
->delta_disks
= -1;
5928 mddev
->raid_disks
-= 1;
5929 return setup_conf(mddev
);
5933 static int raid5_check_reshape(struct mddev
*mddev
)
5935 /* For a 2-drive array, the layout and chunk size can be changed
5936 * immediately as not restriping is needed.
5937 * For larger arrays we record the new value - after validation
5938 * to be used by a reshape pass.
5940 struct r5conf
*conf
= mddev
->private;
5941 int new_chunk
= mddev
->new_chunk_sectors
;
5943 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
5945 if (new_chunk
> 0) {
5946 if (!is_power_of_2(new_chunk
))
5948 if (new_chunk
< (PAGE_SIZE
>>9))
5950 if (mddev
->array_sectors
& (new_chunk
-1))
5951 /* not factor of array size */
5955 /* They look valid */
5957 if (mddev
->raid_disks
== 2) {
5958 /* can make the change immediately */
5959 if (mddev
->new_layout
>= 0) {
5960 conf
->algorithm
= mddev
->new_layout
;
5961 mddev
->layout
= mddev
->new_layout
;
5963 if (new_chunk
> 0) {
5964 conf
->chunk_sectors
= new_chunk
;
5965 mddev
->chunk_sectors
= new_chunk
;
5967 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5968 md_wakeup_thread(mddev
->thread
);
5970 return check_reshape(mddev
);
5973 static int raid6_check_reshape(struct mddev
*mddev
)
5975 int new_chunk
= mddev
->new_chunk_sectors
;
5977 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
5979 if (new_chunk
> 0) {
5980 if (!is_power_of_2(new_chunk
))
5982 if (new_chunk
< (PAGE_SIZE
>> 9))
5984 if (mddev
->array_sectors
& (new_chunk
-1))
5985 /* not factor of array size */
5989 /* They look valid */
5990 return check_reshape(mddev
);
5993 static void *raid5_takeover(struct mddev
*mddev
)
5995 /* raid5 can take over:
5996 * raid0 - if there is only one strip zone - make it a raid4 layout
5997 * raid1 - if there are two drives. We need to know the chunk size
5998 * raid4 - trivial - just use a raid4 layout.
5999 * raid6 - Providing it is a *_6 layout
6001 if (mddev
->level
== 0)
6002 return raid45_takeover_raid0(mddev
, 5);
6003 if (mddev
->level
== 1)
6004 return raid5_takeover_raid1(mddev
);
6005 if (mddev
->level
== 4) {
6006 mddev
->new_layout
= ALGORITHM_PARITY_N
;
6007 mddev
->new_level
= 5;
6008 return setup_conf(mddev
);
6010 if (mddev
->level
== 6)
6011 return raid5_takeover_raid6(mddev
);
6013 return ERR_PTR(-EINVAL
);
6016 static void *raid4_takeover(struct mddev
*mddev
)
6018 /* raid4 can take over:
6019 * raid0 - if there is only one strip zone
6020 * raid5 - if layout is right
6022 if (mddev
->level
== 0)
6023 return raid45_takeover_raid0(mddev
, 4);
6024 if (mddev
->level
== 5 &&
6025 mddev
->layout
== ALGORITHM_PARITY_N
) {
6026 mddev
->new_layout
= 0;
6027 mddev
->new_level
= 4;
6028 return setup_conf(mddev
);
6030 return ERR_PTR(-EINVAL
);
6033 static struct md_personality raid5_personality
;
6035 static void *raid6_takeover(struct mddev
*mddev
)
6037 /* Currently can only take over a raid5. We map the
6038 * personality to an equivalent raid6 personality
6039 * with the Q block at the end.
6043 if (mddev
->pers
!= &raid5_personality
)
6044 return ERR_PTR(-EINVAL
);
6045 if (mddev
->degraded
> 1)
6046 return ERR_PTR(-EINVAL
);
6047 if (mddev
->raid_disks
> 253)
6048 return ERR_PTR(-EINVAL
);
6049 if (mddev
->raid_disks
< 3)
6050 return ERR_PTR(-EINVAL
);
6052 switch (mddev
->layout
) {
6053 case ALGORITHM_LEFT_ASYMMETRIC
:
6054 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
6056 case ALGORITHM_RIGHT_ASYMMETRIC
:
6057 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
6059 case ALGORITHM_LEFT_SYMMETRIC
:
6060 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
6062 case ALGORITHM_RIGHT_SYMMETRIC
:
6063 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
6065 case ALGORITHM_PARITY_0
:
6066 new_layout
= ALGORITHM_PARITY_0_6
;
6068 case ALGORITHM_PARITY_N
:
6069 new_layout
= ALGORITHM_PARITY_N
;
6072 return ERR_PTR(-EINVAL
);
6074 mddev
->new_level
= 6;
6075 mddev
->new_layout
= new_layout
;
6076 mddev
->delta_disks
= 1;
6077 mddev
->raid_disks
+= 1;
6078 return setup_conf(mddev
);
6082 static struct md_personality raid6_personality
=
6086 .owner
= THIS_MODULE
,
6087 .make_request
= make_request
,
6091 .error_handler
= error
,
6092 .hot_add_disk
= raid5_add_disk
,
6093 .hot_remove_disk
= raid5_remove_disk
,
6094 .spare_active
= raid5_spare_active
,
6095 .sync_request
= sync_request
,
6096 .resize
= raid5_resize
,
6098 .check_reshape
= raid6_check_reshape
,
6099 .start_reshape
= raid5_start_reshape
,
6100 .finish_reshape
= raid5_finish_reshape
,
6101 .quiesce
= raid5_quiesce
,
6102 .takeover
= raid6_takeover
,
6104 static struct md_personality raid5_personality
=
6108 .owner
= THIS_MODULE
,
6109 .make_request
= make_request
,
6113 .error_handler
= error
,
6114 .hot_add_disk
= raid5_add_disk
,
6115 .hot_remove_disk
= raid5_remove_disk
,
6116 .spare_active
= raid5_spare_active
,
6117 .sync_request
= sync_request
,
6118 .resize
= raid5_resize
,
6120 .check_reshape
= raid5_check_reshape
,
6121 .start_reshape
= raid5_start_reshape
,
6122 .finish_reshape
= raid5_finish_reshape
,
6123 .quiesce
= raid5_quiesce
,
6124 .takeover
= raid5_takeover
,
6127 static struct md_personality raid4_personality
=
6131 .owner
= THIS_MODULE
,
6132 .make_request
= make_request
,
6136 .error_handler
= error
,
6137 .hot_add_disk
= raid5_add_disk
,
6138 .hot_remove_disk
= raid5_remove_disk
,
6139 .spare_active
= raid5_spare_active
,
6140 .sync_request
= sync_request
,
6141 .resize
= raid5_resize
,
6143 .check_reshape
= raid5_check_reshape
,
6144 .start_reshape
= raid5_start_reshape
,
6145 .finish_reshape
= raid5_finish_reshape
,
6146 .quiesce
= raid5_quiesce
,
6147 .takeover
= raid4_takeover
,
6150 static int __init
raid5_init(void)
6152 register_md_personality(&raid6_personality
);
6153 register_md_personality(&raid5_personality
);
6154 register_md_personality(&raid4_personality
);
6158 static void raid5_exit(void)
6160 unregister_md_personality(&raid6_personality
);
6161 unregister_md_personality(&raid5_personality
);
6162 unregister_md_personality(&raid4_personality
);
6165 module_init(raid5_init
);
6166 module_exit(raid5_exit
);
6167 MODULE_LICENSE("GPL");
6168 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6169 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6170 MODULE_ALIAS("md-raid5");
6171 MODULE_ALIAS("md-raid4");
6172 MODULE_ALIAS("md-level-5");
6173 MODULE_ALIAS("md-level-4");
6174 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6175 MODULE_ALIAS("md-raid6");
6176 MODULE_ALIAS("md-level-6");
6178 /* This used to be two separate modules, they were: */
6179 MODULE_ALIAS("raid5");
6180 MODULE_ALIAS("raid6");