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
);
492 raid5_end_read_request(struct bio
*bi
, int error
);
494 raid5_end_write_request(struct bio
*bi
, int error
);
496 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
498 struct r5conf
*conf
= sh
->raid_conf
;
499 int i
, disks
= sh
->disks
;
503 for (i
= disks
; i
--; ) {
505 int replace_only
= 0;
506 struct bio
*bi
, *rbi
;
507 struct md_rdev
*rdev
, *rrdev
= NULL
;
508 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
509 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
513 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
515 else if (test_and_clear_bit(R5_WantReplace
,
516 &sh
->dev
[i
].flags
)) {
522 bi
= &sh
->dev
[i
].req
;
523 rbi
= &sh
->dev
[i
].rreq
; /* For writing to replacement */
528 bi
->bi_end_io
= raid5_end_write_request
;
529 rbi
->bi_end_io
= raid5_end_write_request
;
531 bi
->bi_end_io
= raid5_end_read_request
;
534 rrdev
= rcu_dereference(conf
->disks
[i
].replacement
);
535 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
536 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
545 /* We raced and saw duplicates */
548 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
) && rrdev
)
553 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
556 atomic_inc(&rdev
->nr_pending
);
557 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
560 atomic_inc(&rrdev
->nr_pending
);
563 /* We have already checked bad blocks for reads. Now
564 * need to check for writes. We never accept write errors
565 * on the replacement, so we don't to check rrdev.
567 while ((rw
& WRITE
) && rdev
&&
568 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
571 int bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
572 &first_bad
, &bad_sectors
);
577 set_bit(BlockedBadBlocks
, &rdev
->flags
);
578 if (!conf
->mddev
->external
&&
579 conf
->mddev
->flags
) {
580 /* It is very unlikely, but we might
581 * still need to write out the
582 * bad block log - better give it
584 md_check_recovery(conf
->mddev
);
586 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
588 /* Acknowledged bad block - skip the write */
589 rdev_dec_pending(rdev
, conf
->mddev
);
595 if (s
->syncing
|| s
->expanding
|| s
->expanded
597 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
599 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
601 bi
->bi_bdev
= rdev
->bdev
;
602 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
603 __func__
, (unsigned long long)sh
->sector
,
605 atomic_inc(&sh
->count
);
606 bi
->bi_sector
= sh
->sector
+ rdev
->data_offset
;
607 bi
->bi_flags
= 1 << BIO_UPTODATE
;
609 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
610 bi
->bi_io_vec
[0].bv_offset
= 0;
611 bi
->bi_size
= STRIPE_SIZE
;
614 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
615 generic_make_request(bi
);
618 if (s
->syncing
|| s
->expanding
|| s
->expanded
620 md_sync_acct(rrdev
->bdev
, STRIPE_SECTORS
);
622 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
624 rbi
->bi_bdev
= rrdev
->bdev
;
625 pr_debug("%s: for %llu schedule op %ld on "
626 "replacement disc %d\n",
627 __func__
, (unsigned long long)sh
->sector
,
629 atomic_inc(&sh
->count
);
630 rbi
->bi_sector
= sh
->sector
+ rrdev
->data_offset
;
631 rbi
->bi_flags
= 1 << BIO_UPTODATE
;
633 rbi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
634 rbi
->bi_io_vec
[0].bv_offset
= 0;
635 rbi
->bi_size
= STRIPE_SIZE
;
637 generic_make_request(rbi
);
639 if (!rdev
&& !rrdev
) {
641 set_bit(STRIPE_DEGRADED
, &sh
->state
);
642 pr_debug("skip op %ld on disc %d for sector %llu\n",
643 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
644 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
645 set_bit(STRIPE_HANDLE
, &sh
->state
);
650 static struct dma_async_tx_descriptor
*
651 async_copy_data(int frombio
, struct bio
*bio
, struct page
*page
,
652 sector_t sector
, struct dma_async_tx_descriptor
*tx
)
655 struct page
*bio_page
;
658 struct async_submit_ctl submit
;
659 enum async_tx_flags flags
= 0;
661 if (bio
->bi_sector
>= sector
)
662 page_offset
= (signed)(bio
->bi_sector
- sector
) * 512;
664 page_offset
= (signed)(sector
- bio
->bi_sector
) * -512;
667 flags
|= ASYNC_TX_FENCE
;
668 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
670 bio_for_each_segment(bvl
, bio
, i
) {
671 int len
= bvl
->bv_len
;
675 if (page_offset
< 0) {
676 b_offset
= -page_offset
;
677 page_offset
+= b_offset
;
681 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
682 clen
= STRIPE_SIZE
- page_offset
;
687 b_offset
+= bvl
->bv_offset
;
688 bio_page
= bvl
->bv_page
;
690 tx
= async_memcpy(page
, bio_page
, page_offset
,
691 b_offset
, clen
, &submit
);
693 tx
= async_memcpy(bio_page
, page
, b_offset
,
694 page_offset
, clen
, &submit
);
696 /* chain the operations */
697 submit
.depend_tx
= tx
;
699 if (clen
< len
) /* hit end of page */
707 static void ops_complete_biofill(void *stripe_head_ref
)
709 struct stripe_head
*sh
= stripe_head_ref
;
710 struct bio
*return_bi
= NULL
;
711 struct r5conf
*conf
= sh
->raid_conf
;
714 pr_debug("%s: stripe %llu\n", __func__
,
715 (unsigned long long)sh
->sector
);
717 /* clear completed biofills */
718 spin_lock_irq(&conf
->device_lock
);
719 for (i
= sh
->disks
; i
--; ) {
720 struct r5dev
*dev
= &sh
->dev
[i
];
722 /* acknowledge completion of a biofill operation */
723 /* and check if we need to reply to a read request,
724 * new R5_Wantfill requests are held off until
725 * !STRIPE_BIOFILL_RUN
727 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
728 struct bio
*rbi
, *rbi2
;
733 while (rbi
&& rbi
->bi_sector
<
734 dev
->sector
+ STRIPE_SECTORS
) {
735 rbi2
= r5_next_bio(rbi
, dev
->sector
);
736 if (!raid5_dec_bi_phys_segments(rbi
)) {
737 rbi
->bi_next
= return_bi
;
744 spin_unlock_irq(&conf
->device_lock
);
745 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
747 return_io(return_bi
);
749 set_bit(STRIPE_HANDLE
, &sh
->state
);
753 static void ops_run_biofill(struct stripe_head
*sh
)
755 struct dma_async_tx_descriptor
*tx
= NULL
;
756 struct r5conf
*conf
= sh
->raid_conf
;
757 struct async_submit_ctl submit
;
760 pr_debug("%s: stripe %llu\n", __func__
,
761 (unsigned long long)sh
->sector
);
763 for (i
= sh
->disks
; i
--; ) {
764 struct r5dev
*dev
= &sh
->dev
[i
];
765 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
767 spin_lock_irq(&conf
->device_lock
);
768 dev
->read
= rbi
= dev
->toread
;
770 spin_unlock_irq(&conf
->device_lock
);
771 while (rbi
&& rbi
->bi_sector
<
772 dev
->sector
+ STRIPE_SECTORS
) {
773 tx
= async_copy_data(0, rbi
, dev
->page
,
775 rbi
= r5_next_bio(rbi
, dev
->sector
);
780 atomic_inc(&sh
->count
);
781 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
782 async_trigger_callback(&submit
);
785 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
792 tgt
= &sh
->dev
[target
];
793 set_bit(R5_UPTODATE
, &tgt
->flags
);
794 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
795 clear_bit(R5_Wantcompute
, &tgt
->flags
);
798 static void ops_complete_compute(void *stripe_head_ref
)
800 struct stripe_head
*sh
= stripe_head_ref
;
802 pr_debug("%s: stripe %llu\n", __func__
,
803 (unsigned long long)sh
->sector
);
805 /* mark the computed target(s) as uptodate */
806 mark_target_uptodate(sh
, sh
->ops
.target
);
807 mark_target_uptodate(sh
, sh
->ops
.target2
);
809 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
810 if (sh
->check_state
== check_state_compute_run
)
811 sh
->check_state
= check_state_compute_result
;
812 set_bit(STRIPE_HANDLE
, &sh
->state
);
816 /* return a pointer to the address conversion region of the scribble buffer */
817 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
818 struct raid5_percpu
*percpu
)
820 return percpu
->scribble
+ sizeof(struct page
*) * (sh
->disks
+ 2);
823 static struct dma_async_tx_descriptor
*
824 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
826 int disks
= sh
->disks
;
827 struct page
**xor_srcs
= percpu
->scribble
;
828 int target
= sh
->ops
.target
;
829 struct r5dev
*tgt
= &sh
->dev
[target
];
830 struct page
*xor_dest
= tgt
->page
;
832 struct dma_async_tx_descriptor
*tx
;
833 struct async_submit_ctl submit
;
836 pr_debug("%s: stripe %llu block: %d\n",
837 __func__
, (unsigned long long)sh
->sector
, target
);
838 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
840 for (i
= disks
; i
--; )
842 xor_srcs
[count
++] = sh
->dev
[i
].page
;
844 atomic_inc(&sh
->count
);
846 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
847 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
));
848 if (unlikely(count
== 1))
849 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
851 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
856 /* set_syndrome_sources - populate source buffers for gen_syndrome
857 * @srcs - (struct page *) array of size sh->disks
858 * @sh - stripe_head to parse
860 * Populates srcs in proper layout order for the stripe and returns the
861 * 'count' of sources to be used in a call to async_gen_syndrome. The P
862 * destination buffer is recorded in srcs[count] and the Q destination
863 * is recorded in srcs[count+1]].
865 static int set_syndrome_sources(struct page
**srcs
, struct stripe_head
*sh
)
867 int disks
= sh
->disks
;
868 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
869 int d0_idx
= raid6_d0(sh
);
873 for (i
= 0; i
< disks
; i
++)
879 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
881 srcs
[slot
] = sh
->dev
[i
].page
;
882 i
= raid6_next_disk(i
, disks
);
883 } while (i
!= d0_idx
);
885 return syndrome_disks
;
888 static struct dma_async_tx_descriptor
*
889 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
891 int disks
= sh
->disks
;
892 struct page
**blocks
= percpu
->scribble
;
894 int qd_idx
= sh
->qd_idx
;
895 struct dma_async_tx_descriptor
*tx
;
896 struct async_submit_ctl submit
;
902 if (sh
->ops
.target
< 0)
903 target
= sh
->ops
.target2
;
904 else if (sh
->ops
.target2
< 0)
905 target
= sh
->ops
.target
;
907 /* we should only have one valid target */
910 pr_debug("%s: stripe %llu block: %d\n",
911 __func__
, (unsigned long long)sh
->sector
, target
);
913 tgt
= &sh
->dev
[target
];
914 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
917 atomic_inc(&sh
->count
);
919 if (target
== qd_idx
) {
920 count
= set_syndrome_sources(blocks
, sh
);
921 blocks
[count
] = NULL
; /* regenerating p is not necessary */
922 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
923 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
924 ops_complete_compute
, sh
,
925 to_addr_conv(sh
, percpu
));
926 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
928 /* Compute any data- or p-drive using XOR */
930 for (i
= disks
; i
-- ; ) {
931 if (i
== target
|| i
== qd_idx
)
933 blocks
[count
++] = sh
->dev
[i
].page
;
936 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
937 NULL
, ops_complete_compute
, sh
,
938 to_addr_conv(sh
, percpu
));
939 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
945 static struct dma_async_tx_descriptor
*
946 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
948 int i
, count
, disks
= sh
->disks
;
949 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
950 int d0_idx
= raid6_d0(sh
);
951 int faila
= -1, failb
= -1;
952 int target
= sh
->ops
.target
;
953 int target2
= sh
->ops
.target2
;
954 struct r5dev
*tgt
= &sh
->dev
[target
];
955 struct r5dev
*tgt2
= &sh
->dev
[target2
];
956 struct dma_async_tx_descriptor
*tx
;
957 struct page
**blocks
= percpu
->scribble
;
958 struct async_submit_ctl submit
;
960 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
961 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
962 BUG_ON(target
< 0 || target2
< 0);
963 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
964 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
966 /* we need to open-code set_syndrome_sources to handle the
967 * slot number conversion for 'faila' and 'failb'
969 for (i
= 0; i
< disks
; i
++)
974 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
976 blocks
[slot
] = sh
->dev
[i
].page
;
982 i
= raid6_next_disk(i
, disks
);
983 } while (i
!= d0_idx
);
985 BUG_ON(faila
== failb
);
988 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
989 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
991 atomic_inc(&sh
->count
);
993 if (failb
== syndrome_disks
+1) {
994 /* Q disk is one of the missing disks */
995 if (faila
== syndrome_disks
) {
996 /* Missing P+Q, just recompute */
997 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
998 ops_complete_compute
, sh
,
999 to_addr_conv(sh
, percpu
));
1000 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
1001 STRIPE_SIZE
, &submit
);
1005 int qd_idx
= sh
->qd_idx
;
1007 /* Missing D+Q: recompute D from P, then recompute Q */
1008 if (target
== qd_idx
)
1009 data_target
= target2
;
1011 data_target
= target
;
1014 for (i
= disks
; i
-- ; ) {
1015 if (i
== data_target
|| i
== qd_idx
)
1017 blocks
[count
++] = sh
->dev
[i
].page
;
1019 dest
= sh
->dev
[data_target
].page
;
1020 init_async_submit(&submit
,
1021 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1023 to_addr_conv(sh
, percpu
));
1024 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
1027 count
= set_syndrome_sources(blocks
, sh
);
1028 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1029 ops_complete_compute
, sh
,
1030 to_addr_conv(sh
, percpu
));
1031 return async_gen_syndrome(blocks
, 0, count
+2,
1032 STRIPE_SIZE
, &submit
);
1035 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1036 ops_complete_compute
, sh
,
1037 to_addr_conv(sh
, percpu
));
1038 if (failb
== syndrome_disks
) {
1039 /* We're missing D+P. */
1040 return async_raid6_datap_recov(syndrome_disks
+2,
1044 /* We're missing D+D. */
1045 return async_raid6_2data_recov(syndrome_disks
+2,
1046 STRIPE_SIZE
, faila
, failb
,
1053 static void ops_complete_prexor(void *stripe_head_ref
)
1055 struct stripe_head
*sh
= stripe_head_ref
;
1057 pr_debug("%s: stripe %llu\n", __func__
,
1058 (unsigned long long)sh
->sector
);
1061 static struct dma_async_tx_descriptor
*
1062 ops_run_prexor(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1063 struct dma_async_tx_descriptor
*tx
)
1065 int disks
= sh
->disks
;
1066 struct page
**xor_srcs
= percpu
->scribble
;
1067 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1068 struct async_submit_ctl submit
;
1070 /* existing parity data subtracted */
1071 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1073 pr_debug("%s: stripe %llu\n", __func__
,
1074 (unsigned long long)sh
->sector
);
1076 for (i
= disks
; i
--; ) {
1077 struct r5dev
*dev
= &sh
->dev
[i
];
1078 /* Only process blocks that are known to be uptodate */
1079 if (test_bit(R5_Wantdrain
, &dev
->flags
))
1080 xor_srcs
[count
++] = dev
->page
;
1083 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1084 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
));
1085 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1090 static struct dma_async_tx_descriptor
*
1091 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1093 int disks
= sh
->disks
;
1096 pr_debug("%s: stripe %llu\n", __func__
,
1097 (unsigned long long)sh
->sector
);
1099 for (i
= disks
; i
--; ) {
1100 struct r5dev
*dev
= &sh
->dev
[i
];
1103 if (test_and_clear_bit(R5_Wantdrain
, &dev
->flags
)) {
1106 spin_lock_irq(&sh
->raid_conf
->device_lock
);
1107 chosen
= dev
->towrite
;
1108 dev
->towrite
= NULL
;
1109 BUG_ON(dev
->written
);
1110 wbi
= dev
->written
= chosen
;
1111 spin_unlock_irq(&sh
->raid_conf
->device_lock
);
1113 while (wbi
&& wbi
->bi_sector
<
1114 dev
->sector
+ STRIPE_SECTORS
) {
1115 if (wbi
->bi_rw
& REQ_FUA
)
1116 set_bit(R5_WantFUA
, &dev
->flags
);
1117 tx
= async_copy_data(1, wbi
, dev
->page
,
1119 wbi
= r5_next_bio(wbi
, dev
->sector
);
1127 static void ops_complete_reconstruct(void *stripe_head_ref
)
1129 struct stripe_head
*sh
= stripe_head_ref
;
1130 int disks
= sh
->disks
;
1131 int pd_idx
= sh
->pd_idx
;
1132 int qd_idx
= sh
->qd_idx
;
1136 pr_debug("%s: stripe %llu\n", __func__
,
1137 (unsigned long long)sh
->sector
);
1139 for (i
= disks
; i
--; )
1140 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1142 for (i
= disks
; i
--; ) {
1143 struct r5dev
*dev
= &sh
->dev
[i
];
1145 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1146 set_bit(R5_UPTODATE
, &dev
->flags
);
1148 set_bit(R5_WantFUA
, &dev
->flags
);
1152 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1153 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1154 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1155 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1157 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1158 sh
->reconstruct_state
= reconstruct_state_result
;
1161 set_bit(STRIPE_HANDLE
, &sh
->state
);
1166 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1167 struct dma_async_tx_descriptor
*tx
)
1169 int disks
= sh
->disks
;
1170 struct page
**xor_srcs
= percpu
->scribble
;
1171 struct async_submit_ctl submit
;
1172 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1173 struct page
*xor_dest
;
1175 unsigned long flags
;
1177 pr_debug("%s: stripe %llu\n", __func__
,
1178 (unsigned long long)sh
->sector
);
1180 /* check if prexor is active which means only process blocks
1181 * that are part of a read-modify-write (written)
1183 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1185 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1186 for (i
= disks
; i
--; ) {
1187 struct r5dev
*dev
= &sh
->dev
[i
];
1189 xor_srcs
[count
++] = dev
->page
;
1192 xor_dest
= sh
->dev
[pd_idx
].page
;
1193 for (i
= disks
; i
--; ) {
1194 struct r5dev
*dev
= &sh
->dev
[i
];
1196 xor_srcs
[count
++] = dev
->page
;
1200 /* 1/ if we prexor'd then the dest is reused as a source
1201 * 2/ if we did not prexor then we are redoing the parity
1202 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1203 * for the synchronous xor case
1205 flags
= ASYNC_TX_ACK
|
1206 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1208 atomic_inc(&sh
->count
);
1210 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, sh
,
1211 to_addr_conv(sh
, percpu
));
1212 if (unlikely(count
== 1))
1213 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1215 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1219 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1220 struct dma_async_tx_descriptor
*tx
)
1222 struct async_submit_ctl submit
;
1223 struct page
**blocks
= percpu
->scribble
;
1226 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1228 count
= set_syndrome_sources(blocks
, sh
);
1230 atomic_inc(&sh
->count
);
1232 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_reconstruct
,
1233 sh
, to_addr_conv(sh
, percpu
));
1234 async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1237 static void ops_complete_check(void *stripe_head_ref
)
1239 struct stripe_head
*sh
= stripe_head_ref
;
1241 pr_debug("%s: stripe %llu\n", __func__
,
1242 (unsigned long long)sh
->sector
);
1244 sh
->check_state
= check_state_check_result
;
1245 set_bit(STRIPE_HANDLE
, &sh
->state
);
1249 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1251 int disks
= sh
->disks
;
1252 int pd_idx
= sh
->pd_idx
;
1253 int qd_idx
= sh
->qd_idx
;
1254 struct page
*xor_dest
;
1255 struct page
**xor_srcs
= percpu
->scribble
;
1256 struct dma_async_tx_descriptor
*tx
;
1257 struct async_submit_ctl submit
;
1261 pr_debug("%s: stripe %llu\n", __func__
,
1262 (unsigned long long)sh
->sector
);
1265 xor_dest
= sh
->dev
[pd_idx
].page
;
1266 xor_srcs
[count
++] = xor_dest
;
1267 for (i
= disks
; i
--; ) {
1268 if (i
== pd_idx
|| i
== qd_idx
)
1270 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1273 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1274 to_addr_conv(sh
, percpu
));
1275 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1276 &sh
->ops
.zero_sum_result
, &submit
);
1278 atomic_inc(&sh
->count
);
1279 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1280 tx
= async_trigger_callback(&submit
);
1283 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1285 struct page
**srcs
= percpu
->scribble
;
1286 struct async_submit_ctl submit
;
1289 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1290 (unsigned long long)sh
->sector
, checkp
);
1292 count
= set_syndrome_sources(srcs
, sh
);
1296 atomic_inc(&sh
->count
);
1297 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1298 sh
, to_addr_conv(sh
, percpu
));
1299 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1300 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1303 static void __raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1305 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1306 struct dma_async_tx_descriptor
*tx
= NULL
;
1307 struct r5conf
*conf
= sh
->raid_conf
;
1308 int level
= conf
->level
;
1309 struct raid5_percpu
*percpu
;
1313 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1314 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1315 ops_run_biofill(sh
);
1319 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1321 tx
= ops_run_compute5(sh
, percpu
);
1323 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
1324 tx
= ops_run_compute6_1(sh
, percpu
);
1326 tx
= ops_run_compute6_2(sh
, percpu
);
1328 /* terminate the chain if reconstruct is not set to be run */
1329 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
1333 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
))
1334 tx
= ops_run_prexor(sh
, percpu
, tx
);
1336 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
1337 tx
= ops_run_biodrain(sh
, tx
);
1341 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
1343 ops_run_reconstruct5(sh
, percpu
, tx
);
1345 ops_run_reconstruct6(sh
, percpu
, tx
);
1348 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
1349 if (sh
->check_state
== check_state_run
)
1350 ops_run_check_p(sh
, percpu
);
1351 else if (sh
->check_state
== check_state_run_q
)
1352 ops_run_check_pq(sh
, percpu
, 0);
1353 else if (sh
->check_state
== check_state_run_pq
)
1354 ops_run_check_pq(sh
, percpu
, 1);
1360 for (i
= disks
; i
--; ) {
1361 struct r5dev
*dev
= &sh
->dev
[i
];
1362 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
1363 wake_up(&sh
->raid_conf
->wait_for_overlap
);
1368 #ifdef CONFIG_MULTICORE_RAID456
1369 static void async_run_ops(void *param
, async_cookie_t cookie
)
1371 struct stripe_head
*sh
= param
;
1372 unsigned long ops_request
= sh
->ops
.request
;
1374 clear_bit_unlock(STRIPE_OPS_REQ_PENDING
, &sh
->state
);
1375 wake_up(&sh
->ops
.wait_for_ops
);
1377 __raid_run_ops(sh
, ops_request
);
1381 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1383 /* since handle_stripe can be called outside of raid5d context
1384 * we need to ensure sh->ops.request is de-staged before another
1387 wait_event(sh
->ops
.wait_for_ops
,
1388 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING
, &sh
->state
));
1389 sh
->ops
.request
= ops_request
;
1391 atomic_inc(&sh
->count
);
1392 async_schedule(async_run_ops
, sh
);
1395 #define raid_run_ops __raid_run_ops
1398 static int grow_one_stripe(struct r5conf
*conf
)
1400 struct stripe_head
*sh
;
1401 sh
= kmem_cache_zalloc(conf
->slab_cache
, GFP_KERNEL
);
1405 sh
->raid_conf
= conf
;
1406 #ifdef CONFIG_MULTICORE_RAID456
1407 init_waitqueue_head(&sh
->ops
.wait_for_ops
);
1410 if (grow_buffers(sh
)) {
1412 kmem_cache_free(conf
->slab_cache
, sh
);
1415 /* we just created an active stripe so... */
1416 atomic_set(&sh
->count
, 1);
1417 atomic_inc(&conf
->active_stripes
);
1418 INIT_LIST_HEAD(&sh
->lru
);
1423 static int grow_stripes(struct r5conf
*conf
, int num
)
1425 struct kmem_cache
*sc
;
1426 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
1428 if (conf
->mddev
->gendisk
)
1429 sprintf(conf
->cache_name
[0],
1430 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
1432 sprintf(conf
->cache_name
[0],
1433 "raid%d-%p", conf
->level
, conf
->mddev
);
1434 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
1436 conf
->active_name
= 0;
1437 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
1438 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
1442 conf
->slab_cache
= sc
;
1443 conf
->pool_size
= devs
;
1445 if (!grow_one_stripe(conf
))
1451 * scribble_len - return the required size of the scribble region
1452 * @num - total number of disks in the array
1454 * The size must be enough to contain:
1455 * 1/ a struct page pointer for each device in the array +2
1456 * 2/ room to convert each entry in (1) to its corresponding dma
1457 * (dma_map_page()) or page (page_address()) address.
1459 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1460 * calculate over all devices (not just the data blocks), using zeros in place
1461 * of the P and Q blocks.
1463 static size_t scribble_len(int num
)
1467 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
1472 static int resize_stripes(struct r5conf
*conf
, int newsize
)
1474 /* Make all the stripes able to hold 'newsize' devices.
1475 * New slots in each stripe get 'page' set to a new page.
1477 * This happens in stages:
1478 * 1/ create a new kmem_cache and allocate the required number of
1480 * 2/ gather all the old stripe_heads and tranfer the pages across
1481 * to the new stripe_heads. This will have the side effect of
1482 * freezing the array as once all stripe_heads have been collected,
1483 * no IO will be possible. Old stripe heads are freed once their
1484 * pages have been transferred over, and the old kmem_cache is
1485 * freed when all stripes are done.
1486 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1487 * we simple return a failre status - no need to clean anything up.
1488 * 4/ allocate new pages for the new slots in the new stripe_heads.
1489 * If this fails, we don't bother trying the shrink the
1490 * stripe_heads down again, we just leave them as they are.
1491 * As each stripe_head is processed the new one is released into
1494 * Once step2 is started, we cannot afford to wait for a write,
1495 * so we use GFP_NOIO allocations.
1497 struct stripe_head
*osh
, *nsh
;
1498 LIST_HEAD(newstripes
);
1499 struct disk_info
*ndisks
;
1502 struct kmem_cache
*sc
;
1505 if (newsize
<= conf
->pool_size
)
1506 return 0; /* never bother to shrink */
1508 err
= md_allow_write(conf
->mddev
);
1513 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
1514 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
1519 for (i
= conf
->max_nr_stripes
; i
; i
--) {
1520 nsh
= kmem_cache_zalloc(sc
, GFP_KERNEL
);
1524 nsh
->raid_conf
= conf
;
1525 #ifdef CONFIG_MULTICORE_RAID456
1526 init_waitqueue_head(&nsh
->ops
.wait_for_ops
);
1529 list_add(&nsh
->lru
, &newstripes
);
1532 /* didn't get enough, give up */
1533 while (!list_empty(&newstripes
)) {
1534 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1535 list_del(&nsh
->lru
);
1536 kmem_cache_free(sc
, nsh
);
1538 kmem_cache_destroy(sc
);
1541 /* Step 2 - Must use GFP_NOIO now.
1542 * OK, we have enough stripes, start collecting inactive
1543 * stripes and copying them over
1545 list_for_each_entry(nsh
, &newstripes
, lru
) {
1546 spin_lock_irq(&conf
->device_lock
);
1547 wait_event_lock_irq(conf
->wait_for_stripe
,
1548 !list_empty(&conf
->inactive_list
),
1551 osh
= get_free_stripe(conf
);
1552 spin_unlock_irq(&conf
->device_lock
);
1553 atomic_set(&nsh
->count
, 1);
1554 for(i
=0; i
<conf
->pool_size
; i
++)
1555 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
1556 for( ; i
<newsize
; i
++)
1557 nsh
->dev
[i
].page
= NULL
;
1558 kmem_cache_free(conf
->slab_cache
, osh
);
1560 kmem_cache_destroy(conf
->slab_cache
);
1563 * At this point, we are holding all the stripes so the array
1564 * is completely stalled, so now is a good time to resize
1565 * conf->disks and the scribble region
1567 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
1569 for (i
=0; i
<conf
->raid_disks
; i
++)
1570 ndisks
[i
] = conf
->disks
[i
];
1572 conf
->disks
= ndisks
;
1577 conf
->scribble_len
= scribble_len(newsize
);
1578 for_each_present_cpu(cpu
) {
1579 struct raid5_percpu
*percpu
;
1582 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1583 scribble
= kmalloc(conf
->scribble_len
, GFP_NOIO
);
1586 kfree(percpu
->scribble
);
1587 percpu
->scribble
= scribble
;
1595 /* Step 4, return new stripes to service */
1596 while(!list_empty(&newstripes
)) {
1597 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1598 list_del_init(&nsh
->lru
);
1600 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
1601 if (nsh
->dev
[i
].page
== NULL
) {
1602 struct page
*p
= alloc_page(GFP_NOIO
);
1603 nsh
->dev
[i
].page
= p
;
1607 release_stripe(nsh
);
1609 /* critical section pass, GFP_NOIO no longer needed */
1611 conf
->slab_cache
= sc
;
1612 conf
->active_name
= 1-conf
->active_name
;
1613 conf
->pool_size
= newsize
;
1617 static int drop_one_stripe(struct r5conf
*conf
)
1619 struct stripe_head
*sh
;
1621 spin_lock_irq(&conf
->device_lock
);
1622 sh
= get_free_stripe(conf
);
1623 spin_unlock_irq(&conf
->device_lock
);
1626 BUG_ON(atomic_read(&sh
->count
));
1628 kmem_cache_free(conf
->slab_cache
, sh
);
1629 atomic_dec(&conf
->active_stripes
);
1633 static void shrink_stripes(struct r5conf
*conf
)
1635 while (drop_one_stripe(conf
))
1638 if (conf
->slab_cache
)
1639 kmem_cache_destroy(conf
->slab_cache
);
1640 conf
->slab_cache
= NULL
;
1643 static void raid5_end_read_request(struct bio
* bi
, int error
)
1645 struct stripe_head
*sh
= bi
->bi_private
;
1646 struct r5conf
*conf
= sh
->raid_conf
;
1647 int disks
= sh
->disks
, i
;
1648 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1649 char b
[BDEVNAME_SIZE
];
1650 struct md_rdev
*rdev
= NULL
;
1653 for (i
=0 ; i
<disks
; i
++)
1654 if (bi
== &sh
->dev
[i
].req
)
1657 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1658 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1664 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
1665 /* If replacement finished while this request was outstanding,
1666 * 'replacement' might be NULL already.
1667 * In that case it moved down to 'rdev'.
1668 * rdev is not removed until all requests are finished.
1670 rdev
= conf
->disks
[i
].replacement
;
1672 rdev
= conf
->disks
[i
].rdev
;
1675 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1676 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
1677 /* Note that this cannot happen on a
1678 * replacement device. We just fail those on
1683 "md/raid:%s: read error corrected"
1684 " (%lu sectors at %llu on %s)\n",
1685 mdname(conf
->mddev
), STRIPE_SECTORS
,
1686 (unsigned long long)(sh
->sector
1687 + rdev
->data_offset
),
1688 bdevname(rdev
->bdev
, b
));
1689 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
1690 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1691 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1693 if (atomic_read(&rdev
->read_errors
))
1694 atomic_set(&rdev
->read_errors
, 0);
1696 const char *bdn
= bdevname(rdev
->bdev
, b
);
1699 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1700 atomic_inc(&rdev
->read_errors
);
1701 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
1704 "md/raid:%s: read error on replacement device "
1705 "(sector %llu on %s).\n",
1706 mdname(conf
->mddev
),
1707 (unsigned long long)(sh
->sector
1708 + rdev
->data_offset
),
1710 else if (conf
->mddev
->degraded
>= conf
->max_degraded
)
1713 "md/raid:%s: read error not correctable "
1714 "(sector %llu on %s).\n",
1715 mdname(conf
->mddev
),
1716 (unsigned long long)(sh
->sector
1717 + rdev
->data_offset
),
1719 else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
))
1723 "md/raid:%s: read error NOT corrected!! "
1724 "(sector %llu on %s).\n",
1725 mdname(conf
->mddev
),
1726 (unsigned long long)(sh
->sector
1727 + rdev
->data_offset
),
1729 else if (atomic_read(&rdev
->read_errors
)
1730 > conf
->max_nr_stripes
)
1732 "md/raid:%s: Too many read errors, failing device %s.\n",
1733 mdname(conf
->mddev
), bdn
);
1737 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1739 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1740 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1741 md_error(conf
->mddev
, rdev
);
1744 rdev_dec_pending(rdev
, conf
->mddev
);
1745 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1746 set_bit(STRIPE_HANDLE
, &sh
->state
);
1750 static void raid5_end_write_request(struct bio
*bi
, int error
)
1752 struct stripe_head
*sh
= bi
->bi_private
;
1753 struct r5conf
*conf
= sh
->raid_conf
;
1754 int disks
= sh
->disks
, i
;
1755 struct md_rdev
*uninitialized_var(rdev
);
1756 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1759 int replacement
= 0;
1761 for (i
= 0 ; i
< disks
; i
++) {
1762 if (bi
== &sh
->dev
[i
].req
) {
1763 rdev
= conf
->disks
[i
].rdev
;
1766 if (bi
== &sh
->dev
[i
].rreq
) {
1767 rdev
= conf
->disks
[i
].replacement
;
1771 /* rdev was removed and 'replacement'
1772 * replaced it. rdev is not removed
1773 * until all requests are finished.
1775 rdev
= conf
->disks
[i
].rdev
;
1779 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1780 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1789 md_error(conf
->mddev
, rdev
);
1790 else if (is_badblock(rdev
, sh
->sector
,
1792 &first_bad
, &bad_sectors
))
1793 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
1796 set_bit(WriteErrorSeen
, &rdev
->flags
);
1797 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
1798 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1799 set_bit(MD_RECOVERY_NEEDED
,
1800 &rdev
->mddev
->recovery
);
1801 } else if (is_badblock(rdev
, sh
->sector
,
1803 &first_bad
, &bad_sectors
))
1804 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
1806 rdev_dec_pending(rdev
, conf
->mddev
);
1808 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
1809 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1810 set_bit(STRIPE_HANDLE
, &sh
->state
);
1814 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
);
1816 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
1818 struct r5dev
*dev
= &sh
->dev
[i
];
1820 bio_init(&dev
->req
);
1821 dev
->req
.bi_io_vec
= &dev
->vec
;
1823 dev
->req
.bi_max_vecs
++;
1824 dev
->req
.bi_private
= sh
;
1825 dev
->vec
.bv_page
= dev
->page
;
1827 bio_init(&dev
->rreq
);
1828 dev
->rreq
.bi_io_vec
= &dev
->rvec
;
1829 dev
->rreq
.bi_vcnt
++;
1830 dev
->rreq
.bi_max_vecs
++;
1831 dev
->rreq
.bi_private
= sh
;
1832 dev
->rvec
.bv_page
= dev
->page
;
1835 dev
->sector
= compute_blocknr(sh
, i
, previous
);
1838 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1840 char b
[BDEVNAME_SIZE
];
1841 struct r5conf
*conf
= mddev
->private;
1842 unsigned long flags
;
1843 pr_debug("raid456: error called\n");
1845 spin_lock_irqsave(&conf
->device_lock
, flags
);
1846 clear_bit(In_sync
, &rdev
->flags
);
1847 mddev
->degraded
= calc_degraded(conf
);
1848 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1849 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1851 set_bit(Blocked
, &rdev
->flags
);
1852 set_bit(Faulty
, &rdev
->flags
);
1853 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1855 "md/raid:%s: Disk failure on %s, disabling device.\n"
1856 "md/raid:%s: Operation continuing on %d devices.\n",
1858 bdevname(rdev
->bdev
, b
),
1860 conf
->raid_disks
- mddev
->degraded
);
1864 * Input: a 'big' sector number,
1865 * Output: index of the data and parity disk, and the sector # in them.
1867 static sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
1868 int previous
, int *dd_idx
,
1869 struct stripe_head
*sh
)
1871 sector_t stripe
, stripe2
;
1872 sector_t chunk_number
;
1873 unsigned int chunk_offset
;
1876 sector_t new_sector
;
1877 int algorithm
= previous
? conf
->prev_algo
1879 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
1880 : conf
->chunk_sectors
;
1881 int raid_disks
= previous
? conf
->previous_raid_disks
1883 int data_disks
= raid_disks
- conf
->max_degraded
;
1885 /* First compute the information on this sector */
1888 * Compute the chunk number and the sector offset inside the chunk
1890 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
1891 chunk_number
= r_sector
;
1894 * Compute the stripe number
1896 stripe
= chunk_number
;
1897 *dd_idx
= sector_div(stripe
, data_disks
);
1900 * Select the parity disk based on the user selected algorithm.
1902 pd_idx
= qd_idx
= -1;
1903 switch(conf
->level
) {
1905 pd_idx
= data_disks
;
1908 switch (algorithm
) {
1909 case ALGORITHM_LEFT_ASYMMETRIC
:
1910 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
1911 if (*dd_idx
>= pd_idx
)
1914 case ALGORITHM_RIGHT_ASYMMETRIC
:
1915 pd_idx
= sector_div(stripe2
, raid_disks
);
1916 if (*dd_idx
>= pd_idx
)
1919 case ALGORITHM_LEFT_SYMMETRIC
:
1920 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
1921 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1923 case ALGORITHM_RIGHT_SYMMETRIC
:
1924 pd_idx
= sector_div(stripe2
, raid_disks
);
1925 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1927 case ALGORITHM_PARITY_0
:
1931 case ALGORITHM_PARITY_N
:
1932 pd_idx
= data_disks
;
1940 switch (algorithm
) {
1941 case ALGORITHM_LEFT_ASYMMETRIC
:
1942 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1943 qd_idx
= pd_idx
+ 1;
1944 if (pd_idx
== raid_disks
-1) {
1945 (*dd_idx
)++; /* Q D D D P */
1947 } else if (*dd_idx
>= pd_idx
)
1948 (*dd_idx
) += 2; /* D D P Q D */
1950 case ALGORITHM_RIGHT_ASYMMETRIC
:
1951 pd_idx
= sector_div(stripe2
, raid_disks
);
1952 qd_idx
= pd_idx
+ 1;
1953 if (pd_idx
== raid_disks
-1) {
1954 (*dd_idx
)++; /* Q D D D P */
1956 } else if (*dd_idx
>= pd_idx
)
1957 (*dd_idx
) += 2; /* D D P Q D */
1959 case ALGORITHM_LEFT_SYMMETRIC
:
1960 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1961 qd_idx
= (pd_idx
+ 1) % raid_disks
;
1962 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
1964 case ALGORITHM_RIGHT_SYMMETRIC
:
1965 pd_idx
= sector_div(stripe2
, raid_disks
);
1966 qd_idx
= (pd_idx
+ 1) % raid_disks
;
1967 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
1970 case ALGORITHM_PARITY_0
:
1975 case ALGORITHM_PARITY_N
:
1976 pd_idx
= data_disks
;
1977 qd_idx
= data_disks
+ 1;
1980 case ALGORITHM_ROTATING_ZERO_RESTART
:
1981 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1982 * of blocks for computing Q is different.
1984 pd_idx
= sector_div(stripe2
, raid_disks
);
1985 qd_idx
= pd_idx
+ 1;
1986 if (pd_idx
== raid_disks
-1) {
1987 (*dd_idx
)++; /* Q D D D P */
1989 } else if (*dd_idx
>= pd_idx
)
1990 (*dd_idx
) += 2; /* D D P Q D */
1994 case ALGORITHM_ROTATING_N_RESTART
:
1995 /* Same a left_asymmetric, by first stripe is
1996 * D D D P Q rather than
2000 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2001 qd_idx
= pd_idx
+ 1;
2002 if (pd_idx
== raid_disks
-1) {
2003 (*dd_idx
)++; /* Q D D D P */
2005 } else if (*dd_idx
>= pd_idx
)
2006 (*dd_idx
) += 2; /* D D P Q D */
2010 case ALGORITHM_ROTATING_N_CONTINUE
:
2011 /* Same as left_symmetric but Q is before P */
2012 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2013 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
2014 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2018 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2019 /* RAID5 left_asymmetric, with Q on last device */
2020 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2021 if (*dd_idx
>= pd_idx
)
2023 qd_idx
= raid_disks
- 1;
2026 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2027 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2028 if (*dd_idx
>= pd_idx
)
2030 qd_idx
= raid_disks
- 1;
2033 case ALGORITHM_LEFT_SYMMETRIC_6
:
2034 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2035 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2036 qd_idx
= raid_disks
- 1;
2039 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2040 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2041 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2042 qd_idx
= raid_disks
- 1;
2045 case ALGORITHM_PARITY_0_6
:
2048 qd_idx
= raid_disks
- 1;
2058 sh
->pd_idx
= pd_idx
;
2059 sh
->qd_idx
= qd_idx
;
2060 sh
->ddf_layout
= ddf_layout
;
2063 * Finally, compute the new sector number
2065 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
2070 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
2072 struct r5conf
*conf
= sh
->raid_conf
;
2073 int raid_disks
= sh
->disks
;
2074 int data_disks
= raid_disks
- conf
->max_degraded
;
2075 sector_t new_sector
= sh
->sector
, check
;
2076 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2077 : conf
->chunk_sectors
;
2078 int algorithm
= previous
? conf
->prev_algo
2082 sector_t chunk_number
;
2083 int dummy1
, dd_idx
= i
;
2085 struct stripe_head sh2
;
2088 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
2089 stripe
= new_sector
;
2091 if (i
== sh
->pd_idx
)
2093 switch(conf
->level
) {
2096 switch (algorithm
) {
2097 case ALGORITHM_LEFT_ASYMMETRIC
:
2098 case ALGORITHM_RIGHT_ASYMMETRIC
:
2102 case ALGORITHM_LEFT_SYMMETRIC
:
2103 case ALGORITHM_RIGHT_SYMMETRIC
:
2106 i
-= (sh
->pd_idx
+ 1);
2108 case ALGORITHM_PARITY_0
:
2111 case ALGORITHM_PARITY_N
:
2118 if (i
== sh
->qd_idx
)
2119 return 0; /* It is the Q disk */
2120 switch (algorithm
) {
2121 case ALGORITHM_LEFT_ASYMMETRIC
:
2122 case ALGORITHM_RIGHT_ASYMMETRIC
:
2123 case ALGORITHM_ROTATING_ZERO_RESTART
:
2124 case ALGORITHM_ROTATING_N_RESTART
:
2125 if (sh
->pd_idx
== raid_disks
-1)
2126 i
--; /* Q D D D P */
2127 else if (i
> sh
->pd_idx
)
2128 i
-= 2; /* D D P Q D */
2130 case ALGORITHM_LEFT_SYMMETRIC
:
2131 case ALGORITHM_RIGHT_SYMMETRIC
:
2132 if (sh
->pd_idx
== raid_disks
-1)
2133 i
--; /* Q D D D P */
2138 i
-= (sh
->pd_idx
+ 2);
2141 case ALGORITHM_PARITY_0
:
2144 case ALGORITHM_PARITY_N
:
2146 case ALGORITHM_ROTATING_N_CONTINUE
:
2147 /* Like left_symmetric, but P is before Q */
2148 if (sh
->pd_idx
== 0)
2149 i
--; /* P D D D Q */
2154 i
-= (sh
->pd_idx
+ 1);
2157 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2158 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2162 case ALGORITHM_LEFT_SYMMETRIC_6
:
2163 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2165 i
+= data_disks
+ 1;
2166 i
-= (sh
->pd_idx
+ 1);
2168 case ALGORITHM_PARITY_0_6
:
2177 chunk_number
= stripe
* data_disks
+ i
;
2178 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2180 check
= raid5_compute_sector(conf
, r_sector
,
2181 previous
, &dummy1
, &sh2
);
2182 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2183 || sh2
.qd_idx
!= sh
->qd_idx
) {
2184 printk(KERN_ERR
"md/raid:%s: compute_blocknr: map not correct\n",
2185 mdname(conf
->mddev
));
2193 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2194 int rcw
, int expand
)
2196 int i
, pd_idx
= sh
->pd_idx
, disks
= sh
->disks
;
2197 struct r5conf
*conf
= sh
->raid_conf
;
2198 int level
= conf
->level
;
2201 /* if we are not expanding this is a proper write request, and
2202 * there will be bios with new data to be drained into the
2206 sh
->reconstruct_state
= reconstruct_state_drain_run
;
2207 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2209 sh
->reconstruct_state
= reconstruct_state_run
;
2211 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2213 for (i
= disks
; i
--; ) {
2214 struct r5dev
*dev
= &sh
->dev
[i
];
2217 set_bit(R5_LOCKED
, &dev
->flags
);
2218 set_bit(R5_Wantdrain
, &dev
->flags
);
2220 clear_bit(R5_UPTODATE
, &dev
->flags
);
2224 if (s
->locked
+ conf
->max_degraded
== disks
)
2225 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2226 atomic_inc(&conf
->pending_full_writes
);
2229 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
2230 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
2232 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
2233 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
2234 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2235 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2237 for (i
= disks
; i
--; ) {
2238 struct r5dev
*dev
= &sh
->dev
[i
];
2243 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2244 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2245 set_bit(R5_Wantdrain
, &dev
->flags
);
2246 set_bit(R5_LOCKED
, &dev
->flags
);
2247 clear_bit(R5_UPTODATE
, &dev
->flags
);
2253 /* keep the parity disk(s) locked while asynchronous operations
2256 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
2257 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2261 int qd_idx
= sh
->qd_idx
;
2262 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
2264 set_bit(R5_LOCKED
, &dev
->flags
);
2265 clear_bit(R5_UPTODATE
, &dev
->flags
);
2269 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2270 __func__
, (unsigned long long)sh
->sector
,
2271 s
->locked
, s
->ops_request
);
2275 * Each stripe/dev can have one or more bion attached.
2276 * toread/towrite point to the first in a chain.
2277 * The bi_next chain must be in order.
2279 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
, int forwrite
)
2282 struct r5conf
*conf
= sh
->raid_conf
;
2285 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2286 (unsigned long long)bi
->bi_sector
,
2287 (unsigned long long)sh
->sector
);
2290 spin_lock_irq(&conf
->device_lock
);
2292 bip
= &sh
->dev
[dd_idx
].towrite
;
2293 if (*bip
== NULL
&& sh
->dev
[dd_idx
].written
== NULL
)
2296 bip
= &sh
->dev
[dd_idx
].toread
;
2297 while (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
) {
2298 if ((*bip
)->bi_sector
+ ((*bip
)->bi_size
>> 9) > bi
->bi_sector
)
2300 bip
= & (*bip
)->bi_next
;
2302 if (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
+ ((bi
->bi_size
)>>9))
2305 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
2309 bi
->bi_phys_segments
++;
2312 /* check if page is covered */
2313 sector_t sector
= sh
->dev
[dd_idx
].sector
;
2314 for (bi
=sh
->dev
[dd_idx
].towrite
;
2315 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
2316 bi
&& bi
->bi_sector
<= sector
;
2317 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
2318 if (bi
->bi_sector
+ (bi
->bi_size
>>9) >= sector
)
2319 sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
2321 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
2322 set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
);
2324 spin_unlock_irq(&conf
->device_lock
);
2326 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2327 (unsigned long long)(*bip
)->bi_sector
,
2328 (unsigned long long)sh
->sector
, dd_idx
);
2330 if (conf
->mddev
->bitmap
&& firstwrite
) {
2331 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
2333 sh
->bm_seq
= conf
->seq_flush
+1;
2334 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
2339 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
2340 spin_unlock_irq(&conf
->device_lock
);
2344 static void end_reshape(struct r5conf
*conf
);
2346 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
2347 struct stripe_head
*sh
)
2349 int sectors_per_chunk
=
2350 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
2352 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
2353 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
2355 raid5_compute_sector(conf
,
2356 stripe
* (disks
- conf
->max_degraded
)
2357 *sectors_per_chunk
+ chunk_offset
,
2363 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
2364 struct stripe_head_state
*s
, int disks
,
2365 struct bio
**return_bi
)
2368 for (i
= disks
; i
--; ) {
2372 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2373 struct md_rdev
*rdev
;
2375 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2376 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
2377 atomic_inc(&rdev
->nr_pending
);
2382 if (!rdev_set_badblocks(
2386 md_error(conf
->mddev
, rdev
);
2387 rdev_dec_pending(rdev
, conf
->mddev
);
2390 spin_lock_irq(&conf
->device_lock
);
2391 /* fail all writes first */
2392 bi
= sh
->dev
[i
].towrite
;
2393 sh
->dev
[i
].towrite
= NULL
;
2399 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2400 wake_up(&conf
->wait_for_overlap
);
2402 while (bi
&& bi
->bi_sector
<
2403 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2404 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2405 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2406 if (!raid5_dec_bi_phys_segments(bi
)) {
2407 md_write_end(conf
->mddev
);
2408 bi
->bi_next
= *return_bi
;
2413 /* and fail all 'written' */
2414 bi
= sh
->dev
[i
].written
;
2415 sh
->dev
[i
].written
= NULL
;
2416 if (bi
) bitmap_end
= 1;
2417 while (bi
&& bi
->bi_sector
<
2418 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2419 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2420 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2421 if (!raid5_dec_bi_phys_segments(bi
)) {
2422 md_write_end(conf
->mddev
);
2423 bi
->bi_next
= *return_bi
;
2429 /* fail any reads if this device is non-operational and
2430 * the data has not reached the cache yet.
2432 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
2433 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
2434 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
2435 bi
= sh
->dev
[i
].toread
;
2436 sh
->dev
[i
].toread
= NULL
;
2437 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2438 wake_up(&conf
->wait_for_overlap
);
2439 if (bi
) s
->to_read
--;
2440 while (bi
&& bi
->bi_sector
<
2441 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2442 struct bio
*nextbi
=
2443 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 bi
->bi_next
= *return_bi
;
2452 spin_unlock_irq(&conf
->device_lock
);
2454 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2455 STRIPE_SECTORS
, 0, 0);
2456 /* If we were in the middle of a write the parity block might
2457 * still be locked - so just clear all R5_LOCKED flags
2459 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2462 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2463 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2464 md_wakeup_thread(conf
->mddev
->thread
);
2468 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
2469 struct stripe_head_state
*s
)
2474 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 0);
2475 clear_bit(STRIPE_SYNCING
, &sh
->state
);
2478 /* There is nothing more to do for sync/check/repair.
2479 * For recover/replace we need to record a bad block on all
2480 * non-sync devices, or abort the recovery
2482 if (!test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
))
2484 /* During recovery devices cannot be removed, so locking and
2485 * refcounting of rdevs is not needed
2487 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2488 struct md_rdev
*rdev
= conf
->disks
[i
].rdev
;
2490 && !test_bit(Faulty
, &rdev
->flags
)
2491 && !test_bit(In_sync
, &rdev
->flags
)
2492 && !rdev_set_badblocks(rdev
, sh
->sector
,
2495 rdev
= conf
->disks
[i
].replacement
;
2497 && !test_bit(Faulty
, &rdev
->flags
)
2498 && !test_bit(In_sync
, &rdev
->flags
)
2499 && !rdev_set_badblocks(rdev
, sh
->sector
,
2504 conf
->recovery_disabled
= conf
->mddev
->recovery_disabled
;
2505 set_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
);
2509 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
2511 struct md_rdev
*rdev
;
2513 /* Doing recovery so rcu locking not required */
2514 rdev
= sh
->raid_conf
->disks
[disk_idx
].replacement
;
2516 && !test_bit(Faulty
, &rdev
->flags
)
2517 && !test_bit(In_sync
, &rdev
->flags
)
2518 && (rdev
->recovery_offset
<= sh
->sector
2519 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
2525 /* fetch_block - checks the given member device to see if its data needs
2526 * to be read or computed to satisfy a request.
2528 * Returns 1 when no more member devices need to be checked, otherwise returns
2529 * 0 to tell the loop in handle_stripe_fill to continue
2531 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2532 int disk_idx
, int disks
)
2534 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
2535 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
2536 &sh
->dev
[s
->failed_num
[1]] };
2538 /* is the data in this block needed, and can we get it? */
2539 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2540 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2542 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2543 s
->syncing
|| s
->expanding
||
2544 (s
->replacing
&& want_replace(sh
, disk_idx
)) ||
2545 (s
->failed
>= 1 && fdev
[0]->toread
) ||
2546 (s
->failed
>= 2 && fdev
[1]->toread
) ||
2547 (sh
->raid_conf
->level
<= 5 && s
->failed
&& fdev
[0]->towrite
&&
2548 !test_bit(R5_OVERWRITE
, &fdev
[0]->flags
)) ||
2549 (sh
->raid_conf
->level
== 6 && s
->failed
&& s
->to_write
))) {
2550 /* we would like to get this block, possibly by computing it,
2551 * otherwise read it if the backing disk is insync
2553 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
2554 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
2555 if ((s
->uptodate
== disks
- 1) &&
2556 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
2557 disk_idx
== s
->failed_num
[1]))) {
2558 /* have disk failed, and we're requested to fetch it;
2561 pr_debug("Computing stripe %llu block %d\n",
2562 (unsigned long long)sh
->sector
, disk_idx
);
2563 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2564 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2565 set_bit(R5_Wantcompute
, &dev
->flags
);
2566 sh
->ops
.target
= disk_idx
;
2567 sh
->ops
.target2
= -1; /* no 2nd target */
2569 /* Careful: from this point on 'uptodate' is in the eye
2570 * of raid_run_ops which services 'compute' operations
2571 * before writes. R5_Wantcompute flags a block that will
2572 * be R5_UPTODATE by the time it is needed for a
2573 * subsequent operation.
2577 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
2578 /* Computing 2-failure is *very* expensive; only
2579 * do it if failed >= 2
2582 for (other
= disks
; other
--; ) {
2583 if (other
== disk_idx
)
2585 if (!test_bit(R5_UPTODATE
,
2586 &sh
->dev
[other
].flags
))
2590 pr_debug("Computing stripe %llu blocks %d,%d\n",
2591 (unsigned long long)sh
->sector
,
2593 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2594 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2595 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
2596 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
2597 sh
->ops
.target
= disk_idx
;
2598 sh
->ops
.target2
= other
;
2602 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2603 set_bit(R5_LOCKED
, &dev
->flags
);
2604 set_bit(R5_Wantread
, &dev
->flags
);
2606 pr_debug("Reading block %d (sync=%d)\n",
2607 disk_idx
, s
->syncing
);
2615 * handle_stripe_fill - read or compute data to satisfy pending requests.
2617 static void handle_stripe_fill(struct stripe_head
*sh
,
2618 struct stripe_head_state
*s
,
2623 /* look for blocks to read/compute, skip this if a compute
2624 * is already in flight, or if the stripe contents are in the
2625 * midst of changing due to a write
2627 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2628 !sh
->reconstruct_state
)
2629 for (i
= disks
; i
--; )
2630 if (fetch_block(sh
, s
, i
, disks
))
2632 set_bit(STRIPE_HANDLE
, &sh
->state
);
2636 /* handle_stripe_clean_event
2637 * any written block on an uptodate or failed drive can be returned.
2638 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2639 * never LOCKED, so we don't need to test 'failed' directly.
2641 static void handle_stripe_clean_event(struct r5conf
*conf
,
2642 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
2647 for (i
= disks
; i
--; )
2648 if (sh
->dev
[i
].written
) {
2650 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2651 test_bit(R5_UPTODATE
, &dev
->flags
)) {
2652 /* We can return any write requests */
2653 struct bio
*wbi
, *wbi2
;
2655 pr_debug("Return write for disc %d\n", i
);
2656 spin_lock_irq(&conf
->device_lock
);
2658 dev
->written
= NULL
;
2659 while (wbi
&& wbi
->bi_sector
<
2660 dev
->sector
+ STRIPE_SECTORS
) {
2661 wbi2
= r5_next_bio(wbi
, dev
->sector
);
2662 if (!raid5_dec_bi_phys_segments(wbi
)) {
2663 md_write_end(conf
->mddev
);
2664 wbi
->bi_next
= *return_bi
;
2669 if (dev
->towrite
== NULL
)
2671 spin_unlock_irq(&conf
->device_lock
);
2673 bitmap_endwrite(conf
->mddev
->bitmap
,
2676 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
2681 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2682 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2683 md_wakeup_thread(conf
->mddev
->thread
);
2686 static void handle_stripe_dirtying(struct r5conf
*conf
,
2687 struct stripe_head
*sh
,
2688 struct stripe_head_state
*s
,
2691 int rmw
= 0, rcw
= 0, i
;
2692 if (conf
->max_degraded
== 2) {
2693 /* RAID6 requires 'rcw' in current implementation
2694 * Calculate the real rcw later - for now fake it
2695 * look like rcw is cheaper
2698 } else for (i
= disks
; i
--; ) {
2699 /* would I have to read this buffer for read_modify_write */
2700 struct r5dev
*dev
= &sh
->dev
[i
];
2701 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2702 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2703 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2704 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2705 if (test_bit(R5_Insync
, &dev
->flags
))
2708 rmw
+= 2*disks
; /* cannot read it */
2710 /* Would I have to read this buffer for reconstruct_write */
2711 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) && i
!= sh
->pd_idx
&&
2712 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2713 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2714 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2715 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2720 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2721 (unsigned long long)sh
->sector
, rmw
, rcw
);
2722 set_bit(STRIPE_HANDLE
, &sh
->state
);
2723 if (rmw
< rcw
&& rmw
> 0)
2724 /* prefer read-modify-write, but need to get some data */
2725 for (i
= disks
; i
--; ) {
2726 struct r5dev
*dev
= &sh
->dev
[i
];
2727 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2728 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2729 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2730 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2731 test_bit(R5_Insync
, &dev
->flags
)) {
2733 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2734 pr_debug("Read_old block "
2735 "%d for r-m-w\n", i
);
2736 set_bit(R5_LOCKED
, &dev
->flags
);
2737 set_bit(R5_Wantread
, &dev
->flags
);
2740 set_bit(STRIPE_DELAYED
, &sh
->state
);
2741 set_bit(STRIPE_HANDLE
, &sh
->state
);
2745 if (rcw
<= rmw
&& rcw
> 0) {
2746 /* want reconstruct write, but need to get some data */
2748 for (i
= disks
; i
--; ) {
2749 struct r5dev
*dev
= &sh
->dev
[i
];
2750 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2751 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
2752 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2753 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2754 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2756 if (!test_bit(R5_Insync
, &dev
->flags
))
2757 continue; /* it's a failed drive */
2759 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2760 pr_debug("Read_old block "
2761 "%d for Reconstruct\n", i
);
2762 set_bit(R5_LOCKED
, &dev
->flags
);
2763 set_bit(R5_Wantread
, &dev
->flags
);
2766 set_bit(STRIPE_DELAYED
, &sh
->state
);
2767 set_bit(STRIPE_HANDLE
, &sh
->state
);
2772 /* now if nothing is locked, and if we have enough data,
2773 * we can start a write request
2775 /* since handle_stripe can be called at any time we need to handle the
2776 * case where a compute block operation has been submitted and then a
2777 * subsequent call wants to start a write request. raid_run_ops only
2778 * handles the case where compute block and reconstruct are requested
2779 * simultaneously. If this is not the case then new writes need to be
2780 * held off until the compute completes.
2782 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
2783 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
2784 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
2785 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
2788 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
2789 struct stripe_head_state
*s
, int disks
)
2791 struct r5dev
*dev
= NULL
;
2793 set_bit(STRIPE_HANDLE
, &sh
->state
);
2795 switch (sh
->check_state
) {
2796 case check_state_idle
:
2797 /* start a new check operation if there are no failures */
2798 if (s
->failed
== 0) {
2799 BUG_ON(s
->uptodate
!= disks
);
2800 sh
->check_state
= check_state_run
;
2801 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2802 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2806 dev
= &sh
->dev
[s
->failed_num
[0]];
2808 case check_state_compute_result
:
2809 sh
->check_state
= check_state_idle
;
2811 dev
= &sh
->dev
[sh
->pd_idx
];
2813 /* check that a write has not made the stripe insync */
2814 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2817 /* either failed parity check, or recovery is happening */
2818 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
2819 BUG_ON(s
->uptodate
!= disks
);
2821 set_bit(R5_LOCKED
, &dev
->flags
);
2823 set_bit(R5_Wantwrite
, &dev
->flags
);
2825 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2826 set_bit(STRIPE_INSYNC
, &sh
->state
);
2828 case check_state_run
:
2829 break; /* we will be called again upon completion */
2830 case check_state_check_result
:
2831 sh
->check_state
= check_state_idle
;
2833 /* if a failure occurred during the check operation, leave
2834 * STRIPE_INSYNC not set and let the stripe be handled again
2839 /* handle a successful check operation, if parity is correct
2840 * we are done. Otherwise update the mismatch count and repair
2841 * parity if !MD_RECOVERY_CHECK
2843 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
2844 /* parity is correct (on disc,
2845 * not in buffer any more)
2847 set_bit(STRIPE_INSYNC
, &sh
->state
);
2849 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2850 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2851 /* don't try to repair!! */
2852 set_bit(STRIPE_INSYNC
, &sh
->state
);
2854 sh
->check_state
= check_state_compute_run
;
2855 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2856 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2857 set_bit(R5_Wantcompute
,
2858 &sh
->dev
[sh
->pd_idx
].flags
);
2859 sh
->ops
.target
= sh
->pd_idx
;
2860 sh
->ops
.target2
= -1;
2865 case check_state_compute_run
:
2868 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
2869 __func__
, sh
->check_state
,
2870 (unsigned long long) sh
->sector
);
2876 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
2877 struct stripe_head_state
*s
,
2880 int pd_idx
= sh
->pd_idx
;
2881 int qd_idx
= sh
->qd_idx
;
2884 set_bit(STRIPE_HANDLE
, &sh
->state
);
2886 BUG_ON(s
->failed
> 2);
2888 /* Want to check and possibly repair P and Q.
2889 * However there could be one 'failed' device, in which
2890 * case we can only check one of them, possibly using the
2891 * other to generate missing data
2894 switch (sh
->check_state
) {
2895 case check_state_idle
:
2896 /* start a new check operation if there are < 2 failures */
2897 if (s
->failed
== s
->q_failed
) {
2898 /* The only possible failed device holds Q, so it
2899 * makes sense to check P (If anything else were failed,
2900 * we would have used P to recreate it).
2902 sh
->check_state
= check_state_run
;
2904 if (!s
->q_failed
&& s
->failed
< 2) {
2905 /* Q is not failed, and we didn't use it to generate
2906 * anything, so it makes sense to check it
2908 if (sh
->check_state
== check_state_run
)
2909 sh
->check_state
= check_state_run_pq
;
2911 sh
->check_state
= check_state_run_q
;
2914 /* discard potentially stale zero_sum_result */
2915 sh
->ops
.zero_sum_result
= 0;
2917 if (sh
->check_state
== check_state_run
) {
2918 /* async_xor_zero_sum destroys the contents of P */
2919 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2922 if (sh
->check_state
>= check_state_run
&&
2923 sh
->check_state
<= check_state_run_pq
) {
2924 /* async_syndrome_zero_sum preserves P and Q, so
2925 * no need to mark them !uptodate here
2927 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2931 /* we have 2-disk failure */
2932 BUG_ON(s
->failed
!= 2);
2934 case check_state_compute_result
:
2935 sh
->check_state
= check_state_idle
;
2937 /* check that a write has not made the stripe insync */
2938 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2941 /* now write out any block on a failed drive,
2942 * or P or Q if they were recomputed
2944 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
2945 if (s
->failed
== 2) {
2946 dev
= &sh
->dev
[s
->failed_num
[1]];
2948 set_bit(R5_LOCKED
, &dev
->flags
);
2949 set_bit(R5_Wantwrite
, &dev
->flags
);
2951 if (s
->failed
>= 1) {
2952 dev
= &sh
->dev
[s
->failed_num
[0]];
2954 set_bit(R5_LOCKED
, &dev
->flags
);
2955 set_bit(R5_Wantwrite
, &dev
->flags
);
2957 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
2958 dev
= &sh
->dev
[pd_idx
];
2960 set_bit(R5_LOCKED
, &dev
->flags
);
2961 set_bit(R5_Wantwrite
, &dev
->flags
);
2963 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
2964 dev
= &sh
->dev
[qd_idx
];
2966 set_bit(R5_LOCKED
, &dev
->flags
);
2967 set_bit(R5_Wantwrite
, &dev
->flags
);
2969 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2971 set_bit(STRIPE_INSYNC
, &sh
->state
);
2973 case check_state_run
:
2974 case check_state_run_q
:
2975 case check_state_run_pq
:
2976 break; /* we will be called again upon completion */
2977 case check_state_check_result
:
2978 sh
->check_state
= check_state_idle
;
2980 /* handle a successful check operation, if parity is correct
2981 * we are done. Otherwise update the mismatch count and repair
2982 * parity if !MD_RECOVERY_CHECK
2984 if (sh
->ops
.zero_sum_result
== 0) {
2985 /* both parities are correct */
2987 set_bit(STRIPE_INSYNC
, &sh
->state
);
2989 /* in contrast to the raid5 case we can validate
2990 * parity, but still have a failure to write
2993 sh
->check_state
= check_state_compute_result
;
2994 /* Returning at this point means that we may go
2995 * off and bring p and/or q uptodate again so
2996 * we make sure to check zero_sum_result again
2997 * to verify if p or q need writeback
3001 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
3002 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3003 /* don't try to repair!! */
3004 set_bit(STRIPE_INSYNC
, &sh
->state
);
3006 int *target
= &sh
->ops
.target
;
3008 sh
->ops
.target
= -1;
3009 sh
->ops
.target2
= -1;
3010 sh
->check_state
= check_state_compute_run
;
3011 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3012 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3013 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3014 set_bit(R5_Wantcompute
,
3015 &sh
->dev
[pd_idx
].flags
);
3017 target
= &sh
->ops
.target2
;
3020 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3021 set_bit(R5_Wantcompute
,
3022 &sh
->dev
[qd_idx
].flags
);
3029 case check_state_compute_run
:
3032 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3033 __func__
, sh
->check_state
,
3034 (unsigned long long) sh
->sector
);
3039 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
3043 /* We have read all the blocks in this stripe and now we need to
3044 * copy some of them into a target stripe for expand.
3046 struct dma_async_tx_descriptor
*tx
= NULL
;
3047 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3048 for (i
= 0; i
< sh
->disks
; i
++)
3049 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
3051 struct stripe_head
*sh2
;
3052 struct async_submit_ctl submit
;
3054 sector_t bn
= compute_blocknr(sh
, i
, 1);
3055 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
3057 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
3059 /* so far only the early blocks of this stripe
3060 * have been requested. When later blocks
3061 * get requested, we will try again
3064 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
3065 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
3066 /* must have already done this block */
3067 release_stripe(sh2
);
3071 /* place all the copies on one channel */
3072 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
3073 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
3074 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
3077 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
3078 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
3079 for (j
= 0; j
< conf
->raid_disks
; j
++)
3080 if (j
!= sh2
->pd_idx
&&
3082 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
3084 if (j
== conf
->raid_disks
) {
3085 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
3086 set_bit(STRIPE_HANDLE
, &sh2
->state
);
3088 release_stripe(sh2
);
3091 /* done submitting copies, wait for them to complete */
3094 dma_wait_for_async_tx(tx
);
3099 * handle_stripe - do things to a stripe.
3101 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3102 * state of various bits to see what needs to be done.
3104 * return some read requests which now have data
3105 * return some write requests which are safely on storage
3106 * schedule a read on some buffers
3107 * schedule a write of some buffers
3108 * return confirmation of parity correctness
3112 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
3114 struct r5conf
*conf
= sh
->raid_conf
;
3115 int disks
= sh
->disks
;
3118 int do_recovery
= 0;
3120 memset(s
, 0, sizeof(*s
));
3122 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3123 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3124 s
->failed_num
[0] = -1;
3125 s
->failed_num
[1] = -1;
3127 /* Now to look around and see what can be done */
3129 spin_lock_irq(&conf
->device_lock
);
3130 for (i
=disks
; i
--; ) {
3131 struct md_rdev
*rdev
;
3138 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3140 dev
->toread
, dev
->towrite
, dev
->written
);
3141 /* maybe we can reply to a read
3143 * new wantfill requests are only permitted while
3144 * ops_complete_biofill is guaranteed to be inactive
3146 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
3147 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
3148 set_bit(R5_Wantfill
, &dev
->flags
);
3150 /* now count some things */
3151 if (test_bit(R5_LOCKED
, &dev
->flags
))
3153 if (test_bit(R5_UPTODATE
, &dev
->flags
))
3155 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
3157 BUG_ON(s
->compute
> 2);
3160 if (test_bit(R5_Wantfill
, &dev
->flags
))
3162 else if (dev
->toread
)
3166 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3171 /* Prefer to use the replacement for reads, but only
3172 * if it is recovered enough and has no bad blocks.
3174 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
3175 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
3176 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
3177 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3178 &first_bad
, &bad_sectors
))
3179 set_bit(R5_ReadRepl
, &dev
->flags
);
3182 set_bit(R5_NeedReplace
, &dev
->flags
);
3183 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3184 clear_bit(R5_ReadRepl
, &dev
->flags
);
3186 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
3189 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3190 &first_bad
, &bad_sectors
);
3191 if (s
->blocked_rdev
== NULL
3192 && (test_bit(Blocked
, &rdev
->flags
)
3195 set_bit(BlockedBadBlocks
,
3197 s
->blocked_rdev
= rdev
;
3198 atomic_inc(&rdev
->nr_pending
);
3201 clear_bit(R5_Insync
, &dev
->flags
);
3205 /* also not in-sync */
3206 if (!test_bit(WriteErrorSeen
, &rdev
->flags
)) {
3207 /* treat as in-sync, but with a read error
3208 * which we can now try to correct
3210 set_bit(R5_Insync
, &dev
->flags
);
3211 set_bit(R5_ReadError
, &dev
->flags
);
3213 } else if (test_bit(In_sync
, &rdev
->flags
))
3214 set_bit(R5_Insync
, &dev
->flags
);
3215 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
3216 /* in sync if before recovery_offset */
3217 set_bit(R5_Insync
, &dev
->flags
);
3218 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
3219 test_bit(R5_Expanded
, &dev
->flags
))
3220 /* If we've reshaped into here, we assume it is Insync.
3221 * We will shortly update recovery_offset to make
3224 set_bit(R5_Insync
, &dev
->flags
);
3226 if (rdev
&& test_bit(R5_WriteError
, &dev
->flags
)) {
3227 /* This flag does not apply to '.replacement'
3228 * only to .rdev, so make sure to check that*/
3229 struct md_rdev
*rdev2
= rcu_dereference(
3230 conf
->disks
[i
].rdev
);
3232 clear_bit(R5_Insync
, &dev
->flags
);
3233 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3234 s
->handle_bad_blocks
= 1;
3235 atomic_inc(&rdev2
->nr_pending
);
3237 clear_bit(R5_WriteError
, &dev
->flags
);
3239 if (rdev
&& test_bit(R5_MadeGood
, &dev
->flags
)) {
3240 /* This flag does not apply to '.replacement'
3241 * only to .rdev, so make sure to check that*/
3242 struct md_rdev
*rdev2
= rcu_dereference(
3243 conf
->disks
[i
].rdev
);
3244 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3245 s
->handle_bad_blocks
= 1;
3246 atomic_inc(&rdev2
->nr_pending
);
3248 clear_bit(R5_MadeGood
, &dev
->flags
);
3250 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
3251 struct md_rdev
*rdev2
= rcu_dereference(
3252 conf
->disks
[i
].replacement
);
3253 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3254 s
->handle_bad_blocks
= 1;
3255 atomic_inc(&rdev2
->nr_pending
);
3257 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
3259 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3260 /* The ReadError flag will just be confusing now */
3261 clear_bit(R5_ReadError
, &dev
->flags
);
3262 clear_bit(R5_ReWrite
, &dev
->flags
);
3264 if (test_bit(R5_ReadError
, &dev
->flags
))
3265 clear_bit(R5_Insync
, &dev
->flags
);
3266 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3268 s
->failed_num
[s
->failed
] = i
;
3270 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
3274 spin_unlock_irq(&conf
->device_lock
);
3275 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
3276 /* If there is a failed device being replaced,
3277 * we must be recovering.
3278 * else if we are after recovery_cp, we must be syncing
3279 * else we can only be replacing
3280 * sync and recovery both need to read all devices, and so
3281 * use the same flag.
3284 sh
->sector
>= conf
->mddev
->recovery_cp
)
3292 static void handle_stripe(struct stripe_head
*sh
)
3294 struct stripe_head_state s
;
3295 struct r5conf
*conf
= sh
->raid_conf
;
3298 int disks
= sh
->disks
;
3299 struct r5dev
*pdev
, *qdev
;
3301 clear_bit(STRIPE_HANDLE
, &sh
->state
);
3302 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
3303 /* already being handled, ensure it gets handled
3304 * again when current action finishes */
3305 set_bit(STRIPE_HANDLE
, &sh
->state
);
3309 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
3310 set_bit(STRIPE_SYNCING
, &sh
->state
);
3311 clear_bit(STRIPE_INSYNC
, &sh
->state
);
3313 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3315 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3316 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3317 (unsigned long long)sh
->sector
, sh
->state
,
3318 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
3319 sh
->check_state
, sh
->reconstruct_state
);
3321 analyse_stripe(sh
, &s
);
3323 if (s
.handle_bad_blocks
) {
3324 set_bit(STRIPE_HANDLE
, &sh
->state
);
3328 if (unlikely(s
.blocked_rdev
)) {
3329 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
3330 s
.replacing
|| s
.to_write
|| s
.written
) {
3331 set_bit(STRIPE_HANDLE
, &sh
->state
);
3334 /* There is nothing for the blocked_rdev to block */
3335 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
3336 s
.blocked_rdev
= NULL
;
3339 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
3340 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
3341 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
3344 pr_debug("locked=%d uptodate=%d to_read=%d"
3345 " to_write=%d failed=%d failed_num=%d,%d\n",
3346 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
3347 s
.failed_num
[0], s
.failed_num
[1]);
3348 /* check if the array has lost more than max_degraded devices and,
3349 * if so, some requests might need to be failed.
3351 if (s
.failed
> conf
->max_degraded
) {
3352 sh
->check_state
= 0;
3353 sh
->reconstruct_state
= 0;
3354 if (s
.to_read
+s
.to_write
+s
.written
)
3355 handle_failed_stripe(conf
, sh
, &s
, disks
, &s
.return_bi
);
3356 if (s
.syncing
+ s
.replacing
)
3357 handle_failed_sync(conf
, sh
, &s
);
3361 * might be able to return some write requests if the parity blocks
3362 * are safe, or on a failed drive
3364 pdev
= &sh
->dev
[sh
->pd_idx
];
3365 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
3366 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
3367 qdev
= &sh
->dev
[sh
->qd_idx
];
3368 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
3369 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
3373 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
3374 && !test_bit(R5_LOCKED
, &pdev
->flags
)
3375 && test_bit(R5_UPTODATE
, &pdev
->flags
)))) &&
3376 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
3377 && !test_bit(R5_LOCKED
, &qdev
->flags
)
3378 && test_bit(R5_UPTODATE
, &qdev
->flags
)))))
3379 handle_stripe_clean_event(conf
, sh
, disks
, &s
.return_bi
);
3381 /* Now we might consider reading some blocks, either to check/generate
3382 * parity, or to satisfy requests
3383 * or to load a block that is being partially written.
3385 if (s
.to_read
|| s
.non_overwrite
3386 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
3387 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
3390 handle_stripe_fill(sh
, &s
, disks
);
3392 /* Now we check to see if any write operations have recently
3396 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
3398 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
3399 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
3400 sh
->reconstruct_state
= reconstruct_state_idle
;
3402 /* All the 'written' buffers and the parity block are ready to
3403 * be written back to disk
3405 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
));
3406 BUG_ON(sh
->qd_idx
>= 0 &&
3407 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
));
3408 for (i
= disks
; i
--; ) {
3409 struct r5dev
*dev
= &sh
->dev
[i
];
3410 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
3411 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3413 pr_debug("Writing block %d\n", i
);
3414 set_bit(R5_Wantwrite
, &dev
->flags
);
3417 if (!test_bit(R5_Insync
, &dev
->flags
) ||
3418 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3420 set_bit(STRIPE_INSYNC
, &sh
->state
);
3423 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3424 s
.dec_preread_active
= 1;
3427 /* Now to consider new write requests and what else, if anything
3428 * should be read. We do not handle new writes when:
3429 * 1/ A 'write' operation (copy+xor) is already in flight.
3430 * 2/ A 'check' operation is in flight, as it may clobber the parity
3433 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
3434 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
3436 /* maybe we need to check and possibly fix the parity for this stripe
3437 * Any reads will already have been scheduled, so we just see if enough
3438 * data is available. The parity check is held off while parity
3439 * dependent operations are in flight.
3441 if (sh
->check_state
||
3442 (s
.syncing
&& s
.locked
== 0 &&
3443 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
3444 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
3445 if (conf
->level
== 6)
3446 handle_parity_checks6(conf
, sh
, &s
, disks
);
3448 handle_parity_checks5(conf
, sh
, &s
, disks
);
3451 if (s
.replacing
&& s
.locked
== 0
3452 && !test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3453 /* Write out to replacement devices where possible */
3454 for (i
= 0; i
< conf
->raid_disks
; i
++)
3455 if (test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
) &&
3456 test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
3457 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
3458 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3461 set_bit(STRIPE_INSYNC
, &sh
->state
);
3463 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
3464 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3465 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3466 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3469 /* If the failed drives are just a ReadError, then we might need
3470 * to progress the repair/check process
3472 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
3473 for (i
= 0; i
< s
.failed
; i
++) {
3474 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
3475 if (test_bit(R5_ReadError
, &dev
->flags
)
3476 && !test_bit(R5_LOCKED
, &dev
->flags
)
3477 && test_bit(R5_UPTODATE
, &dev
->flags
)
3479 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3480 set_bit(R5_Wantwrite
, &dev
->flags
);
3481 set_bit(R5_ReWrite
, &dev
->flags
);
3482 set_bit(R5_LOCKED
, &dev
->flags
);
3485 /* let's read it back */
3486 set_bit(R5_Wantread
, &dev
->flags
);
3487 set_bit(R5_LOCKED
, &dev
->flags
);
3494 /* Finish reconstruct operations initiated by the expansion process */
3495 if (sh
->reconstruct_state
== reconstruct_state_result
) {
3496 struct stripe_head
*sh_src
3497 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
3498 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
3499 /* sh cannot be written until sh_src has been read.
3500 * so arrange for sh to be delayed a little
3502 set_bit(STRIPE_DELAYED
, &sh
->state
);
3503 set_bit(STRIPE_HANDLE
, &sh
->state
);
3504 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
3506 atomic_inc(&conf
->preread_active_stripes
);
3507 release_stripe(sh_src
);
3511 release_stripe(sh_src
);
3513 sh
->reconstruct_state
= reconstruct_state_idle
;
3514 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3515 for (i
= conf
->raid_disks
; i
--; ) {
3516 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3517 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3522 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
3523 !sh
->reconstruct_state
) {
3524 /* Need to write out all blocks after computing parity */
3525 sh
->disks
= conf
->raid_disks
;
3526 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
3527 schedule_reconstruction(sh
, &s
, 1, 1);
3528 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
3529 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3530 atomic_dec(&conf
->reshape_stripes
);
3531 wake_up(&conf
->wait_for_overlap
);
3532 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3535 if (s
.expanding
&& s
.locked
== 0 &&
3536 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
3537 handle_stripe_expansion(conf
, sh
);
3540 /* wait for this device to become unblocked */
3541 if (conf
->mddev
->external
&& unlikely(s
.blocked_rdev
))
3542 md_wait_for_blocked_rdev(s
.blocked_rdev
, conf
->mddev
);
3544 if (s
.handle_bad_blocks
)
3545 for (i
= disks
; i
--; ) {
3546 struct md_rdev
*rdev
;
3547 struct r5dev
*dev
= &sh
->dev
[i
];
3548 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
3549 /* We own a safe reference to the rdev */
3550 rdev
= conf
->disks
[i
].rdev
;
3551 if (!rdev_set_badblocks(rdev
, sh
->sector
,
3553 md_error(conf
->mddev
, rdev
);
3554 rdev_dec_pending(rdev
, conf
->mddev
);
3556 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
3557 rdev
= conf
->disks
[i
].rdev
;
3558 rdev_clear_badblocks(rdev
, sh
->sector
,
3560 rdev_dec_pending(rdev
, conf
->mddev
);
3562 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
3563 rdev
= conf
->disks
[i
].replacement
;
3565 /* rdev have been moved down */
3566 rdev
= conf
->disks
[i
].rdev
;
3567 rdev_clear_badblocks(rdev
, sh
->sector
,
3569 rdev_dec_pending(rdev
, conf
->mddev
);
3574 raid_run_ops(sh
, s
.ops_request
);
3578 if (s
.dec_preread_active
) {
3579 /* We delay this until after ops_run_io so that if make_request
3580 * is waiting on a flush, it won't continue until the writes
3581 * have actually been submitted.
3583 atomic_dec(&conf
->preread_active_stripes
);
3584 if (atomic_read(&conf
->preread_active_stripes
) <
3586 md_wakeup_thread(conf
->mddev
->thread
);
3589 return_io(s
.return_bi
);
3591 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
3594 static void raid5_activate_delayed(struct r5conf
*conf
)
3596 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
3597 while (!list_empty(&conf
->delayed_list
)) {
3598 struct list_head
*l
= conf
->delayed_list
.next
;
3599 struct stripe_head
*sh
;
3600 sh
= list_entry(l
, struct stripe_head
, lru
);
3602 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3603 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3604 atomic_inc(&conf
->preread_active_stripes
);
3605 list_add_tail(&sh
->lru
, &conf
->hold_list
);
3610 static void activate_bit_delay(struct r5conf
*conf
)
3612 /* device_lock is held */
3613 struct list_head head
;
3614 list_add(&head
, &conf
->bitmap_list
);
3615 list_del_init(&conf
->bitmap_list
);
3616 while (!list_empty(&head
)) {
3617 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
3618 list_del_init(&sh
->lru
);
3619 atomic_inc(&sh
->count
);
3620 __release_stripe(conf
, sh
);
3624 int md_raid5_congested(struct mddev
*mddev
, int bits
)
3626 struct r5conf
*conf
= mddev
->private;
3628 /* No difference between reads and writes. Just check
3629 * how busy the stripe_cache is
3632 if (conf
->inactive_blocked
)
3636 if (list_empty_careful(&conf
->inactive_list
))
3641 EXPORT_SYMBOL_GPL(md_raid5_congested
);
3643 static int raid5_congested(void *data
, int bits
)
3645 struct mddev
*mddev
= data
;
3647 return mddev_congested(mddev
, bits
) ||
3648 md_raid5_congested(mddev
, bits
);
3651 /* We want read requests to align with chunks where possible,
3652 * but write requests don't need to.
3654 static int raid5_mergeable_bvec(struct request_queue
*q
,
3655 struct bvec_merge_data
*bvm
,
3656 struct bio_vec
*biovec
)
3658 struct mddev
*mddev
= q
->queuedata
;
3659 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
3661 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3662 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
3664 if ((bvm
->bi_rw
& 1) == WRITE
)
3665 return biovec
->bv_len
; /* always allow writes to be mergeable */
3667 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3668 chunk_sectors
= mddev
->new_chunk_sectors
;
3669 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
3670 if (max
< 0) max
= 0;
3671 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
3672 return biovec
->bv_len
;
3678 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
3680 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
3681 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3682 unsigned int bio_sectors
= bio
->bi_size
>> 9;
3684 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3685 chunk_sectors
= mddev
->new_chunk_sectors
;
3686 return chunk_sectors
>=
3687 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
3691 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3692 * later sampled by raid5d.
3694 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
3696 unsigned long flags
;
3698 spin_lock_irqsave(&conf
->device_lock
, flags
);
3700 bi
->bi_next
= conf
->retry_read_aligned_list
;
3701 conf
->retry_read_aligned_list
= bi
;
3703 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3704 md_wakeup_thread(conf
->mddev
->thread
);
3708 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
)
3712 bi
= conf
->retry_read_aligned
;
3714 conf
->retry_read_aligned
= NULL
;
3717 bi
= conf
->retry_read_aligned_list
;
3719 conf
->retry_read_aligned_list
= bi
->bi_next
;
3722 * this sets the active strip count to 1 and the processed
3723 * strip count to zero (upper 8 bits)
3725 bi
->bi_phys_segments
= 1; /* biased count of active stripes */
3733 * The "raid5_align_endio" should check if the read succeeded and if it
3734 * did, call bio_endio on the original bio (having bio_put the new bio
3736 * If the read failed..
3738 static void raid5_align_endio(struct bio
*bi
, int error
)
3740 struct bio
* raid_bi
= bi
->bi_private
;
3741 struct mddev
*mddev
;
3742 struct r5conf
*conf
;
3743 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3744 struct md_rdev
*rdev
;
3748 rdev
= (void*)raid_bi
->bi_next
;
3749 raid_bi
->bi_next
= NULL
;
3750 mddev
= rdev
->mddev
;
3751 conf
= mddev
->private;
3753 rdev_dec_pending(rdev
, conf
->mddev
);
3755 if (!error
&& uptodate
) {
3756 bio_endio(raid_bi
, 0);
3757 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
3758 wake_up(&conf
->wait_for_stripe
);
3763 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3765 add_bio_to_retry(raid_bi
, conf
);
3768 static int bio_fits_rdev(struct bio
*bi
)
3770 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
3772 if ((bi
->bi_size
>>9) > queue_max_sectors(q
))
3774 blk_recount_segments(q
, bi
);
3775 if (bi
->bi_phys_segments
> queue_max_segments(q
))
3778 if (q
->merge_bvec_fn
)
3779 /* it's too hard to apply the merge_bvec_fn at this stage,
3788 static int chunk_aligned_read(struct mddev
*mddev
, struct bio
* raid_bio
)
3790 struct r5conf
*conf
= mddev
->private;
3792 struct bio
* align_bi
;
3793 struct md_rdev
*rdev
;
3794 sector_t end_sector
;
3796 if (!in_chunk_boundary(mddev
, raid_bio
)) {
3797 pr_debug("chunk_aligned_read : non aligned\n");
3801 * use bio_clone_mddev to make a copy of the bio
3803 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
3807 * set bi_end_io to a new function, and set bi_private to the
3810 align_bi
->bi_end_io
= raid5_align_endio
;
3811 align_bi
->bi_private
= raid_bio
;
3815 align_bi
->bi_sector
= raid5_compute_sector(conf
, raid_bio
->bi_sector
,
3819 end_sector
= align_bi
->bi_sector
+ (align_bi
->bi_size
>> 9);
3821 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
3822 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
3823 rdev
->recovery_offset
< end_sector
) {
3824 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
3826 (test_bit(Faulty
, &rdev
->flags
) ||
3827 !(test_bit(In_sync
, &rdev
->flags
) ||
3828 rdev
->recovery_offset
>= end_sector
)))
3835 atomic_inc(&rdev
->nr_pending
);
3837 raid_bio
->bi_next
= (void*)rdev
;
3838 align_bi
->bi_bdev
= rdev
->bdev
;
3839 align_bi
->bi_flags
&= ~(1 << BIO_SEG_VALID
);
3840 align_bi
->bi_sector
+= rdev
->data_offset
;
3842 if (!bio_fits_rdev(align_bi
) ||
3843 is_badblock(rdev
, align_bi
->bi_sector
, align_bi
->bi_size
>>9,
3844 &first_bad
, &bad_sectors
)) {
3845 /* too big in some way, or has a known bad block */
3847 rdev_dec_pending(rdev
, mddev
);
3851 spin_lock_irq(&conf
->device_lock
);
3852 wait_event_lock_irq(conf
->wait_for_stripe
,
3854 conf
->device_lock
, /* nothing */);
3855 atomic_inc(&conf
->active_aligned_reads
);
3856 spin_unlock_irq(&conf
->device_lock
);
3858 generic_make_request(align_bi
);
3867 /* __get_priority_stripe - get the next stripe to process
3869 * Full stripe writes are allowed to pass preread active stripes up until
3870 * the bypass_threshold is exceeded. In general the bypass_count
3871 * increments when the handle_list is handled before the hold_list; however, it
3872 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3873 * stripe with in flight i/o. The bypass_count will be reset when the
3874 * head of the hold_list has changed, i.e. the head was promoted to the
3877 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
)
3879 struct stripe_head
*sh
;
3881 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3883 list_empty(&conf
->handle_list
) ? "empty" : "busy",
3884 list_empty(&conf
->hold_list
) ? "empty" : "busy",
3885 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
3887 if (!list_empty(&conf
->handle_list
)) {
3888 sh
= list_entry(conf
->handle_list
.next
, typeof(*sh
), lru
);
3890 if (list_empty(&conf
->hold_list
))
3891 conf
->bypass_count
= 0;
3892 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
3893 if (conf
->hold_list
.next
== conf
->last_hold
)
3894 conf
->bypass_count
++;
3896 conf
->last_hold
= conf
->hold_list
.next
;
3897 conf
->bypass_count
-= conf
->bypass_threshold
;
3898 if (conf
->bypass_count
< 0)
3899 conf
->bypass_count
= 0;
3902 } else if (!list_empty(&conf
->hold_list
) &&
3903 ((conf
->bypass_threshold
&&
3904 conf
->bypass_count
> conf
->bypass_threshold
) ||
3905 atomic_read(&conf
->pending_full_writes
) == 0)) {
3906 sh
= list_entry(conf
->hold_list
.next
,
3908 conf
->bypass_count
-= conf
->bypass_threshold
;
3909 if (conf
->bypass_count
< 0)
3910 conf
->bypass_count
= 0;
3914 list_del_init(&sh
->lru
);
3915 atomic_inc(&sh
->count
);
3916 BUG_ON(atomic_read(&sh
->count
) != 1);
3920 static void make_request(struct mddev
*mddev
, struct bio
* bi
)
3922 struct r5conf
*conf
= mddev
->private;
3924 sector_t new_sector
;
3925 sector_t logical_sector
, last_sector
;
3926 struct stripe_head
*sh
;
3927 const int rw
= bio_data_dir(bi
);
3931 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
3932 md_flush_request(mddev
, bi
);
3936 md_write_start(mddev
, bi
);
3939 mddev
->reshape_position
== MaxSector
&&
3940 chunk_aligned_read(mddev
,bi
))
3943 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
3944 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
3946 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
3948 plugged
= mddev_check_plugged(mddev
);
3949 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
3951 int disks
, data_disks
;
3956 disks
= conf
->raid_disks
;
3957 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
3958 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
3959 /* spinlock is needed as reshape_progress may be
3960 * 64bit on a 32bit platform, and so it might be
3961 * possible to see a half-updated value
3962 * Of course reshape_progress could change after
3963 * the lock is dropped, so once we get a reference
3964 * to the stripe that we think it is, we will have
3967 spin_lock_irq(&conf
->device_lock
);
3968 if (mddev
->delta_disks
< 0
3969 ? logical_sector
< conf
->reshape_progress
3970 : logical_sector
>= conf
->reshape_progress
) {
3971 disks
= conf
->previous_raid_disks
;
3974 if (mddev
->delta_disks
< 0
3975 ? logical_sector
< conf
->reshape_safe
3976 : logical_sector
>= conf
->reshape_safe
) {
3977 spin_unlock_irq(&conf
->device_lock
);
3982 spin_unlock_irq(&conf
->device_lock
);
3984 data_disks
= disks
- conf
->max_degraded
;
3986 new_sector
= raid5_compute_sector(conf
, logical_sector
,
3989 pr_debug("raid456: make_request, sector %llu logical %llu\n",
3990 (unsigned long long)new_sector
,
3991 (unsigned long long)logical_sector
);
3993 sh
= get_active_stripe(conf
, new_sector
, previous
,
3994 (bi
->bi_rw
&RWA_MASK
), 0);
3996 if (unlikely(previous
)) {
3997 /* expansion might have moved on while waiting for a
3998 * stripe, so we must do the range check again.
3999 * Expansion could still move past after this
4000 * test, but as we are holding a reference to
4001 * 'sh', we know that if that happens,
4002 * STRIPE_EXPANDING will get set and the expansion
4003 * won't proceed until we finish with the stripe.
4006 spin_lock_irq(&conf
->device_lock
);
4007 if (mddev
->delta_disks
< 0
4008 ? logical_sector
>= conf
->reshape_progress
4009 : logical_sector
< conf
->reshape_progress
)
4010 /* mismatch, need to try again */
4012 spin_unlock_irq(&conf
->device_lock
);
4021 logical_sector
>= mddev
->suspend_lo
&&
4022 logical_sector
< mddev
->suspend_hi
) {
4024 /* As the suspend_* range is controlled by
4025 * userspace, we want an interruptible
4028 flush_signals(current
);
4029 prepare_to_wait(&conf
->wait_for_overlap
,
4030 &w
, TASK_INTERRUPTIBLE
);
4031 if (logical_sector
>= mddev
->suspend_lo
&&
4032 logical_sector
< mddev
->suspend_hi
)
4037 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
4038 !add_stripe_bio(sh
, bi
, dd_idx
, rw
)) {
4039 /* Stripe is busy expanding or
4040 * add failed due to overlap. Flush everything
4043 md_wakeup_thread(mddev
->thread
);
4048 finish_wait(&conf
->wait_for_overlap
, &w
);
4049 set_bit(STRIPE_HANDLE
, &sh
->state
);
4050 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4051 if ((bi
->bi_rw
& REQ_SYNC
) &&
4052 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4053 atomic_inc(&conf
->preread_active_stripes
);
4056 /* cannot get stripe for read-ahead, just give-up */
4057 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
4058 finish_wait(&conf
->wait_for_overlap
, &w
);
4064 md_wakeup_thread(mddev
->thread
);
4066 spin_lock_irq(&conf
->device_lock
);
4067 remaining
= raid5_dec_bi_phys_segments(bi
);
4068 spin_unlock_irq(&conf
->device_lock
);
4069 if (remaining
== 0) {
4072 md_write_end(mddev
);
4078 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
4080 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
4082 /* reshaping is quite different to recovery/resync so it is
4083 * handled quite separately ... here.
4085 * On each call to sync_request, we gather one chunk worth of
4086 * destination stripes and flag them as expanding.
4087 * Then we find all the source stripes and request reads.
4088 * As the reads complete, handle_stripe will copy the data
4089 * into the destination stripe and release that stripe.
4091 struct r5conf
*conf
= mddev
->private;
4092 struct stripe_head
*sh
;
4093 sector_t first_sector
, last_sector
;
4094 int raid_disks
= conf
->previous_raid_disks
;
4095 int data_disks
= raid_disks
- conf
->max_degraded
;
4096 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
4099 sector_t writepos
, readpos
, safepos
;
4100 sector_t stripe_addr
;
4101 int reshape_sectors
;
4102 struct list_head stripes
;
4104 if (sector_nr
== 0) {
4105 /* If restarting in the middle, skip the initial sectors */
4106 if (mddev
->delta_disks
< 0 &&
4107 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
4108 sector_nr
= raid5_size(mddev
, 0, 0)
4109 - conf
->reshape_progress
;
4110 } else if (mddev
->delta_disks
>= 0 &&
4111 conf
->reshape_progress
> 0)
4112 sector_nr
= conf
->reshape_progress
;
4113 sector_div(sector_nr
, new_data_disks
);
4115 mddev
->curr_resync_completed
= sector_nr
;
4116 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4122 /* We need to process a full chunk at a time.
4123 * If old and new chunk sizes differ, we need to process the
4126 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
)
4127 reshape_sectors
= mddev
->new_chunk_sectors
;
4129 reshape_sectors
= mddev
->chunk_sectors
;
4131 /* we update the metadata when there is more than 3Meg
4132 * in the block range (that is rather arbitrary, should
4133 * probably be time based) or when the data about to be
4134 * copied would over-write the source of the data at
4135 * the front of the range.
4136 * i.e. one new_stripe along from reshape_progress new_maps
4137 * to after where reshape_safe old_maps to
4139 writepos
= conf
->reshape_progress
;
4140 sector_div(writepos
, new_data_disks
);
4141 readpos
= conf
->reshape_progress
;
4142 sector_div(readpos
, data_disks
);
4143 safepos
= conf
->reshape_safe
;
4144 sector_div(safepos
, data_disks
);
4145 if (mddev
->delta_disks
< 0) {
4146 writepos
-= min_t(sector_t
, reshape_sectors
, writepos
);
4147 readpos
+= reshape_sectors
;
4148 safepos
+= reshape_sectors
;
4150 writepos
+= reshape_sectors
;
4151 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
4152 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
4155 /* 'writepos' is the most advanced device address we might write.
4156 * 'readpos' is the least advanced device address we might read.
4157 * 'safepos' is the least address recorded in the metadata as having
4159 * If 'readpos' is behind 'writepos', then there is no way that we can
4160 * ensure safety in the face of a crash - that must be done by userspace
4161 * making a backup of the data. So in that case there is no particular
4162 * rush to update metadata.
4163 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4164 * update the metadata to advance 'safepos' to match 'readpos' so that
4165 * we can be safe in the event of a crash.
4166 * So we insist on updating metadata if safepos is behind writepos and
4167 * readpos is beyond writepos.
4168 * In any case, update the metadata every 10 seconds.
4169 * Maybe that number should be configurable, but I'm not sure it is
4170 * worth it.... maybe it could be a multiple of safemode_delay???
4172 if ((mddev
->delta_disks
< 0
4173 ? (safepos
> writepos
&& readpos
< writepos
)
4174 : (safepos
< writepos
&& readpos
> writepos
)) ||
4175 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4176 /* Cannot proceed until we've updated the superblock... */
4177 wait_event(conf
->wait_for_overlap
,
4178 atomic_read(&conf
->reshape_stripes
)==0);
4179 mddev
->reshape_position
= conf
->reshape_progress
;
4180 mddev
->curr_resync_completed
= sector_nr
;
4181 conf
->reshape_checkpoint
= jiffies
;
4182 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4183 md_wakeup_thread(mddev
->thread
);
4184 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4185 kthread_should_stop());
4186 spin_lock_irq(&conf
->device_lock
);
4187 conf
->reshape_safe
= mddev
->reshape_position
;
4188 spin_unlock_irq(&conf
->device_lock
);
4189 wake_up(&conf
->wait_for_overlap
);
4190 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4193 if (mddev
->delta_disks
< 0) {
4194 BUG_ON(conf
->reshape_progress
== 0);
4195 stripe_addr
= writepos
;
4196 BUG_ON((mddev
->dev_sectors
&
4197 ~((sector_t
)reshape_sectors
- 1))
4198 - reshape_sectors
- stripe_addr
4201 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
4202 stripe_addr
= sector_nr
;
4204 INIT_LIST_HEAD(&stripes
);
4205 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
4207 int skipped_disk
= 0;
4208 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
4209 set_bit(STRIPE_EXPANDING
, &sh
->state
);
4210 atomic_inc(&conf
->reshape_stripes
);
4211 /* If any of this stripe is beyond the end of the old
4212 * array, then we need to zero those blocks
4214 for (j
=sh
->disks
; j
--;) {
4216 if (j
== sh
->pd_idx
)
4218 if (conf
->level
== 6 &&
4221 s
= compute_blocknr(sh
, j
, 0);
4222 if (s
< raid5_size(mddev
, 0, 0)) {
4226 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
4227 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
4228 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
4230 if (!skipped_disk
) {
4231 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4232 set_bit(STRIPE_HANDLE
, &sh
->state
);
4234 list_add(&sh
->lru
, &stripes
);
4236 spin_lock_irq(&conf
->device_lock
);
4237 if (mddev
->delta_disks
< 0)
4238 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
4240 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
4241 spin_unlock_irq(&conf
->device_lock
);
4242 /* Ok, those stripe are ready. We can start scheduling
4243 * reads on the source stripes.
4244 * The source stripes are determined by mapping the first and last
4245 * block on the destination stripes.
4248 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
4251 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
4252 * new_data_disks
- 1),
4254 if (last_sector
>= mddev
->dev_sectors
)
4255 last_sector
= mddev
->dev_sectors
- 1;
4256 while (first_sector
<= last_sector
) {
4257 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
4258 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
4259 set_bit(STRIPE_HANDLE
, &sh
->state
);
4261 first_sector
+= STRIPE_SECTORS
;
4263 /* Now that the sources are clearly marked, we can release
4264 * the destination stripes
4266 while (!list_empty(&stripes
)) {
4267 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
4268 list_del_init(&sh
->lru
);
4271 /* If this takes us to the resync_max point where we have to pause,
4272 * then we need to write out the superblock.
4274 sector_nr
+= reshape_sectors
;
4275 if ((sector_nr
- mddev
->curr_resync_completed
) * 2
4276 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
4277 /* Cannot proceed until we've updated the superblock... */
4278 wait_event(conf
->wait_for_overlap
,
4279 atomic_read(&conf
->reshape_stripes
) == 0);
4280 mddev
->reshape_position
= conf
->reshape_progress
;
4281 mddev
->curr_resync_completed
= sector_nr
;
4282 conf
->reshape_checkpoint
= jiffies
;
4283 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4284 md_wakeup_thread(mddev
->thread
);
4285 wait_event(mddev
->sb_wait
,
4286 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
4287 || kthread_should_stop());
4288 spin_lock_irq(&conf
->device_lock
);
4289 conf
->reshape_safe
= mddev
->reshape_position
;
4290 spin_unlock_irq(&conf
->device_lock
);
4291 wake_up(&conf
->wait_for_overlap
);
4292 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4294 return reshape_sectors
;
4297 /* FIXME go_faster isn't used */
4298 static inline sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
4300 struct r5conf
*conf
= mddev
->private;
4301 struct stripe_head
*sh
;
4302 sector_t max_sector
= mddev
->dev_sectors
;
4303 sector_t sync_blocks
;
4304 int still_degraded
= 0;
4307 if (sector_nr
>= max_sector
) {
4308 /* just being told to finish up .. nothing much to do */
4310 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
4315 if (mddev
->curr_resync
< max_sector
) /* aborted */
4316 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
4318 else /* completed sync */
4320 bitmap_close_sync(mddev
->bitmap
);
4325 /* Allow raid5_quiesce to complete */
4326 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
4328 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
4329 return reshape_request(mddev
, sector_nr
, skipped
);
4331 /* No need to check resync_max as we never do more than one
4332 * stripe, and as resync_max will always be on a chunk boundary,
4333 * if the check in md_do_sync didn't fire, there is no chance
4334 * of overstepping resync_max here
4337 /* if there is too many failed drives and we are trying
4338 * to resync, then assert that we are finished, because there is
4339 * nothing we can do.
4341 if (mddev
->degraded
>= conf
->max_degraded
&&
4342 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
4343 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
4347 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
4348 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
4349 !conf
->fullsync
&& sync_blocks
>= STRIPE_SECTORS
) {
4350 /* we can skip this block, and probably more */
4351 sync_blocks
/= STRIPE_SECTORS
;
4353 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
4356 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
4358 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
4360 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
4361 /* make sure we don't swamp the stripe cache if someone else
4362 * is trying to get access
4364 schedule_timeout_uninterruptible(1);
4366 /* Need to check if array will still be degraded after recovery/resync
4367 * We don't need to check the 'failed' flag as when that gets set,
4370 for (i
= 0; i
< conf
->raid_disks
; i
++)
4371 if (conf
->disks
[i
].rdev
== NULL
)
4374 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
4376 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
4381 return STRIPE_SECTORS
;
4384 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
)
4386 /* We may not be able to submit a whole bio at once as there
4387 * may not be enough stripe_heads available.
4388 * We cannot pre-allocate enough stripe_heads as we may need
4389 * more than exist in the cache (if we allow ever large chunks).
4390 * So we do one stripe head at a time and record in
4391 * ->bi_hw_segments how many have been done.
4393 * We *know* that this entire raid_bio is in one chunk, so
4394 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4396 struct stripe_head
*sh
;
4398 sector_t sector
, logical_sector
, last_sector
;
4403 logical_sector
= raid_bio
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4404 sector
= raid5_compute_sector(conf
, logical_sector
,
4406 last_sector
= raid_bio
->bi_sector
+ (raid_bio
->bi_size
>>9);
4408 for (; logical_sector
< last_sector
;
4409 logical_sector
+= STRIPE_SECTORS
,
4410 sector
+= STRIPE_SECTORS
,
4413 if (scnt
< raid5_bi_hw_segments(raid_bio
))
4414 /* already done this stripe */
4417 sh
= get_active_stripe(conf
, sector
, 0, 1, 0);
4420 /* failed to get a stripe - must wait */
4421 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4422 conf
->retry_read_aligned
= raid_bio
;
4426 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0)) {
4428 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4429 conf
->retry_read_aligned
= raid_bio
;
4437 spin_lock_irq(&conf
->device_lock
);
4438 remaining
= raid5_dec_bi_phys_segments(raid_bio
);
4439 spin_unlock_irq(&conf
->device_lock
);
4441 bio_endio(raid_bio
, 0);
4442 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4443 wake_up(&conf
->wait_for_stripe
);
4449 * This is our raid5 kernel thread.
4451 * We scan the hash table for stripes which can be handled now.
4452 * During the scan, completed stripes are saved for us by the interrupt
4453 * handler, so that they will not have to wait for our next wakeup.
4455 static void raid5d(struct mddev
*mddev
)
4457 struct stripe_head
*sh
;
4458 struct r5conf
*conf
= mddev
->private;
4460 struct blk_plug plug
;
4462 pr_debug("+++ raid5d active\n");
4464 md_check_recovery(mddev
);
4466 blk_start_plug(&plug
);
4468 spin_lock_irq(&conf
->device_lock
);
4472 if (atomic_read(&mddev
->plug_cnt
) == 0 &&
4473 !list_empty(&conf
->bitmap_list
)) {
4474 /* Now is a good time to flush some bitmap updates */
4476 spin_unlock_irq(&conf
->device_lock
);
4477 bitmap_unplug(mddev
->bitmap
);
4478 spin_lock_irq(&conf
->device_lock
);
4479 conf
->seq_write
= conf
->seq_flush
;
4480 activate_bit_delay(conf
);
4482 if (atomic_read(&mddev
->plug_cnt
) == 0)
4483 raid5_activate_delayed(conf
);
4485 while ((bio
= remove_bio_from_retry(conf
))) {
4487 spin_unlock_irq(&conf
->device_lock
);
4488 ok
= retry_aligned_read(conf
, bio
);
4489 spin_lock_irq(&conf
->device_lock
);
4495 sh
= __get_priority_stripe(conf
);
4499 spin_unlock_irq(&conf
->device_lock
);
4506 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
4507 md_check_recovery(mddev
);
4509 spin_lock_irq(&conf
->device_lock
);
4511 pr_debug("%d stripes handled\n", handled
);
4513 spin_unlock_irq(&conf
->device_lock
);
4515 async_tx_issue_pending_all();
4516 blk_finish_plug(&plug
);
4518 pr_debug("--- raid5d inactive\n");
4522 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
4524 struct r5conf
*conf
= mddev
->private;
4526 return sprintf(page
, "%d\n", conf
->max_nr_stripes
);
4532 raid5_set_cache_size(struct mddev
*mddev
, int size
)
4534 struct r5conf
*conf
= mddev
->private;
4537 if (size
<= 16 || size
> 32768)
4539 while (size
< conf
->max_nr_stripes
) {
4540 if (drop_one_stripe(conf
))
4541 conf
->max_nr_stripes
--;
4545 err
= md_allow_write(mddev
);
4548 while (size
> conf
->max_nr_stripes
) {
4549 if (grow_one_stripe(conf
))
4550 conf
->max_nr_stripes
++;
4555 EXPORT_SYMBOL(raid5_set_cache_size
);
4558 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
4560 struct r5conf
*conf
= mddev
->private;
4564 if (len
>= PAGE_SIZE
)
4569 if (strict_strtoul(page
, 10, &new))
4571 err
= raid5_set_cache_size(mddev
, new);
4577 static struct md_sysfs_entry
4578 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
4579 raid5_show_stripe_cache_size
,
4580 raid5_store_stripe_cache_size
);
4583 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
4585 struct r5conf
*conf
= mddev
->private;
4587 return sprintf(page
, "%d\n", conf
->bypass_threshold
);
4593 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
4595 struct r5conf
*conf
= mddev
->private;
4597 if (len
>= PAGE_SIZE
)
4602 if (strict_strtoul(page
, 10, &new))
4604 if (new > conf
->max_nr_stripes
)
4606 conf
->bypass_threshold
= new;
4610 static struct md_sysfs_entry
4611 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
4613 raid5_show_preread_threshold
,
4614 raid5_store_preread_threshold
);
4617 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
4619 struct r5conf
*conf
= mddev
->private;
4621 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
4626 static struct md_sysfs_entry
4627 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
4629 static struct attribute
*raid5_attrs
[] = {
4630 &raid5_stripecache_size
.attr
,
4631 &raid5_stripecache_active
.attr
,
4632 &raid5_preread_bypass_threshold
.attr
,
4635 static struct attribute_group raid5_attrs_group
= {
4637 .attrs
= raid5_attrs
,
4641 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
4643 struct r5conf
*conf
= mddev
->private;
4646 sectors
= mddev
->dev_sectors
;
4648 /* size is defined by the smallest of previous and new size */
4649 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
4651 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
4652 sectors
&= ~((sector_t
)mddev
->new_chunk_sectors
- 1);
4653 return sectors
* (raid_disks
- conf
->max_degraded
);
4656 static void raid5_free_percpu(struct r5conf
*conf
)
4658 struct raid5_percpu
*percpu
;
4665 for_each_possible_cpu(cpu
) {
4666 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4667 safe_put_page(percpu
->spare_page
);
4668 kfree(percpu
->scribble
);
4670 #ifdef CONFIG_HOTPLUG_CPU
4671 unregister_cpu_notifier(&conf
->cpu_notify
);
4675 free_percpu(conf
->percpu
);
4678 static void free_conf(struct r5conf
*conf
)
4680 shrink_stripes(conf
);
4681 raid5_free_percpu(conf
);
4683 kfree(conf
->stripe_hashtbl
);
4687 #ifdef CONFIG_HOTPLUG_CPU
4688 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
4691 struct r5conf
*conf
= container_of(nfb
, struct r5conf
, cpu_notify
);
4692 long cpu
= (long)hcpu
;
4693 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4696 case CPU_UP_PREPARE
:
4697 case CPU_UP_PREPARE_FROZEN
:
4698 if (conf
->level
== 6 && !percpu
->spare_page
)
4699 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
4700 if (!percpu
->scribble
)
4701 percpu
->scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
4703 if (!percpu
->scribble
||
4704 (conf
->level
== 6 && !percpu
->spare_page
)) {
4705 safe_put_page(percpu
->spare_page
);
4706 kfree(percpu
->scribble
);
4707 pr_err("%s: failed memory allocation for cpu%ld\n",
4709 return notifier_from_errno(-ENOMEM
);
4713 case CPU_DEAD_FROZEN
:
4714 safe_put_page(percpu
->spare_page
);
4715 kfree(percpu
->scribble
);
4716 percpu
->spare_page
= NULL
;
4717 percpu
->scribble
= NULL
;
4726 static int raid5_alloc_percpu(struct r5conf
*conf
)
4729 struct page
*spare_page
;
4730 struct raid5_percpu __percpu
*allcpus
;
4734 allcpus
= alloc_percpu(struct raid5_percpu
);
4737 conf
->percpu
= allcpus
;
4741 for_each_present_cpu(cpu
) {
4742 if (conf
->level
== 6) {
4743 spare_page
= alloc_page(GFP_KERNEL
);
4748 per_cpu_ptr(conf
->percpu
, cpu
)->spare_page
= spare_page
;
4750 scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
4755 per_cpu_ptr(conf
->percpu
, cpu
)->scribble
= scribble
;
4757 #ifdef CONFIG_HOTPLUG_CPU
4758 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
4759 conf
->cpu_notify
.priority
= 0;
4761 err
= register_cpu_notifier(&conf
->cpu_notify
);
4768 static struct r5conf
*setup_conf(struct mddev
*mddev
)
4770 struct r5conf
*conf
;
4771 int raid_disk
, memory
, max_disks
;
4772 struct md_rdev
*rdev
;
4773 struct disk_info
*disk
;
4775 if (mddev
->new_level
!= 5
4776 && mddev
->new_level
!= 4
4777 && mddev
->new_level
!= 6) {
4778 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4779 mdname(mddev
), mddev
->new_level
);
4780 return ERR_PTR(-EIO
);
4782 if ((mddev
->new_level
== 5
4783 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
4784 (mddev
->new_level
== 6
4785 && !algorithm_valid_raid6(mddev
->new_layout
))) {
4786 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
4787 mdname(mddev
), mddev
->new_layout
);
4788 return ERR_PTR(-EIO
);
4790 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
4791 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4792 mdname(mddev
), mddev
->raid_disks
);
4793 return ERR_PTR(-EINVAL
);
4796 if (!mddev
->new_chunk_sectors
||
4797 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
4798 !is_power_of_2(mddev
->new_chunk_sectors
)) {
4799 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
4800 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
4801 return ERR_PTR(-EINVAL
);
4804 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
4807 spin_lock_init(&conf
->device_lock
);
4808 init_waitqueue_head(&conf
->wait_for_stripe
);
4809 init_waitqueue_head(&conf
->wait_for_overlap
);
4810 INIT_LIST_HEAD(&conf
->handle_list
);
4811 INIT_LIST_HEAD(&conf
->hold_list
);
4812 INIT_LIST_HEAD(&conf
->delayed_list
);
4813 INIT_LIST_HEAD(&conf
->bitmap_list
);
4814 INIT_LIST_HEAD(&conf
->inactive_list
);
4815 atomic_set(&conf
->active_stripes
, 0);
4816 atomic_set(&conf
->preread_active_stripes
, 0);
4817 atomic_set(&conf
->active_aligned_reads
, 0);
4818 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
4819 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
4821 conf
->raid_disks
= mddev
->raid_disks
;
4822 if (mddev
->reshape_position
== MaxSector
)
4823 conf
->previous_raid_disks
= mddev
->raid_disks
;
4825 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4826 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
4827 conf
->scribble_len
= scribble_len(max_disks
);
4829 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
4834 conf
->mddev
= mddev
;
4836 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
4839 conf
->level
= mddev
->new_level
;
4840 if (raid5_alloc_percpu(conf
) != 0)
4843 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
4845 rdev_for_each(rdev
, mddev
) {
4846 raid_disk
= rdev
->raid_disk
;
4847 if (raid_disk
>= max_disks
4850 disk
= conf
->disks
+ raid_disk
;
4852 if (test_bit(Replacement
, &rdev
->flags
)) {
4853 if (disk
->replacement
)
4855 disk
->replacement
= rdev
;
4862 if (test_bit(In_sync
, &rdev
->flags
)) {
4863 char b
[BDEVNAME_SIZE
];
4864 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
4866 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
4867 } else if (rdev
->saved_raid_disk
!= raid_disk
)
4868 /* Cannot rely on bitmap to complete recovery */
4872 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
4873 conf
->level
= mddev
->new_level
;
4874 if (conf
->level
== 6)
4875 conf
->max_degraded
= 2;
4877 conf
->max_degraded
= 1;
4878 conf
->algorithm
= mddev
->new_layout
;
4879 conf
->max_nr_stripes
= NR_STRIPES
;
4880 conf
->reshape_progress
= mddev
->reshape_position
;
4881 if (conf
->reshape_progress
!= MaxSector
) {
4882 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
4883 conf
->prev_algo
= mddev
->layout
;
4886 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
4887 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
4888 if (grow_stripes(conf
, conf
->max_nr_stripes
)) {
4890 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4891 mdname(mddev
), memory
);
4894 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
4895 mdname(mddev
), memory
);
4897 conf
->thread
= md_register_thread(raid5d
, mddev
, NULL
);
4898 if (!conf
->thread
) {
4900 "md/raid:%s: couldn't allocate thread.\n",
4910 return ERR_PTR(-EIO
);
4912 return ERR_PTR(-ENOMEM
);
4916 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
4919 case ALGORITHM_PARITY_0
:
4920 if (raid_disk
< max_degraded
)
4923 case ALGORITHM_PARITY_N
:
4924 if (raid_disk
>= raid_disks
- max_degraded
)
4927 case ALGORITHM_PARITY_0_6
:
4928 if (raid_disk
== 0 ||
4929 raid_disk
== raid_disks
- 1)
4932 case ALGORITHM_LEFT_ASYMMETRIC_6
:
4933 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
4934 case ALGORITHM_LEFT_SYMMETRIC_6
:
4935 case ALGORITHM_RIGHT_SYMMETRIC_6
:
4936 if (raid_disk
== raid_disks
- 1)
4942 static int run(struct mddev
*mddev
)
4944 struct r5conf
*conf
;
4945 int working_disks
= 0;
4946 int dirty_parity_disks
= 0;
4947 struct md_rdev
*rdev
;
4948 sector_t reshape_offset
= 0;
4951 if (mddev
->recovery_cp
!= MaxSector
)
4952 printk(KERN_NOTICE
"md/raid:%s: not clean"
4953 " -- starting background reconstruction\n",
4955 if (mddev
->reshape_position
!= MaxSector
) {
4956 /* Check that we can continue the reshape.
4957 * Currently only disks can change, it must
4958 * increase, and we must be past the point where
4959 * a stripe over-writes itself
4961 sector_t here_new
, here_old
;
4963 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
4965 if (mddev
->new_level
!= mddev
->level
) {
4966 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
4967 "required - aborting.\n",
4971 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4972 /* reshape_position must be on a new-stripe boundary, and one
4973 * further up in new geometry must map after here in old
4976 here_new
= mddev
->reshape_position
;
4977 if (sector_div(here_new
, mddev
->new_chunk_sectors
*
4978 (mddev
->raid_disks
- max_degraded
))) {
4979 printk(KERN_ERR
"md/raid:%s: reshape_position not "
4980 "on a stripe boundary\n", mdname(mddev
));
4983 reshape_offset
= here_new
* mddev
->new_chunk_sectors
;
4984 /* here_new is the stripe we will write to */
4985 here_old
= mddev
->reshape_position
;
4986 sector_div(here_old
, mddev
->chunk_sectors
*
4987 (old_disks
-max_degraded
));
4988 /* here_old is the first stripe that we might need to read
4990 if (mddev
->delta_disks
== 0) {
4991 /* We cannot be sure it is safe to start an in-place
4992 * reshape. It is only safe if user-space if monitoring
4993 * and taking constant backups.
4994 * mdadm always starts a situation like this in
4995 * readonly mode so it can take control before
4996 * allowing any writes. So just check for that.
4998 if ((here_new
* mddev
->new_chunk_sectors
!=
4999 here_old
* mddev
->chunk_sectors
) ||
5001 printk(KERN_ERR
"md/raid:%s: in-place reshape must be started"
5002 " in read-only mode - aborting\n",
5006 } else if (mddev
->delta_disks
< 0
5007 ? (here_new
* mddev
->new_chunk_sectors
<=
5008 here_old
* mddev
->chunk_sectors
)
5009 : (here_new
* mddev
->new_chunk_sectors
>=
5010 here_old
* mddev
->chunk_sectors
)) {
5011 /* Reading from the same stripe as writing to - bad */
5012 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
5013 "auto-recovery - aborting.\n",
5017 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
5019 /* OK, we should be able to continue; */
5021 BUG_ON(mddev
->level
!= mddev
->new_level
);
5022 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
5023 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
5024 BUG_ON(mddev
->delta_disks
!= 0);
5027 if (mddev
->private == NULL
)
5028 conf
= setup_conf(mddev
);
5030 conf
= mddev
->private;
5033 return PTR_ERR(conf
);
5035 mddev
->thread
= conf
->thread
;
5036 conf
->thread
= NULL
;
5037 mddev
->private = conf
;
5039 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
5041 rdev
= conf
->disks
[i
].rdev
;
5042 if (!rdev
&& conf
->disks
[i
].replacement
) {
5043 /* The replacement is all we have yet */
5044 rdev
= conf
->disks
[i
].replacement
;
5045 conf
->disks
[i
].replacement
= NULL
;
5046 clear_bit(Replacement
, &rdev
->flags
);
5047 conf
->disks
[i
].rdev
= rdev
;
5051 if (conf
->disks
[i
].replacement
&&
5052 conf
->reshape_progress
!= MaxSector
) {
5053 /* replacements and reshape simply do not mix. */
5054 printk(KERN_ERR
"md: cannot handle concurrent "
5055 "replacement and reshape.\n");
5058 if (test_bit(In_sync
, &rdev
->flags
)) {
5062 /* This disc is not fully in-sync. However if it
5063 * just stored parity (beyond the recovery_offset),
5064 * when we don't need to be concerned about the
5065 * array being dirty.
5066 * When reshape goes 'backwards', we never have
5067 * partially completed devices, so we only need
5068 * to worry about reshape going forwards.
5070 /* Hack because v0.91 doesn't store recovery_offset properly. */
5071 if (mddev
->major_version
== 0 &&
5072 mddev
->minor_version
> 90)
5073 rdev
->recovery_offset
= reshape_offset
;
5075 if (rdev
->recovery_offset
< reshape_offset
) {
5076 /* We need to check old and new layout */
5077 if (!only_parity(rdev
->raid_disk
,
5080 conf
->max_degraded
))
5083 if (!only_parity(rdev
->raid_disk
,
5085 conf
->previous_raid_disks
,
5086 conf
->max_degraded
))
5088 dirty_parity_disks
++;
5092 * 0 for a fully functional array, 1 or 2 for a degraded array.
5094 mddev
->degraded
= calc_degraded(conf
);
5096 if (has_failed(conf
)) {
5097 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
5098 " (%d/%d failed)\n",
5099 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
5103 /* device size must be a multiple of chunk size */
5104 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
5105 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
5107 if (mddev
->degraded
> dirty_parity_disks
&&
5108 mddev
->recovery_cp
!= MaxSector
) {
5109 if (mddev
->ok_start_degraded
)
5111 "md/raid:%s: starting dirty degraded array"
5112 " - data corruption possible.\n",
5116 "md/raid:%s: cannot start dirty degraded array.\n",
5122 if (mddev
->degraded
== 0)
5123 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
5124 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
5125 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
5128 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
5129 " out of %d devices, algorithm %d\n",
5130 mdname(mddev
), conf
->level
,
5131 mddev
->raid_disks
- mddev
->degraded
,
5132 mddev
->raid_disks
, mddev
->new_layout
);
5134 print_raid5_conf(conf
);
5136 if (conf
->reshape_progress
!= MaxSector
) {
5137 conf
->reshape_safe
= conf
->reshape_progress
;
5138 atomic_set(&conf
->reshape_stripes
, 0);
5139 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5140 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5141 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5142 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5143 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5148 /* Ok, everything is just fine now */
5149 if (mddev
->to_remove
== &raid5_attrs_group
)
5150 mddev
->to_remove
= NULL
;
5151 else if (mddev
->kobj
.sd
&&
5152 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
5154 "raid5: failed to create sysfs attributes for %s\n",
5156 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5160 /* read-ahead size must cover two whole stripes, which
5161 * is 2 * (datadisks) * chunksize where 'n' is the
5162 * number of raid devices
5164 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
5165 int stripe
= data_disks
*
5166 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
5167 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
5168 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
5170 blk_queue_merge_bvec(mddev
->queue
, raid5_mergeable_bvec
);
5172 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
5173 mddev
->queue
->backing_dev_info
.congested_fn
= raid5_congested
;
5175 chunk_size
= mddev
->chunk_sectors
<< 9;
5176 blk_queue_io_min(mddev
->queue
, chunk_size
);
5177 blk_queue_io_opt(mddev
->queue
, chunk_size
*
5178 (conf
->raid_disks
- conf
->max_degraded
));
5180 rdev_for_each(rdev
, mddev
)
5181 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
5182 rdev
->data_offset
<< 9);
5187 md_unregister_thread(&mddev
->thread
);
5188 print_raid5_conf(conf
);
5190 mddev
->private = NULL
;
5191 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
5195 static int stop(struct mddev
*mddev
)
5197 struct r5conf
*conf
= mddev
->private;
5199 md_unregister_thread(&mddev
->thread
);
5201 mddev
->queue
->backing_dev_info
.congested_fn
= NULL
;
5203 mddev
->private = NULL
;
5204 mddev
->to_remove
= &raid5_attrs_group
;
5208 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
5210 struct r5conf
*conf
= mddev
->private;
5213 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
5214 mddev
->chunk_sectors
/ 2, mddev
->layout
);
5215 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
5216 for (i
= 0; i
< conf
->raid_disks
; i
++)
5217 seq_printf (seq
, "%s",
5218 conf
->disks
[i
].rdev
&&
5219 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
5220 seq_printf (seq
, "]");
5223 static void print_raid5_conf (struct r5conf
*conf
)
5226 struct disk_info
*tmp
;
5228 printk(KERN_DEBUG
"RAID conf printout:\n");
5230 printk("(conf==NULL)\n");
5233 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
5235 conf
->raid_disks
- conf
->mddev
->degraded
);
5237 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5238 char b
[BDEVNAME_SIZE
];
5239 tmp
= conf
->disks
+ i
;
5241 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
5242 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
5243 bdevname(tmp
->rdev
->bdev
, b
));
5247 static int raid5_spare_active(struct mddev
*mddev
)
5250 struct r5conf
*conf
= mddev
->private;
5251 struct disk_info
*tmp
;
5253 unsigned long flags
;
5255 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5256 tmp
= conf
->disks
+ i
;
5257 if (tmp
->replacement
5258 && tmp
->replacement
->recovery_offset
== MaxSector
5259 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
5260 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
5261 /* Replacement has just become active. */
5263 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
5266 /* Replaced device not technically faulty,
5267 * but we need to be sure it gets removed
5268 * and never re-added.
5270 set_bit(Faulty
, &tmp
->rdev
->flags
);
5271 sysfs_notify_dirent_safe(
5272 tmp
->rdev
->sysfs_state
);
5274 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
5275 } else if (tmp
->rdev
5276 && tmp
->rdev
->recovery_offset
== MaxSector
5277 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
5278 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
5280 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
5283 spin_lock_irqsave(&conf
->device_lock
, flags
);
5284 mddev
->degraded
= calc_degraded(conf
);
5285 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5286 print_raid5_conf(conf
);
5290 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
5292 struct r5conf
*conf
= mddev
->private;
5294 int number
= rdev
->raid_disk
;
5295 struct md_rdev
**rdevp
;
5296 struct disk_info
*p
= conf
->disks
+ number
;
5298 print_raid5_conf(conf
);
5299 if (rdev
== p
->rdev
)
5301 else if (rdev
== p
->replacement
)
5302 rdevp
= &p
->replacement
;
5306 if (number
>= conf
->raid_disks
&&
5307 conf
->reshape_progress
== MaxSector
)
5308 clear_bit(In_sync
, &rdev
->flags
);
5310 if (test_bit(In_sync
, &rdev
->flags
) ||
5311 atomic_read(&rdev
->nr_pending
)) {
5315 /* Only remove non-faulty devices if recovery
5318 if (!test_bit(Faulty
, &rdev
->flags
) &&
5319 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
5320 !has_failed(conf
) &&
5321 (!p
->replacement
|| p
->replacement
== rdev
) &&
5322 number
< conf
->raid_disks
) {
5328 if (atomic_read(&rdev
->nr_pending
)) {
5329 /* lost the race, try later */
5332 } else if (p
->replacement
) {
5333 /* We must have just cleared 'rdev' */
5334 p
->rdev
= p
->replacement
;
5335 clear_bit(Replacement
, &p
->replacement
->flags
);
5336 smp_mb(); /* Make sure other CPUs may see both as identical
5337 * but will never see neither - if they are careful
5339 p
->replacement
= NULL
;
5340 clear_bit(WantReplacement
, &rdev
->flags
);
5342 /* We might have just removed the Replacement as faulty-
5343 * clear the bit just in case
5345 clear_bit(WantReplacement
, &rdev
->flags
);
5348 print_raid5_conf(conf
);
5352 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
5354 struct r5conf
*conf
= mddev
->private;
5357 struct disk_info
*p
;
5359 int last
= conf
->raid_disks
- 1;
5361 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
5364 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
5365 /* no point adding a device */
5368 if (rdev
->raid_disk
>= 0)
5369 first
= last
= rdev
->raid_disk
;
5372 * find the disk ... but prefer rdev->saved_raid_disk
5375 if (rdev
->saved_raid_disk
>= 0 &&
5376 rdev
->saved_raid_disk
>= first
&&
5377 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
5378 disk
= rdev
->saved_raid_disk
;
5381 for ( ; disk
<= last
; disk
++) {
5382 p
= conf
->disks
+ disk
;
5383 if (p
->rdev
== NULL
) {
5384 clear_bit(In_sync
, &rdev
->flags
);
5385 rdev
->raid_disk
= disk
;
5387 if (rdev
->saved_raid_disk
!= disk
)
5389 rcu_assign_pointer(p
->rdev
, rdev
);
5392 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
5393 p
->replacement
== NULL
) {
5394 clear_bit(In_sync
, &rdev
->flags
);
5395 set_bit(Replacement
, &rdev
->flags
);
5396 rdev
->raid_disk
= disk
;
5399 rcu_assign_pointer(p
->replacement
, rdev
);
5403 print_raid5_conf(conf
);
5407 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
5409 /* no resync is happening, and there is enough space
5410 * on all devices, so we can resize.
5411 * We need to make sure resync covers any new space.
5412 * If the array is shrinking we should possibly wait until
5413 * any io in the removed space completes, but it hardly seems
5416 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
5417 md_set_array_sectors(mddev
, raid5_size(mddev
, sectors
,
5418 mddev
->raid_disks
));
5419 if (mddev
->array_sectors
>
5420 raid5_size(mddev
, sectors
, mddev
->raid_disks
))
5422 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5423 revalidate_disk(mddev
->gendisk
);
5424 if (sectors
> mddev
->dev_sectors
&&
5425 mddev
->recovery_cp
> mddev
->dev_sectors
) {
5426 mddev
->recovery_cp
= mddev
->dev_sectors
;
5427 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
5429 mddev
->dev_sectors
= sectors
;
5430 mddev
->resync_max_sectors
= sectors
;
5434 static int check_stripe_cache(struct mddev
*mddev
)
5436 /* Can only proceed if there are plenty of stripe_heads.
5437 * We need a minimum of one full stripe,, and for sensible progress
5438 * it is best to have about 4 times that.
5439 * If we require 4 times, then the default 256 4K stripe_heads will
5440 * allow for chunk sizes up to 256K, which is probably OK.
5441 * If the chunk size is greater, user-space should request more
5442 * stripe_heads first.
5444 struct r5conf
*conf
= mddev
->private;
5445 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5446 > conf
->max_nr_stripes
||
5447 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5448 > conf
->max_nr_stripes
) {
5449 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5451 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
5458 static int check_reshape(struct mddev
*mddev
)
5460 struct r5conf
*conf
= mddev
->private;
5462 if (mddev
->delta_disks
== 0 &&
5463 mddev
->new_layout
== mddev
->layout
&&
5464 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
5465 return 0; /* nothing to do */
5467 /* Cannot grow a bitmap yet */
5469 if (has_failed(conf
))
5471 if (mddev
->delta_disks
< 0) {
5472 /* We might be able to shrink, but the devices must
5473 * be made bigger first.
5474 * For raid6, 4 is the minimum size.
5475 * Otherwise 2 is the minimum
5478 if (mddev
->level
== 6)
5480 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
5484 if (!check_stripe_cache(mddev
))
5487 return resize_stripes(conf
, conf
->raid_disks
+ mddev
->delta_disks
);
5490 static int raid5_start_reshape(struct mddev
*mddev
)
5492 struct r5conf
*conf
= mddev
->private;
5493 struct md_rdev
*rdev
;
5495 unsigned long flags
;
5497 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
5500 if (!check_stripe_cache(mddev
))
5503 rdev_for_each(rdev
, mddev
)
5504 if (!test_bit(In_sync
, &rdev
->flags
)
5505 && !test_bit(Faulty
, &rdev
->flags
))
5508 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
5509 /* Not enough devices even to make a degraded array
5514 /* Refuse to reduce size of the array. Any reductions in
5515 * array size must be through explicit setting of array_size
5518 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
5519 < mddev
->array_sectors
) {
5520 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
5521 "before number of disks\n", mdname(mddev
));
5525 atomic_set(&conf
->reshape_stripes
, 0);
5526 spin_lock_irq(&conf
->device_lock
);
5527 conf
->previous_raid_disks
= conf
->raid_disks
;
5528 conf
->raid_disks
+= mddev
->delta_disks
;
5529 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
5530 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
5531 conf
->prev_algo
= conf
->algorithm
;
5532 conf
->algorithm
= mddev
->new_layout
;
5533 if (mddev
->delta_disks
< 0)
5534 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
5536 conf
->reshape_progress
= 0;
5537 conf
->reshape_safe
= conf
->reshape_progress
;
5539 spin_unlock_irq(&conf
->device_lock
);
5541 /* Add some new drives, as many as will fit.
5542 * We know there are enough to make the newly sized array work.
5543 * Don't add devices if we are reducing the number of
5544 * devices in the array. This is because it is not possible
5545 * to correctly record the "partially reconstructed" state of
5546 * such devices during the reshape and confusion could result.
5548 if (mddev
->delta_disks
>= 0) {
5549 rdev_for_each(rdev
, mddev
)
5550 if (rdev
->raid_disk
< 0 &&
5551 !test_bit(Faulty
, &rdev
->flags
)) {
5552 if (raid5_add_disk(mddev
, rdev
) == 0) {
5554 >= conf
->previous_raid_disks
)
5555 set_bit(In_sync
, &rdev
->flags
);
5557 rdev
->recovery_offset
= 0;
5559 if (sysfs_link_rdev(mddev
, rdev
))
5560 /* Failure here is OK */;
5562 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
5563 && !test_bit(Faulty
, &rdev
->flags
)) {
5564 /* This is a spare that was manually added */
5565 set_bit(In_sync
, &rdev
->flags
);
5568 /* When a reshape changes the number of devices,
5569 * ->degraded is measured against the larger of the
5570 * pre and post number of devices.
5572 spin_lock_irqsave(&conf
->device_lock
, flags
);
5573 mddev
->degraded
= calc_degraded(conf
);
5574 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5576 mddev
->raid_disks
= conf
->raid_disks
;
5577 mddev
->reshape_position
= conf
->reshape_progress
;
5578 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5580 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5581 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5582 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5583 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5584 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5586 if (!mddev
->sync_thread
) {
5587 mddev
->recovery
= 0;
5588 spin_lock_irq(&conf
->device_lock
);
5589 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
5590 conf
->reshape_progress
= MaxSector
;
5591 mddev
->reshape_position
= MaxSector
;
5592 spin_unlock_irq(&conf
->device_lock
);
5595 conf
->reshape_checkpoint
= jiffies
;
5596 md_wakeup_thread(mddev
->sync_thread
);
5597 md_new_event(mddev
);
5601 /* This is called from the reshape thread and should make any
5602 * changes needed in 'conf'
5604 static void end_reshape(struct r5conf
*conf
)
5607 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
5609 spin_lock_irq(&conf
->device_lock
);
5610 conf
->previous_raid_disks
= conf
->raid_disks
;
5611 conf
->reshape_progress
= MaxSector
;
5612 spin_unlock_irq(&conf
->device_lock
);
5613 wake_up(&conf
->wait_for_overlap
);
5615 /* read-ahead size must cover two whole stripes, which is
5616 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5618 if (conf
->mddev
->queue
) {
5619 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5620 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
5622 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
5623 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
5628 /* This is called from the raid5d thread with mddev_lock held.
5629 * It makes config changes to the device.
5631 static void raid5_finish_reshape(struct mddev
*mddev
)
5633 struct r5conf
*conf
= mddev
->private;
5635 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
5637 if (mddev
->delta_disks
> 0) {
5638 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5639 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5640 revalidate_disk(mddev
->gendisk
);
5643 spin_lock_irq(&conf
->device_lock
);
5644 mddev
->degraded
= calc_degraded(conf
);
5645 spin_unlock_irq(&conf
->device_lock
);
5646 for (d
= conf
->raid_disks
;
5647 d
< conf
->raid_disks
- mddev
->delta_disks
;
5649 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
5651 raid5_remove_disk(mddev
, rdev
) == 0) {
5652 sysfs_unlink_rdev(mddev
, rdev
);
5653 rdev
->raid_disk
= -1;
5657 mddev
->layout
= conf
->algorithm
;
5658 mddev
->chunk_sectors
= conf
->chunk_sectors
;
5659 mddev
->reshape_position
= MaxSector
;
5660 mddev
->delta_disks
= 0;
5664 static void raid5_quiesce(struct mddev
*mddev
, int state
)
5666 struct r5conf
*conf
= mddev
->private;
5669 case 2: /* resume for a suspend */
5670 wake_up(&conf
->wait_for_overlap
);
5673 case 1: /* stop all writes */
5674 spin_lock_irq(&conf
->device_lock
);
5675 /* '2' tells resync/reshape to pause so that all
5676 * active stripes can drain
5679 wait_event_lock_irq(conf
->wait_for_stripe
,
5680 atomic_read(&conf
->active_stripes
) == 0 &&
5681 atomic_read(&conf
->active_aligned_reads
) == 0,
5682 conf
->device_lock
, /* nothing */);
5684 spin_unlock_irq(&conf
->device_lock
);
5685 /* allow reshape to continue */
5686 wake_up(&conf
->wait_for_overlap
);
5689 case 0: /* re-enable writes */
5690 spin_lock_irq(&conf
->device_lock
);
5692 wake_up(&conf
->wait_for_stripe
);
5693 wake_up(&conf
->wait_for_overlap
);
5694 spin_unlock_irq(&conf
->device_lock
);
5700 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
5702 struct r0conf
*raid0_conf
= mddev
->private;
5705 /* for raid0 takeover only one zone is supported */
5706 if (raid0_conf
->nr_strip_zones
> 1) {
5707 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5709 return ERR_PTR(-EINVAL
);
5712 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
5713 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
5714 mddev
->dev_sectors
= sectors
;
5715 mddev
->new_level
= level
;
5716 mddev
->new_layout
= ALGORITHM_PARITY_N
;
5717 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
5718 mddev
->raid_disks
+= 1;
5719 mddev
->delta_disks
= 1;
5720 /* make sure it will be not marked as dirty */
5721 mddev
->recovery_cp
= MaxSector
;
5723 return setup_conf(mddev
);
5727 static void *raid5_takeover_raid1(struct mddev
*mddev
)
5731 if (mddev
->raid_disks
!= 2 ||
5732 mddev
->degraded
> 1)
5733 return ERR_PTR(-EINVAL
);
5735 /* Should check if there are write-behind devices? */
5737 chunksect
= 64*2; /* 64K by default */
5739 /* The array must be an exact multiple of chunksize */
5740 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
5743 if ((chunksect
<<9) < STRIPE_SIZE
)
5744 /* array size does not allow a suitable chunk size */
5745 return ERR_PTR(-EINVAL
);
5747 mddev
->new_level
= 5;
5748 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5749 mddev
->new_chunk_sectors
= chunksect
;
5751 return setup_conf(mddev
);
5754 static void *raid5_takeover_raid6(struct mddev
*mddev
)
5758 switch (mddev
->layout
) {
5759 case ALGORITHM_LEFT_ASYMMETRIC_6
:
5760 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
5762 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
5763 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
5765 case ALGORITHM_LEFT_SYMMETRIC_6
:
5766 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5768 case ALGORITHM_RIGHT_SYMMETRIC_6
:
5769 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
5771 case ALGORITHM_PARITY_0_6
:
5772 new_layout
= ALGORITHM_PARITY_0
;
5774 case ALGORITHM_PARITY_N
:
5775 new_layout
= ALGORITHM_PARITY_N
;
5778 return ERR_PTR(-EINVAL
);
5780 mddev
->new_level
= 5;
5781 mddev
->new_layout
= new_layout
;
5782 mddev
->delta_disks
= -1;
5783 mddev
->raid_disks
-= 1;
5784 return setup_conf(mddev
);
5788 static int raid5_check_reshape(struct mddev
*mddev
)
5790 /* For a 2-drive array, the layout and chunk size can be changed
5791 * immediately as not restriping is needed.
5792 * For larger arrays we record the new value - after validation
5793 * to be used by a reshape pass.
5795 struct r5conf
*conf
= mddev
->private;
5796 int new_chunk
= mddev
->new_chunk_sectors
;
5798 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
5800 if (new_chunk
> 0) {
5801 if (!is_power_of_2(new_chunk
))
5803 if (new_chunk
< (PAGE_SIZE
>>9))
5805 if (mddev
->array_sectors
& (new_chunk
-1))
5806 /* not factor of array size */
5810 /* They look valid */
5812 if (mddev
->raid_disks
== 2) {
5813 /* can make the change immediately */
5814 if (mddev
->new_layout
>= 0) {
5815 conf
->algorithm
= mddev
->new_layout
;
5816 mddev
->layout
= mddev
->new_layout
;
5818 if (new_chunk
> 0) {
5819 conf
->chunk_sectors
= new_chunk
;
5820 mddev
->chunk_sectors
= new_chunk
;
5822 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5823 md_wakeup_thread(mddev
->thread
);
5825 return check_reshape(mddev
);
5828 static int raid6_check_reshape(struct mddev
*mddev
)
5830 int new_chunk
= mddev
->new_chunk_sectors
;
5832 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
5834 if (new_chunk
> 0) {
5835 if (!is_power_of_2(new_chunk
))
5837 if (new_chunk
< (PAGE_SIZE
>> 9))
5839 if (mddev
->array_sectors
& (new_chunk
-1))
5840 /* not factor of array size */
5844 /* They look valid */
5845 return check_reshape(mddev
);
5848 static void *raid5_takeover(struct mddev
*mddev
)
5850 /* raid5 can take over:
5851 * raid0 - if there is only one strip zone - make it a raid4 layout
5852 * raid1 - if there are two drives. We need to know the chunk size
5853 * raid4 - trivial - just use a raid4 layout.
5854 * raid6 - Providing it is a *_6 layout
5856 if (mddev
->level
== 0)
5857 return raid45_takeover_raid0(mddev
, 5);
5858 if (mddev
->level
== 1)
5859 return raid5_takeover_raid1(mddev
);
5860 if (mddev
->level
== 4) {
5861 mddev
->new_layout
= ALGORITHM_PARITY_N
;
5862 mddev
->new_level
= 5;
5863 return setup_conf(mddev
);
5865 if (mddev
->level
== 6)
5866 return raid5_takeover_raid6(mddev
);
5868 return ERR_PTR(-EINVAL
);
5871 static void *raid4_takeover(struct mddev
*mddev
)
5873 /* raid4 can take over:
5874 * raid0 - if there is only one strip zone
5875 * raid5 - if layout is right
5877 if (mddev
->level
== 0)
5878 return raid45_takeover_raid0(mddev
, 4);
5879 if (mddev
->level
== 5 &&
5880 mddev
->layout
== ALGORITHM_PARITY_N
) {
5881 mddev
->new_layout
= 0;
5882 mddev
->new_level
= 4;
5883 return setup_conf(mddev
);
5885 return ERR_PTR(-EINVAL
);
5888 static struct md_personality raid5_personality
;
5890 static void *raid6_takeover(struct mddev
*mddev
)
5892 /* Currently can only take over a raid5. We map the
5893 * personality to an equivalent raid6 personality
5894 * with the Q block at the end.
5898 if (mddev
->pers
!= &raid5_personality
)
5899 return ERR_PTR(-EINVAL
);
5900 if (mddev
->degraded
> 1)
5901 return ERR_PTR(-EINVAL
);
5902 if (mddev
->raid_disks
> 253)
5903 return ERR_PTR(-EINVAL
);
5904 if (mddev
->raid_disks
< 3)
5905 return ERR_PTR(-EINVAL
);
5907 switch (mddev
->layout
) {
5908 case ALGORITHM_LEFT_ASYMMETRIC
:
5909 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
5911 case ALGORITHM_RIGHT_ASYMMETRIC
:
5912 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
5914 case ALGORITHM_LEFT_SYMMETRIC
:
5915 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
5917 case ALGORITHM_RIGHT_SYMMETRIC
:
5918 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
5920 case ALGORITHM_PARITY_0
:
5921 new_layout
= ALGORITHM_PARITY_0_6
;
5923 case ALGORITHM_PARITY_N
:
5924 new_layout
= ALGORITHM_PARITY_N
;
5927 return ERR_PTR(-EINVAL
);
5929 mddev
->new_level
= 6;
5930 mddev
->new_layout
= new_layout
;
5931 mddev
->delta_disks
= 1;
5932 mddev
->raid_disks
+= 1;
5933 return setup_conf(mddev
);
5937 static struct md_personality raid6_personality
=
5941 .owner
= THIS_MODULE
,
5942 .make_request
= make_request
,
5946 .error_handler
= error
,
5947 .hot_add_disk
= raid5_add_disk
,
5948 .hot_remove_disk
= raid5_remove_disk
,
5949 .spare_active
= raid5_spare_active
,
5950 .sync_request
= sync_request
,
5951 .resize
= raid5_resize
,
5953 .check_reshape
= raid6_check_reshape
,
5954 .start_reshape
= raid5_start_reshape
,
5955 .finish_reshape
= raid5_finish_reshape
,
5956 .quiesce
= raid5_quiesce
,
5957 .takeover
= raid6_takeover
,
5959 static struct md_personality raid5_personality
=
5963 .owner
= THIS_MODULE
,
5964 .make_request
= make_request
,
5968 .error_handler
= error
,
5969 .hot_add_disk
= raid5_add_disk
,
5970 .hot_remove_disk
= raid5_remove_disk
,
5971 .spare_active
= raid5_spare_active
,
5972 .sync_request
= sync_request
,
5973 .resize
= raid5_resize
,
5975 .check_reshape
= raid5_check_reshape
,
5976 .start_reshape
= raid5_start_reshape
,
5977 .finish_reshape
= raid5_finish_reshape
,
5978 .quiesce
= raid5_quiesce
,
5979 .takeover
= raid5_takeover
,
5982 static struct md_personality raid4_personality
=
5986 .owner
= THIS_MODULE
,
5987 .make_request
= make_request
,
5991 .error_handler
= error
,
5992 .hot_add_disk
= raid5_add_disk
,
5993 .hot_remove_disk
= raid5_remove_disk
,
5994 .spare_active
= raid5_spare_active
,
5995 .sync_request
= sync_request
,
5996 .resize
= raid5_resize
,
5998 .check_reshape
= raid5_check_reshape
,
5999 .start_reshape
= raid5_start_reshape
,
6000 .finish_reshape
= raid5_finish_reshape
,
6001 .quiesce
= raid5_quiesce
,
6002 .takeover
= raid4_takeover
,
6005 static int __init
raid5_init(void)
6007 register_md_personality(&raid6_personality
);
6008 register_md_personality(&raid5_personality
);
6009 register_md_personality(&raid4_personality
);
6013 static void raid5_exit(void)
6015 unregister_md_personality(&raid6_personality
);
6016 unregister_md_personality(&raid5_personality
);
6017 unregister_md_personality(&raid4_personality
);
6020 module_init(raid5_init
);
6021 module_exit(raid5_exit
);
6022 MODULE_LICENSE("GPL");
6023 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6024 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6025 MODULE_ALIAS("md-raid5");
6026 MODULE_ALIAS("md-raid4");
6027 MODULE_ALIAS("md-level-5");
6028 MODULE_ALIAS("md-level-4");
6029 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6030 MODULE_ALIAS("md-raid6");
6031 MODULE_ALIAS("md-level-6");
6033 /* This used to be two separate modules, they were: */
6034 MODULE_ALIAS("raid5");
6035 MODULE_ALIAS("raid6");