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_phys_segments(struct bio
*bio
)
104 return bio
->bi_phys_segments
& 0xffff;
107 static inline int raid5_bi_hw_segments(struct bio
*bio
)
109 return (bio
->bi_phys_segments
>> 16) & 0xffff;
112 static inline int raid5_dec_bi_phys_segments(struct bio
*bio
)
114 --bio
->bi_phys_segments
;
115 return raid5_bi_phys_segments(bio
);
118 static inline int raid5_dec_bi_hw_segments(struct bio
*bio
)
120 unsigned short val
= raid5_bi_hw_segments(bio
);
123 bio
->bi_phys_segments
= (val
<< 16) | raid5_bi_phys_segments(bio
);
127 static inline void raid5_set_bi_hw_segments(struct bio
*bio
, unsigned int cnt
)
129 bio
->bi_phys_segments
= raid5_bi_phys_segments(bio
) | (cnt
<< 16);
132 /* Find first data disk in a raid6 stripe */
133 static inline int raid6_d0(struct stripe_head
*sh
)
136 /* ddf always start from first device */
138 /* md starts just after Q block */
139 if (sh
->qd_idx
== sh
->disks
- 1)
142 return sh
->qd_idx
+ 1;
144 static inline int raid6_next_disk(int disk
, int raid_disks
)
147 return (disk
< raid_disks
) ? disk
: 0;
150 /* When walking through the disks in a raid5, starting at raid6_d0,
151 * We need to map each disk to a 'slot', where the data disks are slot
152 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
153 * is raid_disks-1. This help does that mapping.
155 static int raid6_idx_to_slot(int idx
, struct stripe_head
*sh
,
156 int *count
, int syndrome_disks
)
162 if (idx
== sh
->pd_idx
)
163 return syndrome_disks
;
164 if (idx
== sh
->qd_idx
)
165 return syndrome_disks
+ 1;
171 static void return_io(struct bio
*return_bi
)
173 struct bio
*bi
= return_bi
;
176 return_bi
= bi
->bi_next
;
184 static void print_raid5_conf (struct r5conf
*conf
);
186 static int stripe_operations_active(struct stripe_head
*sh
)
188 return sh
->check_state
|| sh
->reconstruct_state
||
189 test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
) ||
190 test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
193 static void __release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
)
195 if (atomic_dec_and_test(&sh
->count
)) {
196 BUG_ON(!list_empty(&sh
->lru
));
197 BUG_ON(atomic_read(&conf
->active_stripes
)==0);
198 if (test_bit(STRIPE_HANDLE
, &sh
->state
)) {
199 if (test_bit(STRIPE_DELAYED
, &sh
->state
))
200 list_add_tail(&sh
->lru
, &conf
->delayed_list
);
201 else if (test_bit(STRIPE_BIT_DELAY
, &sh
->state
) &&
202 sh
->bm_seq
- conf
->seq_write
> 0)
203 list_add_tail(&sh
->lru
, &conf
->bitmap_list
);
205 clear_bit(STRIPE_BIT_DELAY
, &sh
->state
);
206 list_add_tail(&sh
->lru
, &conf
->handle_list
);
208 md_wakeup_thread(conf
->mddev
->thread
);
210 BUG_ON(stripe_operations_active(sh
));
211 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
212 if (atomic_dec_return(&conf
->preread_active_stripes
)
214 md_wakeup_thread(conf
->mddev
->thread
);
215 atomic_dec(&conf
->active_stripes
);
216 if (!test_bit(STRIPE_EXPANDING
, &sh
->state
)) {
217 list_add_tail(&sh
->lru
, &conf
->inactive_list
);
218 wake_up(&conf
->wait_for_stripe
);
219 if (conf
->retry_read_aligned
)
220 md_wakeup_thread(conf
->mddev
->thread
);
226 static void release_stripe(struct stripe_head
*sh
)
228 struct r5conf
*conf
= sh
->raid_conf
;
231 spin_lock_irqsave(&conf
->device_lock
, flags
);
232 __release_stripe(conf
, sh
);
233 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
236 static inline void remove_hash(struct stripe_head
*sh
)
238 pr_debug("remove_hash(), stripe %llu\n",
239 (unsigned long long)sh
->sector
);
241 hlist_del_init(&sh
->hash
);
244 static inline void insert_hash(struct r5conf
*conf
, struct stripe_head
*sh
)
246 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
248 pr_debug("insert_hash(), stripe %llu\n",
249 (unsigned long long)sh
->sector
);
251 hlist_add_head(&sh
->hash
, hp
);
255 /* find an idle stripe, make sure it is unhashed, and return it. */
256 static struct stripe_head
*get_free_stripe(struct r5conf
*conf
)
258 struct stripe_head
*sh
= NULL
;
259 struct list_head
*first
;
261 if (list_empty(&conf
->inactive_list
))
263 first
= conf
->inactive_list
.next
;
264 sh
= list_entry(first
, struct stripe_head
, lru
);
265 list_del_init(first
);
267 atomic_inc(&conf
->active_stripes
);
272 static void shrink_buffers(struct stripe_head
*sh
)
276 int num
= sh
->raid_conf
->pool_size
;
278 for (i
= 0; i
< num
; i
++) {
282 sh
->dev
[i
].page
= NULL
;
287 static int grow_buffers(struct stripe_head
*sh
)
290 int num
= sh
->raid_conf
->pool_size
;
292 for (i
= 0; i
< num
; i
++) {
295 if (!(page
= alloc_page(GFP_KERNEL
))) {
298 sh
->dev
[i
].page
= page
;
303 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
);
304 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
305 struct stripe_head
*sh
);
307 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int previous
)
309 struct r5conf
*conf
= sh
->raid_conf
;
312 BUG_ON(atomic_read(&sh
->count
) != 0);
313 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
314 BUG_ON(stripe_operations_active(sh
));
316 pr_debug("init_stripe called, stripe %llu\n",
317 (unsigned long long)sh
->sector
);
321 sh
->generation
= conf
->generation
- previous
;
322 sh
->disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
324 stripe_set_idx(sector
, conf
, previous
, sh
);
328 for (i
= sh
->disks
; i
--; ) {
329 struct r5dev
*dev
= &sh
->dev
[i
];
331 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
332 test_bit(R5_LOCKED
, &dev
->flags
)) {
333 printk(KERN_ERR
"sector=%llx i=%d %p %p %p %p %d\n",
334 (unsigned long long)sh
->sector
, i
, dev
->toread
,
335 dev
->read
, dev
->towrite
, dev
->written
,
336 test_bit(R5_LOCKED
, &dev
->flags
));
340 raid5_build_block(sh
, i
, previous
);
342 insert_hash(conf
, sh
);
345 static struct stripe_head
*__find_stripe(struct r5conf
*conf
, sector_t sector
,
348 struct stripe_head
*sh
;
349 struct hlist_node
*hn
;
351 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
352 hlist_for_each_entry(sh
, hn
, stripe_hash(conf
, sector
), hash
)
353 if (sh
->sector
== sector
&& sh
->generation
== generation
)
355 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
360 * Need to check if array has failed when deciding whether to:
362 * - remove non-faulty devices
365 * This determination is simple when no reshape is happening.
366 * However if there is a reshape, we need to carefully check
367 * both the before and after sections.
368 * This is because some failed devices may only affect one
369 * of the two sections, and some non-in_sync devices may
370 * be insync in the section most affected by failed devices.
372 static int calc_degraded(struct r5conf
*conf
)
374 int degraded
, degraded2
;
379 for (i
= 0; i
< conf
->previous_raid_disks
; i
++) {
380 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
381 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
383 else if (test_bit(In_sync
, &rdev
->flags
))
386 /* not in-sync or faulty.
387 * If the reshape increases the number of devices,
388 * this is being recovered by the reshape, so
389 * this 'previous' section is not in_sync.
390 * If the number of devices is being reduced however,
391 * the device can only be part of the array if
392 * we are reverting a reshape, so this section will
395 if (conf
->raid_disks
>= conf
->previous_raid_disks
)
399 if (conf
->raid_disks
== conf
->previous_raid_disks
)
403 for (i
= 0; i
< conf
->raid_disks
; i
++) {
404 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
405 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
407 else if (test_bit(In_sync
, &rdev
->flags
))
410 /* not in-sync or faulty.
411 * If reshape increases the number of devices, this
412 * section has already been recovered, else it
413 * almost certainly hasn't.
415 if (conf
->raid_disks
<= conf
->previous_raid_disks
)
419 if (degraded2
> degraded
)
424 static int has_failed(struct r5conf
*conf
)
428 if (conf
->mddev
->reshape_position
== MaxSector
)
429 return conf
->mddev
->degraded
> conf
->max_degraded
;
431 degraded
= calc_degraded(conf
);
432 if (degraded
> conf
->max_degraded
)
437 static struct stripe_head
*
438 get_active_stripe(struct r5conf
*conf
, sector_t sector
,
439 int previous
, int noblock
, int noquiesce
)
441 struct stripe_head
*sh
;
443 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
445 spin_lock_irq(&conf
->device_lock
);
448 wait_event_lock_irq(conf
->wait_for_stripe
,
449 conf
->quiesce
== 0 || noquiesce
,
450 conf
->device_lock
, /* nothing */);
451 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
453 if (!conf
->inactive_blocked
)
454 sh
= get_free_stripe(conf
);
455 if (noblock
&& sh
== NULL
)
458 conf
->inactive_blocked
= 1;
459 wait_event_lock_irq(conf
->wait_for_stripe
,
460 !list_empty(&conf
->inactive_list
) &&
461 (atomic_read(&conf
->active_stripes
)
462 < (conf
->max_nr_stripes
*3/4)
463 || !conf
->inactive_blocked
),
466 conf
->inactive_blocked
= 0;
468 init_stripe(sh
, sector
, previous
);
470 if (atomic_read(&sh
->count
)) {
471 BUG_ON(!list_empty(&sh
->lru
)
472 && !test_bit(STRIPE_EXPANDING
, &sh
->state
));
474 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
475 atomic_inc(&conf
->active_stripes
);
476 if (list_empty(&sh
->lru
) &&
477 !test_bit(STRIPE_EXPANDING
, &sh
->state
))
479 list_del_init(&sh
->lru
);
482 } while (sh
== NULL
);
485 atomic_inc(&sh
->count
);
487 spin_unlock_irq(&conf
->device_lock
);
491 /* Determine if 'data_offset' or 'new_data_offset' should be used
492 * in this stripe_head.
494 static int use_new_offset(struct r5conf
*conf
, struct stripe_head
*sh
)
496 sector_t progress
= conf
->reshape_progress
;
497 /* Need a memory barrier to make sure we see the value
498 * of conf->generation, or ->data_offset that was set before
499 * reshape_progress was updated.
502 if (progress
== MaxSector
)
504 if (sh
->generation
== conf
->generation
- 1)
506 /* We are in a reshape, and this is a new-generation stripe,
507 * so use new_data_offset.
513 raid5_end_read_request(struct bio
*bi
, int error
);
515 raid5_end_write_request(struct bio
*bi
, int error
);
517 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
519 struct r5conf
*conf
= sh
->raid_conf
;
520 int i
, disks
= sh
->disks
;
524 for (i
= disks
; i
--; ) {
526 int replace_only
= 0;
527 struct bio
*bi
, *rbi
;
528 struct md_rdev
*rdev
, *rrdev
= NULL
;
529 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
530 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
534 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
536 else if (test_and_clear_bit(R5_WantReplace
,
537 &sh
->dev
[i
].flags
)) {
542 if (test_and_clear_bit(R5_SyncIO
, &sh
->dev
[i
].flags
))
545 bi
= &sh
->dev
[i
].req
;
546 rbi
= &sh
->dev
[i
].rreq
; /* For writing to replacement */
551 bi
->bi_end_io
= raid5_end_write_request
;
552 rbi
->bi_end_io
= raid5_end_write_request
;
554 bi
->bi_end_io
= raid5_end_read_request
;
557 rrdev
= rcu_dereference(conf
->disks
[i
].replacement
);
558 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
559 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
568 /* We raced and saw duplicates */
571 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
) && rrdev
)
576 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
579 atomic_inc(&rdev
->nr_pending
);
580 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
583 atomic_inc(&rrdev
->nr_pending
);
586 /* We have already checked bad blocks for reads. Now
587 * need to check for writes. We never accept write errors
588 * on the replacement, so we don't to check rrdev.
590 while ((rw
& WRITE
) && rdev
&&
591 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
594 int bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
595 &first_bad
, &bad_sectors
);
600 set_bit(BlockedBadBlocks
, &rdev
->flags
);
601 if (!conf
->mddev
->external
&&
602 conf
->mddev
->flags
) {
603 /* It is very unlikely, but we might
604 * still need to write out the
605 * bad block log - better give it
607 md_check_recovery(conf
->mddev
);
609 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
611 /* Acknowledged bad block - skip the write */
612 rdev_dec_pending(rdev
, conf
->mddev
);
618 if (s
->syncing
|| s
->expanding
|| s
->expanded
620 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
622 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
624 bi
->bi_bdev
= rdev
->bdev
;
625 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
626 __func__
, (unsigned long long)sh
->sector
,
628 atomic_inc(&sh
->count
);
629 if (use_new_offset(conf
, sh
))
630 bi
->bi_sector
= (sh
->sector
631 + rdev
->new_data_offset
);
633 bi
->bi_sector
= (sh
->sector
634 + rdev
->data_offset
);
635 bi
->bi_flags
= 1 << BIO_UPTODATE
;
637 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
638 bi
->bi_io_vec
[0].bv_offset
= 0;
639 bi
->bi_size
= STRIPE_SIZE
;
642 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
643 generic_make_request(bi
);
646 if (s
->syncing
|| s
->expanding
|| s
->expanded
648 md_sync_acct(rrdev
->bdev
, STRIPE_SECTORS
);
650 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
652 rbi
->bi_bdev
= rrdev
->bdev
;
653 pr_debug("%s: for %llu schedule op %ld on "
654 "replacement disc %d\n",
655 __func__
, (unsigned long long)sh
->sector
,
657 atomic_inc(&sh
->count
);
658 if (use_new_offset(conf
, sh
))
659 rbi
->bi_sector
= (sh
->sector
660 + rrdev
->new_data_offset
);
662 rbi
->bi_sector
= (sh
->sector
663 + rrdev
->data_offset
);
664 rbi
->bi_flags
= 1 << BIO_UPTODATE
;
666 rbi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
667 rbi
->bi_io_vec
[0].bv_offset
= 0;
668 rbi
->bi_size
= STRIPE_SIZE
;
670 generic_make_request(rbi
);
672 if (!rdev
&& !rrdev
) {
674 set_bit(STRIPE_DEGRADED
, &sh
->state
);
675 pr_debug("skip op %ld on disc %d for sector %llu\n",
676 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
677 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
678 set_bit(STRIPE_HANDLE
, &sh
->state
);
683 static struct dma_async_tx_descriptor
*
684 async_copy_data(int frombio
, struct bio
*bio
, struct page
*page
,
685 sector_t sector
, struct dma_async_tx_descriptor
*tx
)
688 struct page
*bio_page
;
691 struct async_submit_ctl submit
;
692 enum async_tx_flags flags
= 0;
694 if (bio
->bi_sector
>= sector
)
695 page_offset
= (signed)(bio
->bi_sector
- sector
) * 512;
697 page_offset
= (signed)(sector
- bio
->bi_sector
) * -512;
700 flags
|= ASYNC_TX_FENCE
;
701 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
703 bio_for_each_segment(bvl
, bio
, i
) {
704 int len
= bvl
->bv_len
;
708 if (page_offset
< 0) {
709 b_offset
= -page_offset
;
710 page_offset
+= b_offset
;
714 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
715 clen
= STRIPE_SIZE
- page_offset
;
720 b_offset
+= bvl
->bv_offset
;
721 bio_page
= bvl
->bv_page
;
723 tx
= async_memcpy(page
, bio_page
, page_offset
,
724 b_offset
, clen
, &submit
);
726 tx
= async_memcpy(bio_page
, page
, b_offset
,
727 page_offset
, clen
, &submit
);
729 /* chain the operations */
730 submit
.depend_tx
= tx
;
732 if (clen
< len
) /* hit end of page */
740 static void ops_complete_biofill(void *stripe_head_ref
)
742 struct stripe_head
*sh
= stripe_head_ref
;
743 struct bio
*return_bi
= NULL
;
744 struct r5conf
*conf
= sh
->raid_conf
;
747 pr_debug("%s: stripe %llu\n", __func__
,
748 (unsigned long long)sh
->sector
);
750 /* clear completed biofills */
751 spin_lock_irq(&conf
->device_lock
);
752 for (i
= sh
->disks
; i
--; ) {
753 struct r5dev
*dev
= &sh
->dev
[i
];
755 /* acknowledge completion of a biofill operation */
756 /* and check if we need to reply to a read request,
757 * new R5_Wantfill requests are held off until
758 * !STRIPE_BIOFILL_RUN
760 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
761 struct bio
*rbi
, *rbi2
;
766 while (rbi
&& rbi
->bi_sector
<
767 dev
->sector
+ STRIPE_SECTORS
) {
768 rbi2
= r5_next_bio(rbi
, dev
->sector
);
769 if (!raid5_dec_bi_phys_segments(rbi
)) {
770 rbi
->bi_next
= return_bi
;
777 spin_unlock_irq(&conf
->device_lock
);
778 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
780 return_io(return_bi
);
782 set_bit(STRIPE_HANDLE
, &sh
->state
);
786 static void ops_run_biofill(struct stripe_head
*sh
)
788 struct dma_async_tx_descriptor
*tx
= NULL
;
789 struct r5conf
*conf
= sh
->raid_conf
;
790 struct async_submit_ctl submit
;
793 pr_debug("%s: stripe %llu\n", __func__
,
794 (unsigned long long)sh
->sector
);
796 for (i
= sh
->disks
; i
--; ) {
797 struct r5dev
*dev
= &sh
->dev
[i
];
798 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
800 spin_lock_irq(&conf
->device_lock
);
801 dev
->read
= rbi
= dev
->toread
;
803 spin_unlock_irq(&conf
->device_lock
);
804 while (rbi
&& rbi
->bi_sector
<
805 dev
->sector
+ STRIPE_SECTORS
) {
806 tx
= async_copy_data(0, rbi
, dev
->page
,
808 rbi
= r5_next_bio(rbi
, dev
->sector
);
813 atomic_inc(&sh
->count
);
814 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
815 async_trigger_callback(&submit
);
818 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
825 tgt
= &sh
->dev
[target
];
826 set_bit(R5_UPTODATE
, &tgt
->flags
);
827 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
828 clear_bit(R5_Wantcompute
, &tgt
->flags
);
831 static void ops_complete_compute(void *stripe_head_ref
)
833 struct stripe_head
*sh
= stripe_head_ref
;
835 pr_debug("%s: stripe %llu\n", __func__
,
836 (unsigned long long)sh
->sector
);
838 /* mark the computed target(s) as uptodate */
839 mark_target_uptodate(sh
, sh
->ops
.target
);
840 mark_target_uptodate(sh
, sh
->ops
.target2
);
842 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
843 if (sh
->check_state
== check_state_compute_run
)
844 sh
->check_state
= check_state_compute_result
;
845 set_bit(STRIPE_HANDLE
, &sh
->state
);
849 /* return a pointer to the address conversion region of the scribble buffer */
850 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
851 struct raid5_percpu
*percpu
)
853 return percpu
->scribble
+ sizeof(struct page
*) * (sh
->disks
+ 2);
856 static struct dma_async_tx_descriptor
*
857 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
859 int disks
= sh
->disks
;
860 struct page
**xor_srcs
= percpu
->scribble
;
861 int target
= sh
->ops
.target
;
862 struct r5dev
*tgt
= &sh
->dev
[target
];
863 struct page
*xor_dest
= tgt
->page
;
865 struct dma_async_tx_descriptor
*tx
;
866 struct async_submit_ctl submit
;
869 pr_debug("%s: stripe %llu block: %d\n",
870 __func__
, (unsigned long long)sh
->sector
, target
);
871 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
873 for (i
= disks
; i
--; )
875 xor_srcs
[count
++] = sh
->dev
[i
].page
;
877 atomic_inc(&sh
->count
);
879 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
880 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
));
881 if (unlikely(count
== 1))
882 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
884 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
889 /* set_syndrome_sources - populate source buffers for gen_syndrome
890 * @srcs - (struct page *) array of size sh->disks
891 * @sh - stripe_head to parse
893 * Populates srcs in proper layout order for the stripe and returns the
894 * 'count' of sources to be used in a call to async_gen_syndrome. The P
895 * destination buffer is recorded in srcs[count] and the Q destination
896 * is recorded in srcs[count+1]].
898 static int set_syndrome_sources(struct page
**srcs
, struct stripe_head
*sh
)
900 int disks
= sh
->disks
;
901 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
902 int d0_idx
= raid6_d0(sh
);
906 for (i
= 0; i
< disks
; i
++)
912 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
914 srcs
[slot
] = sh
->dev
[i
].page
;
915 i
= raid6_next_disk(i
, disks
);
916 } while (i
!= d0_idx
);
918 return syndrome_disks
;
921 static struct dma_async_tx_descriptor
*
922 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
924 int disks
= sh
->disks
;
925 struct page
**blocks
= percpu
->scribble
;
927 int qd_idx
= sh
->qd_idx
;
928 struct dma_async_tx_descriptor
*tx
;
929 struct async_submit_ctl submit
;
935 if (sh
->ops
.target
< 0)
936 target
= sh
->ops
.target2
;
937 else if (sh
->ops
.target2
< 0)
938 target
= sh
->ops
.target
;
940 /* we should only have one valid target */
943 pr_debug("%s: stripe %llu block: %d\n",
944 __func__
, (unsigned long long)sh
->sector
, target
);
946 tgt
= &sh
->dev
[target
];
947 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
950 atomic_inc(&sh
->count
);
952 if (target
== qd_idx
) {
953 count
= set_syndrome_sources(blocks
, sh
);
954 blocks
[count
] = NULL
; /* regenerating p is not necessary */
955 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
956 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
957 ops_complete_compute
, sh
,
958 to_addr_conv(sh
, percpu
));
959 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
961 /* Compute any data- or p-drive using XOR */
963 for (i
= disks
; i
-- ; ) {
964 if (i
== target
|| i
== qd_idx
)
966 blocks
[count
++] = sh
->dev
[i
].page
;
969 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
970 NULL
, ops_complete_compute
, sh
,
971 to_addr_conv(sh
, percpu
));
972 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
978 static struct dma_async_tx_descriptor
*
979 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
981 int i
, count
, disks
= sh
->disks
;
982 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
983 int d0_idx
= raid6_d0(sh
);
984 int faila
= -1, failb
= -1;
985 int target
= sh
->ops
.target
;
986 int target2
= sh
->ops
.target2
;
987 struct r5dev
*tgt
= &sh
->dev
[target
];
988 struct r5dev
*tgt2
= &sh
->dev
[target2
];
989 struct dma_async_tx_descriptor
*tx
;
990 struct page
**blocks
= percpu
->scribble
;
991 struct async_submit_ctl submit
;
993 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
994 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
995 BUG_ON(target
< 0 || target2
< 0);
996 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
997 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
999 /* we need to open-code set_syndrome_sources to handle the
1000 * slot number conversion for 'faila' and 'failb'
1002 for (i
= 0; i
< disks
; i
++)
1007 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1009 blocks
[slot
] = sh
->dev
[i
].page
;
1015 i
= raid6_next_disk(i
, disks
);
1016 } while (i
!= d0_idx
);
1018 BUG_ON(faila
== failb
);
1021 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1022 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1024 atomic_inc(&sh
->count
);
1026 if (failb
== syndrome_disks
+1) {
1027 /* Q disk is one of the missing disks */
1028 if (faila
== syndrome_disks
) {
1029 /* Missing P+Q, just recompute */
1030 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1031 ops_complete_compute
, sh
,
1032 to_addr_conv(sh
, percpu
));
1033 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
1034 STRIPE_SIZE
, &submit
);
1038 int qd_idx
= sh
->qd_idx
;
1040 /* Missing D+Q: recompute D from P, then recompute Q */
1041 if (target
== qd_idx
)
1042 data_target
= target2
;
1044 data_target
= target
;
1047 for (i
= disks
; i
-- ; ) {
1048 if (i
== data_target
|| i
== qd_idx
)
1050 blocks
[count
++] = sh
->dev
[i
].page
;
1052 dest
= sh
->dev
[data_target
].page
;
1053 init_async_submit(&submit
,
1054 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1056 to_addr_conv(sh
, percpu
));
1057 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
1060 count
= set_syndrome_sources(blocks
, sh
);
1061 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1062 ops_complete_compute
, sh
,
1063 to_addr_conv(sh
, percpu
));
1064 return async_gen_syndrome(blocks
, 0, count
+2,
1065 STRIPE_SIZE
, &submit
);
1068 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1069 ops_complete_compute
, sh
,
1070 to_addr_conv(sh
, percpu
));
1071 if (failb
== syndrome_disks
) {
1072 /* We're missing D+P. */
1073 return async_raid6_datap_recov(syndrome_disks
+2,
1077 /* We're missing D+D. */
1078 return async_raid6_2data_recov(syndrome_disks
+2,
1079 STRIPE_SIZE
, faila
, failb
,
1086 static void ops_complete_prexor(void *stripe_head_ref
)
1088 struct stripe_head
*sh
= stripe_head_ref
;
1090 pr_debug("%s: stripe %llu\n", __func__
,
1091 (unsigned long long)sh
->sector
);
1094 static struct dma_async_tx_descriptor
*
1095 ops_run_prexor(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1096 struct dma_async_tx_descriptor
*tx
)
1098 int disks
= sh
->disks
;
1099 struct page
**xor_srcs
= percpu
->scribble
;
1100 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1101 struct async_submit_ctl submit
;
1103 /* existing parity data subtracted */
1104 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1106 pr_debug("%s: stripe %llu\n", __func__
,
1107 (unsigned long long)sh
->sector
);
1109 for (i
= disks
; i
--; ) {
1110 struct r5dev
*dev
= &sh
->dev
[i
];
1111 /* Only process blocks that are known to be uptodate */
1112 if (test_bit(R5_Wantdrain
, &dev
->flags
))
1113 xor_srcs
[count
++] = dev
->page
;
1116 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1117 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
));
1118 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1123 static struct dma_async_tx_descriptor
*
1124 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1126 int disks
= sh
->disks
;
1129 pr_debug("%s: stripe %llu\n", __func__
,
1130 (unsigned long long)sh
->sector
);
1132 for (i
= disks
; i
--; ) {
1133 struct r5dev
*dev
= &sh
->dev
[i
];
1136 if (test_and_clear_bit(R5_Wantdrain
, &dev
->flags
)) {
1139 spin_lock_irq(&sh
->raid_conf
->device_lock
);
1140 chosen
= dev
->towrite
;
1141 dev
->towrite
= NULL
;
1142 BUG_ON(dev
->written
);
1143 wbi
= dev
->written
= chosen
;
1144 spin_unlock_irq(&sh
->raid_conf
->device_lock
);
1146 while (wbi
&& wbi
->bi_sector
<
1147 dev
->sector
+ STRIPE_SECTORS
) {
1148 if (wbi
->bi_rw
& REQ_FUA
)
1149 set_bit(R5_WantFUA
, &dev
->flags
);
1150 if (wbi
->bi_rw
& REQ_SYNC
)
1151 set_bit(R5_SyncIO
, &dev
->flags
);
1152 tx
= async_copy_data(1, wbi
, dev
->page
,
1154 wbi
= r5_next_bio(wbi
, dev
->sector
);
1162 static void ops_complete_reconstruct(void *stripe_head_ref
)
1164 struct stripe_head
*sh
= stripe_head_ref
;
1165 int disks
= sh
->disks
;
1166 int pd_idx
= sh
->pd_idx
;
1167 int qd_idx
= sh
->qd_idx
;
1169 bool fua
= false, sync
= false;
1171 pr_debug("%s: stripe %llu\n", __func__
,
1172 (unsigned long long)sh
->sector
);
1174 for (i
= disks
; i
--; ) {
1175 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1176 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
1179 for (i
= disks
; i
--; ) {
1180 struct r5dev
*dev
= &sh
->dev
[i
];
1182 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1183 set_bit(R5_UPTODATE
, &dev
->flags
);
1185 set_bit(R5_WantFUA
, &dev
->flags
);
1187 set_bit(R5_SyncIO
, &dev
->flags
);
1191 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1192 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1193 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1194 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1196 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1197 sh
->reconstruct_state
= reconstruct_state_result
;
1200 set_bit(STRIPE_HANDLE
, &sh
->state
);
1205 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1206 struct dma_async_tx_descriptor
*tx
)
1208 int disks
= sh
->disks
;
1209 struct page
**xor_srcs
= percpu
->scribble
;
1210 struct async_submit_ctl submit
;
1211 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1212 struct page
*xor_dest
;
1214 unsigned long flags
;
1216 pr_debug("%s: stripe %llu\n", __func__
,
1217 (unsigned long long)sh
->sector
);
1219 /* check if prexor is active which means only process blocks
1220 * that are part of a read-modify-write (written)
1222 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1224 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1225 for (i
= disks
; i
--; ) {
1226 struct r5dev
*dev
= &sh
->dev
[i
];
1228 xor_srcs
[count
++] = dev
->page
;
1231 xor_dest
= sh
->dev
[pd_idx
].page
;
1232 for (i
= disks
; i
--; ) {
1233 struct r5dev
*dev
= &sh
->dev
[i
];
1235 xor_srcs
[count
++] = dev
->page
;
1239 /* 1/ if we prexor'd then the dest is reused as a source
1240 * 2/ if we did not prexor then we are redoing the parity
1241 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1242 * for the synchronous xor case
1244 flags
= ASYNC_TX_ACK
|
1245 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1247 atomic_inc(&sh
->count
);
1249 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, sh
,
1250 to_addr_conv(sh
, percpu
));
1251 if (unlikely(count
== 1))
1252 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1254 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1258 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1259 struct dma_async_tx_descriptor
*tx
)
1261 struct async_submit_ctl submit
;
1262 struct page
**blocks
= percpu
->scribble
;
1265 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1267 count
= set_syndrome_sources(blocks
, sh
);
1269 atomic_inc(&sh
->count
);
1271 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_reconstruct
,
1272 sh
, to_addr_conv(sh
, percpu
));
1273 async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1276 static void ops_complete_check(void *stripe_head_ref
)
1278 struct stripe_head
*sh
= stripe_head_ref
;
1280 pr_debug("%s: stripe %llu\n", __func__
,
1281 (unsigned long long)sh
->sector
);
1283 sh
->check_state
= check_state_check_result
;
1284 set_bit(STRIPE_HANDLE
, &sh
->state
);
1288 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1290 int disks
= sh
->disks
;
1291 int pd_idx
= sh
->pd_idx
;
1292 int qd_idx
= sh
->qd_idx
;
1293 struct page
*xor_dest
;
1294 struct page
**xor_srcs
= percpu
->scribble
;
1295 struct dma_async_tx_descriptor
*tx
;
1296 struct async_submit_ctl submit
;
1300 pr_debug("%s: stripe %llu\n", __func__
,
1301 (unsigned long long)sh
->sector
);
1304 xor_dest
= sh
->dev
[pd_idx
].page
;
1305 xor_srcs
[count
++] = xor_dest
;
1306 for (i
= disks
; i
--; ) {
1307 if (i
== pd_idx
|| i
== qd_idx
)
1309 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1312 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1313 to_addr_conv(sh
, percpu
));
1314 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1315 &sh
->ops
.zero_sum_result
, &submit
);
1317 atomic_inc(&sh
->count
);
1318 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1319 tx
= async_trigger_callback(&submit
);
1322 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1324 struct page
**srcs
= percpu
->scribble
;
1325 struct async_submit_ctl submit
;
1328 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1329 (unsigned long long)sh
->sector
, checkp
);
1331 count
= set_syndrome_sources(srcs
, sh
);
1335 atomic_inc(&sh
->count
);
1336 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1337 sh
, to_addr_conv(sh
, percpu
));
1338 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1339 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1342 static void __raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1344 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1345 struct dma_async_tx_descriptor
*tx
= NULL
;
1346 struct r5conf
*conf
= sh
->raid_conf
;
1347 int level
= conf
->level
;
1348 struct raid5_percpu
*percpu
;
1352 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1353 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1354 ops_run_biofill(sh
);
1358 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1360 tx
= ops_run_compute5(sh
, percpu
);
1362 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
1363 tx
= ops_run_compute6_1(sh
, percpu
);
1365 tx
= ops_run_compute6_2(sh
, percpu
);
1367 /* terminate the chain if reconstruct is not set to be run */
1368 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
1372 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
))
1373 tx
= ops_run_prexor(sh
, percpu
, tx
);
1375 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
1376 tx
= ops_run_biodrain(sh
, tx
);
1380 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
1382 ops_run_reconstruct5(sh
, percpu
, tx
);
1384 ops_run_reconstruct6(sh
, percpu
, tx
);
1387 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
1388 if (sh
->check_state
== check_state_run
)
1389 ops_run_check_p(sh
, percpu
);
1390 else if (sh
->check_state
== check_state_run_q
)
1391 ops_run_check_pq(sh
, percpu
, 0);
1392 else if (sh
->check_state
== check_state_run_pq
)
1393 ops_run_check_pq(sh
, percpu
, 1);
1399 for (i
= disks
; i
--; ) {
1400 struct r5dev
*dev
= &sh
->dev
[i
];
1401 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
1402 wake_up(&sh
->raid_conf
->wait_for_overlap
);
1407 #ifdef CONFIG_MULTICORE_RAID456
1408 static void async_run_ops(void *param
, async_cookie_t cookie
)
1410 struct stripe_head
*sh
= param
;
1411 unsigned long ops_request
= sh
->ops
.request
;
1413 clear_bit_unlock(STRIPE_OPS_REQ_PENDING
, &sh
->state
);
1414 wake_up(&sh
->ops
.wait_for_ops
);
1416 __raid_run_ops(sh
, ops_request
);
1420 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1422 /* since handle_stripe can be called outside of raid5d context
1423 * we need to ensure sh->ops.request is de-staged before another
1426 wait_event(sh
->ops
.wait_for_ops
,
1427 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING
, &sh
->state
));
1428 sh
->ops
.request
= ops_request
;
1430 atomic_inc(&sh
->count
);
1431 async_schedule(async_run_ops
, sh
);
1434 #define raid_run_ops __raid_run_ops
1437 static int grow_one_stripe(struct r5conf
*conf
)
1439 struct stripe_head
*sh
;
1440 sh
= kmem_cache_zalloc(conf
->slab_cache
, GFP_KERNEL
);
1444 sh
->raid_conf
= conf
;
1445 #ifdef CONFIG_MULTICORE_RAID456
1446 init_waitqueue_head(&sh
->ops
.wait_for_ops
);
1449 if (grow_buffers(sh
)) {
1451 kmem_cache_free(conf
->slab_cache
, sh
);
1454 /* we just created an active stripe so... */
1455 atomic_set(&sh
->count
, 1);
1456 atomic_inc(&conf
->active_stripes
);
1457 INIT_LIST_HEAD(&sh
->lru
);
1462 static int grow_stripes(struct r5conf
*conf
, int num
)
1464 struct kmem_cache
*sc
;
1465 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
1467 if (conf
->mddev
->gendisk
)
1468 sprintf(conf
->cache_name
[0],
1469 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
1471 sprintf(conf
->cache_name
[0],
1472 "raid%d-%p", conf
->level
, conf
->mddev
);
1473 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
1475 conf
->active_name
= 0;
1476 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
1477 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
1481 conf
->slab_cache
= sc
;
1482 conf
->pool_size
= devs
;
1484 if (!grow_one_stripe(conf
))
1490 * scribble_len - return the required size of the scribble region
1491 * @num - total number of disks in the array
1493 * The size must be enough to contain:
1494 * 1/ a struct page pointer for each device in the array +2
1495 * 2/ room to convert each entry in (1) to its corresponding dma
1496 * (dma_map_page()) or page (page_address()) address.
1498 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1499 * calculate over all devices (not just the data blocks), using zeros in place
1500 * of the P and Q blocks.
1502 static size_t scribble_len(int num
)
1506 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
1511 static int resize_stripes(struct r5conf
*conf
, int newsize
)
1513 /* Make all the stripes able to hold 'newsize' devices.
1514 * New slots in each stripe get 'page' set to a new page.
1516 * This happens in stages:
1517 * 1/ create a new kmem_cache and allocate the required number of
1519 * 2/ gather all the old stripe_heads and tranfer the pages across
1520 * to the new stripe_heads. This will have the side effect of
1521 * freezing the array as once all stripe_heads have been collected,
1522 * no IO will be possible. Old stripe heads are freed once their
1523 * pages have been transferred over, and the old kmem_cache is
1524 * freed when all stripes are done.
1525 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1526 * we simple return a failre status - no need to clean anything up.
1527 * 4/ allocate new pages for the new slots in the new stripe_heads.
1528 * If this fails, we don't bother trying the shrink the
1529 * stripe_heads down again, we just leave them as they are.
1530 * As each stripe_head is processed the new one is released into
1533 * Once step2 is started, we cannot afford to wait for a write,
1534 * so we use GFP_NOIO allocations.
1536 struct stripe_head
*osh
, *nsh
;
1537 LIST_HEAD(newstripes
);
1538 struct disk_info
*ndisks
;
1541 struct kmem_cache
*sc
;
1544 if (newsize
<= conf
->pool_size
)
1545 return 0; /* never bother to shrink */
1547 err
= md_allow_write(conf
->mddev
);
1552 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
1553 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
1558 for (i
= conf
->max_nr_stripes
; i
; i
--) {
1559 nsh
= kmem_cache_zalloc(sc
, GFP_KERNEL
);
1563 nsh
->raid_conf
= conf
;
1564 #ifdef CONFIG_MULTICORE_RAID456
1565 init_waitqueue_head(&nsh
->ops
.wait_for_ops
);
1568 list_add(&nsh
->lru
, &newstripes
);
1571 /* didn't get enough, give up */
1572 while (!list_empty(&newstripes
)) {
1573 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1574 list_del(&nsh
->lru
);
1575 kmem_cache_free(sc
, nsh
);
1577 kmem_cache_destroy(sc
);
1580 /* Step 2 - Must use GFP_NOIO now.
1581 * OK, we have enough stripes, start collecting inactive
1582 * stripes and copying them over
1584 list_for_each_entry(nsh
, &newstripes
, lru
) {
1585 spin_lock_irq(&conf
->device_lock
);
1586 wait_event_lock_irq(conf
->wait_for_stripe
,
1587 !list_empty(&conf
->inactive_list
),
1590 osh
= get_free_stripe(conf
);
1591 spin_unlock_irq(&conf
->device_lock
);
1592 atomic_set(&nsh
->count
, 1);
1593 for(i
=0; i
<conf
->pool_size
; i
++)
1594 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
1595 for( ; i
<newsize
; i
++)
1596 nsh
->dev
[i
].page
= NULL
;
1597 kmem_cache_free(conf
->slab_cache
, osh
);
1599 kmem_cache_destroy(conf
->slab_cache
);
1602 * At this point, we are holding all the stripes so the array
1603 * is completely stalled, so now is a good time to resize
1604 * conf->disks and the scribble region
1606 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
1608 for (i
=0; i
<conf
->raid_disks
; i
++)
1609 ndisks
[i
] = conf
->disks
[i
];
1611 conf
->disks
= ndisks
;
1616 conf
->scribble_len
= scribble_len(newsize
);
1617 for_each_present_cpu(cpu
) {
1618 struct raid5_percpu
*percpu
;
1621 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1622 scribble
= kmalloc(conf
->scribble_len
, GFP_NOIO
);
1625 kfree(percpu
->scribble
);
1626 percpu
->scribble
= scribble
;
1634 /* Step 4, return new stripes to service */
1635 while(!list_empty(&newstripes
)) {
1636 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1637 list_del_init(&nsh
->lru
);
1639 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
1640 if (nsh
->dev
[i
].page
== NULL
) {
1641 struct page
*p
= alloc_page(GFP_NOIO
);
1642 nsh
->dev
[i
].page
= p
;
1646 release_stripe(nsh
);
1648 /* critical section pass, GFP_NOIO no longer needed */
1650 conf
->slab_cache
= sc
;
1651 conf
->active_name
= 1-conf
->active_name
;
1652 conf
->pool_size
= newsize
;
1656 static int drop_one_stripe(struct r5conf
*conf
)
1658 struct stripe_head
*sh
;
1660 spin_lock_irq(&conf
->device_lock
);
1661 sh
= get_free_stripe(conf
);
1662 spin_unlock_irq(&conf
->device_lock
);
1665 BUG_ON(atomic_read(&sh
->count
));
1667 kmem_cache_free(conf
->slab_cache
, sh
);
1668 atomic_dec(&conf
->active_stripes
);
1672 static void shrink_stripes(struct r5conf
*conf
)
1674 while (drop_one_stripe(conf
))
1677 if (conf
->slab_cache
)
1678 kmem_cache_destroy(conf
->slab_cache
);
1679 conf
->slab_cache
= NULL
;
1682 static void raid5_end_read_request(struct bio
* bi
, int error
)
1684 struct stripe_head
*sh
= bi
->bi_private
;
1685 struct r5conf
*conf
= sh
->raid_conf
;
1686 int disks
= sh
->disks
, i
;
1687 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1688 char b
[BDEVNAME_SIZE
];
1689 struct md_rdev
*rdev
= NULL
;
1692 for (i
=0 ; i
<disks
; i
++)
1693 if (bi
== &sh
->dev
[i
].req
)
1696 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1697 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1703 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
1704 /* If replacement finished while this request was outstanding,
1705 * 'replacement' might be NULL already.
1706 * In that case it moved down to 'rdev'.
1707 * rdev is not removed until all requests are finished.
1709 rdev
= conf
->disks
[i
].replacement
;
1711 rdev
= conf
->disks
[i
].rdev
;
1713 if (use_new_offset(conf
, sh
))
1714 s
= sh
->sector
+ rdev
->new_data_offset
;
1716 s
= sh
->sector
+ rdev
->data_offset
;
1718 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1719 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
1720 /* Note that this cannot happen on a
1721 * replacement device. We just fail those on
1726 "md/raid:%s: read error corrected"
1727 " (%lu sectors at %llu on %s)\n",
1728 mdname(conf
->mddev
), STRIPE_SECTORS
,
1729 (unsigned long long)s
,
1730 bdevname(rdev
->bdev
, b
));
1731 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
1732 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1733 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1735 if (atomic_read(&rdev
->read_errors
))
1736 atomic_set(&rdev
->read_errors
, 0);
1738 const char *bdn
= bdevname(rdev
->bdev
, b
);
1741 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1742 atomic_inc(&rdev
->read_errors
);
1743 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
1746 "md/raid:%s: read error on replacement device "
1747 "(sector %llu on %s).\n",
1748 mdname(conf
->mddev
),
1749 (unsigned long long)s
,
1751 else if (conf
->mddev
->degraded
>= conf
->max_degraded
)
1754 "md/raid:%s: read error not correctable "
1755 "(sector %llu on %s).\n",
1756 mdname(conf
->mddev
),
1757 (unsigned long long)s
,
1759 else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
))
1763 "md/raid:%s: read error NOT corrected!! "
1764 "(sector %llu on %s).\n",
1765 mdname(conf
->mddev
),
1766 (unsigned long long)s
,
1768 else if (atomic_read(&rdev
->read_errors
)
1769 > conf
->max_nr_stripes
)
1771 "md/raid:%s: Too many read errors, failing device %s.\n",
1772 mdname(conf
->mddev
), bdn
);
1776 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1778 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1779 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1780 md_error(conf
->mddev
, rdev
);
1783 rdev_dec_pending(rdev
, conf
->mddev
);
1784 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1785 set_bit(STRIPE_HANDLE
, &sh
->state
);
1789 static void raid5_end_write_request(struct bio
*bi
, int error
)
1791 struct stripe_head
*sh
= bi
->bi_private
;
1792 struct r5conf
*conf
= sh
->raid_conf
;
1793 int disks
= sh
->disks
, i
;
1794 struct md_rdev
*uninitialized_var(rdev
);
1795 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1798 int replacement
= 0;
1800 for (i
= 0 ; i
< disks
; i
++) {
1801 if (bi
== &sh
->dev
[i
].req
) {
1802 rdev
= conf
->disks
[i
].rdev
;
1805 if (bi
== &sh
->dev
[i
].rreq
) {
1806 rdev
= conf
->disks
[i
].replacement
;
1810 /* rdev was removed and 'replacement'
1811 * replaced it. rdev is not removed
1812 * until all requests are finished.
1814 rdev
= conf
->disks
[i
].rdev
;
1818 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1819 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1828 md_error(conf
->mddev
, rdev
);
1829 else if (is_badblock(rdev
, sh
->sector
,
1831 &first_bad
, &bad_sectors
))
1832 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
1835 set_bit(WriteErrorSeen
, &rdev
->flags
);
1836 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
1837 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1838 set_bit(MD_RECOVERY_NEEDED
,
1839 &rdev
->mddev
->recovery
);
1840 } else if (is_badblock(rdev
, sh
->sector
,
1842 &first_bad
, &bad_sectors
))
1843 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
1845 rdev_dec_pending(rdev
, conf
->mddev
);
1847 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
1848 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1849 set_bit(STRIPE_HANDLE
, &sh
->state
);
1853 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
);
1855 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
1857 struct r5dev
*dev
= &sh
->dev
[i
];
1859 bio_init(&dev
->req
);
1860 dev
->req
.bi_io_vec
= &dev
->vec
;
1862 dev
->req
.bi_max_vecs
++;
1863 dev
->req
.bi_private
= sh
;
1864 dev
->vec
.bv_page
= dev
->page
;
1866 bio_init(&dev
->rreq
);
1867 dev
->rreq
.bi_io_vec
= &dev
->rvec
;
1868 dev
->rreq
.bi_vcnt
++;
1869 dev
->rreq
.bi_max_vecs
++;
1870 dev
->rreq
.bi_private
= sh
;
1871 dev
->rvec
.bv_page
= dev
->page
;
1874 dev
->sector
= compute_blocknr(sh
, i
, previous
);
1877 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1879 char b
[BDEVNAME_SIZE
];
1880 struct r5conf
*conf
= mddev
->private;
1881 unsigned long flags
;
1882 pr_debug("raid456: error called\n");
1884 spin_lock_irqsave(&conf
->device_lock
, flags
);
1885 clear_bit(In_sync
, &rdev
->flags
);
1886 mddev
->degraded
= calc_degraded(conf
);
1887 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1888 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1890 set_bit(Blocked
, &rdev
->flags
);
1891 set_bit(Faulty
, &rdev
->flags
);
1892 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1894 "md/raid:%s: Disk failure on %s, disabling device.\n"
1895 "md/raid:%s: Operation continuing on %d devices.\n",
1897 bdevname(rdev
->bdev
, b
),
1899 conf
->raid_disks
- mddev
->degraded
);
1903 * Input: a 'big' sector number,
1904 * Output: index of the data and parity disk, and the sector # in them.
1906 static sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
1907 int previous
, int *dd_idx
,
1908 struct stripe_head
*sh
)
1910 sector_t stripe
, stripe2
;
1911 sector_t chunk_number
;
1912 unsigned int chunk_offset
;
1915 sector_t new_sector
;
1916 int algorithm
= previous
? conf
->prev_algo
1918 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
1919 : conf
->chunk_sectors
;
1920 int raid_disks
= previous
? conf
->previous_raid_disks
1922 int data_disks
= raid_disks
- conf
->max_degraded
;
1924 /* First compute the information on this sector */
1927 * Compute the chunk number and the sector offset inside the chunk
1929 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
1930 chunk_number
= r_sector
;
1933 * Compute the stripe number
1935 stripe
= chunk_number
;
1936 *dd_idx
= sector_div(stripe
, data_disks
);
1939 * Select the parity disk based on the user selected algorithm.
1941 pd_idx
= qd_idx
= -1;
1942 switch(conf
->level
) {
1944 pd_idx
= data_disks
;
1947 switch (algorithm
) {
1948 case ALGORITHM_LEFT_ASYMMETRIC
:
1949 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
1950 if (*dd_idx
>= pd_idx
)
1953 case ALGORITHM_RIGHT_ASYMMETRIC
:
1954 pd_idx
= sector_div(stripe2
, raid_disks
);
1955 if (*dd_idx
>= pd_idx
)
1958 case ALGORITHM_LEFT_SYMMETRIC
:
1959 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
1960 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1962 case ALGORITHM_RIGHT_SYMMETRIC
:
1963 pd_idx
= sector_div(stripe2
, raid_disks
);
1964 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1966 case ALGORITHM_PARITY_0
:
1970 case ALGORITHM_PARITY_N
:
1971 pd_idx
= data_disks
;
1979 switch (algorithm
) {
1980 case ALGORITHM_LEFT_ASYMMETRIC
:
1981 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1982 qd_idx
= pd_idx
+ 1;
1983 if (pd_idx
== raid_disks
-1) {
1984 (*dd_idx
)++; /* Q D D D P */
1986 } else if (*dd_idx
>= pd_idx
)
1987 (*dd_idx
) += 2; /* D D P Q D */
1989 case ALGORITHM_RIGHT_ASYMMETRIC
:
1990 pd_idx
= sector_div(stripe2
, raid_disks
);
1991 qd_idx
= pd_idx
+ 1;
1992 if (pd_idx
== raid_disks
-1) {
1993 (*dd_idx
)++; /* Q D D D P */
1995 } else if (*dd_idx
>= pd_idx
)
1996 (*dd_idx
) += 2; /* D D P Q D */
1998 case ALGORITHM_LEFT_SYMMETRIC
:
1999 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2000 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2001 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2003 case ALGORITHM_RIGHT_SYMMETRIC
:
2004 pd_idx
= sector_div(stripe2
, raid_disks
);
2005 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2006 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2009 case ALGORITHM_PARITY_0
:
2014 case ALGORITHM_PARITY_N
:
2015 pd_idx
= data_disks
;
2016 qd_idx
= data_disks
+ 1;
2019 case ALGORITHM_ROTATING_ZERO_RESTART
:
2020 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2021 * of blocks for computing Q is different.
2023 pd_idx
= sector_div(stripe2
, raid_disks
);
2024 qd_idx
= pd_idx
+ 1;
2025 if (pd_idx
== raid_disks
-1) {
2026 (*dd_idx
)++; /* Q D D D P */
2028 } else if (*dd_idx
>= pd_idx
)
2029 (*dd_idx
) += 2; /* D D P Q D */
2033 case ALGORITHM_ROTATING_N_RESTART
:
2034 /* Same a left_asymmetric, by first stripe is
2035 * D D D P Q rather than
2039 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2040 qd_idx
= pd_idx
+ 1;
2041 if (pd_idx
== raid_disks
-1) {
2042 (*dd_idx
)++; /* Q D D D P */
2044 } else if (*dd_idx
>= pd_idx
)
2045 (*dd_idx
) += 2; /* D D P Q D */
2049 case ALGORITHM_ROTATING_N_CONTINUE
:
2050 /* Same as left_symmetric but Q is before P */
2051 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2052 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
2053 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2057 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2058 /* RAID5 left_asymmetric, with Q on last device */
2059 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2060 if (*dd_idx
>= pd_idx
)
2062 qd_idx
= raid_disks
- 1;
2065 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2066 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2067 if (*dd_idx
>= pd_idx
)
2069 qd_idx
= raid_disks
- 1;
2072 case ALGORITHM_LEFT_SYMMETRIC_6
:
2073 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2074 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2075 qd_idx
= raid_disks
- 1;
2078 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2079 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2080 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2081 qd_idx
= raid_disks
- 1;
2084 case ALGORITHM_PARITY_0_6
:
2087 qd_idx
= raid_disks
- 1;
2097 sh
->pd_idx
= pd_idx
;
2098 sh
->qd_idx
= qd_idx
;
2099 sh
->ddf_layout
= ddf_layout
;
2102 * Finally, compute the new sector number
2104 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
2109 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
2111 struct r5conf
*conf
= sh
->raid_conf
;
2112 int raid_disks
= sh
->disks
;
2113 int data_disks
= raid_disks
- conf
->max_degraded
;
2114 sector_t new_sector
= sh
->sector
, check
;
2115 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2116 : conf
->chunk_sectors
;
2117 int algorithm
= previous
? conf
->prev_algo
2121 sector_t chunk_number
;
2122 int dummy1
, dd_idx
= i
;
2124 struct stripe_head sh2
;
2127 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
2128 stripe
= new_sector
;
2130 if (i
== sh
->pd_idx
)
2132 switch(conf
->level
) {
2135 switch (algorithm
) {
2136 case ALGORITHM_LEFT_ASYMMETRIC
:
2137 case ALGORITHM_RIGHT_ASYMMETRIC
:
2141 case ALGORITHM_LEFT_SYMMETRIC
:
2142 case ALGORITHM_RIGHT_SYMMETRIC
:
2145 i
-= (sh
->pd_idx
+ 1);
2147 case ALGORITHM_PARITY_0
:
2150 case ALGORITHM_PARITY_N
:
2157 if (i
== sh
->qd_idx
)
2158 return 0; /* It is the Q disk */
2159 switch (algorithm
) {
2160 case ALGORITHM_LEFT_ASYMMETRIC
:
2161 case ALGORITHM_RIGHT_ASYMMETRIC
:
2162 case ALGORITHM_ROTATING_ZERO_RESTART
:
2163 case ALGORITHM_ROTATING_N_RESTART
:
2164 if (sh
->pd_idx
== raid_disks
-1)
2165 i
--; /* Q D D D P */
2166 else if (i
> sh
->pd_idx
)
2167 i
-= 2; /* D D P Q D */
2169 case ALGORITHM_LEFT_SYMMETRIC
:
2170 case ALGORITHM_RIGHT_SYMMETRIC
:
2171 if (sh
->pd_idx
== raid_disks
-1)
2172 i
--; /* Q D D D P */
2177 i
-= (sh
->pd_idx
+ 2);
2180 case ALGORITHM_PARITY_0
:
2183 case ALGORITHM_PARITY_N
:
2185 case ALGORITHM_ROTATING_N_CONTINUE
:
2186 /* Like left_symmetric, but P is before Q */
2187 if (sh
->pd_idx
== 0)
2188 i
--; /* P D D D Q */
2193 i
-= (sh
->pd_idx
+ 1);
2196 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2197 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2201 case ALGORITHM_LEFT_SYMMETRIC_6
:
2202 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2204 i
+= data_disks
+ 1;
2205 i
-= (sh
->pd_idx
+ 1);
2207 case ALGORITHM_PARITY_0_6
:
2216 chunk_number
= stripe
* data_disks
+ i
;
2217 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2219 check
= raid5_compute_sector(conf
, r_sector
,
2220 previous
, &dummy1
, &sh2
);
2221 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2222 || sh2
.qd_idx
!= sh
->qd_idx
) {
2223 printk(KERN_ERR
"md/raid:%s: compute_blocknr: map not correct\n",
2224 mdname(conf
->mddev
));
2232 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2233 int rcw
, int expand
)
2235 int i
, pd_idx
= sh
->pd_idx
, disks
= sh
->disks
;
2236 struct r5conf
*conf
= sh
->raid_conf
;
2237 int level
= conf
->level
;
2240 /* if we are not expanding this is a proper write request, and
2241 * there will be bios with new data to be drained into the
2245 sh
->reconstruct_state
= reconstruct_state_drain_run
;
2246 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2248 sh
->reconstruct_state
= reconstruct_state_run
;
2250 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2252 for (i
= disks
; i
--; ) {
2253 struct r5dev
*dev
= &sh
->dev
[i
];
2256 set_bit(R5_LOCKED
, &dev
->flags
);
2257 set_bit(R5_Wantdrain
, &dev
->flags
);
2259 clear_bit(R5_UPTODATE
, &dev
->flags
);
2263 if (s
->locked
+ conf
->max_degraded
== disks
)
2264 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2265 atomic_inc(&conf
->pending_full_writes
);
2268 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
2269 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
2271 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
2272 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
2273 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2274 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2276 for (i
= disks
; i
--; ) {
2277 struct r5dev
*dev
= &sh
->dev
[i
];
2282 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2283 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2284 set_bit(R5_Wantdrain
, &dev
->flags
);
2285 set_bit(R5_LOCKED
, &dev
->flags
);
2286 clear_bit(R5_UPTODATE
, &dev
->flags
);
2292 /* keep the parity disk(s) locked while asynchronous operations
2295 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
2296 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2300 int qd_idx
= sh
->qd_idx
;
2301 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
2303 set_bit(R5_LOCKED
, &dev
->flags
);
2304 clear_bit(R5_UPTODATE
, &dev
->flags
);
2308 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2309 __func__
, (unsigned long long)sh
->sector
,
2310 s
->locked
, s
->ops_request
);
2314 * Each stripe/dev can have one or more bion attached.
2315 * toread/towrite point to the first in a chain.
2316 * The bi_next chain must be in order.
2318 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
, int forwrite
)
2321 struct r5conf
*conf
= sh
->raid_conf
;
2324 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2325 (unsigned long long)bi
->bi_sector
,
2326 (unsigned long long)sh
->sector
);
2329 spin_lock_irq(&conf
->device_lock
);
2331 bip
= &sh
->dev
[dd_idx
].towrite
;
2332 if (*bip
== NULL
&& sh
->dev
[dd_idx
].written
== NULL
)
2335 bip
= &sh
->dev
[dd_idx
].toread
;
2336 while (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
) {
2337 if ((*bip
)->bi_sector
+ ((*bip
)->bi_size
>> 9) > bi
->bi_sector
)
2339 bip
= & (*bip
)->bi_next
;
2341 if (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
+ ((bi
->bi_size
)>>9))
2344 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
2348 bi
->bi_phys_segments
++;
2351 /* check if page is covered */
2352 sector_t sector
= sh
->dev
[dd_idx
].sector
;
2353 for (bi
=sh
->dev
[dd_idx
].towrite
;
2354 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
2355 bi
&& bi
->bi_sector
<= sector
;
2356 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
2357 if (bi
->bi_sector
+ (bi
->bi_size
>>9) >= sector
)
2358 sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
2360 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
2361 set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
);
2363 spin_unlock_irq(&conf
->device_lock
);
2365 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2366 (unsigned long long)(*bip
)->bi_sector
,
2367 (unsigned long long)sh
->sector
, dd_idx
);
2369 if (conf
->mddev
->bitmap
&& firstwrite
) {
2370 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
2372 sh
->bm_seq
= conf
->seq_flush
+1;
2373 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
2378 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
2379 spin_unlock_irq(&conf
->device_lock
);
2383 static void end_reshape(struct r5conf
*conf
);
2385 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
2386 struct stripe_head
*sh
)
2388 int sectors_per_chunk
=
2389 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
2391 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
2392 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
2394 raid5_compute_sector(conf
,
2395 stripe
* (disks
- conf
->max_degraded
)
2396 *sectors_per_chunk
+ chunk_offset
,
2402 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
2403 struct stripe_head_state
*s
, int disks
,
2404 struct bio
**return_bi
)
2407 for (i
= disks
; i
--; ) {
2411 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2412 struct md_rdev
*rdev
;
2414 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2415 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
2416 atomic_inc(&rdev
->nr_pending
);
2421 if (!rdev_set_badblocks(
2425 md_error(conf
->mddev
, rdev
);
2426 rdev_dec_pending(rdev
, conf
->mddev
);
2429 spin_lock_irq(&conf
->device_lock
);
2430 /* fail all writes first */
2431 bi
= sh
->dev
[i
].towrite
;
2432 sh
->dev
[i
].towrite
= NULL
;
2438 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2439 wake_up(&conf
->wait_for_overlap
);
2441 while (bi
&& bi
->bi_sector
<
2442 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2443 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2444 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2445 if (!raid5_dec_bi_phys_segments(bi
)) {
2446 md_write_end(conf
->mddev
);
2447 bi
->bi_next
= *return_bi
;
2452 /* and fail all 'written' */
2453 bi
= sh
->dev
[i
].written
;
2454 sh
->dev
[i
].written
= NULL
;
2455 if (bi
) bitmap_end
= 1;
2456 while (bi
&& bi
->bi_sector
<
2457 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2458 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2459 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2460 if (!raid5_dec_bi_phys_segments(bi
)) {
2461 md_write_end(conf
->mddev
);
2462 bi
->bi_next
= *return_bi
;
2468 /* fail any reads if this device is non-operational and
2469 * the data has not reached the cache yet.
2471 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
2472 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
2473 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
2474 bi
= sh
->dev
[i
].toread
;
2475 sh
->dev
[i
].toread
= NULL
;
2476 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2477 wake_up(&conf
->wait_for_overlap
);
2478 if (bi
) s
->to_read
--;
2479 while (bi
&& bi
->bi_sector
<
2480 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2481 struct bio
*nextbi
=
2482 r5_next_bio(bi
, sh
->dev
[i
].sector
);
2483 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2484 if (!raid5_dec_bi_phys_segments(bi
)) {
2485 bi
->bi_next
= *return_bi
;
2491 spin_unlock_irq(&conf
->device_lock
);
2493 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2494 STRIPE_SECTORS
, 0, 0);
2495 /* If we were in the middle of a write the parity block might
2496 * still be locked - so just clear all R5_LOCKED flags
2498 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2501 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2502 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2503 md_wakeup_thread(conf
->mddev
->thread
);
2507 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
2508 struct stripe_head_state
*s
)
2513 clear_bit(STRIPE_SYNCING
, &sh
->state
);
2516 /* There is nothing more to do for sync/check/repair.
2517 * Don't even need to abort as that is handled elsewhere
2518 * if needed, and not always wanted e.g. if there is a known
2520 * For recover/replace we need to record a bad block on all
2521 * non-sync devices, or abort the recovery
2523 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
2524 /* During recovery devices cannot be removed, so
2525 * locking and refcounting of rdevs is not needed
2527 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2528 struct md_rdev
*rdev
= conf
->disks
[i
].rdev
;
2530 && !test_bit(Faulty
, &rdev
->flags
)
2531 && !test_bit(In_sync
, &rdev
->flags
)
2532 && !rdev_set_badblocks(rdev
, sh
->sector
,
2535 rdev
= conf
->disks
[i
].replacement
;
2537 && !test_bit(Faulty
, &rdev
->flags
)
2538 && !test_bit(In_sync
, &rdev
->flags
)
2539 && !rdev_set_badblocks(rdev
, sh
->sector
,
2544 conf
->recovery_disabled
=
2545 conf
->mddev
->recovery_disabled
;
2547 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, !abort
);
2550 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
2552 struct md_rdev
*rdev
;
2554 /* Doing recovery so rcu locking not required */
2555 rdev
= sh
->raid_conf
->disks
[disk_idx
].replacement
;
2557 && !test_bit(Faulty
, &rdev
->flags
)
2558 && !test_bit(In_sync
, &rdev
->flags
)
2559 && (rdev
->recovery_offset
<= sh
->sector
2560 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
2566 /* fetch_block - checks the given member device to see if its data needs
2567 * to be read or computed to satisfy a request.
2569 * Returns 1 when no more member devices need to be checked, otherwise returns
2570 * 0 to tell the loop in handle_stripe_fill to continue
2572 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2573 int disk_idx
, int disks
)
2575 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
2576 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
2577 &sh
->dev
[s
->failed_num
[1]] };
2579 /* is the data in this block needed, and can we get it? */
2580 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2581 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2583 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2584 s
->syncing
|| s
->expanding
||
2585 (s
->replacing
&& want_replace(sh
, disk_idx
)) ||
2586 (s
->failed
>= 1 && fdev
[0]->toread
) ||
2587 (s
->failed
>= 2 && fdev
[1]->toread
) ||
2588 (sh
->raid_conf
->level
<= 5 && s
->failed
&& fdev
[0]->towrite
&&
2589 !test_bit(R5_OVERWRITE
, &fdev
[0]->flags
)) ||
2590 (sh
->raid_conf
->level
== 6 && s
->failed
&& s
->to_write
))) {
2591 /* we would like to get this block, possibly by computing it,
2592 * otherwise read it if the backing disk is insync
2594 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
2595 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
2596 if ((s
->uptodate
== disks
- 1) &&
2597 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
2598 disk_idx
== s
->failed_num
[1]))) {
2599 /* have disk failed, and we're requested to fetch it;
2602 pr_debug("Computing stripe %llu block %d\n",
2603 (unsigned long long)sh
->sector
, disk_idx
);
2604 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2605 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2606 set_bit(R5_Wantcompute
, &dev
->flags
);
2607 sh
->ops
.target
= disk_idx
;
2608 sh
->ops
.target2
= -1; /* no 2nd target */
2610 /* Careful: from this point on 'uptodate' is in the eye
2611 * of raid_run_ops which services 'compute' operations
2612 * before writes. R5_Wantcompute flags a block that will
2613 * be R5_UPTODATE by the time it is needed for a
2614 * subsequent operation.
2618 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
2619 /* Computing 2-failure is *very* expensive; only
2620 * do it if failed >= 2
2623 for (other
= disks
; other
--; ) {
2624 if (other
== disk_idx
)
2626 if (!test_bit(R5_UPTODATE
,
2627 &sh
->dev
[other
].flags
))
2631 pr_debug("Computing stripe %llu blocks %d,%d\n",
2632 (unsigned long long)sh
->sector
,
2634 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2635 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2636 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
2637 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
2638 sh
->ops
.target
= disk_idx
;
2639 sh
->ops
.target2
= other
;
2643 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2644 set_bit(R5_LOCKED
, &dev
->flags
);
2645 set_bit(R5_Wantread
, &dev
->flags
);
2647 pr_debug("Reading block %d (sync=%d)\n",
2648 disk_idx
, s
->syncing
);
2656 * handle_stripe_fill - read or compute data to satisfy pending requests.
2658 static void handle_stripe_fill(struct stripe_head
*sh
,
2659 struct stripe_head_state
*s
,
2664 /* look for blocks to read/compute, skip this if a compute
2665 * is already in flight, or if the stripe contents are in the
2666 * midst of changing due to a write
2668 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2669 !sh
->reconstruct_state
)
2670 for (i
= disks
; i
--; )
2671 if (fetch_block(sh
, s
, i
, disks
))
2673 set_bit(STRIPE_HANDLE
, &sh
->state
);
2677 /* handle_stripe_clean_event
2678 * any written block on an uptodate or failed drive can be returned.
2679 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2680 * never LOCKED, so we don't need to test 'failed' directly.
2682 static void handle_stripe_clean_event(struct r5conf
*conf
,
2683 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
2688 for (i
= disks
; i
--; )
2689 if (sh
->dev
[i
].written
) {
2691 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2692 test_bit(R5_UPTODATE
, &dev
->flags
)) {
2693 /* We can return any write requests */
2694 struct bio
*wbi
, *wbi2
;
2696 pr_debug("Return write for disc %d\n", i
);
2697 spin_lock_irq(&conf
->device_lock
);
2699 dev
->written
= NULL
;
2700 while (wbi
&& wbi
->bi_sector
<
2701 dev
->sector
+ STRIPE_SECTORS
) {
2702 wbi2
= r5_next_bio(wbi
, dev
->sector
);
2703 if (!raid5_dec_bi_phys_segments(wbi
)) {
2704 md_write_end(conf
->mddev
);
2705 wbi
->bi_next
= *return_bi
;
2710 if (dev
->towrite
== NULL
)
2712 spin_unlock_irq(&conf
->device_lock
);
2714 bitmap_endwrite(conf
->mddev
->bitmap
,
2717 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
2722 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2723 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2724 md_wakeup_thread(conf
->mddev
->thread
);
2727 static void handle_stripe_dirtying(struct r5conf
*conf
,
2728 struct stripe_head
*sh
,
2729 struct stripe_head_state
*s
,
2732 int rmw
= 0, rcw
= 0, i
;
2733 if (conf
->max_degraded
== 2) {
2734 /* RAID6 requires 'rcw' in current implementation
2735 * Calculate the real rcw later - for now fake it
2736 * look like rcw is cheaper
2739 } else for (i
= disks
; i
--; ) {
2740 /* would I have to read this buffer for read_modify_write */
2741 struct r5dev
*dev
= &sh
->dev
[i
];
2742 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2743 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2744 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2745 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2746 if (test_bit(R5_Insync
, &dev
->flags
))
2749 rmw
+= 2*disks
; /* cannot read it */
2751 /* Would I have to read this buffer for reconstruct_write */
2752 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) && i
!= sh
->pd_idx
&&
2753 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2754 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2755 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2756 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2761 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2762 (unsigned long long)sh
->sector
, rmw
, rcw
);
2763 set_bit(STRIPE_HANDLE
, &sh
->state
);
2764 if (rmw
< rcw
&& rmw
> 0)
2765 /* prefer read-modify-write, but need to get some data */
2766 for (i
= disks
; i
--; ) {
2767 struct r5dev
*dev
= &sh
->dev
[i
];
2768 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2769 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2770 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2771 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2772 test_bit(R5_Insync
, &dev
->flags
)) {
2774 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2775 pr_debug("Read_old block "
2776 "%d for r-m-w\n", i
);
2777 set_bit(R5_LOCKED
, &dev
->flags
);
2778 set_bit(R5_Wantread
, &dev
->flags
);
2781 set_bit(STRIPE_DELAYED
, &sh
->state
);
2782 set_bit(STRIPE_HANDLE
, &sh
->state
);
2786 if (rcw
<= rmw
&& rcw
> 0) {
2787 /* want reconstruct write, but need to get some data */
2789 for (i
= disks
; i
--; ) {
2790 struct r5dev
*dev
= &sh
->dev
[i
];
2791 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2792 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
2793 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2794 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2795 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2797 if (!test_bit(R5_Insync
, &dev
->flags
))
2798 continue; /* it's a failed drive */
2800 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2801 pr_debug("Read_old block "
2802 "%d for Reconstruct\n", i
);
2803 set_bit(R5_LOCKED
, &dev
->flags
);
2804 set_bit(R5_Wantread
, &dev
->flags
);
2807 set_bit(STRIPE_DELAYED
, &sh
->state
);
2808 set_bit(STRIPE_HANDLE
, &sh
->state
);
2813 /* now if nothing is locked, and if we have enough data,
2814 * we can start a write request
2816 /* since handle_stripe can be called at any time we need to handle the
2817 * case where a compute block operation has been submitted and then a
2818 * subsequent call wants to start a write request. raid_run_ops only
2819 * handles the case where compute block and reconstruct are requested
2820 * simultaneously. If this is not the case then new writes need to be
2821 * held off until the compute completes.
2823 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
2824 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
2825 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
2826 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
2829 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
2830 struct stripe_head_state
*s
, int disks
)
2832 struct r5dev
*dev
= NULL
;
2834 set_bit(STRIPE_HANDLE
, &sh
->state
);
2836 switch (sh
->check_state
) {
2837 case check_state_idle
:
2838 /* start a new check operation if there are no failures */
2839 if (s
->failed
== 0) {
2840 BUG_ON(s
->uptodate
!= disks
);
2841 sh
->check_state
= check_state_run
;
2842 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2843 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2847 dev
= &sh
->dev
[s
->failed_num
[0]];
2849 case check_state_compute_result
:
2850 sh
->check_state
= check_state_idle
;
2852 dev
= &sh
->dev
[sh
->pd_idx
];
2854 /* check that a write has not made the stripe insync */
2855 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2858 /* either failed parity check, or recovery is happening */
2859 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
2860 BUG_ON(s
->uptodate
!= disks
);
2862 set_bit(R5_LOCKED
, &dev
->flags
);
2864 set_bit(R5_Wantwrite
, &dev
->flags
);
2866 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2867 set_bit(STRIPE_INSYNC
, &sh
->state
);
2869 case check_state_run
:
2870 break; /* we will be called again upon completion */
2871 case check_state_check_result
:
2872 sh
->check_state
= check_state_idle
;
2874 /* if a failure occurred during the check operation, leave
2875 * STRIPE_INSYNC not set and let the stripe be handled again
2880 /* handle a successful check operation, if parity is correct
2881 * we are done. Otherwise update the mismatch count and repair
2882 * parity if !MD_RECOVERY_CHECK
2884 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
2885 /* parity is correct (on disc,
2886 * not in buffer any more)
2888 set_bit(STRIPE_INSYNC
, &sh
->state
);
2890 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2891 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2892 /* don't try to repair!! */
2893 set_bit(STRIPE_INSYNC
, &sh
->state
);
2895 sh
->check_state
= check_state_compute_run
;
2896 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2897 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2898 set_bit(R5_Wantcompute
,
2899 &sh
->dev
[sh
->pd_idx
].flags
);
2900 sh
->ops
.target
= sh
->pd_idx
;
2901 sh
->ops
.target2
= -1;
2906 case check_state_compute_run
:
2909 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
2910 __func__
, sh
->check_state
,
2911 (unsigned long long) sh
->sector
);
2917 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
2918 struct stripe_head_state
*s
,
2921 int pd_idx
= sh
->pd_idx
;
2922 int qd_idx
= sh
->qd_idx
;
2925 set_bit(STRIPE_HANDLE
, &sh
->state
);
2927 BUG_ON(s
->failed
> 2);
2929 /* Want to check and possibly repair P and Q.
2930 * However there could be one 'failed' device, in which
2931 * case we can only check one of them, possibly using the
2932 * other to generate missing data
2935 switch (sh
->check_state
) {
2936 case check_state_idle
:
2937 /* start a new check operation if there are < 2 failures */
2938 if (s
->failed
== s
->q_failed
) {
2939 /* The only possible failed device holds Q, so it
2940 * makes sense to check P (If anything else were failed,
2941 * we would have used P to recreate it).
2943 sh
->check_state
= check_state_run
;
2945 if (!s
->q_failed
&& s
->failed
< 2) {
2946 /* Q is not failed, and we didn't use it to generate
2947 * anything, so it makes sense to check it
2949 if (sh
->check_state
== check_state_run
)
2950 sh
->check_state
= check_state_run_pq
;
2952 sh
->check_state
= check_state_run_q
;
2955 /* discard potentially stale zero_sum_result */
2956 sh
->ops
.zero_sum_result
= 0;
2958 if (sh
->check_state
== check_state_run
) {
2959 /* async_xor_zero_sum destroys the contents of P */
2960 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2963 if (sh
->check_state
>= check_state_run
&&
2964 sh
->check_state
<= check_state_run_pq
) {
2965 /* async_syndrome_zero_sum preserves P and Q, so
2966 * no need to mark them !uptodate here
2968 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2972 /* we have 2-disk failure */
2973 BUG_ON(s
->failed
!= 2);
2975 case check_state_compute_result
:
2976 sh
->check_state
= check_state_idle
;
2978 /* check that a write has not made the stripe insync */
2979 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2982 /* now write out any block on a failed drive,
2983 * or P or Q if they were recomputed
2985 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
2986 if (s
->failed
== 2) {
2987 dev
= &sh
->dev
[s
->failed_num
[1]];
2989 set_bit(R5_LOCKED
, &dev
->flags
);
2990 set_bit(R5_Wantwrite
, &dev
->flags
);
2992 if (s
->failed
>= 1) {
2993 dev
= &sh
->dev
[s
->failed_num
[0]];
2995 set_bit(R5_LOCKED
, &dev
->flags
);
2996 set_bit(R5_Wantwrite
, &dev
->flags
);
2998 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
2999 dev
= &sh
->dev
[pd_idx
];
3001 set_bit(R5_LOCKED
, &dev
->flags
);
3002 set_bit(R5_Wantwrite
, &dev
->flags
);
3004 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3005 dev
= &sh
->dev
[qd_idx
];
3007 set_bit(R5_LOCKED
, &dev
->flags
);
3008 set_bit(R5_Wantwrite
, &dev
->flags
);
3010 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3012 set_bit(STRIPE_INSYNC
, &sh
->state
);
3014 case check_state_run
:
3015 case check_state_run_q
:
3016 case check_state_run_pq
:
3017 break; /* we will be called again upon completion */
3018 case check_state_check_result
:
3019 sh
->check_state
= check_state_idle
;
3021 /* handle a successful check operation, if parity is correct
3022 * we are done. Otherwise update the mismatch count and repair
3023 * parity if !MD_RECOVERY_CHECK
3025 if (sh
->ops
.zero_sum_result
== 0) {
3026 /* both parities are correct */
3028 set_bit(STRIPE_INSYNC
, &sh
->state
);
3030 /* in contrast to the raid5 case we can validate
3031 * parity, but still have a failure to write
3034 sh
->check_state
= check_state_compute_result
;
3035 /* Returning at this point means that we may go
3036 * off and bring p and/or q uptodate again so
3037 * we make sure to check zero_sum_result again
3038 * to verify if p or q need writeback
3042 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
3043 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3044 /* don't try to repair!! */
3045 set_bit(STRIPE_INSYNC
, &sh
->state
);
3047 int *target
= &sh
->ops
.target
;
3049 sh
->ops
.target
= -1;
3050 sh
->ops
.target2
= -1;
3051 sh
->check_state
= check_state_compute_run
;
3052 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3053 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3054 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3055 set_bit(R5_Wantcompute
,
3056 &sh
->dev
[pd_idx
].flags
);
3058 target
= &sh
->ops
.target2
;
3061 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3062 set_bit(R5_Wantcompute
,
3063 &sh
->dev
[qd_idx
].flags
);
3070 case check_state_compute_run
:
3073 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3074 __func__
, sh
->check_state
,
3075 (unsigned long long) sh
->sector
);
3080 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
3084 /* We have read all the blocks in this stripe and now we need to
3085 * copy some of them into a target stripe for expand.
3087 struct dma_async_tx_descriptor
*tx
= NULL
;
3088 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3089 for (i
= 0; i
< sh
->disks
; i
++)
3090 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
3092 struct stripe_head
*sh2
;
3093 struct async_submit_ctl submit
;
3095 sector_t bn
= compute_blocknr(sh
, i
, 1);
3096 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
3098 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
3100 /* so far only the early blocks of this stripe
3101 * have been requested. When later blocks
3102 * get requested, we will try again
3105 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
3106 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
3107 /* must have already done this block */
3108 release_stripe(sh2
);
3112 /* place all the copies on one channel */
3113 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
3114 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
3115 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
3118 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
3119 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
3120 for (j
= 0; j
< conf
->raid_disks
; j
++)
3121 if (j
!= sh2
->pd_idx
&&
3123 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
3125 if (j
== conf
->raid_disks
) {
3126 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
3127 set_bit(STRIPE_HANDLE
, &sh2
->state
);
3129 release_stripe(sh2
);
3132 /* done submitting copies, wait for them to complete */
3135 dma_wait_for_async_tx(tx
);
3140 * handle_stripe - do things to a stripe.
3142 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3143 * state of various bits to see what needs to be done.
3145 * return some read requests which now have data
3146 * return some write requests which are safely on storage
3147 * schedule a read on some buffers
3148 * schedule a write of some buffers
3149 * return confirmation of parity correctness
3153 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
3155 struct r5conf
*conf
= sh
->raid_conf
;
3156 int disks
= sh
->disks
;
3159 int do_recovery
= 0;
3161 memset(s
, 0, sizeof(*s
));
3163 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3164 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3165 s
->failed_num
[0] = -1;
3166 s
->failed_num
[1] = -1;
3168 /* Now to look around and see what can be done */
3170 spin_lock_irq(&conf
->device_lock
);
3171 for (i
=disks
; i
--; ) {
3172 struct md_rdev
*rdev
;
3179 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3181 dev
->toread
, dev
->towrite
, dev
->written
);
3182 /* maybe we can reply to a read
3184 * new wantfill requests are only permitted while
3185 * ops_complete_biofill is guaranteed to be inactive
3187 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
3188 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
3189 set_bit(R5_Wantfill
, &dev
->flags
);
3191 /* now count some things */
3192 if (test_bit(R5_LOCKED
, &dev
->flags
))
3194 if (test_bit(R5_UPTODATE
, &dev
->flags
))
3196 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
3198 BUG_ON(s
->compute
> 2);
3201 if (test_bit(R5_Wantfill
, &dev
->flags
))
3203 else if (dev
->toread
)
3207 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3212 /* Prefer to use the replacement for reads, but only
3213 * if it is recovered enough and has no bad blocks.
3215 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
3216 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
3217 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
3218 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3219 &first_bad
, &bad_sectors
))
3220 set_bit(R5_ReadRepl
, &dev
->flags
);
3223 set_bit(R5_NeedReplace
, &dev
->flags
);
3224 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3225 clear_bit(R5_ReadRepl
, &dev
->flags
);
3227 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
3230 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3231 &first_bad
, &bad_sectors
);
3232 if (s
->blocked_rdev
== NULL
3233 && (test_bit(Blocked
, &rdev
->flags
)
3236 set_bit(BlockedBadBlocks
,
3238 s
->blocked_rdev
= rdev
;
3239 atomic_inc(&rdev
->nr_pending
);
3242 clear_bit(R5_Insync
, &dev
->flags
);
3246 /* also not in-sync */
3247 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
3248 test_bit(R5_UPTODATE
, &dev
->flags
)) {
3249 /* treat as in-sync, but with a read error
3250 * which we can now try to correct
3252 set_bit(R5_Insync
, &dev
->flags
);
3253 set_bit(R5_ReadError
, &dev
->flags
);
3255 } else if (test_bit(In_sync
, &rdev
->flags
))
3256 set_bit(R5_Insync
, &dev
->flags
);
3257 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
3258 /* in sync if before recovery_offset */
3259 set_bit(R5_Insync
, &dev
->flags
);
3260 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
3261 test_bit(R5_Expanded
, &dev
->flags
))
3262 /* If we've reshaped into here, we assume it is Insync.
3263 * We will shortly update recovery_offset to make
3266 set_bit(R5_Insync
, &dev
->flags
);
3268 if (rdev
&& test_bit(R5_WriteError
, &dev
->flags
)) {
3269 /* This flag does not apply to '.replacement'
3270 * only to .rdev, so make sure to check that*/
3271 struct md_rdev
*rdev2
= rcu_dereference(
3272 conf
->disks
[i
].rdev
);
3274 clear_bit(R5_Insync
, &dev
->flags
);
3275 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3276 s
->handle_bad_blocks
= 1;
3277 atomic_inc(&rdev2
->nr_pending
);
3279 clear_bit(R5_WriteError
, &dev
->flags
);
3281 if (rdev
&& test_bit(R5_MadeGood
, &dev
->flags
)) {
3282 /* This flag does not apply to '.replacement'
3283 * only to .rdev, so make sure to check that*/
3284 struct md_rdev
*rdev2
= rcu_dereference(
3285 conf
->disks
[i
].rdev
);
3286 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3287 s
->handle_bad_blocks
= 1;
3288 atomic_inc(&rdev2
->nr_pending
);
3290 clear_bit(R5_MadeGood
, &dev
->flags
);
3292 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
3293 struct md_rdev
*rdev2
= rcu_dereference(
3294 conf
->disks
[i
].replacement
);
3295 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3296 s
->handle_bad_blocks
= 1;
3297 atomic_inc(&rdev2
->nr_pending
);
3299 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
3301 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3302 /* The ReadError flag will just be confusing now */
3303 clear_bit(R5_ReadError
, &dev
->flags
);
3304 clear_bit(R5_ReWrite
, &dev
->flags
);
3306 if (test_bit(R5_ReadError
, &dev
->flags
))
3307 clear_bit(R5_Insync
, &dev
->flags
);
3308 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3310 s
->failed_num
[s
->failed
] = i
;
3312 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
3316 spin_unlock_irq(&conf
->device_lock
);
3317 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
3318 /* If there is a failed device being replaced,
3319 * we must be recovering.
3320 * else if we are after recovery_cp, we must be syncing
3321 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3322 * else we can only be replacing
3323 * sync and recovery both need to read all devices, and so
3324 * use the same flag.
3327 sh
->sector
>= conf
->mddev
->recovery_cp
||
3328 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
3336 static void handle_stripe(struct stripe_head
*sh
)
3338 struct stripe_head_state s
;
3339 struct r5conf
*conf
= sh
->raid_conf
;
3342 int disks
= sh
->disks
;
3343 struct r5dev
*pdev
, *qdev
;
3345 clear_bit(STRIPE_HANDLE
, &sh
->state
);
3346 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
3347 /* already being handled, ensure it gets handled
3348 * again when current action finishes */
3349 set_bit(STRIPE_HANDLE
, &sh
->state
);
3353 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
3354 set_bit(STRIPE_SYNCING
, &sh
->state
);
3355 clear_bit(STRIPE_INSYNC
, &sh
->state
);
3357 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3359 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3360 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3361 (unsigned long long)sh
->sector
, sh
->state
,
3362 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
3363 sh
->check_state
, sh
->reconstruct_state
);
3365 analyse_stripe(sh
, &s
);
3367 if (s
.handle_bad_blocks
) {
3368 set_bit(STRIPE_HANDLE
, &sh
->state
);
3372 if (unlikely(s
.blocked_rdev
)) {
3373 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
3374 s
.replacing
|| s
.to_write
|| s
.written
) {
3375 set_bit(STRIPE_HANDLE
, &sh
->state
);
3378 /* There is nothing for the blocked_rdev to block */
3379 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
3380 s
.blocked_rdev
= NULL
;
3383 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
3384 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
3385 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
3388 pr_debug("locked=%d uptodate=%d to_read=%d"
3389 " to_write=%d failed=%d failed_num=%d,%d\n",
3390 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
3391 s
.failed_num
[0], s
.failed_num
[1]);
3392 /* check if the array has lost more than max_degraded devices and,
3393 * if so, some requests might need to be failed.
3395 if (s
.failed
> conf
->max_degraded
) {
3396 sh
->check_state
= 0;
3397 sh
->reconstruct_state
= 0;
3398 if (s
.to_read
+s
.to_write
+s
.written
)
3399 handle_failed_stripe(conf
, sh
, &s
, disks
, &s
.return_bi
);
3400 if (s
.syncing
+ s
.replacing
)
3401 handle_failed_sync(conf
, sh
, &s
);
3405 * might be able to return some write requests if the parity blocks
3406 * are safe, or on a failed drive
3408 pdev
= &sh
->dev
[sh
->pd_idx
];
3409 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
3410 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
3411 qdev
= &sh
->dev
[sh
->qd_idx
];
3412 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
3413 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
3417 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
3418 && !test_bit(R5_LOCKED
, &pdev
->flags
)
3419 && test_bit(R5_UPTODATE
, &pdev
->flags
)))) &&
3420 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
3421 && !test_bit(R5_LOCKED
, &qdev
->flags
)
3422 && test_bit(R5_UPTODATE
, &qdev
->flags
)))))
3423 handle_stripe_clean_event(conf
, sh
, disks
, &s
.return_bi
);
3425 /* Now we might consider reading some blocks, either to check/generate
3426 * parity, or to satisfy requests
3427 * or to load a block that is being partially written.
3429 if (s
.to_read
|| s
.non_overwrite
3430 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
3431 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
3434 handle_stripe_fill(sh
, &s
, disks
);
3436 /* Now we check to see if any write operations have recently
3440 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
3442 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
3443 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
3444 sh
->reconstruct_state
= reconstruct_state_idle
;
3446 /* All the 'written' buffers and the parity block are ready to
3447 * be written back to disk
3449 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
));
3450 BUG_ON(sh
->qd_idx
>= 0 &&
3451 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
));
3452 for (i
= disks
; i
--; ) {
3453 struct r5dev
*dev
= &sh
->dev
[i
];
3454 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
3455 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3457 pr_debug("Writing block %d\n", i
);
3458 set_bit(R5_Wantwrite
, &dev
->flags
);
3461 if (!test_bit(R5_Insync
, &dev
->flags
) ||
3462 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3464 set_bit(STRIPE_INSYNC
, &sh
->state
);
3467 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3468 s
.dec_preread_active
= 1;
3471 /* Now to consider new write requests and what else, if anything
3472 * should be read. We do not handle new writes when:
3473 * 1/ A 'write' operation (copy+xor) is already in flight.
3474 * 2/ A 'check' operation is in flight, as it may clobber the parity
3477 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
3478 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
3480 /* maybe we need to check and possibly fix the parity for this stripe
3481 * Any reads will already have been scheduled, so we just see if enough
3482 * data is available. The parity check is held off while parity
3483 * dependent operations are in flight.
3485 if (sh
->check_state
||
3486 (s
.syncing
&& s
.locked
== 0 &&
3487 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
3488 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
3489 if (conf
->level
== 6)
3490 handle_parity_checks6(conf
, sh
, &s
, disks
);
3492 handle_parity_checks5(conf
, sh
, &s
, disks
);
3495 if (s
.replacing
&& s
.locked
== 0
3496 && !test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3497 /* Write out to replacement devices where possible */
3498 for (i
= 0; i
< conf
->raid_disks
; i
++)
3499 if (test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
) &&
3500 test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
3501 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
3502 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3505 set_bit(STRIPE_INSYNC
, &sh
->state
);
3507 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
3508 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3509 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3510 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3513 /* If the failed drives are just a ReadError, then we might need
3514 * to progress the repair/check process
3516 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
3517 for (i
= 0; i
< s
.failed
; i
++) {
3518 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
3519 if (test_bit(R5_ReadError
, &dev
->flags
)
3520 && !test_bit(R5_LOCKED
, &dev
->flags
)
3521 && test_bit(R5_UPTODATE
, &dev
->flags
)
3523 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3524 set_bit(R5_Wantwrite
, &dev
->flags
);
3525 set_bit(R5_ReWrite
, &dev
->flags
);
3526 set_bit(R5_LOCKED
, &dev
->flags
);
3529 /* let's read it back */
3530 set_bit(R5_Wantread
, &dev
->flags
);
3531 set_bit(R5_LOCKED
, &dev
->flags
);
3538 /* Finish reconstruct operations initiated by the expansion process */
3539 if (sh
->reconstruct_state
== reconstruct_state_result
) {
3540 struct stripe_head
*sh_src
3541 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
3542 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
3543 /* sh cannot be written until sh_src has been read.
3544 * so arrange for sh to be delayed a little
3546 set_bit(STRIPE_DELAYED
, &sh
->state
);
3547 set_bit(STRIPE_HANDLE
, &sh
->state
);
3548 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
3550 atomic_inc(&conf
->preread_active_stripes
);
3551 release_stripe(sh_src
);
3555 release_stripe(sh_src
);
3557 sh
->reconstruct_state
= reconstruct_state_idle
;
3558 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3559 for (i
= conf
->raid_disks
; i
--; ) {
3560 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3561 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3566 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
3567 !sh
->reconstruct_state
) {
3568 /* Need to write out all blocks after computing parity */
3569 sh
->disks
= conf
->raid_disks
;
3570 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
3571 schedule_reconstruction(sh
, &s
, 1, 1);
3572 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
3573 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3574 atomic_dec(&conf
->reshape_stripes
);
3575 wake_up(&conf
->wait_for_overlap
);
3576 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3579 if (s
.expanding
&& s
.locked
== 0 &&
3580 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
3581 handle_stripe_expansion(conf
, sh
);
3584 /* wait for this device to become unblocked */
3585 if (conf
->mddev
->external
&& unlikely(s
.blocked_rdev
))
3586 md_wait_for_blocked_rdev(s
.blocked_rdev
, conf
->mddev
);
3588 if (s
.handle_bad_blocks
)
3589 for (i
= disks
; i
--; ) {
3590 struct md_rdev
*rdev
;
3591 struct r5dev
*dev
= &sh
->dev
[i
];
3592 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
3593 /* We own a safe reference to the rdev */
3594 rdev
= conf
->disks
[i
].rdev
;
3595 if (!rdev_set_badblocks(rdev
, sh
->sector
,
3597 md_error(conf
->mddev
, rdev
);
3598 rdev_dec_pending(rdev
, conf
->mddev
);
3600 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
3601 rdev
= conf
->disks
[i
].rdev
;
3602 rdev_clear_badblocks(rdev
, sh
->sector
,
3604 rdev_dec_pending(rdev
, conf
->mddev
);
3606 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
3607 rdev
= conf
->disks
[i
].replacement
;
3609 /* rdev have been moved down */
3610 rdev
= conf
->disks
[i
].rdev
;
3611 rdev_clear_badblocks(rdev
, sh
->sector
,
3613 rdev_dec_pending(rdev
, conf
->mddev
);
3618 raid_run_ops(sh
, s
.ops_request
);
3622 if (s
.dec_preread_active
) {
3623 /* We delay this until after ops_run_io so that if make_request
3624 * is waiting on a flush, it won't continue until the writes
3625 * have actually been submitted.
3627 atomic_dec(&conf
->preread_active_stripes
);
3628 if (atomic_read(&conf
->preread_active_stripes
) <
3630 md_wakeup_thread(conf
->mddev
->thread
);
3633 return_io(s
.return_bi
);
3635 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
3638 static void raid5_activate_delayed(struct r5conf
*conf
)
3640 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
3641 while (!list_empty(&conf
->delayed_list
)) {
3642 struct list_head
*l
= conf
->delayed_list
.next
;
3643 struct stripe_head
*sh
;
3644 sh
= list_entry(l
, struct stripe_head
, lru
);
3646 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3647 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3648 atomic_inc(&conf
->preread_active_stripes
);
3649 list_add_tail(&sh
->lru
, &conf
->hold_list
);
3654 static void activate_bit_delay(struct r5conf
*conf
)
3656 /* device_lock is held */
3657 struct list_head head
;
3658 list_add(&head
, &conf
->bitmap_list
);
3659 list_del_init(&conf
->bitmap_list
);
3660 while (!list_empty(&head
)) {
3661 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
3662 list_del_init(&sh
->lru
);
3663 atomic_inc(&sh
->count
);
3664 __release_stripe(conf
, sh
);
3668 int md_raid5_congested(struct mddev
*mddev
, int bits
)
3670 struct r5conf
*conf
= mddev
->private;
3672 /* No difference between reads and writes. Just check
3673 * how busy the stripe_cache is
3676 if (conf
->inactive_blocked
)
3680 if (list_empty_careful(&conf
->inactive_list
))
3685 EXPORT_SYMBOL_GPL(md_raid5_congested
);
3687 static int raid5_congested(void *data
, int bits
)
3689 struct mddev
*mddev
= data
;
3691 return mddev_congested(mddev
, bits
) ||
3692 md_raid5_congested(mddev
, bits
);
3695 /* We want read requests to align with chunks where possible,
3696 * but write requests don't need to.
3698 static int raid5_mergeable_bvec(struct request_queue
*q
,
3699 struct bvec_merge_data
*bvm
,
3700 struct bio_vec
*biovec
)
3702 struct mddev
*mddev
= q
->queuedata
;
3703 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
3705 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3706 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
3708 if ((bvm
->bi_rw
& 1) == WRITE
)
3709 return biovec
->bv_len
; /* always allow writes to be mergeable */
3711 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3712 chunk_sectors
= mddev
->new_chunk_sectors
;
3713 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
3714 if (max
< 0) max
= 0;
3715 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
3716 return biovec
->bv_len
;
3722 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
3724 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
3725 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3726 unsigned int bio_sectors
= bio
->bi_size
>> 9;
3728 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3729 chunk_sectors
= mddev
->new_chunk_sectors
;
3730 return chunk_sectors
>=
3731 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
3735 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3736 * later sampled by raid5d.
3738 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
3740 unsigned long flags
;
3742 spin_lock_irqsave(&conf
->device_lock
, flags
);
3744 bi
->bi_next
= conf
->retry_read_aligned_list
;
3745 conf
->retry_read_aligned_list
= bi
;
3747 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3748 md_wakeup_thread(conf
->mddev
->thread
);
3752 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
)
3756 bi
= conf
->retry_read_aligned
;
3758 conf
->retry_read_aligned
= NULL
;
3761 bi
= conf
->retry_read_aligned_list
;
3763 conf
->retry_read_aligned_list
= bi
->bi_next
;
3766 * this sets the active strip count to 1 and the processed
3767 * strip count to zero (upper 8 bits)
3769 bi
->bi_phys_segments
= 1; /* biased count of active stripes */
3777 * The "raid5_align_endio" should check if the read succeeded and if it
3778 * did, call bio_endio on the original bio (having bio_put the new bio
3780 * If the read failed..
3782 static void raid5_align_endio(struct bio
*bi
, int error
)
3784 struct bio
* raid_bi
= bi
->bi_private
;
3785 struct mddev
*mddev
;
3786 struct r5conf
*conf
;
3787 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3788 struct md_rdev
*rdev
;
3792 rdev
= (void*)raid_bi
->bi_next
;
3793 raid_bi
->bi_next
= NULL
;
3794 mddev
= rdev
->mddev
;
3795 conf
= mddev
->private;
3797 rdev_dec_pending(rdev
, conf
->mddev
);
3799 if (!error
&& uptodate
) {
3800 bio_endio(raid_bi
, 0);
3801 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
3802 wake_up(&conf
->wait_for_stripe
);
3807 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3809 add_bio_to_retry(raid_bi
, conf
);
3812 static int bio_fits_rdev(struct bio
*bi
)
3814 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
3816 if ((bi
->bi_size
>>9) > queue_max_sectors(q
))
3818 blk_recount_segments(q
, bi
);
3819 if (bi
->bi_phys_segments
> queue_max_segments(q
))
3822 if (q
->merge_bvec_fn
)
3823 /* it's too hard to apply the merge_bvec_fn at this stage,
3832 static int chunk_aligned_read(struct mddev
*mddev
, struct bio
* raid_bio
)
3834 struct r5conf
*conf
= mddev
->private;
3836 struct bio
* align_bi
;
3837 struct md_rdev
*rdev
;
3838 sector_t end_sector
;
3840 if (!in_chunk_boundary(mddev
, raid_bio
)) {
3841 pr_debug("chunk_aligned_read : non aligned\n");
3845 * use bio_clone_mddev to make a copy of the bio
3847 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
3851 * set bi_end_io to a new function, and set bi_private to the
3854 align_bi
->bi_end_io
= raid5_align_endio
;
3855 align_bi
->bi_private
= raid_bio
;
3859 align_bi
->bi_sector
= raid5_compute_sector(conf
, raid_bio
->bi_sector
,
3863 end_sector
= align_bi
->bi_sector
+ (align_bi
->bi_size
>> 9);
3865 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
3866 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
3867 rdev
->recovery_offset
< end_sector
) {
3868 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
3870 (test_bit(Faulty
, &rdev
->flags
) ||
3871 !(test_bit(In_sync
, &rdev
->flags
) ||
3872 rdev
->recovery_offset
>= end_sector
)))
3879 atomic_inc(&rdev
->nr_pending
);
3881 raid_bio
->bi_next
= (void*)rdev
;
3882 align_bi
->bi_bdev
= rdev
->bdev
;
3883 align_bi
->bi_flags
&= ~(1 << BIO_SEG_VALID
);
3884 /* No reshape active, so we can trust rdev->data_offset */
3885 align_bi
->bi_sector
+= rdev
->data_offset
;
3887 if (!bio_fits_rdev(align_bi
) ||
3888 is_badblock(rdev
, align_bi
->bi_sector
, align_bi
->bi_size
>>9,
3889 &first_bad
, &bad_sectors
)) {
3890 /* too big in some way, or has a known bad block */
3892 rdev_dec_pending(rdev
, mddev
);
3896 spin_lock_irq(&conf
->device_lock
);
3897 wait_event_lock_irq(conf
->wait_for_stripe
,
3899 conf
->device_lock
, /* nothing */);
3900 atomic_inc(&conf
->active_aligned_reads
);
3901 spin_unlock_irq(&conf
->device_lock
);
3903 generic_make_request(align_bi
);
3912 /* __get_priority_stripe - get the next stripe to process
3914 * Full stripe writes are allowed to pass preread active stripes up until
3915 * the bypass_threshold is exceeded. In general the bypass_count
3916 * increments when the handle_list is handled before the hold_list; however, it
3917 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3918 * stripe with in flight i/o. The bypass_count will be reset when the
3919 * head of the hold_list has changed, i.e. the head was promoted to the
3922 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
)
3924 struct stripe_head
*sh
;
3926 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3928 list_empty(&conf
->handle_list
) ? "empty" : "busy",
3929 list_empty(&conf
->hold_list
) ? "empty" : "busy",
3930 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
3932 if (!list_empty(&conf
->handle_list
)) {
3933 sh
= list_entry(conf
->handle_list
.next
, typeof(*sh
), lru
);
3935 if (list_empty(&conf
->hold_list
))
3936 conf
->bypass_count
= 0;
3937 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
3938 if (conf
->hold_list
.next
== conf
->last_hold
)
3939 conf
->bypass_count
++;
3941 conf
->last_hold
= conf
->hold_list
.next
;
3942 conf
->bypass_count
-= conf
->bypass_threshold
;
3943 if (conf
->bypass_count
< 0)
3944 conf
->bypass_count
= 0;
3947 } else if (!list_empty(&conf
->hold_list
) &&
3948 ((conf
->bypass_threshold
&&
3949 conf
->bypass_count
> conf
->bypass_threshold
) ||
3950 atomic_read(&conf
->pending_full_writes
) == 0)) {
3951 sh
= list_entry(conf
->hold_list
.next
,
3953 conf
->bypass_count
-= conf
->bypass_threshold
;
3954 if (conf
->bypass_count
< 0)
3955 conf
->bypass_count
= 0;
3959 list_del_init(&sh
->lru
);
3960 atomic_inc(&sh
->count
);
3961 BUG_ON(atomic_read(&sh
->count
) != 1);
3965 static void make_request(struct mddev
*mddev
, struct bio
* bi
)
3967 struct r5conf
*conf
= mddev
->private;
3969 sector_t new_sector
;
3970 sector_t logical_sector
, last_sector
;
3971 struct stripe_head
*sh
;
3972 const int rw
= bio_data_dir(bi
);
3976 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
3977 md_flush_request(mddev
, bi
);
3981 md_write_start(mddev
, bi
);
3984 mddev
->reshape_position
== MaxSector
&&
3985 chunk_aligned_read(mddev
,bi
))
3988 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
3989 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
3991 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
3993 plugged
= mddev_check_plugged(mddev
);
3994 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
4000 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
4001 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
4002 /* spinlock is needed as reshape_progress may be
4003 * 64bit on a 32bit platform, and so it might be
4004 * possible to see a half-updated value
4005 * Of course reshape_progress could change after
4006 * the lock is dropped, so once we get a reference
4007 * to the stripe that we think it is, we will have
4010 spin_lock_irq(&conf
->device_lock
);
4011 if (mddev
->reshape_backwards
4012 ? logical_sector
< conf
->reshape_progress
4013 : logical_sector
>= conf
->reshape_progress
) {
4016 if (mddev
->reshape_backwards
4017 ? logical_sector
< conf
->reshape_safe
4018 : logical_sector
>= conf
->reshape_safe
) {
4019 spin_unlock_irq(&conf
->device_lock
);
4024 spin_unlock_irq(&conf
->device_lock
);
4027 new_sector
= raid5_compute_sector(conf
, logical_sector
,
4030 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4031 (unsigned long long)new_sector
,
4032 (unsigned long long)logical_sector
);
4034 sh
= get_active_stripe(conf
, new_sector
, previous
,
4035 (bi
->bi_rw
&RWA_MASK
), 0);
4037 if (unlikely(previous
)) {
4038 /* expansion might have moved on while waiting for a
4039 * stripe, so we must do the range check again.
4040 * Expansion could still move past after this
4041 * test, but as we are holding a reference to
4042 * 'sh', we know that if that happens,
4043 * STRIPE_EXPANDING will get set and the expansion
4044 * won't proceed until we finish with the stripe.
4047 spin_lock_irq(&conf
->device_lock
);
4048 if (mddev
->reshape_backwards
4049 ? logical_sector
>= conf
->reshape_progress
4050 : logical_sector
< conf
->reshape_progress
)
4051 /* mismatch, need to try again */
4053 spin_unlock_irq(&conf
->device_lock
);
4062 logical_sector
>= mddev
->suspend_lo
&&
4063 logical_sector
< mddev
->suspend_hi
) {
4065 /* As the suspend_* range is controlled by
4066 * userspace, we want an interruptible
4069 flush_signals(current
);
4070 prepare_to_wait(&conf
->wait_for_overlap
,
4071 &w
, TASK_INTERRUPTIBLE
);
4072 if (logical_sector
>= mddev
->suspend_lo
&&
4073 logical_sector
< mddev
->suspend_hi
)
4078 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
4079 !add_stripe_bio(sh
, bi
, dd_idx
, rw
)) {
4080 /* Stripe is busy expanding or
4081 * add failed due to overlap. Flush everything
4084 md_wakeup_thread(mddev
->thread
);
4089 finish_wait(&conf
->wait_for_overlap
, &w
);
4090 set_bit(STRIPE_HANDLE
, &sh
->state
);
4091 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4092 if ((bi
->bi_rw
& REQ_SYNC
) &&
4093 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4094 atomic_inc(&conf
->preread_active_stripes
);
4097 /* cannot get stripe for read-ahead, just give-up */
4098 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
4099 finish_wait(&conf
->wait_for_overlap
, &w
);
4105 md_wakeup_thread(mddev
->thread
);
4107 spin_lock_irq(&conf
->device_lock
);
4108 remaining
= raid5_dec_bi_phys_segments(bi
);
4109 spin_unlock_irq(&conf
->device_lock
);
4110 if (remaining
== 0) {
4113 md_write_end(mddev
);
4119 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
4121 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
4123 /* reshaping is quite different to recovery/resync so it is
4124 * handled quite separately ... here.
4126 * On each call to sync_request, we gather one chunk worth of
4127 * destination stripes and flag them as expanding.
4128 * Then we find all the source stripes and request reads.
4129 * As the reads complete, handle_stripe will copy the data
4130 * into the destination stripe and release that stripe.
4132 struct r5conf
*conf
= mddev
->private;
4133 struct stripe_head
*sh
;
4134 sector_t first_sector
, last_sector
;
4135 int raid_disks
= conf
->previous_raid_disks
;
4136 int data_disks
= raid_disks
- conf
->max_degraded
;
4137 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
4140 sector_t writepos
, readpos
, safepos
;
4141 sector_t stripe_addr
;
4142 int reshape_sectors
;
4143 struct list_head stripes
;
4145 if (sector_nr
== 0) {
4146 /* If restarting in the middle, skip the initial sectors */
4147 if (mddev
->reshape_backwards
&&
4148 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
4149 sector_nr
= raid5_size(mddev
, 0, 0)
4150 - conf
->reshape_progress
;
4151 } else if (!mddev
->reshape_backwards
&&
4152 conf
->reshape_progress
> 0)
4153 sector_nr
= conf
->reshape_progress
;
4154 sector_div(sector_nr
, new_data_disks
);
4156 mddev
->curr_resync_completed
= sector_nr
;
4157 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4163 /* We need to process a full chunk at a time.
4164 * If old and new chunk sizes differ, we need to process the
4167 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
)
4168 reshape_sectors
= mddev
->new_chunk_sectors
;
4170 reshape_sectors
= mddev
->chunk_sectors
;
4172 /* We update the metadata at least every 10 seconds, or when
4173 * the data about to be copied would over-write the source of
4174 * the data at the front of the range. i.e. one new_stripe
4175 * along from reshape_progress new_maps to after where
4176 * reshape_safe old_maps to
4178 writepos
= conf
->reshape_progress
;
4179 sector_div(writepos
, new_data_disks
);
4180 readpos
= conf
->reshape_progress
;
4181 sector_div(readpos
, data_disks
);
4182 safepos
= conf
->reshape_safe
;
4183 sector_div(safepos
, data_disks
);
4184 if (mddev
->reshape_backwards
) {
4185 writepos
-= min_t(sector_t
, reshape_sectors
, writepos
);
4186 readpos
+= reshape_sectors
;
4187 safepos
+= reshape_sectors
;
4189 writepos
+= reshape_sectors
;
4190 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
4191 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
4194 /* Having calculated the 'writepos' possibly use it
4195 * to set 'stripe_addr' which is where we will write to.
4197 if (mddev
->reshape_backwards
) {
4198 BUG_ON(conf
->reshape_progress
== 0);
4199 stripe_addr
= writepos
;
4200 BUG_ON((mddev
->dev_sectors
&
4201 ~((sector_t
)reshape_sectors
- 1))
4202 - reshape_sectors
- stripe_addr
4205 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
4206 stripe_addr
= sector_nr
;
4209 /* 'writepos' is the most advanced device address we might write.
4210 * 'readpos' is the least advanced device address we might read.
4211 * 'safepos' is the least address recorded in the metadata as having
4213 * If there is a min_offset_diff, these are adjusted either by
4214 * increasing the safepos/readpos if diff is negative, or
4215 * increasing writepos if diff is positive.
4216 * If 'readpos' is then behind 'writepos', there is no way that we can
4217 * ensure safety in the face of a crash - that must be done by userspace
4218 * making a backup of the data. So in that case there is no particular
4219 * rush to update metadata.
4220 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4221 * update the metadata to advance 'safepos' to match 'readpos' so that
4222 * we can be safe in the event of a crash.
4223 * So we insist on updating metadata if safepos is behind writepos and
4224 * readpos is beyond writepos.
4225 * In any case, update the metadata every 10 seconds.
4226 * Maybe that number should be configurable, but I'm not sure it is
4227 * worth it.... maybe it could be a multiple of safemode_delay???
4229 if (conf
->min_offset_diff
< 0) {
4230 safepos
+= -conf
->min_offset_diff
;
4231 readpos
+= -conf
->min_offset_diff
;
4233 writepos
+= conf
->min_offset_diff
;
4235 if ((mddev
->reshape_backwards
4236 ? (safepos
> writepos
&& readpos
< writepos
)
4237 : (safepos
< writepos
&& readpos
> writepos
)) ||
4238 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4239 /* Cannot proceed until we've updated the superblock... */
4240 wait_event(conf
->wait_for_overlap
,
4241 atomic_read(&conf
->reshape_stripes
)==0);
4242 mddev
->reshape_position
= conf
->reshape_progress
;
4243 mddev
->curr_resync_completed
= sector_nr
;
4244 conf
->reshape_checkpoint
= jiffies
;
4245 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4246 md_wakeup_thread(mddev
->thread
);
4247 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4248 kthread_should_stop());
4249 spin_lock_irq(&conf
->device_lock
);
4250 conf
->reshape_safe
= mddev
->reshape_position
;
4251 spin_unlock_irq(&conf
->device_lock
);
4252 wake_up(&conf
->wait_for_overlap
);
4253 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4256 INIT_LIST_HEAD(&stripes
);
4257 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
4259 int skipped_disk
= 0;
4260 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
4261 set_bit(STRIPE_EXPANDING
, &sh
->state
);
4262 atomic_inc(&conf
->reshape_stripes
);
4263 /* If any of this stripe is beyond the end of the old
4264 * array, then we need to zero those blocks
4266 for (j
=sh
->disks
; j
--;) {
4268 if (j
== sh
->pd_idx
)
4270 if (conf
->level
== 6 &&
4273 s
= compute_blocknr(sh
, j
, 0);
4274 if (s
< raid5_size(mddev
, 0, 0)) {
4278 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
4279 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
4280 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
4282 if (!skipped_disk
) {
4283 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4284 set_bit(STRIPE_HANDLE
, &sh
->state
);
4286 list_add(&sh
->lru
, &stripes
);
4288 spin_lock_irq(&conf
->device_lock
);
4289 if (mddev
->reshape_backwards
)
4290 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
4292 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
4293 spin_unlock_irq(&conf
->device_lock
);
4294 /* Ok, those stripe are ready. We can start scheduling
4295 * reads on the source stripes.
4296 * The source stripes are determined by mapping the first and last
4297 * block on the destination stripes.
4300 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
4303 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
4304 * new_data_disks
- 1),
4306 if (last_sector
>= mddev
->dev_sectors
)
4307 last_sector
= mddev
->dev_sectors
- 1;
4308 while (first_sector
<= last_sector
) {
4309 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
4310 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
4311 set_bit(STRIPE_HANDLE
, &sh
->state
);
4313 first_sector
+= STRIPE_SECTORS
;
4315 /* Now that the sources are clearly marked, we can release
4316 * the destination stripes
4318 while (!list_empty(&stripes
)) {
4319 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
4320 list_del_init(&sh
->lru
);
4323 /* If this takes us to the resync_max point where we have to pause,
4324 * then we need to write out the superblock.
4326 sector_nr
+= reshape_sectors
;
4327 if ((sector_nr
- mddev
->curr_resync_completed
) * 2
4328 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
4329 /* Cannot proceed until we've updated the superblock... */
4330 wait_event(conf
->wait_for_overlap
,
4331 atomic_read(&conf
->reshape_stripes
) == 0);
4332 mddev
->reshape_position
= conf
->reshape_progress
;
4333 mddev
->curr_resync_completed
= sector_nr
;
4334 conf
->reshape_checkpoint
= jiffies
;
4335 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4336 md_wakeup_thread(mddev
->thread
);
4337 wait_event(mddev
->sb_wait
,
4338 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
4339 || kthread_should_stop());
4340 spin_lock_irq(&conf
->device_lock
);
4341 conf
->reshape_safe
= mddev
->reshape_position
;
4342 spin_unlock_irq(&conf
->device_lock
);
4343 wake_up(&conf
->wait_for_overlap
);
4344 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4346 return reshape_sectors
;
4349 /* FIXME go_faster isn't used */
4350 static inline sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
4352 struct r5conf
*conf
= mddev
->private;
4353 struct stripe_head
*sh
;
4354 sector_t max_sector
= mddev
->dev_sectors
;
4355 sector_t sync_blocks
;
4356 int still_degraded
= 0;
4359 if (sector_nr
>= max_sector
) {
4360 /* just being told to finish up .. nothing much to do */
4362 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
4367 if (mddev
->curr_resync
< max_sector
) /* aborted */
4368 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
4370 else /* completed sync */
4372 bitmap_close_sync(mddev
->bitmap
);
4377 /* Allow raid5_quiesce to complete */
4378 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
4380 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
4381 return reshape_request(mddev
, sector_nr
, skipped
);
4383 /* No need to check resync_max as we never do more than one
4384 * stripe, and as resync_max will always be on a chunk boundary,
4385 * if the check in md_do_sync didn't fire, there is no chance
4386 * of overstepping resync_max here
4389 /* if there is too many failed drives and we are trying
4390 * to resync, then assert that we are finished, because there is
4391 * nothing we can do.
4393 if (mddev
->degraded
>= conf
->max_degraded
&&
4394 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
4395 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
4399 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
4400 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
4401 !conf
->fullsync
&& sync_blocks
>= STRIPE_SECTORS
) {
4402 /* we can skip this block, and probably more */
4403 sync_blocks
/= STRIPE_SECTORS
;
4405 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
4408 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
4410 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
4412 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
4413 /* make sure we don't swamp the stripe cache if someone else
4414 * is trying to get access
4416 schedule_timeout_uninterruptible(1);
4418 /* Need to check if array will still be degraded after recovery/resync
4419 * We don't need to check the 'failed' flag as when that gets set,
4422 for (i
= 0; i
< conf
->raid_disks
; i
++)
4423 if (conf
->disks
[i
].rdev
== NULL
)
4426 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
4428 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
4433 return STRIPE_SECTORS
;
4436 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
)
4438 /* We may not be able to submit a whole bio at once as there
4439 * may not be enough stripe_heads available.
4440 * We cannot pre-allocate enough stripe_heads as we may need
4441 * more than exist in the cache (if we allow ever large chunks).
4442 * So we do one stripe head at a time and record in
4443 * ->bi_hw_segments how many have been done.
4445 * We *know* that this entire raid_bio is in one chunk, so
4446 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4448 struct stripe_head
*sh
;
4450 sector_t sector
, logical_sector
, last_sector
;
4455 logical_sector
= raid_bio
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4456 sector
= raid5_compute_sector(conf
, logical_sector
,
4458 last_sector
= raid_bio
->bi_sector
+ (raid_bio
->bi_size
>>9);
4460 for (; logical_sector
< last_sector
;
4461 logical_sector
+= STRIPE_SECTORS
,
4462 sector
+= STRIPE_SECTORS
,
4465 if (scnt
< raid5_bi_hw_segments(raid_bio
))
4466 /* already done this stripe */
4469 sh
= get_active_stripe(conf
, sector
, 0, 1, 0);
4472 /* failed to get a stripe - must wait */
4473 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4474 conf
->retry_read_aligned
= raid_bio
;
4478 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0)) {
4480 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4481 conf
->retry_read_aligned
= raid_bio
;
4489 spin_lock_irq(&conf
->device_lock
);
4490 remaining
= raid5_dec_bi_phys_segments(raid_bio
);
4491 spin_unlock_irq(&conf
->device_lock
);
4493 bio_endio(raid_bio
, 0);
4494 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4495 wake_up(&conf
->wait_for_stripe
);
4501 * This is our raid5 kernel thread.
4503 * We scan the hash table for stripes which can be handled now.
4504 * During the scan, completed stripes are saved for us by the interrupt
4505 * handler, so that they will not have to wait for our next wakeup.
4507 static void raid5d(struct mddev
*mddev
)
4509 struct stripe_head
*sh
;
4510 struct r5conf
*conf
= mddev
->private;
4512 struct blk_plug plug
;
4514 pr_debug("+++ raid5d active\n");
4516 md_check_recovery(mddev
);
4518 blk_start_plug(&plug
);
4520 spin_lock_irq(&conf
->device_lock
);
4524 if (atomic_read(&mddev
->plug_cnt
) == 0 &&
4525 !list_empty(&conf
->bitmap_list
)) {
4526 /* Now is a good time to flush some bitmap updates */
4528 spin_unlock_irq(&conf
->device_lock
);
4529 bitmap_unplug(mddev
->bitmap
);
4530 spin_lock_irq(&conf
->device_lock
);
4531 conf
->seq_write
= conf
->seq_flush
;
4532 activate_bit_delay(conf
);
4534 if (atomic_read(&mddev
->plug_cnt
) == 0)
4535 raid5_activate_delayed(conf
);
4537 while ((bio
= remove_bio_from_retry(conf
))) {
4539 spin_unlock_irq(&conf
->device_lock
);
4540 ok
= retry_aligned_read(conf
, bio
);
4541 spin_lock_irq(&conf
->device_lock
);
4547 sh
= __get_priority_stripe(conf
);
4551 spin_unlock_irq(&conf
->device_lock
);
4558 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
4559 md_check_recovery(mddev
);
4561 spin_lock_irq(&conf
->device_lock
);
4563 pr_debug("%d stripes handled\n", handled
);
4565 spin_unlock_irq(&conf
->device_lock
);
4567 async_tx_issue_pending_all();
4568 blk_finish_plug(&plug
);
4570 pr_debug("--- raid5d inactive\n");
4574 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
4576 struct r5conf
*conf
= mddev
->private;
4578 return sprintf(page
, "%d\n", conf
->max_nr_stripes
);
4584 raid5_set_cache_size(struct mddev
*mddev
, int size
)
4586 struct r5conf
*conf
= mddev
->private;
4589 if (size
<= 16 || size
> 32768)
4591 while (size
< conf
->max_nr_stripes
) {
4592 if (drop_one_stripe(conf
))
4593 conf
->max_nr_stripes
--;
4597 err
= md_allow_write(mddev
);
4600 while (size
> conf
->max_nr_stripes
) {
4601 if (grow_one_stripe(conf
))
4602 conf
->max_nr_stripes
++;
4607 EXPORT_SYMBOL(raid5_set_cache_size
);
4610 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
4612 struct r5conf
*conf
= mddev
->private;
4616 if (len
>= PAGE_SIZE
)
4621 if (strict_strtoul(page
, 10, &new))
4623 err
= raid5_set_cache_size(mddev
, new);
4629 static struct md_sysfs_entry
4630 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
4631 raid5_show_stripe_cache_size
,
4632 raid5_store_stripe_cache_size
);
4635 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
4637 struct r5conf
*conf
= mddev
->private;
4639 return sprintf(page
, "%d\n", conf
->bypass_threshold
);
4645 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
4647 struct r5conf
*conf
= mddev
->private;
4649 if (len
>= PAGE_SIZE
)
4654 if (strict_strtoul(page
, 10, &new))
4656 if (new > conf
->max_nr_stripes
)
4658 conf
->bypass_threshold
= new;
4662 static struct md_sysfs_entry
4663 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
4665 raid5_show_preread_threshold
,
4666 raid5_store_preread_threshold
);
4669 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
4671 struct r5conf
*conf
= mddev
->private;
4673 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
4678 static struct md_sysfs_entry
4679 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
4681 static struct attribute
*raid5_attrs
[] = {
4682 &raid5_stripecache_size
.attr
,
4683 &raid5_stripecache_active
.attr
,
4684 &raid5_preread_bypass_threshold
.attr
,
4687 static struct attribute_group raid5_attrs_group
= {
4689 .attrs
= raid5_attrs
,
4693 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
4695 struct r5conf
*conf
= mddev
->private;
4698 sectors
= mddev
->dev_sectors
;
4700 /* size is defined by the smallest of previous and new size */
4701 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
4703 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
4704 sectors
&= ~((sector_t
)mddev
->new_chunk_sectors
- 1);
4705 return sectors
* (raid_disks
- conf
->max_degraded
);
4708 static void raid5_free_percpu(struct r5conf
*conf
)
4710 struct raid5_percpu
*percpu
;
4717 for_each_possible_cpu(cpu
) {
4718 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4719 safe_put_page(percpu
->spare_page
);
4720 kfree(percpu
->scribble
);
4722 #ifdef CONFIG_HOTPLUG_CPU
4723 unregister_cpu_notifier(&conf
->cpu_notify
);
4727 free_percpu(conf
->percpu
);
4730 static void free_conf(struct r5conf
*conf
)
4732 shrink_stripes(conf
);
4733 raid5_free_percpu(conf
);
4735 kfree(conf
->stripe_hashtbl
);
4739 #ifdef CONFIG_HOTPLUG_CPU
4740 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
4743 struct r5conf
*conf
= container_of(nfb
, struct r5conf
, cpu_notify
);
4744 long cpu
= (long)hcpu
;
4745 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4748 case CPU_UP_PREPARE
:
4749 case CPU_UP_PREPARE_FROZEN
:
4750 if (conf
->level
== 6 && !percpu
->spare_page
)
4751 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
4752 if (!percpu
->scribble
)
4753 percpu
->scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
4755 if (!percpu
->scribble
||
4756 (conf
->level
== 6 && !percpu
->spare_page
)) {
4757 safe_put_page(percpu
->spare_page
);
4758 kfree(percpu
->scribble
);
4759 pr_err("%s: failed memory allocation for cpu%ld\n",
4761 return notifier_from_errno(-ENOMEM
);
4765 case CPU_DEAD_FROZEN
:
4766 safe_put_page(percpu
->spare_page
);
4767 kfree(percpu
->scribble
);
4768 percpu
->spare_page
= NULL
;
4769 percpu
->scribble
= NULL
;
4778 static int raid5_alloc_percpu(struct r5conf
*conf
)
4781 struct page
*spare_page
;
4782 struct raid5_percpu __percpu
*allcpus
;
4786 allcpus
= alloc_percpu(struct raid5_percpu
);
4789 conf
->percpu
= allcpus
;
4793 for_each_present_cpu(cpu
) {
4794 if (conf
->level
== 6) {
4795 spare_page
= alloc_page(GFP_KERNEL
);
4800 per_cpu_ptr(conf
->percpu
, cpu
)->spare_page
= spare_page
;
4802 scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
4807 per_cpu_ptr(conf
->percpu
, cpu
)->scribble
= scribble
;
4809 #ifdef CONFIG_HOTPLUG_CPU
4810 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
4811 conf
->cpu_notify
.priority
= 0;
4813 err
= register_cpu_notifier(&conf
->cpu_notify
);
4820 static struct r5conf
*setup_conf(struct mddev
*mddev
)
4822 struct r5conf
*conf
;
4823 int raid_disk
, memory
, max_disks
;
4824 struct md_rdev
*rdev
;
4825 struct disk_info
*disk
;
4827 if (mddev
->new_level
!= 5
4828 && mddev
->new_level
!= 4
4829 && mddev
->new_level
!= 6) {
4830 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4831 mdname(mddev
), mddev
->new_level
);
4832 return ERR_PTR(-EIO
);
4834 if ((mddev
->new_level
== 5
4835 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
4836 (mddev
->new_level
== 6
4837 && !algorithm_valid_raid6(mddev
->new_layout
))) {
4838 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
4839 mdname(mddev
), mddev
->new_layout
);
4840 return ERR_PTR(-EIO
);
4842 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
4843 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4844 mdname(mddev
), mddev
->raid_disks
);
4845 return ERR_PTR(-EINVAL
);
4848 if (!mddev
->new_chunk_sectors
||
4849 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
4850 !is_power_of_2(mddev
->new_chunk_sectors
)) {
4851 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
4852 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
4853 return ERR_PTR(-EINVAL
);
4856 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
4859 spin_lock_init(&conf
->device_lock
);
4860 init_waitqueue_head(&conf
->wait_for_stripe
);
4861 init_waitqueue_head(&conf
->wait_for_overlap
);
4862 INIT_LIST_HEAD(&conf
->handle_list
);
4863 INIT_LIST_HEAD(&conf
->hold_list
);
4864 INIT_LIST_HEAD(&conf
->delayed_list
);
4865 INIT_LIST_HEAD(&conf
->bitmap_list
);
4866 INIT_LIST_HEAD(&conf
->inactive_list
);
4867 atomic_set(&conf
->active_stripes
, 0);
4868 atomic_set(&conf
->preread_active_stripes
, 0);
4869 atomic_set(&conf
->active_aligned_reads
, 0);
4870 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
4871 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
4873 conf
->raid_disks
= mddev
->raid_disks
;
4874 if (mddev
->reshape_position
== MaxSector
)
4875 conf
->previous_raid_disks
= mddev
->raid_disks
;
4877 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4878 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
4879 conf
->scribble_len
= scribble_len(max_disks
);
4881 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
4886 conf
->mddev
= mddev
;
4888 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
4891 conf
->level
= mddev
->new_level
;
4892 if (raid5_alloc_percpu(conf
) != 0)
4895 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
4897 rdev_for_each(rdev
, mddev
) {
4898 raid_disk
= rdev
->raid_disk
;
4899 if (raid_disk
>= max_disks
4902 disk
= conf
->disks
+ raid_disk
;
4904 if (test_bit(Replacement
, &rdev
->flags
)) {
4905 if (disk
->replacement
)
4907 disk
->replacement
= rdev
;
4914 if (test_bit(In_sync
, &rdev
->flags
)) {
4915 char b
[BDEVNAME_SIZE
];
4916 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
4918 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
4919 } else if (rdev
->saved_raid_disk
!= raid_disk
)
4920 /* Cannot rely on bitmap to complete recovery */
4924 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
4925 conf
->level
= mddev
->new_level
;
4926 if (conf
->level
== 6)
4927 conf
->max_degraded
= 2;
4929 conf
->max_degraded
= 1;
4930 conf
->algorithm
= mddev
->new_layout
;
4931 conf
->max_nr_stripes
= NR_STRIPES
;
4932 conf
->reshape_progress
= mddev
->reshape_position
;
4933 if (conf
->reshape_progress
!= MaxSector
) {
4934 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
4935 conf
->prev_algo
= mddev
->layout
;
4938 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
4939 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
4940 if (grow_stripes(conf
, conf
->max_nr_stripes
)) {
4942 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4943 mdname(mddev
), memory
);
4946 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
4947 mdname(mddev
), memory
);
4949 conf
->thread
= md_register_thread(raid5d
, mddev
, NULL
);
4950 if (!conf
->thread
) {
4952 "md/raid:%s: couldn't allocate thread.\n",
4962 return ERR_PTR(-EIO
);
4964 return ERR_PTR(-ENOMEM
);
4968 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
4971 case ALGORITHM_PARITY_0
:
4972 if (raid_disk
< max_degraded
)
4975 case ALGORITHM_PARITY_N
:
4976 if (raid_disk
>= raid_disks
- max_degraded
)
4979 case ALGORITHM_PARITY_0_6
:
4980 if (raid_disk
== 0 ||
4981 raid_disk
== raid_disks
- 1)
4984 case ALGORITHM_LEFT_ASYMMETRIC_6
:
4985 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
4986 case ALGORITHM_LEFT_SYMMETRIC_6
:
4987 case ALGORITHM_RIGHT_SYMMETRIC_6
:
4988 if (raid_disk
== raid_disks
- 1)
4994 static int run(struct mddev
*mddev
)
4996 struct r5conf
*conf
;
4997 int working_disks
= 0;
4998 int dirty_parity_disks
= 0;
4999 struct md_rdev
*rdev
;
5000 sector_t reshape_offset
= 0;
5002 long long min_offset_diff
= 0;
5005 if (mddev
->recovery_cp
!= MaxSector
)
5006 printk(KERN_NOTICE
"md/raid:%s: not clean"
5007 " -- starting background reconstruction\n",
5010 rdev_for_each(rdev
, mddev
) {
5012 if (rdev
->raid_disk
< 0)
5014 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
5016 min_offset_diff
= diff
;
5018 } else if (mddev
->reshape_backwards
&&
5019 diff
< min_offset_diff
)
5020 min_offset_diff
= diff
;
5021 else if (!mddev
->reshape_backwards
&&
5022 diff
> min_offset_diff
)
5023 min_offset_diff
= diff
;
5026 if (mddev
->reshape_position
!= MaxSector
) {
5027 /* Check that we can continue the reshape.
5028 * Difficulties arise if the stripe we would write to
5029 * next is at or after the stripe we would read from next.
5030 * For a reshape that changes the number of devices, this
5031 * is only possible for a very short time, and mdadm makes
5032 * sure that time appears to have past before assembling
5033 * the array. So we fail if that time hasn't passed.
5034 * For a reshape that keeps the number of devices the same
5035 * mdadm must be monitoring the reshape can keeping the
5036 * critical areas read-only and backed up. It will start
5037 * the array in read-only mode, so we check for that.
5039 sector_t here_new
, here_old
;
5041 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
5043 if (mddev
->new_level
!= mddev
->level
) {
5044 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
5045 "required - aborting.\n",
5049 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
5050 /* reshape_position must be on a new-stripe boundary, and one
5051 * further up in new geometry must map after here in old
5054 here_new
= mddev
->reshape_position
;
5055 if (sector_div(here_new
, mddev
->new_chunk_sectors
*
5056 (mddev
->raid_disks
- max_degraded
))) {
5057 printk(KERN_ERR
"md/raid:%s: reshape_position not "
5058 "on a stripe boundary\n", mdname(mddev
));
5061 reshape_offset
= here_new
* mddev
->new_chunk_sectors
;
5062 /* here_new is the stripe we will write to */
5063 here_old
= mddev
->reshape_position
;
5064 sector_div(here_old
, mddev
->chunk_sectors
*
5065 (old_disks
-max_degraded
));
5066 /* here_old is the first stripe that we might need to read
5068 if (mddev
->delta_disks
== 0) {
5069 if ((here_new
* mddev
->new_chunk_sectors
!=
5070 here_old
* mddev
->chunk_sectors
)) {
5071 printk(KERN_ERR
"md/raid:%s: reshape position is"
5072 " confused - aborting\n", mdname(mddev
));
5075 /* We cannot be sure it is safe to start an in-place
5076 * reshape. It is only safe if user-space is monitoring
5077 * and taking constant backups.
5078 * mdadm always starts a situation like this in
5079 * readonly mode so it can take control before
5080 * allowing any writes. So just check for that.
5082 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
5083 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
5084 /* not really in-place - so OK */;
5085 else if (mddev
->ro
== 0) {
5086 printk(KERN_ERR
"md/raid:%s: in-place reshape "
5087 "must be started in read-only mode "
5092 } else if (mddev
->reshape_backwards
5093 ? (here_new
* mddev
->new_chunk_sectors
+ min_offset_diff
<=
5094 here_old
* mddev
->chunk_sectors
)
5095 : (here_new
* mddev
->new_chunk_sectors
>=
5096 here_old
* mddev
->chunk_sectors
+ (-min_offset_diff
))) {
5097 /* Reading from the same stripe as writing to - bad */
5098 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
5099 "auto-recovery - aborting.\n",
5103 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
5105 /* OK, we should be able to continue; */
5107 BUG_ON(mddev
->level
!= mddev
->new_level
);
5108 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
5109 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
5110 BUG_ON(mddev
->delta_disks
!= 0);
5113 if (mddev
->private == NULL
)
5114 conf
= setup_conf(mddev
);
5116 conf
= mddev
->private;
5119 return PTR_ERR(conf
);
5121 conf
->min_offset_diff
= min_offset_diff
;
5122 mddev
->thread
= conf
->thread
;
5123 conf
->thread
= NULL
;
5124 mddev
->private = conf
;
5126 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
5128 rdev
= conf
->disks
[i
].rdev
;
5129 if (!rdev
&& conf
->disks
[i
].replacement
) {
5130 /* The replacement is all we have yet */
5131 rdev
= conf
->disks
[i
].replacement
;
5132 conf
->disks
[i
].replacement
= NULL
;
5133 clear_bit(Replacement
, &rdev
->flags
);
5134 conf
->disks
[i
].rdev
= rdev
;
5138 if (conf
->disks
[i
].replacement
&&
5139 conf
->reshape_progress
!= MaxSector
) {
5140 /* replacements and reshape simply do not mix. */
5141 printk(KERN_ERR
"md: cannot handle concurrent "
5142 "replacement and reshape.\n");
5145 if (test_bit(In_sync
, &rdev
->flags
)) {
5149 /* This disc is not fully in-sync. However if it
5150 * just stored parity (beyond the recovery_offset),
5151 * when we don't need to be concerned about the
5152 * array being dirty.
5153 * When reshape goes 'backwards', we never have
5154 * partially completed devices, so we only need
5155 * to worry about reshape going forwards.
5157 /* Hack because v0.91 doesn't store recovery_offset properly. */
5158 if (mddev
->major_version
== 0 &&
5159 mddev
->minor_version
> 90)
5160 rdev
->recovery_offset
= reshape_offset
;
5162 if (rdev
->recovery_offset
< reshape_offset
) {
5163 /* We need to check old and new layout */
5164 if (!only_parity(rdev
->raid_disk
,
5167 conf
->max_degraded
))
5170 if (!only_parity(rdev
->raid_disk
,
5172 conf
->previous_raid_disks
,
5173 conf
->max_degraded
))
5175 dirty_parity_disks
++;
5179 * 0 for a fully functional array, 1 or 2 for a degraded array.
5181 mddev
->degraded
= calc_degraded(conf
);
5183 if (has_failed(conf
)) {
5184 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
5185 " (%d/%d failed)\n",
5186 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
5190 /* device size must be a multiple of chunk size */
5191 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
5192 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
5194 if (mddev
->degraded
> dirty_parity_disks
&&
5195 mddev
->recovery_cp
!= MaxSector
) {
5196 if (mddev
->ok_start_degraded
)
5198 "md/raid:%s: starting dirty degraded array"
5199 " - data corruption possible.\n",
5203 "md/raid:%s: cannot start dirty degraded array.\n",
5209 if (mddev
->degraded
== 0)
5210 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
5211 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
5212 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
5215 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
5216 " out of %d devices, algorithm %d\n",
5217 mdname(mddev
), conf
->level
,
5218 mddev
->raid_disks
- mddev
->degraded
,
5219 mddev
->raid_disks
, mddev
->new_layout
);
5221 print_raid5_conf(conf
);
5223 if (conf
->reshape_progress
!= MaxSector
) {
5224 conf
->reshape_safe
= conf
->reshape_progress
;
5225 atomic_set(&conf
->reshape_stripes
, 0);
5226 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5227 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5228 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5229 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5230 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5235 /* Ok, everything is just fine now */
5236 if (mddev
->to_remove
== &raid5_attrs_group
)
5237 mddev
->to_remove
= NULL
;
5238 else if (mddev
->kobj
.sd
&&
5239 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
5241 "raid5: failed to create sysfs attributes for %s\n",
5243 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5247 /* read-ahead size must cover two whole stripes, which
5248 * is 2 * (datadisks) * chunksize where 'n' is the
5249 * number of raid devices
5251 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
5252 int stripe
= data_disks
*
5253 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
5254 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
5255 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
5257 blk_queue_merge_bvec(mddev
->queue
, raid5_mergeable_bvec
);
5259 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
5260 mddev
->queue
->backing_dev_info
.congested_fn
= raid5_congested
;
5262 chunk_size
= mddev
->chunk_sectors
<< 9;
5263 blk_queue_io_min(mddev
->queue
, chunk_size
);
5264 blk_queue_io_opt(mddev
->queue
, chunk_size
*
5265 (conf
->raid_disks
- conf
->max_degraded
));
5267 rdev_for_each(rdev
, mddev
) {
5268 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
5269 rdev
->data_offset
<< 9);
5270 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
5271 rdev
->new_data_offset
<< 9);
5277 md_unregister_thread(&mddev
->thread
);
5278 print_raid5_conf(conf
);
5280 mddev
->private = NULL
;
5281 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
5285 static int stop(struct mddev
*mddev
)
5287 struct r5conf
*conf
= mddev
->private;
5289 md_unregister_thread(&mddev
->thread
);
5291 mddev
->queue
->backing_dev_info
.congested_fn
= NULL
;
5293 mddev
->private = NULL
;
5294 mddev
->to_remove
= &raid5_attrs_group
;
5298 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
5300 struct r5conf
*conf
= mddev
->private;
5303 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
5304 mddev
->chunk_sectors
/ 2, mddev
->layout
);
5305 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
5306 for (i
= 0; i
< conf
->raid_disks
; i
++)
5307 seq_printf (seq
, "%s",
5308 conf
->disks
[i
].rdev
&&
5309 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
5310 seq_printf (seq
, "]");
5313 static void print_raid5_conf (struct r5conf
*conf
)
5316 struct disk_info
*tmp
;
5318 printk(KERN_DEBUG
"RAID conf printout:\n");
5320 printk("(conf==NULL)\n");
5323 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
5325 conf
->raid_disks
- conf
->mddev
->degraded
);
5327 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5328 char b
[BDEVNAME_SIZE
];
5329 tmp
= conf
->disks
+ i
;
5331 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
5332 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
5333 bdevname(tmp
->rdev
->bdev
, b
));
5337 static int raid5_spare_active(struct mddev
*mddev
)
5340 struct r5conf
*conf
= mddev
->private;
5341 struct disk_info
*tmp
;
5343 unsigned long flags
;
5345 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5346 tmp
= conf
->disks
+ i
;
5347 if (tmp
->replacement
5348 && tmp
->replacement
->recovery_offset
== MaxSector
5349 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
5350 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
5351 /* Replacement has just become active. */
5353 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
5356 /* Replaced device not technically faulty,
5357 * but we need to be sure it gets removed
5358 * and never re-added.
5360 set_bit(Faulty
, &tmp
->rdev
->flags
);
5361 sysfs_notify_dirent_safe(
5362 tmp
->rdev
->sysfs_state
);
5364 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
5365 } else if (tmp
->rdev
5366 && tmp
->rdev
->recovery_offset
== MaxSector
5367 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
5368 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
5370 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
5373 spin_lock_irqsave(&conf
->device_lock
, flags
);
5374 mddev
->degraded
= calc_degraded(conf
);
5375 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5376 print_raid5_conf(conf
);
5380 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
5382 struct r5conf
*conf
= mddev
->private;
5384 int number
= rdev
->raid_disk
;
5385 struct md_rdev
**rdevp
;
5386 struct disk_info
*p
= conf
->disks
+ number
;
5388 print_raid5_conf(conf
);
5389 if (rdev
== p
->rdev
)
5391 else if (rdev
== p
->replacement
)
5392 rdevp
= &p
->replacement
;
5396 if (number
>= conf
->raid_disks
&&
5397 conf
->reshape_progress
== MaxSector
)
5398 clear_bit(In_sync
, &rdev
->flags
);
5400 if (test_bit(In_sync
, &rdev
->flags
) ||
5401 atomic_read(&rdev
->nr_pending
)) {
5405 /* Only remove non-faulty devices if recovery
5408 if (!test_bit(Faulty
, &rdev
->flags
) &&
5409 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
5410 !has_failed(conf
) &&
5411 (!p
->replacement
|| p
->replacement
== rdev
) &&
5412 number
< conf
->raid_disks
) {
5418 if (atomic_read(&rdev
->nr_pending
)) {
5419 /* lost the race, try later */
5422 } else if (p
->replacement
) {
5423 /* We must have just cleared 'rdev' */
5424 p
->rdev
= p
->replacement
;
5425 clear_bit(Replacement
, &p
->replacement
->flags
);
5426 smp_mb(); /* Make sure other CPUs may see both as identical
5427 * but will never see neither - if they are careful
5429 p
->replacement
= NULL
;
5430 clear_bit(WantReplacement
, &rdev
->flags
);
5432 /* We might have just removed the Replacement as faulty-
5433 * clear the bit just in case
5435 clear_bit(WantReplacement
, &rdev
->flags
);
5438 print_raid5_conf(conf
);
5442 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
5444 struct r5conf
*conf
= mddev
->private;
5447 struct disk_info
*p
;
5449 int last
= conf
->raid_disks
- 1;
5451 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
5454 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
5455 /* no point adding a device */
5458 if (rdev
->raid_disk
>= 0)
5459 first
= last
= rdev
->raid_disk
;
5462 * find the disk ... but prefer rdev->saved_raid_disk
5465 if (rdev
->saved_raid_disk
>= 0 &&
5466 rdev
->saved_raid_disk
>= first
&&
5467 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
5468 disk
= rdev
->saved_raid_disk
;
5471 for ( ; disk
<= last
; disk
++) {
5472 p
= conf
->disks
+ disk
;
5473 if (p
->rdev
== NULL
) {
5474 clear_bit(In_sync
, &rdev
->flags
);
5475 rdev
->raid_disk
= disk
;
5477 if (rdev
->saved_raid_disk
!= disk
)
5479 rcu_assign_pointer(p
->rdev
, rdev
);
5482 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
5483 p
->replacement
== NULL
) {
5484 clear_bit(In_sync
, &rdev
->flags
);
5485 set_bit(Replacement
, &rdev
->flags
);
5486 rdev
->raid_disk
= disk
;
5489 rcu_assign_pointer(p
->replacement
, rdev
);
5493 print_raid5_conf(conf
);
5497 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
5499 /* no resync is happening, and there is enough space
5500 * on all devices, so we can resize.
5501 * We need to make sure resync covers any new space.
5502 * If the array is shrinking we should possibly wait until
5503 * any io in the removed space completes, but it hardly seems
5507 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
5508 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
5509 if (mddev
->external_size
&&
5510 mddev
->array_sectors
> newsize
)
5512 if (mddev
->bitmap
) {
5513 int ret
= bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
5517 md_set_array_sectors(mddev
, newsize
);
5518 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5519 revalidate_disk(mddev
->gendisk
);
5520 if (sectors
> mddev
->dev_sectors
&&
5521 mddev
->recovery_cp
> mddev
->dev_sectors
) {
5522 mddev
->recovery_cp
= mddev
->dev_sectors
;
5523 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
5525 mddev
->dev_sectors
= sectors
;
5526 mddev
->resync_max_sectors
= sectors
;
5530 static int check_stripe_cache(struct mddev
*mddev
)
5532 /* Can only proceed if there are plenty of stripe_heads.
5533 * We need a minimum of one full stripe,, and for sensible progress
5534 * it is best to have about 4 times that.
5535 * If we require 4 times, then the default 256 4K stripe_heads will
5536 * allow for chunk sizes up to 256K, which is probably OK.
5537 * If the chunk size is greater, user-space should request more
5538 * stripe_heads first.
5540 struct r5conf
*conf
= mddev
->private;
5541 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5542 > conf
->max_nr_stripes
||
5543 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5544 > conf
->max_nr_stripes
) {
5545 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5547 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
5554 static int check_reshape(struct mddev
*mddev
)
5556 struct r5conf
*conf
= mddev
->private;
5558 if (mddev
->delta_disks
== 0 &&
5559 mddev
->new_layout
== mddev
->layout
&&
5560 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
5561 return 0; /* nothing to do */
5562 if (has_failed(conf
))
5564 if (mddev
->delta_disks
< 0) {
5565 /* We might be able to shrink, but the devices must
5566 * be made bigger first.
5567 * For raid6, 4 is the minimum size.
5568 * Otherwise 2 is the minimum
5571 if (mddev
->level
== 6)
5573 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
5577 if (!check_stripe_cache(mddev
))
5580 return resize_stripes(conf
, conf
->raid_disks
+ mddev
->delta_disks
);
5583 static int raid5_start_reshape(struct mddev
*mddev
)
5585 struct r5conf
*conf
= mddev
->private;
5586 struct md_rdev
*rdev
;
5588 unsigned long flags
;
5590 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
5593 if (!check_stripe_cache(mddev
))
5596 if (has_failed(conf
))
5599 rdev_for_each(rdev
, mddev
) {
5600 if (!test_bit(In_sync
, &rdev
->flags
)
5601 && !test_bit(Faulty
, &rdev
->flags
))
5605 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
5606 /* Not enough devices even to make a degraded array
5611 /* Refuse to reduce size of the array. Any reductions in
5612 * array size must be through explicit setting of array_size
5615 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
5616 < mddev
->array_sectors
) {
5617 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
5618 "before number of disks\n", mdname(mddev
));
5622 atomic_set(&conf
->reshape_stripes
, 0);
5623 spin_lock_irq(&conf
->device_lock
);
5624 conf
->previous_raid_disks
= conf
->raid_disks
;
5625 conf
->raid_disks
+= mddev
->delta_disks
;
5626 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
5627 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
5628 conf
->prev_algo
= conf
->algorithm
;
5629 conf
->algorithm
= mddev
->new_layout
;
5631 /* Code that selects data_offset needs to see the generation update
5632 * if reshape_progress has been set - so a memory barrier needed.
5635 if (mddev
->reshape_backwards
)
5636 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
5638 conf
->reshape_progress
= 0;
5639 conf
->reshape_safe
= conf
->reshape_progress
;
5640 spin_unlock_irq(&conf
->device_lock
);
5642 /* Add some new drives, as many as will fit.
5643 * We know there are enough to make the newly sized array work.
5644 * Don't add devices if we are reducing the number of
5645 * devices in the array. This is because it is not possible
5646 * to correctly record the "partially reconstructed" state of
5647 * such devices during the reshape and confusion could result.
5649 if (mddev
->delta_disks
>= 0) {
5650 rdev_for_each(rdev
, mddev
)
5651 if (rdev
->raid_disk
< 0 &&
5652 !test_bit(Faulty
, &rdev
->flags
)) {
5653 if (raid5_add_disk(mddev
, rdev
) == 0) {
5655 >= conf
->previous_raid_disks
)
5656 set_bit(In_sync
, &rdev
->flags
);
5658 rdev
->recovery_offset
= 0;
5660 if (sysfs_link_rdev(mddev
, rdev
))
5661 /* Failure here is OK */;
5663 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
5664 && !test_bit(Faulty
, &rdev
->flags
)) {
5665 /* This is a spare that was manually added */
5666 set_bit(In_sync
, &rdev
->flags
);
5669 /* When a reshape changes the number of devices,
5670 * ->degraded is measured against the larger of the
5671 * pre and post number of devices.
5673 spin_lock_irqsave(&conf
->device_lock
, flags
);
5674 mddev
->degraded
= calc_degraded(conf
);
5675 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5677 mddev
->raid_disks
= conf
->raid_disks
;
5678 mddev
->reshape_position
= conf
->reshape_progress
;
5679 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5681 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5682 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5683 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5684 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5685 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5687 if (!mddev
->sync_thread
) {
5688 mddev
->recovery
= 0;
5689 spin_lock_irq(&conf
->device_lock
);
5690 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
5691 rdev_for_each(rdev
, mddev
)
5692 rdev
->new_data_offset
= rdev
->data_offset
;
5694 conf
->reshape_progress
= MaxSector
;
5695 mddev
->reshape_position
= MaxSector
;
5696 spin_unlock_irq(&conf
->device_lock
);
5699 conf
->reshape_checkpoint
= jiffies
;
5700 md_wakeup_thread(mddev
->sync_thread
);
5701 md_new_event(mddev
);
5705 /* This is called from the reshape thread and should make any
5706 * changes needed in 'conf'
5708 static void end_reshape(struct r5conf
*conf
)
5711 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
5712 struct md_rdev
*rdev
;
5714 spin_lock_irq(&conf
->device_lock
);
5715 conf
->previous_raid_disks
= conf
->raid_disks
;
5716 rdev_for_each(rdev
, conf
->mddev
)
5717 rdev
->data_offset
= rdev
->new_data_offset
;
5719 conf
->reshape_progress
= MaxSector
;
5720 spin_unlock_irq(&conf
->device_lock
);
5721 wake_up(&conf
->wait_for_overlap
);
5723 /* read-ahead size must cover two whole stripes, which is
5724 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5726 if (conf
->mddev
->queue
) {
5727 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5728 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
5730 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
5731 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
5736 /* This is called from the raid5d thread with mddev_lock held.
5737 * It makes config changes to the device.
5739 static void raid5_finish_reshape(struct mddev
*mddev
)
5741 struct r5conf
*conf
= mddev
->private;
5743 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
5745 if (mddev
->delta_disks
> 0) {
5746 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5747 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5748 revalidate_disk(mddev
->gendisk
);
5751 spin_lock_irq(&conf
->device_lock
);
5752 mddev
->degraded
= calc_degraded(conf
);
5753 spin_unlock_irq(&conf
->device_lock
);
5754 for (d
= conf
->raid_disks
;
5755 d
< conf
->raid_disks
- mddev
->delta_disks
;
5757 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
5759 clear_bit(In_sync
, &rdev
->flags
);
5760 rdev
= conf
->disks
[d
].replacement
;
5762 clear_bit(In_sync
, &rdev
->flags
);
5765 mddev
->layout
= conf
->algorithm
;
5766 mddev
->chunk_sectors
= conf
->chunk_sectors
;
5767 mddev
->reshape_position
= MaxSector
;
5768 mddev
->delta_disks
= 0;
5769 mddev
->reshape_backwards
= 0;
5773 static void raid5_quiesce(struct mddev
*mddev
, int state
)
5775 struct r5conf
*conf
= mddev
->private;
5778 case 2: /* resume for a suspend */
5779 wake_up(&conf
->wait_for_overlap
);
5782 case 1: /* stop all writes */
5783 spin_lock_irq(&conf
->device_lock
);
5784 /* '2' tells resync/reshape to pause so that all
5785 * active stripes can drain
5788 wait_event_lock_irq(conf
->wait_for_stripe
,
5789 atomic_read(&conf
->active_stripes
) == 0 &&
5790 atomic_read(&conf
->active_aligned_reads
) == 0,
5791 conf
->device_lock
, /* nothing */);
5793 spin_unlock_irq(&conf
->device_lock
);
5794 /* allow reshape to continue */
5795 wake_up(&conf
->wait_for_overlap
);
5798 case 0: /* re-enable writes */
5799 spin_lock_irq(&conf
->device_lock
);
5801 wake_up(&conf
->wait_for_stripe
);
5802 wake_up(&conf
->wait_for_overlap
);
5803 spin_unlock_irq(&conf
->device_lock
);
5809 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
5811 struct r0conf
*raid0_conf
= mddev
->private;
5814 /* for raid0 takeover only one zone is supported */
5815 if (raid0_conf
->nr_strip_zones
> 1) {
5816 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5818 return ERR_PTR(-EINVAL
);
5821 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
5822 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
5823 mddev
->dev_sectors
= sectors
;
5824 mddev
->new_level
= level
;
5825 mddev
->new_layout
= ALGORITHM_PARITY_N
;
5826 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
5827 mddev
->raid_disks
+= 1;
5828 mddev
->delta_disks
= 1;
5829 /* make sure it will be not marked as dirty */
5830 mddev
->recovery_cp
= MaxSector
;
5832 return setup_conf(mddev
);
5836 static void *raid5_takeover_raid1(struct mddev
*mddev
)
5840 if (mddev
->raid_disks
!= 2 ||
5841 mddev
->degraded
> 1)
5842 return ERR_PTR(-EINVAL
);
5844 /* Should check if there are write-behind devices? */
5846 chunksect
= 64*2; /* 64K by default */
5848 /* The array must be an exact multiple of chunksize */
5849 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
5852 if ((chunksect
<<9) < STRIPE_SIZE
)
5853 /* array size does not allow a suitable chunk size */
5854 return ERR_PTR(-EINVAL
);
5856 mddev
->new_level
= 5;
5857 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5858 mddev
->new_chunk_sectors
= chunksect
;
5860 return setup_conf(mddev
);
5863 static void *raid5_takeover_raid6(struct mddev
*mddev
)
5867 switch (mddev
->layout
) {
5868 case ALGORITHM_LEFT_ASYMMETRIC_6
:
5869 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
5871 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
5872 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
5874 case ALGORITHM_LEFT_SYMMETRIC_6
:
5875 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5877 case ALGORITHM_RIGHT_SYMMETRIC_6
:
5878 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
5880 case ALGORITHM_PARITY_0_6
:
5881 new_layout
= ALGORITHM_PARITY_0
;
5883 case ALGORITHM_PARITY_N
:
5884 new_layout
= ALGORITHM_PARITY_N
;
5887 return ERR_PTR(-EINVAL
);
5889 mddev
->new_level
= 5;
5890 mddev
->new_layout
= new_layout
;
5891 mddev
->delta_disks
= -1;
5892 mddev
->raid_disks
-= 1;
5893 return setup_conf(mddev
);
5897 static int raid5_check_reshape(struct mddev
*mddev
)
5899 /* For a 2-drive array, the layout and chunk size can be changed
5900 * immediately as not restriping is needed.
5901 * For larger arrays we record the new value - after validation
5902 * to be used by a reshape pass.
5904 struct r5conf
*conf
= mddev
->private;
5905 int new_chunk
= mddev
->new_chunk_sectors
;
5907 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
5909 if (new_chunk
> 0) {
5910 if (!is_power_of_2(new_chunk
))
5912 if (new_chunk
< (PAGE_SIZE
>>9))
5914 if (mddev
->array_sectors
& (new_chunk
-1))
5915 /* not factor of array size */
5919 /* They look valid */
5921 if (mddev
->raid_disks
== 2) {
5922 /* can make the change immediately */
5923 if (mddev
->new_layout
>= 0) {
5924 conf
->algorithm
= mddev
->new_layout
;
5925 mddev
->layout
= mddev
->new_layout
;
5927 if (new_chunk
> 0) {
5928 conf
->chunk_sectors
= new_chunk
;
5929 mddev
->chunk_sectors
= new_chunk
;
5931 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5932 md_wakeup_thread(mddev
->thread
);
5934 return check_reshape(mddev
);
5937 static int raid6_check_reshape(struct mddev
*mddev
)
5939 int new_chunk
= mddev
->new_chunk_sectors
;
5941 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
5943 if (new_chunk
> 0) {
5944 if (!is_power_of_2(new_chunk
))
5946 if (new_chunk
< (PAGE_SIZE
>> 9))
5948 if (mddev
->array_sectors
& (new_chunk
-1))
5949 /* not factor of array size */
5953 /* They look valid */
5954 return check_reshape(mddev
);
5957 static void *raid5_takeover(struct mddev
*mddev
)
5959 /* raid5 can take over:
5960 * raid0 - if there is only one strip zone - make it a raid4 layout
5961 * raid1 - if there are two drives. We need to know the chunk size
5962 * raid4 - trivial - just use a raid4 layout.
5963 * raid6 - Providing it is a *_6 layout
5965 if (mddev
->level
== 0)
5966 return raid45_takeover_raid0(mddev
, 5);
5967 if (mddev
->level
== 1)
5968 return raid5_takeover_raid1(mddev
);
5969 if (mddev
->level
== 4) {
5970 mddev
->new_layout
= ALGORITHM_PARITY_N
;
5971 mddev
->new_level
= 5;
5972 return setup_conf(mddev
);
5974 if (mddev
->level
== 6)
5975 return raid5_takeover_raid6(mddev
);
5977 return ERR_PTR(-EINVAL
);
5980 static void *raid4_takeover(struct mddev
*mddev
)
5982 /* raid4 can take over:
5983 * raid0 - if there is only one strip zone
5984 * raid5 - if layout is right
5986 if (mddev
->level
== 0)
5987 return raid45_takeover_raid0(mddev
, 4);
5988 if (mddev
->level
== 5 &&
5989 mddev
->layout
== ALGORITHM_PARITY_N
) {
5990 mddev
->new_layout
= 0;
5991 mddev
->new_level
= 4;
5992 return setup_conf(mddev
);
5994 return ERR_PTR(-EINVAL
);
5997 static struct md_personality raid5_personality
;
5999 static void *raid6_takeover(struct mddev
*mddev
)
6001 /* Currently can only take over a raid5. We map the
6002 * personality to an equivalent raid6 personality
6003 * with the Q block at the end.
6007 if (mddev
->pers
!= &raid5_personality
)
6008 return ERR_PTR(-EINVAL
);
6009 if (mddev
->degraded
> 1)
6010 return ERR_PTR(-EINVAL
);
6011 if (mddev
->raid_disks
> 253)
6012 return ERR_PTR(-EINVAL
);
6013 if (mddev
->raid_disks
< 3)
6014 return ERR_PTR(-EINVAL
);
6016 switch (mddev
->layout
) {
6017 case ALGORITHM_LEFT_ASYMMETRIC
:
6018 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
6020 case ALGORITHM_RIGHT_ASYMMETRIC
:
6021 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
6023 case ALGORITHM_LEFT_SYMMETRIC
:
6024 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
6026 case ALGORITHM_RIGHT_SYMMETRIC
:
6027 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
6029 case ALGORITHM_PARITY_0
:
6030 new_layout
= ALGORITHM_PARITY_0_6
;
6032 case ALGORITHM_PARITY_N
:
6033 new_layout
= ALGORITHM_PARITY_N
;
6036 return ERR_PTR(-EINVAL
);
6038 mddev
->new_level
= 6;
6039 mddev
->new_layout
= new_layout
;
6040 mddev
->delta_disks
= 1;
6041 mddev
->raid_disks
+= 1;
6042 return setup_conf(mddev
);
6046 static struct md_personality raid6_personality
=
6050 .owner
= THIS_MODULE
,
6051 .make_request
= make_request
,
6055 .error_handler
= error
,
6056 .hot_add_disk
= raid5_add_disk
,
6057 .hot_remove_disk
= raid5_remove_disk
,
6058 .spare_active
= raid5_spare_active
,
6059 .sync_request
= sync_request
,
6060 .resize
= raid5_resize
,
6062 .check_reshape
= raid6_check_reshape
,
6063 .start_reshape
= raid5_start_reshape
,
6064 .finish_reshape
= raid5_finish_reshape
,
6065 .quiesce
= raid5_quiesce
,
6066 .takeover
= raid6_takeover
,
6068 static struct md_personality raid5_personality
=
6072 .owner
= THIS_MODULE
,
6073 .make_request
= make_request
,
6077 .error_handler
= error
,
6078 .hot_add_disk
= raid5_add_disk
,
6079 .hot_remove_disk
= raid5_remove_disk
,
6080 .spare_active
= raid5_spare_active
,
6081 .sync_request
= sync_request
,
6082 .resize
= raid5_resize
,
6084 .check_reshape
= raid5_check_reshape
,
6085 .start_reshape
= raid5_start_reshape
,
6086 .finish_reshape
= raid5_finish_reshape
,
6087 .quiesce
= raid5_quiesce
,
6088 .takeover
= raid5_takeover
,
6091 static struct md_personality raid4_personality
=
6095 .owner
= THIS_MODULE
,
6096 .make_request
= make_request
,
6100 .error_handler
= error
,
6101 .hot_add_disk
= raid5_add_disk
,
6102 .hot_remove_disk
= raid5_remove_disk
,
6103 .spare_active
= raid5_spare_active
,
6104 .sync_request
= sync_request
,
6105 .resize
= raid5_resize
,
6107 .check_reshape
= raid5_check_reshape
,
6108 .start_reshape
= raid5_start_reshape
,
6109 .finish_reshape
= raid5_finish_reshape
,
6110 .quiesce
= raid5_quiesce
,
6111 .takeover
= raid4_takeover
,
6114 static int __init
raid5_init(void)
6116 register_md_personality(&raid6_personality
);
6117 register_md_personality(&raid5_personality
);
6118 register_md_personality(&raid4_personality
);
6122 static void raid5_exit(void)
6124 unregister_md_personality(&raid6_personality
);
6125 unregister_md_personality(&raid5_personality
);
6126 unregister_md_personality(&raid4_personality
);
6129 module_init(raid5_init
);
6130 module_exit(raid5_exit
);
6131 MODULE_LICENSE("GPL");
6132 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6133 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6134 MODULE_ALIAS("md-raid5");
6135 MODULE_ALIAS("md-raid4");
6136 MODULE_ALIAS("md-level-5");
6137 MODULE_ALIAS("md-level-4");
6138 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6139 MODULE_ALIAS("md-raid6");
6140 MODULE_ALIAS("md-level-6");
6142 /* This used to be two separate modules, they were: */
6143 MODULE_ALIAS("raid5");
6144 MODULE_ALIAS("raid6");