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/async.h>
51 #include <linux/seq_file.h>
52 #include <linux/cpu.h>
53 #include <linux/slab.h>
54 #include <linux/ratelimit.h>
64 #define NR_STRIPES 256
65 #define STRIPE_SIZE PAGE_SIZE
66 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
67 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
68 #define IO_THRESHOLD 1
69 #define BYPASS_THRESHOLD 1
70 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
71 #define HASH_MASK (NR_HASH - 1)
73 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
75 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
76 * order without overlap. There may be several bio's per stripe+device, and
77 * a bio could span several devices.
78 * When walking this list for a particular stripe+device, we must never proceed
79 * beyond a bio that extends past this device, as the next bio might no longer
81 * This macro is used to determine the 'next' bio in the list, given the sector
82 * of the current stripe+device
84 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
86 * The following can be used to debug the driver
88 #define RAID5_PARANOIA 1
89 #if RAID5_PARANOIA && defined(CONFIG_SMP)
90 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
92 # define CHECK_DEVLOCK()
101 * We maintain a biased count of active stripes in the bottom 16 bits of
102 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
104 static inline int raid5_bi_phys_segments(struct bio
*bio
)
106 return bio
->bi_phys_segments
& 0xffff;
109 static inline int raid5_bi_hw_segments(struct bio
*bio
)
111 return (bio
->bi_phys_segments
>> 16) & 0xffff;
114 static inline int raid5_dec_bi_phys_segments(struct bio
*bio
)
116 --bio
->bi_phys_segments
;
117 return raid5_bi_phys_segments(bio
);
120 static inline int raid5_dec_bi_hw_segments(struct bio
*bio
)
122 unsigned short val
= raid5_bi_hw_segments(bio
);
125 bio
->bi_phys_segments
= (val
<< 16) | raid5_bi_phys_segments(bio
);
129 static inline void raid5_set_bi_hw_segments(struct bio
*bio
, unsigned int cnt
)
131 bio
->bi_phys_segments
= raid5_bi_phys_segments(bio
) | (cnt
<< 16);
134 /* Find first data disk in a raid6 stripe */
135 static inline int raid6_d0(struct stripe_head
*sh
)
138 /* ddf always start from first device */
140 /* md starts just after Q block */
141 if (sh
->qd_idx
== sh
->disks
- 1)
144 return sh
->qd_idx
+ 1;
146 static inline int raid6_next_disk(int disk
, int raid_disks
)
149 return (disk
< raid_disks
) ? disk
: 0;
152 /* When walking through the disks in a raid5, starting at raid6_d0,
153 * We need to map each disk to a 'slot', where the data disks are slot
154 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
155 * is raid_disks-1. This help does that mapping.
157 static int raid6_idx_to_slot(int idx
, struct stripe_head
*sh
,
158 int *count
, int syndrome_disks
)
164 if (idx
== sh
->pd_idx
)
165 return syndrome_disks
;
166 if (idx
== sh
->qd_idx
)
167 return syndrome_disks
+ 1;
173 static void return_io(struct bio
*return_bi
)
175 struct bio
*bi
= return_bi
;
178 return_bi
= bi
->bi_next
;
186 static void print_raid5_conf (raid5_conf_t
*conf
);
188 static int stripe_operations_active(struct stripe_head
*sh
)
190 return sh
->check_state
|| sh
->reconstruct_state
||
191 test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
) ||
192 test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
195 static void __release_stripe(raid5_conf_t
*conf
, struct stripe_head
*sh
)
197 if (atomic_dec_and_test(&sh
->count
)) {
198 BUG_ON(!list_empty(&sh
->lru
));
199 BUG_ON(atomic_read(&conf
->active_stripes
)==0);
200 if (test_bit(STRIPE_HANDLE
, &sh
->state
)) {
201 if (test_bit(STRIPE_DELAYED
, &sh
->state
))
202 list_add_tail(&sh
->lru
, &conf
->delayed_list
);
203 else if (test_bit(STRIPE_BIT_DELAY
, &sh
->state
) &&
204 sh
->bm_seq
- conf
->seq_write
> 0)
205 list_add_tail(&sh
->lru
, &conf
->bitmap_list
);
207 clear_bit(STRIPE_BIT_DELAY
, &sh
->state
);
208 list_add_tail(&sh
->lru
, &conf
->handle_list
);
210 md_wakeup_thread(conf
->mddev
->thread
);
212 BUG_ON(stripe_operations_active(sh
));
213 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
214 atomic_dec(&conf
->preread_active_stripes
);
215 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
)
216 md_wakeup_thread(conf
->mddev
->thread
);
218 atomic_dec(&conf
->active_stripes
);
219 if (!test_bit(STRIPE_EXPANDING
, &sh
->state
)) {
220 list_add_tail(&sh
->lru
, &conf
->inactive_list
);
221 wake_up(&conf
->wait_for_stripe
);
222 if (conf
->retry_read_aligned
)
223 md_wakeup_thread(conf
->mddev
->thread
);
229 static void release_stripe(struct stripe_head
*sh
)
231 raid5_conf_t
*conf
= sh
->raid_conf
;
234 spin_lock_irqsave(&conf
->device_lock
, flags
);
235 __release_stripe(conf
, sh
);
236 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
239 static inline void remove_hash(struct stripe_head
*sh
)
241 pr_debug("remove_hash(), stripe %llu\n",
242 (unsigned long long)sh
->sector
);
244 hlist_del_init(&sh
->hash
);
247 static inline void insert_hash(raid5_conf_t
*conf
, struct stripe_head
*sh
)
249 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
251 pr_debug("insert_hash(), stripe %llu\n",
252 (unsigned long long)sh
->sector
);
255 hlist_add_head(&sh
->hash
, hp
);
259 /* find an idle stripe, make sure it is unhashed, and return it. */
260 static struct stripe_head
*get_free_stripe(raid5_conf_t
*conf
)
262 struct stripe_head
*sh
= NULL
;
263 struct list_head
*first
;
266 if (list_empty(&conf
->inactive_list
))
268 first
= conf
->inactive_list
.next
;
269 sh
= list_entry(first
, struct stripe_head
, lru
);
270 list_del_init(first
);
272 atomic_inc(&conf
->active_stripes
);
277 static void shrink_buffers(struct stripe_head
*sh
)
281 int num
= sh
->raid_conf
->pool_size
;
283 for (i
= 0; i
< num
; i
++) {
287 sh
->dev
[i
].page
= NULL
;
292 static int grow_buffers(struct stripe_head
*sh
)
295 int num
= sh
->raid_conf
->pool_size
;
297 for (i
= 0; i
< num
; i
++) {
300 if (!(page
= alloc_page(GFP_KERNEL
))) {
303 sh
->dev
[i
].page
= page
;
308 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
);
309 static void stripe_set_idx(sector_t stripe
, raid5_conf_t
*conf
, int previous
,
310 struct stripe_head
*sh
);
312 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int previous
)
314 raid5_conf_t
*conf
= sh
->raid_conf
;
317 BUG_ON(atomic_read(&sh
->count
) != 0);
318 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
319 BUG_ON(stripe_operations_active(sh
));
322 pr_debug("init_stripe called, stripe %llu\n",
323 (unsigned long long)sh
->sector
);
327 sh
->generation
= conf
->generation
- previous
;
328 sh
->disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
330 stripe_set_idx(sector
, conf
, previous
, sh
);
334 for (i
= sh
->disks
; i
--; ) {
335 struct r5dev
*dev
= &sh
->dev
[i
];
337 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
338 test_bit(R5_LOCKED
, &dev
->flags
)) {
339 printk(KERN_ERR
"sector=%llx i=%d %p %p %p %p %d\n",
340 (unsigned long long)sh
->sector
, i
, dev
->toread
,
341 dev
->read
, dev
->towrite
, dev
->written
,
342 test_bit(R5_LOCKED
, &dev
->flags
));
346 raid5_build_block(sh
, i
, previous
);
348 insert_hash(conf
, sh
);
351 static struct stripe_head
*__find_stripe(raid5_conf_t
*conf
, sector_t sector
,
354 struct stripe_head
*sh
;
355 struct hlist_node
*hn
;
358 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
359 hlist_for_each_entry(sh
, hn
, stripe_hash(conf
, sector
), hash
)
360 if (sh
->sector
== sector
&& sh
->generation
== generation
)
362 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
367 * Need to check if array has failed when deciding whether to:
369 * - remove non-faulty devices
372 * This determination is simple when no reshape is happening.
373 * However if there is a reshape, we need to carefully check
374 * both the before and after sections.
375 * This is because some failed devices may only affect one
376 * of the two sections, and some non-in_sync devices may
377 * be insync in the section most affected by failed devices.
379 static int has_failed(raid5_conf_t
*conf
)
383 if (conf
->mddev
->reshape_position
== MaxSector
)
384 return conf
->mddev
->degraded
> conf
->max_degraded
;
388 for (i
= 0; i
< conf
->previous_raid_disks
; i
++) {
389 mdk_rdev_t
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
390 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
392 else if (test_bit(In_sync
, &rdev
->flags
))
395 /* not in-sync or faulty.
396 * If the reshape increases the number of devices,
397 * this is being recovered by the reshape, so
398 * this 'previous' section is not in_sync.
399 * If the number of devices is being reduced however,
400 * the device can only be part of the array if
401 * we are reverting a reshape, so this section will
404 if (conf
->raid_disks
>= conf
->previous_raid_disks
)
408 if (degraded
> conf
->max_degraded
)
412 for (i
= 0; i
< conf
->raid_disks
; i
++) {
413 mdk_rdev_t
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
414 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
416 else if (test_bit(In_sync
, &rdev
->flags
))
419 /* not in-sync or faulty.
420 * If reshape increases the number of devices, this
421 * section has already been recovered, else it
422 * almost certainly hasn't.
424 if (conf
->raid_disks
<= conf
->previous_raid_disks
)
428 if (degraded
> conf
->max_degraded
)
433 static struct stripe_head
*
434 get_active_stripe(raid5_conf_t
*conf
, sector_t sector
,
435 int previous
, int noblock
, int noquiesce
)
437 struct stripe_head
*sh
;
439 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
441 spin_lock_irq(&conf
->device_lock
);
444 wait_event_lock_irq(conf
->wait_for_stripe
,
445 conf
->quiesce
== 0 || noquiesce
,
446 conf
->device_lock
, /* nothing */);
447 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
449 if (!conf
->inactive_blocked
)
450 sh
= get_free_stripe(conf
);
451 if (noblock
&& sh
== NULL
)
454 conf
->inactive_blocked
= 1;
455 wait_event_lock_irq(conf
->wait_for_stripe
,
456 !list_empty(&conf
->inactive_list
) &&
457 (atomic_read(&conf
->active_stripes
)
458 < (conf
->max_nr_stripes
*3/4)
459 || !conf
->inactive_blocked
),
462 conf
->inactive_blocked
= 0;
464 init_stripe(sh
, sector
, previous
);
466 if (atomic_read(&sh
->count
)) {
467 BUG_ON(!list_empty(&sh
->lru
)
468 && !test_bit(STRIPE_EXPANDING
, &sh
->state
));
470 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
471 atomic_inc(&conf
->active_stripes
);
472 if (list_empty(&sh
->lru
) &&
473 !test_bit(STRIPE_EXPANDING
, &sh
->state
))
475 list_del_init(&sh
->lru
);
478 } while (sh
== NULL
);
481 atomic_inc(&sh
->count
);
483 spin_unlock_irq(&conf
->device_lock
);
488 raid5_end_read_request(struct bio
*bi
, int error
);
490 raid5_end_write_request(struct bio
*bi
, int error
);
492 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
494 raid5_conf_t
*conf
= sh
->raid_conf
;
495 int i
, disks
= sh
->disks
;
499 for (i
= disks
; i
--; ) {
503 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
504 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
508 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
513 bi
= &sh
->dev
[i
].req
;
517 bi
->bi_end_io
= raid5_end_write_request
;
519 bi
->bi_end_io
= raid5_end_read_request
;
522 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
523 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
526 atomic_inc(&rdev
->nr_pending
);
529 /* We have already checked bad blocks for reads. Now
530 * need to check for writes.
532 while ((rw
& WRITE
) && rdev
&&
533 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
536 int bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
537 &first_bad
, &bad_sectors
);
542 set_bit(BlockedBadBlocks
, &rdev
->flags
);
543 if (!conf
->mddev
->external
&&
544 conf
->mddev
->flags
) {
545 /* It is very unlikely, but we might
546 * still need to write out the
547 * bad block log - better give it
549 md_check_recovery(conf
->mddev
);
551 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
553 /* Acknowledged bad block - skip the write */
554 rdev_dec_pending(rdev
, conf
->mddev
);
560 if (s
->syncing
|| s
->expanding
|| s
->expanded
)
561 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
563 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
565 bi
->bi_bdev
= rdev
->bdev
;
566 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
567 __func__
, (unsigned long long)sh
->sector
,
569 atomic_inc(&sh
->count
);
570 bi
->bi_sector
= sh
->sector
+ rdev
->data_offset
;
571 bi
->bi_flags
= 1 << BIO_UPTODATE
;
575 bi
->bi_io_vec
= &sh
->dev
[i
].vec
;
576 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
577 bi
->bi_io_vec
[0].bv_offset
= 0;
578 bi
->bi_size
= STRIPE_SIZE
;
580 generic_make_request(bi
);
583 set_bit(STRIPE_DEGRADED
, &sh
->state
);
584 pr_debug("skip op %ld on disc %d for sector %llu\n",
585 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
586 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
587 set_bit(STRIPE_HANDLE
, &sh
->state
);
592 static struct dma_async_tx_descriptor
*
593 async_copy_data(int frombio
, struct bio
*bio
, struct page
*page
,
594 sector_t sector
, struct dma_async_tx_descriptor
*tx
)
597 struct page
*bio_page
;
600 struct async_submit_ctl submit
;
601 enum async_tx_flags flags
= 0;
603 if (bio
->bi_sector
>= sector
)
604 page_offset
= (signed)(bio
->bi_sector
- sector
) * 512;
606 page_offset
= (signed)(sector
- bio
->bi_sector
) * -512;
609 flags
|= ASYNC_TX_FENCE
;
610 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
612 bio_for_each_segment(bvl
, bio
, i
) {
613 int len
= bvl
->bv_len
;
617 if (page_offset
< 0) {
618 b_offset
= -page_offset
;
619 page_offset
+= b_offset
;
623 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
624 clen
= STRIPE_SIZE
- page_offset
;
629 b_offset
+= bvl
->bv_offset
;
630 bio_page
= bvl
->bv_page
;
632 tx
= async_memcpy(page
, bio_page
, page_offset
,
633 b_offset
, clen
, &submit
);
635 tx
= async_memcpy(bio_page
, page
, b_offset
,
636 page_offset
, clen
, &submit
);
638 /* chain the operations */
639 submit
.depend_tx
= tx
;
641 if (clen
< len
) /* hit end of page */
649 static void ops_complete_biofill(void *stripe_head_ref
)
651 struct stripe_head
*sh
= stripe_head_ref
;
652 struct bio
*return_bi
= NULL
;
653 raid5_conf_t
*conf
= sh
->raid_conf
;
656 pr_debug("%s: stripe %llu\n", __func__
,
657 (unsigned long long)sh
->sector
);
659 /* clear completed biofills */
660 spin_lock_irq(&conf
->device_lock
);
661 for (i
= sh
->disks
; i
--; ) {
662 struct r5dev
*dev
= &sh
->dev
[i
];
664 /* acknowledge completion of a biofill operation */
665 /* and check if we need to reply to a read request,
666 * new R5_Wantfill requests are held off until
667 * !STRIPE_BIOFILL_RUN
669 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
670 struct bio
*rbi
, *rbi2
;
675 while (rbi
&& rbi
->bi_sector
<
676 dev
->sector
+ STRIPE_SECTORS
) {
677 rbi2
= r5_next_bio(rbi
, dev
->sector
);
678 if (!raid5_dec_bi_phys_segments(rbi
)) {
679 rbi
->bi_next
= return_bi
;
686 spin_unlock_irq(&conf
->device_lock
);
687 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
689 return_io(return_bi
);
691 set_bit(STRIPE_HANDLE
, &sh
->state
);
695 static void ops_run_biofill(struct stripe_head
*sh
)
697 struct dma_async_tx_descriptor
*tx
= NULL
;
698 raid5_conf_t
*conf
= sh
->raid_conf
;
699 struct async_submit_ctl submit
;
702 pr_debug("%s: stripe %llu\n", __func__
,
703 (unsigned long long)sh
->sector
);
705 for (i
= sh
->disks
; i
--; ) {
706 struct r5dev
*dev
= &sh
->dev
[i
];
707 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
709 spin_lock_irq(&conf
->device_lock
);
710 dev
->read
= rbi
= dev
->toread
;
712 spin_unlock_irq(&conf
->device_lock
);
713 while (rbi
&& rbi
->bi_sector
<
714 dev
->sector
+ STRIPE_SECTORS
) {
715 tx
= async_copy_data(0, rbi
, dev
->page
,
717 rbi
= r5_next_bio(rbi
, dev
->sector
);
722 atomic_inc(&sh
->count
);
723 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
724 async_trigger_callback(&submit
);
727 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
734 tgt
= &sh
->dev
[target
];
735 set_bit(R5_UPTODATE
, &tgt
->flags
);
736 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
737 clear_bit(R5_Wantcompute
, &tgt
->flags
);
740 static void ops_complete_compute(void *stripe_head_ref
)
742 struct stripe_head
*sh
= stripe_head_ref
;
744 pr_debug("%s: stripe %llu\n", __func__
,
745 (unsigned long long)sh
->sector
);
747 /* mark the computed target(s) as uptodate */
748 mark_target_uptodate(sh
, sh
->ops
.target
);
749 mark_target_uptodate(sh
, sh
->ops
.target2
);
751 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
752 if (sh
->check_state
== check_state_compute_run
)
753 sh
->check_state
= check_state_compute_result
;
754 set_bit(STRIPE_HANDLE
, &sh
->state
);
758 /* return a pointer to the address conversion region of the scribble buffer */
759 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
760 struct raid5_percpu
*percpu
)
762 return percpu
->scribble
+ sizeof(struct page
*) * (sh
->disks
+ 2);
765 static struct dma_async_tx_descriptor
*
766 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
768 int disks
= sh
->disks
;
769 struct page
**xor_srcs
= percpu
->scribble
;
770 int target
= sh
->ops
.target
;
771 struct r5dev
*tgt
= &sh
->dev
[target
];
772 struct page
*xor_dest
= tgt
->page
;
774 struct dma_async_tx_descriptor
*tx
;
775 struct async_submit_ctl submit
;
778 pr_debug("%s: stripe %llu block: %d\n",
779 __func__
, (unsigned long long)sh
->sector
, target
);
780 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
782 for (i
= disks
; i
--; )
784 xor_srcs
[count
++] = sh
->dev
[i
].page
;
786 atomic_inc(&sh
->count
);
788 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
789 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
));
790 if (unlikely(count
== 1))
791 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
793 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
798 /* set_syndrome_sources - populate source buffers for gen_syndrome
799 * @srcs - (struct page *) array of size sh->disks
800 * @sh - stripe_head to parse
802 * Populates srcs in proper layout order for the stripe and returns the
803 * 'count' of sources to be used in a call to async_gen_syndrome. The P
804 * destination buffer is recorded in srcs[count] and the Q destination
805 * is recorded in srcs[count+1]].
807 static int set_syndrome_sources(struct page
**srcs
, struct stripe_head
*sh
)
809 int disks
= sh
->disks
;
810 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
811 int d0_idx
= raid6_d0(sh
);
815 for (i
= 0; i
< disks
; i
++)
821 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
823 srcs
[slot
] = sh
->dev
[i
].page
;
824 i
= raid6_next_disk(i
, disks
);
825 } while (i
!= d0_idx
);
827 return syndrome_disks
;
830 static struct dma_async_tx_descriptor
*
831 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
833 int disks
= sh
->disks
;
834 struct page
**blocks
= percpu
->scribble
;
836 int qd_idx
= sh
->qd_idx
;
837 struct dma_async_tx_descriptor
*tx
;
838 struct async_submit_ctl submit
;
844 if (sh
->ops
.target
< 0)
845 target
= sh
->ops
.target2
;
846 else if (sh
->ops
.target2
< 0)
847 target
= sh
->ops
.target
;
849 /* we should only have one valid target */
852 pr_debug("%s: stripe %llu block: %d\n",
853 __func__
, (unsigned long long)sh
->sector
, target
);
855 tgt
= &sh
->dev
[target
];
856 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
859 atomic_inc(&sh
->count
);
861 if (target
== qd_idx
) {
862 count
= set_syndrome_sources(blocks
, sh
);
863 blocks
[count
] = NULL
; /* regenerating p is not necessary */
864 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
865 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
866 ops_complete_compute
, sh
,
867 to_addr_conv(sh
, percpu
));
868 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
870 /* Compute any data- or p-drive using XOR */
872 for (i
= disks
; i
-- ; ) {
873 if (i
== target
|| i
== qd_idx
)
875 blocks
[count
++] = sh
->dev
[i
].page
;
878 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
879 NULL
, ops_complete_compute
, sh
,
880 to_addr_conv(sh
, percpu
));
881 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
887 static struct dma_async_tx_descriptor
*
888 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
890 int i
, count
, disks
= sh
->disks
;
891 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
892 int d0_idx
= raid6_d0(sh
);
893 int faila
= -1, failb
= -1;
894 int target
= sh
->ops
.target
;
895 int target2
= sh
->ops
.target2
;
896 struct r5dev
*tgt
= &sh
->dev
[target
];
897 struct r5dev
*tgt2
= &sh
->dev
[target2
];
898 struct dma_async_tx_descriptor
*tx
;
899 struct page
**blocks
= percpu
->scribble
;
900 struct async_submit_ctl submit
;
902 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
903 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
904 BUG_ON(target
< 0 || target2
< 0);
905 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
906 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
908 /* we need to open-code set_syndrome_sources to handle the
909 * slot number conversion for 'faila' and 'failb'
911 for (i
= 0; i
< disks
; i
++)
916 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
918 blocks
[slot
] = sh
->dev
[i
].page
;
924 i
= raid6_next_disk(i
, disks
);
925 } while (i
!= d0_idx
);
927 BUG_ON(faila
== failb
);
930 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
931 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
933 atomic_inc(&sh
->count
);
935 if (failb
== syndrome_disks
+1) {
936 /* Q disk is one of the missing disks */
937 if (faila
== syndrome_disks
) {
938 /* Missing P+Q, just recompute */
939 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
940 ops_complete_compute
, sh
,
941 to_addr_conv(sh
, percpu
));
942 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
943 STRIPE_SIZE
, &submit
);
947 int qd_idx
= sh
->qd_idx
;
949 /* Missing D+Q: recompute D from P, then recompute Q */
950 if (target
== qd_idx
)
951 data_target
= target2
;
953 data_target
= target
;
956 for (i
= disks
; i
-- ; ) {
957 if (i
== data_target
|| i
== qd_idx
)
959 blocks
[count
++] = sh
->dev
[i
].page
;
961 dest
= sh
->dev
[data_target
].page
;
962 init_async_submit(&submit
,
963 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
965 to_addr_conv(sh
, percpu
));
966 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
969 count
= set_syndrome_sources(blocks
, sh
);
970 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
971 ops_complete_compute
, sh
,
972 to_addr_conv(sh
, percpu
));
973 return async_gen_syndrome(blocks
, 0, count
+2,
974 STRIPE_SIZE
, &submit
);
977 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
978 ops_complete_compute
, sh
,
979 to_addr_conv(sh
, percpu
));
980 if (failb
== syndrome_disks
) {
981 /* We're missing D+P. */
982 return async_raid6_datap_recov(syndrome_disks
+2,
986 /* We're missing D+D. */
987 return async_raid6_2data_recov(syndrome_disks
+2,
988 STRIPE_SIZE
, faila
, failb
,
995 static void ops_complete_prexor(void *stripe_head_ref
)
997 struct stripe_head
*sh
= stripe_head_ref
;
999 pr_debug("%s: stripe %llu\n", __func__
,
1000 (unsigned long long)sh
->sector
);
1003 static struct dma_async_tx_descriptor
*
1004 ops_run_prexor(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1005 struct dma_async_tx_descriptor
*tx
)
1007 int disks
= sh
->disks
;
1008 struct page
**xor_srcs
= percpu
->scribble
;
1009 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1010 struct async_submit_ctl submit
;
1012 /* existing parity data subtracted */
1013 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1015 pr_debug("%s: stripe %llu\n", __func__
,
1016 (unsigned long long)sh
->sector
);
1018 for (i
= disks
; i
--; ) {
1019 struct r5dev
*dev
= &sh
->dev
[i
];
1020 /* Only process blocks that are known to be uptodate */
1021 if (test_bit(R5_Wantdrain
, &dev
->flags
))
1022 xor_srcs
[count
++] = dev
->page
;
1025 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1026 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
));
1027 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1032 static struct dma_async_tx_descriptor
*
1033 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1035 int disks
= sh
->disks
;
1038 pr_debug("%s: stripe %llu\n", __func__
,
1039 (unsigned long long)sh
->sector
);
1041 for (i
= disks
; i
--; ) {
1042 struct r5dev
*dev
= &sh
->dev
[i
];
1045 if (test_and_clear_bit(R5_Wantdrain
, &dev
->flags
)) {
1048 spin_lock_irq(&sh
->raid_conf
->device_lock
);
1049 chosen
= dev
->towrite
;
1050 dev
->towrite
= NULL
;
1051 BUG_ON(dev
->written
);
1052 wbi
= dev
->written
= chosen
;
1053 spin_unlock_irq(&sh
->raid_conf
->device_lock
);
1055 while (wbi
&& wbi
->bi_sector
<
1056 dev
->sector
+ STRIPE_SECTORS
) {
1057 if (wbi
->bi_rw
& REQ_FUA
)
1058 set_bit(R5_WantFUA
, &dev
->flags
);
1059 tx
= async_copy_data(1, wbi
, dev
->page
,
1061 wbi
= r5_next_bio(wbi
, dev
->sector
);
1069 static void ops_complete_reconstruct(void *stripe_head_ref
)
1071 struct stripe_head
*sh
= stripe_head_ref
;
1072 int disks
= sh
->disks
;
1073 int pd_idx
= sh
->pd_idx
;
1074 int qd_idx
= sh
->qd_idx
;
1078 pr_debug("%s: stripe %llu\n", __func__
,
1079 (unsigned long long)sh
->sector
);
1081 for (i
= disks
; i
--; )
1082 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1084 for (i
= disks
; i
--; ) {
1085 struct r5dev
*dev
= &sh
->dev
[i
];
1087 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1088 set_bit(R5_UPTODATE
, &dev
->flags
);
1090 set_bit(R5_WantFUA
, &dev
->flags
);
1094 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1095 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1096 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1097 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1099 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1100 sh
->reconstruct_state
= reconstruct_state_result
;
1103 set_bit(STRIPE_HANDLE
, &sh
->state
);
1108 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1109 struct dma_async_tx_descriptor
*tx
)
1111 int disks
= sh
->disks
;
1112 struct page
**xor_srcs
= percpu
->scribble
;
1113 struct async_submit_ctl submit
;
1114 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1115 struct page
*xor_dest
;
1117 unsigned long flags
;
1119 pr_debug("%s: stripe %llu\n", __func__
,
1120 (unsigned long long)sh
->sector
);
1122 /* check if prexor is active which means only process blocks
1123 * that are part of a read-modify-write (written)
1125 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1127 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1128 for (i
= disks
; i
--; ) {
1129 struct r5dev
*dev
= &sh
->dev
[i
];
1131 xor_srcs
[count
++] = dev
->page
;
1134 xor_dest
= sh
->dev
[pd_idx
].page
;
1135 for (i
= disks
; i
--; ) {
1136 struct r5dev
*dev
= &sh
->dev
[i
];
1138 xor_srcs
[count
++] = dev
->page
;
1142 /* 1/ if we prexor'd then the dest is reused as a source
1143 * 2/ if we did not prexor then we are redoing the parity
1144 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1145 * for the synchronous xor case
1147 flags
= ASYNC_TX_ACK
|
1148 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1150 atomic_inc(&sh
->count
);
1152 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, sh
,
1153 to_addr_conv(sh
, percpu
));
1154 if (unlikely(count
== 1))
1155 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1157 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1161 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1162 struct dma_async_tx_descriptor
*tx
)
1164 struct async_submit_ctl submit
;
1165 struct page
**blocks
= percpu
->scribble
;
1168 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1170 count
= set_syndrome_sources(blocks
, sh
);
1172 atomic_inc(&sh
->count
);
1174 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_reconstruct
,
1175 sh
, to_addr_conv(sh
, percpu
));
1176 async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1179 static void ops_complete_check(void *stripe_head_ref
)
1181 struct stripe_head
*sh
= stripe_head_ref
;
1183 pr_debug("%s: stripe %llu\n", __func__
,
1184 (unsigned long long)sh
->sector
);
1186 sh
->check_state
= check_state_check_result
;
1187 set_bit(STRIPE_HANDLE
, &sh
->state
);
1191 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1193 int disks
= sh
->disks
;
1194 int pd_idx
= sh
->pd_idx
;
1195 int qd_idx
= sh
->qd_idx
;
1196 struct page
*xor_dest
;
1197 struct page
**xor_srcs
= percpu
->scribble
;
1198 struct dma_async_tx_descriptor
*tx
;
1199 struct async_submit_ctl submit
;
1203 pr_debug("%s: stripe %llu\n", __func__
,
1204 (unsigned long long)sh
->sector
);
1207 xor_dest
= sh
->dev
[pd_idx
].page
;
1208 xor_srcs
[count
++] = xor_dest
;
1209 for (i
= disks
; i
--; ) {
1210 if (i
== pd_idx
|| i
== qd_idx
)
1212 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1215 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1216 to_addr_conv(sh
, percpu
));
1217 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1218 &sh
->ops
.zero_sum_result
, &submit
);
1220 atomic_inc(&sh
->count
);
1221 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1222 tx
= async_trigger_callback(&submit
);
1225 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1227 struct page
**srcs
= percpu
->scribble
;
1228 struct async_submit_ctl submit
;
1231 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1232 (unsigned long long)sh
->sector
, checkp
);
1234 count
= set_syndrome_sources(srcs
, sh
);
1238 atomic_inc(&sh
->count
);
1239 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1240 sh
, to_addr_conv(sh
, percpu
));
1241 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1242 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1245 static void __raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1247 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1248 struct dma_async_tx_descriptor
*tx
= NULL
;
1249 raid5_conf_t
*conf
= sh
->raid_conf
;
1250 int level
= conf
->level
;
1251 struct raid5_percpu
*percpu
;
1255 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1256 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1257 ops_run_biofill(sh
);
1261 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1263 tx
= ops_run_compute5(sh
, percpu
);
1265 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
1266 tx
= ops_run_compute6_1(sh
, percpu
);
1268 tx
= ops_run_compute6_2(sh
, percpu
);
1270 /* terminate the chain if reconstruct is not set to be run */
1271 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
1275 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
))
1276 tx
= ops_run_prexor(sh
, percpu
, tx
);
1278 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
1279 tx
= ops_run_biodrain(sh
, tx
);
1283 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
1285 ops_run_reconstruct5(sh
, percpu
, tx
);
1287 ops_run_reconstruct6(sh
, percpu
, tx
);
1290 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
1291 if (sh
->check_state
== check_state_run
)
1292 ops_run_check_p(sh
, percpu
);
1293 else if (sh
->check_state
== check_state_run_q
)
1294 ops_run_check_pq(sh
, percpu
, 0);
1295 else if (sh
->check_state
== check_state_run_pq
)
1296 ops_run_check_pq(sh
, percpu
, 1);
1302 for (i
= disks
; i
--; ) {
1303 struct r5dev
*dev
= &sh
->dev
[i
];
1304 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
1305 wake_up(&sh
->raid_conf
->wait_for_overlap
);
1310 #ifdef CONFIG_MULTICORE_RAID456
1311 static void async_run_ops(void *param
, async_cookie_t cookie
)
1313 struct stripe_head
*sh
= param
;
1314 unsigned long ops_request
= sh
->ops
.request
;
1316 clear_bit_unlock(STRIPE_OPS_REQ_PENDING
, &sh
->state
);
1317 wake_up(&sh
->ops
.wait_for_ops
);
1319 __raid_run_ops(sh
, ops_request
);
1323 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1325 /* since handle_stripe can be called outside of raid5d context
1326 * we need to ensure sh->ops.request is de-staged before another
1329 wait_event(sh
->ops
.wait_for_ops
,
1330 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING
, &sh
->state
));
1331 sh
->ops
.request
= ops_request
;
1333 atomic_inc(&sh
->count
);
1334 async_schedule(async_run_ops
, sh
);
1337 #define raid_run_ops __raid_run_ops
1340 static int grow_one_stripe(raid5_conf_t
*conf
)
1342 struct stripe_head
*sh
;
1343 sh
= kmem_cache_zalloc(conf
->slab_cache
, GFP_KERNEL
);
1347 sh
->raid_conf
= conf
;
1348 #ifdef CONFIG_MULTICORE_RAID456
1349 init_waitqueue_head(&sh
->ops
.wait_for_ops
);
1352 if (grow_buffers(sh
)) {
1354 kmem_cache_free(conf
->slab_cache
, sh
);
1357 /* we just created an active stripe so... */
1358 atomic_set(&sh
->count
, 1);
1359 atomic_inc(&conf
->active_stripes
);
1360 INIT_LIST_HEAD(&sh
->lru
);
1365 static int grow_stripes(raid5_conf_t
*conf
, int num
)
1367 struct kmem_cache
*sc
;
1368 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
1370 if (conf
->mddev
->gendisk
)
1371 sprintf(conf
->cache_name
[0],
1372 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
1374 sprintf(conf
->cache_name
[0],
1375 "raid%d-%p", conf
->level
, conf
->mddev
);
1376 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
1378 conf
->active_name
= 0;
1379 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
1380 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
1384 conf
->slab_cache
= sc
;
1385 conf
->pool_size
= devs
;
1387 if (!grow_one_stripe(conf
))
1393 * scribble_len - return the required size of the scribble region
1394 * @num - total number of disks in the array
1396 * The size must be enough to contain:
1397 * 1/ a struct page pointer for each device in the array +2
1398 * 2/ room to convert each entry in (1) to its corresponding dma
1399 * (dma_map_page()) or page (page_address()) address.
1401 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1402 * calculate over all devices (not just the data blocks), using zeros in place
1403 * of the P and Q blocks.
1405 static size_t scribble_len(int num
)
1409 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
1414 static int resize_stripes(raid5_conf_t
*conf
, int newsize
)
1416 /* Make all the stripes able to hold 'newsize' devices.
1417 * New slots in each stripe get 'page' set to a new page.
1419 * This happens in stages:
1420 * 1/ create a new kmem_cache and allocate the required number of
1422 * 2/ gather all the old stripe_heads and tranfer the pages across
1423 * to the new stripe_heads. This will have the side effect of
1424 * freezing the array as once all stripe_heads have been collected,
1425 * no IO will be possible. Old stripe heads are freed once their
1426 * pages have been transferred over, and the old kmem_cache is
1427 * freed when all stripes are done.
1428 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1429 * we simple return a failre status - no need to clean anything up.
1430 * 4/ allocate new pages for the new slots in the new stripe_heads.
1431 * If this fails, we don't bother trying the shrink the
1432 * stripe_heads down again, we just leave them as they are.
1433 * As each stripe_head is processed the new one is released into
1436 * Once step2 is started, we cannot afford to wait for a write,
1437 * so we use GFP_NOIO allocations.
1439 struct stripe_head
*osh
, *nsh
;
1440 LIST_HEAD(newstripes
);
1441 struct disk_info
*ndisks
;
1444 struct kmem_cache
*sc
;
1447 if (newsize
<= conf
->pool_size
)
1448 return 0; /* never bother to shrink */
1450 err
= md_allow_write(conf
->mddev
);
1455 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
1456 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
1461 for (i
= conf
->max_nr_stripes
; i
; i
--) {
1462 nsh
= kmem_cache_zalloc(sc
, GFP_KERNEL
);
1466 nsh
->raid_conf
= conf
;
1467 #ifdef CONFIG_MULTICORE_RAID456
1468 init_waitqueue_head(&nsh
->ops
.wait_for_ops
);
1471 list_add(&nsh
->lru
, &newstripes
);
1474 /* didn't get enough, give up */
1475 while (!list_empty(&newstripes
)) {
1476 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1477 list_del(&nsh
->lru
);
1478 kmem_cache_free(sc
, nsh
);
1480 kmem_cache_destroy(sc
);
1483 /* Step 2 - Must use GFP_NOIO now.
1484 * OK, we have enough stripes, start collecting inactive
1485 * stripes and copying them over
1487 list_for_each_entry(nsh
, &newstripes
, lru
) {
1488 spin_lock_irq(&conf
->device_lock
);
1489 wait_event_lock_irq(conf
->wait_for_stripe
,
1490 !list_empty(&conf
->inactive_list
),
1493 osh
= get_free_stripe(conf
);
1494 spin_unlock_irq(&conf
->device_lock
);
1495 atomic_set(&nsh
->count
, 1);
1496 for(i
=0; i
<conf
->pool_size
; i
++)
1497 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
1498 for( ; i
<newsize
; i
++)
1499 nsh
->dev
[i
].page
= NULL
;
1500 kmem_cache_free(conf
->slab_cache
, osh
);
1502 kmem_cache_destroy(conf
->slab_cache
);
1505 * At this point, we are holding all the stripes so the array
1506 * is completely stalled, so now is a good time to resize
1507 * conf->disks and the scribble region
1509 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
1511 for (i
=0; i
<conf
->raid_disks
; i
++)
1512 ndisks
[i
] = conf
->disks
[i
];
1514 conf
->disks
= ndisks
;
1519 conf
->scribble_len
= scribble_len(newsize
);
1520 for_each_present_cpu(cpu
) {
1521 struct raid5_percpu
*percpu
;
1524 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1525 scribble
= kmalloc(conf
->scribble_len
, GFP_NOIO
);
1528 kfree(percpu
->scribble
);
1529 percpu
->scribble
= scribble
;
1537 /* Step 4, return new stripes to service */
1538 while(!list_empty(&newstripes
)) {
1539 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1540 list_del_init(&nsh
->lru
);
1542 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
1543 if (nsh
->dev
[i
].page
== NULL
) {
1544 struct page
*p
= alloc_page(GFP_NOIO
);
1545 nsh
->dev
[i
].page
= p
;
1549 release_stripe(nsh
);
1551 /* critical section pass, GFP_NOIO no longer needed */
1553 conf
->slab_cache
= sc
;
1554 conf
->active_name
= 1-conf
->active_name
;
1555 conf
->pool_size
= newsize
;
1559 static int drop_one_stripe(raid5_conf_t
*conf
)
1561 struct stripe_head
*sh
;
1563 spin_lock_irq(&conf
->device_lock
);
1564 sh
= get_free_stripe(conf
);
1565 spin_unlock_irq(&conf
->device_lock
);
1568 BUG_ON(atomic_read(&sh
->count
));
1570 kmem_cache_free(conf
->slab_cache
, sh
);
1571 atomic_dec(&conf
->active_stripes
);
1575 static void shrink_stripes(raid5_conf_t
*conf
)
1577 while (drop_one_stripe(conf
))
1580 if (conf
->slab_cache
)
1581 kmem_cache_destroy(conf
->slab_cache
);
1582 conf
->slab_cache
= NULL
;
1585 static void raid5_end_read_request(struct bio
* bi
, int error
)
1587 struct stripe_head
*sh
= bi
->bi_private
;
1588 raid5_conf_t
*conf
= sh
->raid_conf
;
1589 int disks
= sh
->disks
, i
;
1590 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1591 char b
[BDEVNAME_SIZE
];
1595 for (i
=0 ; i
<disks
; i
++)
1596 if (bi
== &sh
->dev
[i
].req
)
1599 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1600 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1608 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1609 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
1610 rdev
= conf
->disks
[i
].rdev
;
1613 "md/raid:%s: read error corrected"
1614 " (%lu sectors at %llu on %s)\n",
1615 mdname(conf
->mddev
), STRIPE_SECTORS
,
1616 (unsigned long long)(sh
->sector
1617 + rdev
->data_offset
),
1618 bdevname(rdev
->bdev
, b
));
1619 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
1620 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1621 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1623 if (atomic_read(&conf
->disks
[i
].rdev
->read_errors
))
1624 atomic_set(&conf
->disks
[i
].rdev
->read_errors
, 0);
1626 const char *bdn
= bdevname(conf
->disks
[i
].rdev
->bdev
, b
);
1628 rdev
= conf
->disks
[i
].rdev
;
1630 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1631 atomic_inc(&rdev
->read_errors
);
1632 if (conf
->mddev
->degraded
>= conf
->max_degraded
)
1635 "md/raid:%s: read error not correctable "
1636 "(sector %llu on %s).\n",
1637 mdname(conf
->mddev
),
1638 (unsigned long long)(sh
->sector
1639 + rdev
->data_offset
),
1641 else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
))
1645 "md/raid:%s: read error NOT corrected!! "
1646 "(sector %llu on %s).\n",
1647 mdname(conf
->mddev
),
1648 (unsigned long long)(sh
->sector
1649 + rdev
->data_offset
),
1651 else if (atomic_read(&rdev
->read_errors
)
1652 > conf
->max_nr_stripes
)
1654 "md/raid:%s: Too many read errors, failing device %s.\n",
1655 mdname(conf
->mddev
), bdn
);
1659 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1661 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1662 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1663 md_error(conf
->mddev
, rdev
);
1666 rdev_dec_pending(conf
->disks
[i
].rdev
, conf
->mddev
);
1667 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1668 set_bit(STRIPE_HANDLE
, &sh
->state
);
1672 static void raid5_end_write_request(struct bio
*bi
, int error
)
1674 struct stripe_head
*sh
= bi
->bi_private
;
1675 raid5_conf_t
*conf
= sh
->raid_conf
;
1676 int disks
= sh
->disks
, i
;
1677 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1681 for (i
=0 ; i
<disks
; i
++)
1682 if (bi
== &sh
->dev
[i
].req
)
1685 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1686 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1694 set_bit(WriteErrorSeen
, &conf
->disks
[i
].rdev
->flags
);
1695 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
1696 } else if (is_badblock(conf
->disks
[i
].rdev
, sh
->sector
, STRIPE_SECTORS
,
1697 &first_bad
, &bad_sectors
))
1698 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
1700 rdev_dec_pending(conf
->disks
[i
].rdev
, conf
->mddev
);
1702 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1703 set_bit(STRIPE_HANDLE
, &sh
->state
);
1708 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
);
1710 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
1712 struct r5dev
*dev
= &sh
->dev
[i
];
1714 bio_init(&dev
->req
);
1715 dev
->req
.bi_io_vec
= &dev
->vec
;
1717 dev
->req
.bi_max_vecs
++;
1718 dev
->vec
.bv_page
= dev
->page
;
1719 dev
->vec
.bv_len
= STRIPE_SIZE
;
1720 dev
->vec
.bv_offset
= 0;
1722 dev
->req
.bi_sector
= sh
->sector
;
1723 dev
->req
.bi_private
= sh
;
1726 dev
->sector
= compute_blocknr(sh
, i
, previous
);
1729 static void error(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1731 char b
[BDEVNAME_SIZE
];
1732 raid5_conf_t
*conf
= mddev
->private;
1733 pr_debug("raid456: error called\n");
1735 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1736 unsigned long flags
;
1737 spin_lock_irqsave(&conf
->device_lock
, flags
);
1739 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1741 * if recovery was running, make sure it aborts.
1743 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1745 set_bit(Blocked
, &rdev
->flags
);
1746 set_bit(Faulty
, &rdev
->flags
);
1747 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1749 "md/raid:%s: Disk failure on %s, disabling device.\n"
1750 "md/raid:%s: Operation continuing on %d devices.\n",
1752 bdevname(rdev
->bdev
, b
),
1754 conf
->raid_disks
- mddev
->degraded
);
1758 * Input: a 'big' sector number,
1759 * Output: index of the data and parity disk, and the sector # in them.
1761 static sector_t
raid5_compute_sector(raid5_conf_t
*conf
, sector_t r_sector
,
1762 int previous
, int *dd_idx
,
1763 struct stripe_head
*sh
)
1765 sector_t stripe
, stripe2
;
1766 sector_t chunk_number
;
1767 unsigned int chunk_offset
;
1770 sector_t new_sector
;
1771 int algorithm
= previous
? conf
->prev_algo
1773 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
1774 : conf
->chunk_sectors
;
1775 int raid_disks
= previous
? conf
->previous_raid_disks
1777 int data_disks
= raid_disks
- conf
->max_degraded
;
1779 /* First compute the information on this sector */
1782 * Compute the chunk number and the sector offset inside the chunk
1784 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
1785 chunk_number
= r_sector
;
1788 * Compute the stripe number
1790 stripe
= chunk_number
;
1791 *dd_idx
= sector_div(stripe
, data_disks
);
1794 * Select the parity disk based on the user selected algorithm.
1796 pd_idx
= qd_idx
= -1;
1797 switch(conf
->level
) {
1799 pd_idx
= data_disks
;
1802 switch (algorithm
) {
1803 case ALGORITHM_LEFT_ASYMMETRIC
:
1804 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
1805 if (*dd_idx
>= pd_idx
)
1808 case ALGORITHM_RIGHT_ASYMMETRIC
:
1809 pd_idx
= sector_div(stripe2
, raid_disks
);
1810 if (*dd_idx
>= pd_idx
)
1813 case ALGORITHM_LEFT_SYMMETRIC
:
1814 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
1815 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1817 case ALGORITHM_RIGHT_SYMMETRIC
:
1818 pd_idx
= sector_div(stripe2
, raid_disks
);
1819 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1821 case ALGORITHM_PARITY_0
:
1825 case ALGORITHM_PARITY_N
:
1826 pd_idx
= data_disks
;
1834 switch (algorithm
) {
1835 case ALGORITHM_LEFT_ASYMMETRIC
:
1836 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1837 qd_idx
= pd_idx
+ 1;
1838 if (pd_idx
== raid_disks
-1) {
1839 (*dd_idx
)++; /* Q D D D P */
1841 } else if (*dd_idx
>= pd_idx
)
1842 (*dd_idx
) += 2; /* D D P Q D */
1844 case ALGORITHM_RIGHT_ASYMMETRIC
:
1845 pd_idx
= sector_div(stripe2
, raid_disks
);
1846 qd_idx
= pd_idx
+ 1;
1847 if (pd_idx
== raid_disks
-1) {
1848 (*dd_idx
)++; /* Q D D D P */
1850 } else if (*dd_idx
>= pd_idx
)
1851 (*dd_idx
) += 2; /* D D P Q D */
1853 case ALGORITHM_LEFT_SYMMETRIC
:
1854 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1855 qd_idx
= (pd_idx
+ 1) % raid_disks
;
1856 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
1858 case ALGORITHM_RIGHT_SYMMETRIC
:
1859 pd_idx
= sector_div(stripe2
, raid_disks
);
1860 qd_idx
= (pd_idx
+ 1) % raid_disks
;
1861 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
1864 case ALGORITHM_PARITY_0
:
1869 case ALGORITHM_PARITY_N
:
1870 pd_idx
= data_disks
;
1871 qd_idx
= data_disks
+ 1;
1874 case ALGORITHM_ROTATING_ZERO_RESTART
:
1875 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1876 * of blocks for computing Q is different.
1878 pd_idx
= sector_div(stripe2
, raid_disks
);
1879 qd_idx
= pd_idx
+ 1;
1880 if (pd_idx
== raid_disks
-1) {
1881 (*dd_idx
)++; /* Q D D D P */
1883 } else if (*dd_idx
>= pd_idx
)
1884 (*dd_idx
) += 2; /* D D P Q D */
1888 case ALGORITHM_ROTATING_N_RESTART
:
1889 /* Same a left_asymmetric, by first stripe is
1890 * D D D P Q rather than
1894 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1895 qd_idx
= pd_idx
+ 1;
1896 if (pd_idx
== raid_disks
-1) {
1897 (*dd_idx
)++; /* Q D D D P */
1899 } else if (*dd_idx
>= pd_idx
)
1900 (*dd_idx
) += 2; /* D D P Q D */
1904 case ALGORITHM_ROTATING_N_CONTINUE
:
1905 /* Same as left_symmetric but Q is before P */
1906 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1907 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
1908 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1912 case ALGORITHM_LEFT_ASYMMETRIC_6
:
1913 /* RAID5 left_asymmetric, with Q on last device */
1914 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
1915 if (*dd_idx
>= pd_idx
)
1917 qd_idx
= raid_disks
- 1;
1920 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
1921 pd_idx
= sector_div(stripe2
, raid_disks
-1);
1922 if (*dd_idx
>= pd_idx
)
1924 qd_idx
= raid_disks
- 1;
1927 case ALGORITHM_LEFT_SYMMETRIC_6
:
1928 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
1929 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
1930 qd_idx
= raid_disks
- 1;
1933 case ALGORITHM_RIGHT_SYMMETRIC_6
:
1934 pd_idx
= sector_div(stripe2
, raid_disks
-1);
1935 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
1936 qd_idx
= raid_disks
- 1;
1939 case ALGORITHM_PARITY_0_6
:
1942 qd_idx
= raid_disks
- 1;
1952 sh
->pd_idx
= pd_idx
;
1953 sh
->qd_idx
= qd_idx
;
1954 sh
->ddf_layout
= ddf_layout
;
1957 * Finally, compute the new sector number
1959 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
1964 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
1966 raid5_conf_t
*conf
= sh
->raid_conf
;
1967 int raid_disks
= sh
->disks
;
1968 int data_disks
= raid_disks
- conf
->max_degraded
;
1969 sector_t new_sector
= sh
->sector
, check
;
1970 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
1971 : conf
->chunk_sectors
;
1972 int algorithm
= previous
? conf
->prev_algo
1976 sector_t chunk_number
;
1977 int dummy1
, dd_idx
= i
;
1979 struct stripe_head sh2
;
1982 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
1983 stripe
= new_sector
;
1985 if (i
== sh
->pd_idx
)
1987 switch(conf
->level
) {
1990 switch (algorithm
) {
1991 case ALGORITHM_LEFT_ASYMMETRIC
:
1992 case ALGORITHM_RIGHT_ASYMMETRIC
:
1996 case ALGORITHM_LEFT_SYMMETRIC
:
1997 case ALGORITHM_RIGHT_SYMMETRIC
:
2000 i
-= (sh
->pd_idx
+ 1);
2002 case ALGORITHM_PARITY_0
:
2005 case ALGORITHM_PARITY_N
:
2012 if (i
== sh
->qd_idx
)
2013 return 0; /* It is the Q disk */
2014 switch (algorithm
) {
2015 case ALGORITHM_LEFT_ASYMMETRIC
:
2016 case ALGORITHM_RIGHT_ASYMMETRIC
:
2017 case ALGORITHM_ROTATING_ZERO_RESTART
:
2018 case ALGORITHM_ROTATING_N_RESTART
:
2019 if (sh
->pd_idx
== raid_disks
-1)
2020 i
--; /* Q D D D P */
2021 else if (i
> sh
->pd_idx
)
2022 i
-= 2; /* D D P Q D */
2024 case ALGORITHM_LEFT_SYMMETRIC
:
2025 case ALGORITHM_RIGHT_SYMMETRIC
:
2026 if (sh
->pd_idx
== raid_disks
-1)
2027 i
--; /* Q D D D P */
2032 i
-= (sh
->pd_idx
+ 2);
2035 case ALGORITHM_PARITY_0
:
2038 case ALGORITHM_PARITY_N
:
2040 case ALGORITHM_ROTATING_N_CONTINUE
:
2041 /* Like left_symmetric, but P is before Q */
2042 if (sh
->pd_idx
== 0)
2043 i
--; /* P D D D Q */
2048 i
-= (sh
->pd_idx
+ 1);
2051 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2052 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2056 case ALGORITHM_LEFT_SYMMETRIC_6
:
2057 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2059 i
+= data_disks
+ 1;
2060 i
-= (sh
->pd_idx
+ 1);
2062 case ALGORITHM_PARITY_0_6
:
2071 chunk_number
= stripe
* data_disks
+ i
;
2072 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2074 check
= raid5_compute_sector(conf
, r_sector
,
2075 previous
, &dummy1
, &sh2
);
2076 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2077 || sh2
.qd_idx
!= sh
->qd_idx
) {
2078 printk(KERN_ERR
"md/raid:%s: compute_blocknr: map not correct\n",
2079 mdname(conf
->mddev
));
2087 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2088 int rcw
, int expand
)
2090 int i
, pd_idx
= sh
->pd_idx
, disks
= sh
->disks
;
2091 raid5_conf_t
*conf
= sh
->raid_conf
;
2092 int level
= conf
->level
;
2095 /* if we are not expanding this is a proper write request, and
2096 * there will be bios with new data to be drained into the
2100 sh
->reconstruct_state
= reconstruct_state_drain_run
;
2101 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2103 sh
->reconstruct_state
= reconstruct_state_run
;
2105 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2107 for (i
= disks
; i
--; ) {
2108 struct r5dev
*dev
= &sh
->dev
[i
];
2111 set_bit(R5_LOCKED
, &dev
->flags
);
2112 set_bit(R5_Wantdrain
, &dev
->flags
);
2114 clear_bit(R5_UPTODATE
, &dev
->flags
);
2118 if (s
->locked
+ conf
->max_degraded
== disks
)
2119 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2120 atomic_inc(&conf
->pending_full_writes
);
2123 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
2124 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
2126 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
2127 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
2128 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2129 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2131 for (i
= disks
; i
--; ) {
2132 struct r5dev
*dev
= &sh
->dev
[i
];
2137 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2138 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2139 set_bit(R5_Wantdrain
, &dev
->flags
);
2140 set_bit(R5_LOCKED
, &dev
->flags
);
2141 clear_bit(R5_UPTODATE
, &dev
->flags
);
2147 /* keep the parity disk(s) locked while asynchronous operations
2150 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
2151 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2155 int qd_idx
= sh
->qd_idx
;
2156 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
2158 set_bit(R5_LOCKED
, &dev
->flags
);
2159 clear_bit(R5_UPTODATE
, &dev
->flags
);
2163 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2164 __func__
, (unsigned long long)sh
->sector
,
2165 s
->locked
, s
->ops_request
);
2169 * Each stripe/dev can have one or more bion attached.
2170 * toread/towrite point to the first in a chain.
2171 * The bi_next chain must be in order.
2173 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
, int forwrite
)
2176 raid5_conf_t
*conf
= sh
->raid_conf
;
2179 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2180 (unsigned long long)bi
->bi_sector
,
2181 (unsigned long long)sh
->sector
);
2184 spin_lock_irq(&conf
->device_lock
);
2186 bip
= &sh
->dev
[dd_idx
].towrite
;
2187 if (*bip
== NULL
&& sh
->dev
[dd_idx
].written
== NULL
)
2190 bip
= &sh
->dev
[dd_idx
].toread
;
2191 while (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
) {
2192 if ((*bip
)->bi_sector
+ ((*bip
)->bi_size
>> 9) > bi
->bi_sector
)
2194 bip
= & (*bip
)->bi_next
;
2196 if (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
+ ((bi
->bi_size
)>>9))
2199 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
2203 bi
->bi_phys_segments
++;
2206 /* check if page is covered */
2207 sector_t sector
= sh
->dev
[dd_idx
].sector
;
2208 for (bi
=sh
->dev
[dd_idx
].towrite
;
2209 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
2210 bi
&& bi
->bi_sector
<= sector
;
2211 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
2212 if (bi
->bi_sector
+ (bi
->bi_size
>>9) >= sector
)
2213 sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
2215 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
2216 set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
);
2218 spin_unlock_irq(&conf
->device_lock
);
2220 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2221 (unsigned long long)(*bip
)->bi_sector
,
2222 (unsigned long long)sh
->sector
, dd_idx
);
2224 if (conf
->mddev
->bitmap
&& firstwrite
) {
2225 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
2227 sh
->bm_seq
= conf
->seq_flush
+1;
2228 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
2233 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
2234 spin_unlock_irq(&conf
->device_lock
);
2238 static void end_reshape(raid5_conf_t
*conf
);
2240 static void stripe_set_idx(sector_t stripe
, raid5_conf_t
*conf
, int previous
,
2241 struct stripe_head
*sh
)
2243 int sectors_per_chunk
=
2244 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
2246 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
2247 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
2249 raid5_compute_sector(conf
,
2250 stripe
* (disks
- conf
->max_degraded
)
2251 *sectors_per_chunk
+ chunk_offset
,
2257 handle_failed_stripe(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2258 struct stripe_head_state
*s
, int disks
,
2259 struct bio
**return_bi
)
2262 for (i
= disks
; i
--; ) {
2266 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2269 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2270 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
2271 atomic_inc(&rdev
->nr_pending
);
2276 if (!rdev_set_badblocks(
2280 md_error(conf
->mddev
, rdev
);
2281 rdev_dec_pending(rdev
, conf
->mddev
);
2284 spin_lock_irq(&conf
->device_lock
);
2285 /* fail all writes first */
2286 bi
= sh
->dev
[i
].towrite
;
2287 sh
->dev
[i
].towrite
= NULL
;
2293 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2294 wake_up(&conf
->wait_for_overlap
);
2296 while (bi
&& bi
->bi_sector
<
2297 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2298 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2299 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2300 if (!raid5_dec_bi_phys_segments(bi
)) {
2301 md_write_end(conf
->mddev
);
2302 bi
->bi_next
= *return_bi
;
2307 /* and fail all 'written' */
2308 bi
= sh
->dev
[i
].written
;
2309 sh
->dev
[i
].written
= NULL
;
2310 if (bi
) bitmap_end
= 1;
2311 while (bi
&& bi
->bi_sector
<
2312 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2313 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2314 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2315 if (!raid5_dec_bi_phys_segments(bi
)) {
2316 md_write_end(conf
->mddev
);
2317 bi
->bi_next
= *return_bi
;
2323 /* fail any reads if this device is non-operational and
2324 * the data has not reached the cache yet.
2326 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
2327 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
2328 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
2329 bi
= sh
->dev
[i
].toread
;
2330 sh
->dev
[i
].toread
= NULL
;
2331 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2332 wake_up(&conf
->wait_for_overlap
);
2333 if (bi
) s
->to_read
--;
2334 while (bi
&& bi
->bi_sector
<
2335 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2336 struct bio
*nextbi
=
2337 r5_next_bio(bi
, sh
->dev
[i
].sector
);
2338 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2339 if (!raid5_dec_bi_phys_segments(bi
)) {
2340 bi
->bi_next
= *return_bi
;
2346 spin_unlock_irq(&conf
->device_lock
);
2348 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2349 STRIPE_SECTORS
, 0, 0);
2350 /* If we were in the middle of a write the parity block might
2351 * still be locked - so just clear all R5_LOCKED flags
2353 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2356 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2357 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2358 md_wakeup_thread(conf
->mddev
->thread
);
2362 handle_failed_sync(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2363 struct stripe_head_state
*s
)
2368 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 0);
2369 clear_bit(STRIPE_SYNCING
, &sh
->state
);
2371 /* There is nothing more to do for sync/check/repair.
2372 * For recover we need to record a bad block on all
2373 * non-sync devices, or abort the recovery
2375 if (!test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
))
2377 /* During recovery devices cannot be removed, so locking and
2378 * refcounting of rdevs is not needed
2380 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2381 mdk_rdev_t
*rdev
= conf
->disks
[i
].rdev
;
2383 || test_bit(Faulty
, &rdev
->flags
)
2384 || test_bit(In_sync
, &rdev
->flags
))
2386 if (!rdev_set_badblocks(rdev
, sh
->sector
,
2391 conf
->recovery_disabled
= conf
->mddev
->recovery_disabled
;
2392 set_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
);
2396 /* fetch_block - checks the given member device to see if its data needs
2397 * to be read or computed to satisfy a request.
2399 * Returns 1 when no more member devices need to be checked, otherwise returns
2400 * 0 to tell the loop in handle_stripe_fill to continue
2402 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2403 int disk_idx
, int disks
)
2405 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
2406 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
2407 &sh
->dev
[s
->failed_num
[1]] };
2409 /* is the data in this block needed, and can we get it? */
2410 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2411 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2413 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2414 s
->syncing
|| s
->expanding
||
2415 (s
->failed
>= 1 && fdev
[0]->toread
) ||
2416 (s
->failed
>= 2 && fdev
[1]->toread
) ||
2417 (sh
->raid_conf
->level
<= 5 && s
->failed
&& fdev
[0]->towrite
&&
2418 !test_bit(R5_OVERWRITE
, &fdev
[0]->flags
)) ||
2419 (sh
->raid_conf
->level
== 6 && s
->failed
&& s
->to_write
))) {
2420 /* we would like to get this block, possibly by computing it,
2421 * otherwise read it if the backing disk is insync
2423 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
2424 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
2425 if ((s
->uptodate
== disks
- 1) &&
2426 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
2427 disk_idx
== s
->failed_num
[1]))) {
2428 /* have disk failed, and we're requested to fetch it;
2431 pr_debug("Computing stripe %llu block %d\n",
2432 (unsigned long long)sh
->sector
, disk_idx
);
2433 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2434 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2435 set_bit(R5_Wantcompute
, &dev
->flags
);
2436 sh
->ops
.target
= disk_idx
;
2437 sh
->ops
.target2
= -1; /* no 2nd target */
2439 /* Careful: from this point on 'uptodate' is in the eye
2440 * of raid_run_ops which services 'compute' operations
2441 * before writes. R5_Wantcompute flags a block that will
2442 * be R5_UPTODATE by the time it is needed for a
2443 * subsequent operation.
2447 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
2448 /* Computing 2-failure is *very* expensive; only
2449 * do it if failed >= 2
2452 for (other
= disks
; other
--; ) {
2453 if (other
== disk_idx
)
2455 if (!test_bit(R5_UPTODATE
,
2456 &sh
->dev
[other
].flags
))
2460 pr_debug("Computing stripe %llu blocks %d,%d\n",
2461 (unsigned long long)sh
->sector
,
2463 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2464 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2465 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
2466 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
2467 sh
->ops
.target
= disk_idx
;
2468 sh
->ops
.target2
= other
;
2472 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2473 set_bit(R5_LOCKED
, &dev
->flags
);
2474 set_bit(R5_Wantread
, &dev
->flags
);
2476 pr_debug("Reading block %d (sync=%d)\n",
2477 disk_idx
, s
->syncing
);
2485 * handle_stripe_fill - read or compute data to satisfy pending requests.
2487 static void handle_stripe_fill(struct stripe_head
*sh
,
2488 struct stripe_head_state
*s
,
2493 /* look for blocks to read/compute, skip this if a compute
2494 * is already in flight, or if the stripe contents are in the
2495 * midst of changing due to a write
2497 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2498 !sh
->reconstruct_state
)
2499 for (i
= disks
; i
--; )
2500 if (fetch_block(sh
, s
, i
, disks
))
2502 set_bit(STRIPE_HANDLE
, &sh
->state
);
2506 /* handle_stripe_clean_event
2507 * any written block on an uptodate or failed drive can be returned.
2508 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2509 * never LOCKED, so we don't need to test 'failed' directly.
2511 static void handle_stripe_clean_event(raid5_conf_t
*conf
,
2512 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
2517 for (i
= disks
; i
--; )
2518 if (sh
->dev
[i
].written
) {
2520 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2521 test_bit(R5_UPTODATE
, &dev
->flags
)) {
2522 /* We can return any write requests */
2523 struct bio
*wbi
, *wbi2
;
2525 pr_debug("Return write for disc %d\n", i
);
2526 spin_lock_irq(&conf
->device_lock
);
2528 dev
->written
= NULL
;
2529 while (wbi
&& wbi
->bi_sector
<
2530 dev
->sector
+ STRIPE_SECTORS
) {
2531 wbi2
= r5_next_bio(wbi
, dev
->sector
);
2532 if (!raid5_dec_bi_phys_segments(wbi
)) {
2533 md_write_end(conf
->mddev
);
2534 wbi
->bi_next
= *return_bi
;
2539 if (dev
->towrite
== NULL
)
2541 spin_unlock_irq(&conf
->device_lock
);
2543 bitmap_endwrite(conf
->mddev
->bitmap
,
2546 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
2551 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2552 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2553 md_wakeup_thread(conf
->mddev
->thread
);
2556 static void handle_stripe_dirtying(raid5_conf_t
*conf
,
2557 struct stripe_head
*sh
,
2558 struct stripe_head_state
*s
,
2561 int rmw
= 0, rcw
= 0, i
;
2562 if (conf
->max_degraded
== 2) {
2563 /* RAID6 requires 'rcw' in current implementation
2564 * Calculate the real rcw later - for now fake it
2565 * look like rcw is cheaper
2568 } else for (i
= disks
; i
--; ) {
2569 /* would I have to read this buffer for read_modify_write */
2570 struct r5dev
*dev
= &sh
->dev
[i
];
2571 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2572 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2573 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2574 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2575 if (test_bit(R5_Insync
, &dev
->flags
))
2578 rmw
+= 2*disks
; /* cannot read it */
2580 /* Would I have to read this buffer for reconstruct_write */
2581 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) && i
!= sh
->pd_idx
&&
2582 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2583 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2584 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2585 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2590 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2591 (unsigned long long)sh
->sector
, rmw
, rcw
);
2592 set_bit(STRIPE_HANDLE
, &sh
->state
);
2593 if (rmw
< rcw
&& rmw
> 0)
2594 /* prefer read-modify-write, but need to get some data */
2595 for (i
= disks
; i
--; ) {
2596 struct r5dev
*dev
= &sh
->dev
[i
];
2597 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2598 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2599 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2600 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2601 test_bit(R5_Insync
, &dev
->flags
)) {
2603 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2604 pr_debug("Read_old block "
2605 "%d for r-m-w\n", i
);
2606 set_bit(R5_LOCKED
, &dev
->flags
);
2607 set_bit(R5_Wantread
, &dev
->flags
);
2610 set_bit(STRIPE_DELAYED
, &sh
->state
);
2611 set_bit(STRIPE_HANDLE
, &sh
->state
);
2615 if (rcw
<= rmw
&& rcw
> 0) {
2616 /* want reconstruct write, but need to get some data */
2618 for (i
= disks
; i
--; ) {
2619 struct r5dev
*dev
= &sh
->dev
[i
];
2620 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2621 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
2622 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2623 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2624 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2626 if (!test_bit(R5_Insync
, &dev
->flags
))
2627 continue; /* it's a failed drive */
2629 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2630 pr_debug("Read_old block "
2631 "%d for Reconstruct\n", i
);
2632 set_bit(R5_LOCKED
, &dev
->flags
);
2633 set_bit(R5_Wantread
, &dev
->flags
);
2636 set_bit(STRIPE_DELAYED
, &sh
->state
);
2637 set_bit(STRIPE_HANDLE
, &sh
->state
);
2642 /* now if nothing is locked, and if we have enough data,
2643 * we can start a write request
2645 /* since handle_stripe can be called at any time we need to handle the
2646 * case where a compute block operation has been submitted and then a
2647 * subsequent call wants to start a write request. raid_run_ops only
2648 * handles the case where compute block and reconstruct are requested
2649 * simultaneously. If this is not the case then new writes need to be
2650 * held off until the compute completes.
2652 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
2653 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
2654 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
2655 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
2658 static void handle_parity_checks5(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2659 struct stripe_head_state
*s
, int disks
)
2661 struct r5dev
*dev
= NULL
;
2663 set_bit(STRIPE_HANDLE
, &sh
->state
);
2665 switch (sh
->check_state
) {
2666 case check_state_idle
:
2667 /* start a new check operation if there are no failures */
2668 if (s
->failed
== 0) {
2669 BUG_ON(s
->uptodate
!= disks
);
2670 sh
->check_state
= check_state_run
;
2671 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2672 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2676 dev
= &sh
->dev
[s
->failed_num
[0]];
2678 case check_state_compute_result
:
2679 sh
->check_state
= check_state_idle
;
2681 dev
= &sh
->dev
[sh
->pd_idx
];
2683 /* check that a write has not made the stripe insync */
2684 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2687 /* either failed parity check, or recovery is happening */
2688 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
2689 BUG_ON(s
->uptodate
!= disks
);
2691 set_bit(R5_LOCKED
, &dev
->flags
);
2693 set_bit(R5_Wantwrite
, &dev
->flags
);
2695 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2696 set_bit(STRIPE_INSYNC
, &sh
->state
);
2698 case check_state_run
:
2699 break; /* we will be called again upon completion */
2700 case check_state_check_result
:
2701 sh
->check_state
= check_state_idle
;
2703 /* if a failure occurred during the check operation, leave
2704 * STRIPE_INSYNC not set and let the stripe be handled again
2709 /* handle a successful check operation, if parity is correct
2710 * we are done. Otherwise update the mismatch count and repair
2711 * parity if !MD_RECOVERY_CHECK
2713 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
2714 /* parity is correct (on disc,
2715 * not in buffer any more)
2717 set_bit(STRIPE_INSYNC
, &sh
->state
);
2719 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2720 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2721 /* don't try to repair!! */
2722 set_bit(STRIPE_INSYNC
, &sh
->state
);
2724 sh
->check_state
= check_state_compute_run
;
2725 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2726 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2727 set_bit(R5_Wantcompute
,
2728 &sh
->dev
[sh
->pd_idx
].flags
);
2729 sh
->ops
.target
= sh
->pd_idx
;
2730 sh
->ops
.target2
= -1;
2735 case check_state_compute_run
:
2738 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
2739 __func__
, sh
->check_state
,
2740 (unsigned long long) sh
->sector
);
2746 static void handle_parity_checks6(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2747 struct stripe_head_state
*s
,
2750 int pd_idx
= sh
->pd_idx
;
2751 int qd_idx
= sh
->qd_idx
;
2754 set_bit(STRIPE_HANDLE
, &sh
->state
);
2756 BUG_ON(s
->failed
> 2);
2758 /* Want to check and possibly repair P and Q.
2759 * However there could be one 'failed' device, in which
2760 * case we can only check one of them, possibly using the
2761 * other to generate missing data
2764 switch (sh
->check_state
) {
2765 case check_state_idle
:
2766 /* start a new check operation if there are < 2 failures */
2767 if (s
->failed
== s
->q_failed
) {
2768 /* The only possible failed device holds Q, so it
2769 * makes sense to check P (If anything else were failed,
2770 * we would have used P to recreate it).
2772 sh
->check_state
= check_state_run
;
2774 if (!s
->q_failed
&& s
->failed
< 2) {
2775 /* Q is not failed, and we didn't use it to generate
2776 * anything, so it makes sense to check it
2778 if (sh
->check_state
== check_state_run
)
2779 sh
->check_state
= check_state_run_pq
;
2781 sh
->check_state
= check_state_run_q
;
2784 /* discard potentially stale zero_sum_result */
2785 sh
->ops
.zero_sum_result
= 0;
2787 if (sh
->check_state
== check_state_run
) {
2788 /* async_xor_zero_sum destroys the contents of P */
2789 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2792 if (sh
->check_state
>= check_state_run
&&
2793 sh
->check_state
<= check_state_run_pq
) {
2794 /* async_syndrome_zero_sum preserves P and Q, so
2795 * no need to mark them !uptodate here
2797 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2801 /* we have 2-disk failure */
2802 BUG_ON(s
->failed
!= 2);
2804 case check_state_compute_result
:
2805 sh
->check_state
= check_state_idle
;
2807 /* check that a write has not made the stripe insync */
2808 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2811 /* now write out any block on a failed drive,
2812 * or P or Q if they were recomputed
2814 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
2815 if (s
->failed
== 2) {
2816 dev
= &sh
->dev
[s
->failed_num
[1]];
2818 set_bit(R5_LOCKED
, &dev
->flags
);
2819 set_bit(R5_Wantwrite
, &dev
->flags
);
2821 if (s
->failed
>= 1) {
2822 dev
= &sh
->dev
[s
->failed_num
[0]];
2824 set_bit(R5_LOCKED
, &dev
->flags
);
2825 set_bit(R5_Wantwrite
, &dev
->flags
);
2827 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
2828 dev
= &sh
->dev
[pd_idx
];
2830 set_bit(R5_LOCKED
, &dev
->flags
);
2831 set_bit(R5_Wantwrite
, &dev
->flags
);
2833 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
2834 dev
= &sh
->dev
[qd_idx
];
2836 set_bit(R5_LOCKED
, &dev
->flags
);
2837 set_bit(R5_Wantwrite
, &dev
->flags
);
2839 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2841 set_bit(STRIPE_INSYNC
, &sh
->state
);
2843 case check_state_run
:
2844 case check_state_run_q
:
2845 case check_state_run_pq
:
2846 break; /* we will be called again upon completion */
2847 case check_state_check_result
:
2848 sh
->check_state
= check_state_idle
;
2850 /* handle a successful check operation, if parity is correct
2851 * we are done. Otherwise update the mismatch count and repair
2852 * parity if !MD_RECOVERY_CHECK
2854 if (sh
->ops
.zero_sum_result
== 0) {
2855 /* both parities are correct */
2857 set_bit(STRIPE_INSYNC
, &sh
->state
);
2859 /* in contrast to the raid5 case we can validate
2860 * parity, but still have a failure to write
2863 sh
->check_state
= check_state_compute_result
;
2864 /* Returning at this point means that we may go
2865 * off and bring p and/or q uptodate again so
2866 * we make sure to check zero_sum_result again
2867 * to verify if p or q need writeback
2871 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2872 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2873 /* don't try to repair!! */
2874 set_bit(STRIPE_INSYNC
, &sh
->state
);
2876 int *target
= &sh
->ops
.target
;
2878 sh
->ops
.target
= -1;
2879 sh
->ops
.target2
= -1;
2880 sh
->check_state
= check_state_compute_run
;
2881 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2882 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2883 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
2884 set_bit(R5_Wantcompute
,
2885 &sh
->dev
[pd_idx
].flags
);
2887 target
= &sh
->ops
.target2
;
2890 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
2891 set_bit(R5_Wantcompute
,
2892 &sh
->dev
[qd_idx
].flags
);
2899 case check_state_compute_run
:
2902 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
2903 __func__
, sh
->check_state
,
2904 (unsigned long long) sh
->sector
);
2909 static void handle_stripe_expansion(raid5_conf_t
*conf
, struct stripe_head
*sh
)
2913 /* We have read all the blocks in this stripe and now we need to
2914 * copy some of them into a target stripe for expand.
2916 struct dma_async_tx_descriptor
*tx
= NULL
;
2917 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
2918 for (i
= 0; i
< sh
->disks
; i
++)
2919 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
2921 struct stripe_head
*sh2
;
2922 struct async_submit_ctl submit
;
2924 sector_t bn
= compute_blocknr(sh
, i
, 1);
2925 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
2927 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
2929 /* so far only the early blocks of this stripe
2930 * have been requested. When later blocks
2931 * get requested, we will try again
2934 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
2935 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
2936 /* must have already done this block */
2937 release_stripe(sh2
);
2941 /* place all the copies on one channel */
2942 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
2943 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
2944 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
2947 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
2948 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
2949 for (j
= 0; j
< conf
->raid_disks
; j
++)
2950 if (j
!= sh2
->pd_idx
&&
2952 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
2954 if (j
== conf
->raid_disks
) {
2955 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
2956 set_bit(STRIPE_HANDLE
, &sh2
->state
);
2958 release_stripe(sh2
);
2961 /* done submitting copies, wait for them to complete */
2964 dma_wait_for_async_tx(tx
);
2970 * handle_stripe - do things to a stripe.
2972 * We lock the stripe and then examine the state of various bits
2973 * to see what needs to be done.
2975 * return some read request which now have data
2976 * return some write requests which are safely on disc
2977 * schedule a read on some buffers
2978 * schedule a write of some buffers
2979 * return confirmation of parity correctness
2981 * buffers are taken off read_list or write_list, and bh_cache buffers
2982 * get BH_Lock set before the stripe lock is released.
2986 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
2988 raid5_conf_t
*conf
= sh
->raid_conf
;
2989 int disks
= sh
->disks
;
2993 memset(s
, 0, sizeof(*s
));
2995 s
->syncing
= test_bit(STRIPE_SYNCING
, &sh
->state
);
2996 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
2997 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
2998 s
->failed_num
[0] = -1;
2999 s
->failed_num
[1] = -1;
3001 /* Now to look around and see what can be done */
3003 spin_lock_irq(&conf
->device_lock
);
3004 for (i
=disks
; i
--; ) {
3012 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3013 i
, dev
->flags
, dev
->toread
, dev
->towrite
, dev
->written
);
3014 /* maybe we can reply to a read
3016 * new wantfill requests are only permitted while
3017 * ops_complete_biofill is guaranteed to be inactive
3019 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
3020 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
3021 set_bit(R5_Wantfill
, &dev
->flags
);
3023 /* now count some things */
3024 if (test_bit(R5_LOCKED
, &dev
->flags
))
3026 if (test_bit(R5_UPTODATE
, &dev
->flags
))
3028 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
3030 BUG_ON(s
->compute
> 2);
3033 if (test_bit(R5_Wantfill
, &dev
->flags
))
3035 else if (dev
->toread
)
3039 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3044 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3046 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3047 &first_bad
, &bad_sectors
);
3048 if (s
->blocked_rdev
== NULL
3049 && (test_bit(Blocked
, &rdev
->flags
)
3052 set_bit(BlockedBadBlocks
,
3054 s
->blocked_rdev
= rdev
;
3055 atomic_inc(&rdev
->nr_pending
);
3058 clear_bit(R5_Insync
, &dev
->flags
);
3062 /* also not in-sync */
3063 if (!test_bit(WriteErrorSeen
, &rdev
->flags
)) {
3064 /* treat as in-sync, but with a read error
3065 * which we can now try to correct
3067 set_bit(R5_Insync
, &dev
->flags
);
3068 set_bit(R5_ReadError
, &dev
->flags
);
3070 } else if (test_bit(In_sync
, &rdev
->flags
))
3071 set_bit(R5_Insync
, &dev
->flags
);
3072 else if (!test_bit(Faulty
, &rdev
->flags
)) {
3073 /* in sync if before recovery_offset */
3074 if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
3075 set_bit(R5_Insync
, &dev
->flags
);
3077 if (test_bit(R5_WriteError
, &dev
->flags
)) {
3078 clear_bit(R5_Insync
, &dev
->flags
);
3079 if (!test_bit(Faulty
, &rdev
->flags
)) {
3080 s
->handle_bad_blocks
= 1;
3081 atomic_inc(&rdev
->nr_pending
);
3083 clear_bit(R5_WriteError
, &dev
->flags
);
3085 if (test_bit(R5_MadeGood
, &dev
->flags
)) {
3086 if (!test_bit(Faulty
, &rdev
->flags
)) {
3087 s
->handle_bad_blocks
= 1;
3088 atomic_inc(&rdev
->nr_pending
);
3090 clear_bit(R5_MadeGood
, &dev
->flags
);
3092 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3093 /* The ReadError flag will just be confusing now */
3094 clear_bit(R5_ReadError
, &dev
->flags
);
3095 clear_bit(R5_ReWrite
, &dev
->flags
);
3097 if (test_bit(R5_ReadError
, &dev
->flags
))
3098 clear_bit(R5_Insync
, &dev
->flags
);
3099 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3101 s
->failed_num
[s
->failed
] = i
;
3105 spin_unlock_irq(&conf
->device_lock
);
3109 static void handle_stripe(struct stripe_head
*sh
)
3111 struct stripe_head_state s
;
3112 raid5_conf_t
*conf
= sh
->raid_conf
;
3115 int disks
= sh
->disks
;
3116 struct r5dev
*pdev
, *qdev
;
3118 clear_bit(STRIPE_HANDLE
, &sh
->state
);
3119 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
3120 /* already being handled, ensure it gets handled
3121 * again when current action finishes */
3122 set_bit(STRIPE_HANDLE
, &sh
->state
);
3126 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
3127 set_bit(STRIPE_SYNCING
, &sh
->state
);
3128 clear_bit(STRIPE_INSYNC
, &sh
->state
);
3130 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3132 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3133 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3134 (unsigned long long)sh
->sector
, sh
->state
,
3135 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
3136 sh
->check_state
, sh
->reconstruct_state
);
3138 analyse_stripe(sh
, &s
);
3140 if (s
.handle_bad_blocks
) {
3141 set_bit(STRIPE_HANDLE
, &sh
->state
);
3145 if (unlikely(s
.blocked_rdev
)) {
3146 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
3147 s
.to_write
|| s
.written
) {
3148 set_bit(STRIPE_HANDLE
, &sh
->state
);
3151 /* There is nothing for the blocked_rdev to block */
3152 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
3153 s
.blocked_rdev
= NULL
;
3156 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
3157 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
3158 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
3161 pr_debug("locked=%d uptodate=%d to_read=%d"
3162 " to_write=%d failed=%d failed_num=%d,%d\n",
3163 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
3164 s
.failed_num
[0], s
.failed_num
[1]);
3165 /* check if the array has lost more than max_degraded devices and,
3166 * if so, some requests might need to be failed.
3168 if (s
.failed
> conf
->max_degraded
) {
3169 sh
->check_state
= 0;
3170 sh
->reconstruct_state
= 0;
3171 if (s
.to_read
+s
.to_write
+s
.written
)
3172 handle_failed_stripe(conf
, sh
, &s
, disks
, &s
.return_bi
);
3174 handle_failed_sync(conf
, sh
, &s
);
3178 * might be able to return some write requests if the parity blocks
3179 * are safe, or on a failed drive
3181 pdev
= &sh
->dev
[sh
->pd_idx
];
3182 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
3183 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
3184 qdev
= &sh
->dev
[sh
->qd_idx
];
3185 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
3186 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
3190 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
3191 && !test_bit(R5_LOCKED
, &pdev
->flags
)
3192 && test_bit(R5_UPTODATE
, &pdev
->flags
)))) &&
3193 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
3194 && !test_bit(R5_LOCKED
, &qdev
->flags
)
3195 && test_bit(R5_UPTODATE
, &qdev
->flags
)))))
3196 handle_stripe_clean_event(conf
, sh
, disks
, &s
.return_bi
);
3198 /* Now we might consider reading some blocks, either to check/generate
3199 * parity, or to satisfy requests
3200 * or to load a block that is being partially written.
3202 if (s
.to_read
|| s
.non_overwrite
3203 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
3204 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
)) || s
.expanding
)
3205 handle_stripe_fill(sh
, &s
, disks
);
3207 /* Now we check to see if any write operations have recently
3211 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
3213 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
3214 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
3215 sh
->reconstruct_state
= reconstruct_state_idle
;
3217 /* All the 'written' buffers and the parity block are ready to
3218 * be written back to disk
3220 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
));
3221 BUG_ON(sh
->qd_idx
>= 0 &&
3222 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
));
3223 for (i
= disks
; i
--; ) {
3224 struct r5dev
*dev
= &sh
->dev
[i
];
3225 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
3226 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3228 pr_debug("Writing block %d\n", i
);
3229 set_bit(R5_Wantwrite
, &dev
->flags
);
3232 if (!test_bit(R5_Insync
, &dev
->flags
) ||
3233 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3235 set_bit(STRIPE_INSYNC
, &sh
->state
);
3238 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3239 s
.dec_preread_active
= 1;
3242 /* Now to consider new write requests and what else, if anything
3243 * should be read. We do not handle new writes when:
3244 * 1/ A 'write' operation (copy+xor) is already in flight.
3245 * 2/ A 'check' operation is in flight, as it may clobber the parity
3248 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
3249 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
3251 /* maybe we need to check and possibly fix the parity for this stripe
3252 * Any reads will already have been scheduled, so we just see if enough
3253 * data is available. The parity check is held off while parity
3254 * dependent operations are in flight.
3256 if (sh
->check_state
||
3257 (s
.syncing
&& s
.locked
== 0 &&
3258 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
3259 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
3260 if (conf
->level
== 6)
3261 handle_parity_checks6(conf
, sh
, &s
, disks
);
3263 handle_parity_checks5(conf
, sh
, &s
, disks
);
3266 if (s
.syncing
&& s
.locked
== 0 && test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3267 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3268 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3271 /* If the failed drives are just a ReadError, then we might need
3272 * to progress the repair/check process
3274 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
3275 for (i
= 0; i
< s
.failed
; i
++) {
3276 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
3277 if (test_bit(R5_ReadError
, &dev
->flags
)
3278 && !test_bit(R5_LOCKED
, &dev
->flags
)
3279 && test_bit(R5_UPTODATE
, &dev
->flags
)
3281 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3282 set_bit(R5_Wantwrite
, &dev
->flags
);
3283 set_bit(R5_ReWrite
, &dev
->flags
);
3284 set_bit(R5_LOCKED
, &dev
->flags
);
3287 /* let's read it back */
3288 set_bit(R5_Wantread
, &dev
->flags
);
3289 set_bit(R5_LOCKED
, &dev
->flags
);
3296 /* Finish reconstruct operations initiated by the expansion process */
3297 if (sh
->reconstruct_state
== reconstruct_state_result
) {
3298 struct stripe_head
*sh_src
3299 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
3300 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
3301 /* sh cannot be written until sh_src has been read.
3302 * so arrange for sh to be delayed a little
3304 set_bit(STRIPE_DELAYED
, &sh
->state
);
3305 set_bit(STRIPE_HANDLE
, &sh
->state
);
3306 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
3308 atomic_inc(&conf
->preread_active_stripes
);
3309 release_stripe(sh_src
);
3313 release_stripe(sh_src
);
3315 sh
->reconstruct_state
= reconstruct_state_idle
;
3316 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3317 for (i
= conf
->raid_disks
; i
--; ) {
3318 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3319 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3324 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
3325 !sh
->reconstruct_state
) {
3326 /* Need to write out all blocks after computing parity */
3327 sh
->disks
= conf
->raid_disks
;
3328 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
3329 schedule_reconstruction(sh
, &s
, 1, 1);
3330 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
3331 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3332 atomic_dec(&conf
->reshape_stripes
);
3333 wake_up(&conf
->wait_for_overlap
);
3334 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3337 if (s
.expanding
&& s
.locked
== 0 &&
3338 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
3339 handle_stripe_expansion(conf
, sh
);
3342 /* wait for this device to become unblocked */
3343 if (conf
->mddev
->external
&& unlikely(s
.blocked_rdev
))
3344 md_wait_for_blocked_rdev(s
.blocked_rdev
, conf
->mddev
);
3346 if (s
.handle_bad_blocks
)
3347 for (i
= disks
; i
--; ) {
3349 struct r5dev
*dev
= &sh
->dev
[i
];
3350 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
3351 /* We own a safe reference to the rdev */
3352 rdev
= conf
->disks
[i
].rdev
;
3353 if (!rdev_set_badblocks(rdev
, sh
->sector
,
3355 md_error(conf
->mddev
, rdev
);
3356 rdev_dec_pending(rdev
, conf
->mddev
);
3358 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
3359 rdev
= conf
->disks
[i
].rdev
;
3360 rdev_clear_badblocks(rdev
, sh
->sector
,
3362 rdev_dec_pending(rdev
, conf
->mddev
);
3367 raid_run_ops(sh
, s
.ops_request
);
3371 if (s
.dec_preread_active
) {
3372 /* We delay this until after ops_run_io so that if make_request
3373 * is waiting on a flush, it won't continue until the writes
3374 * have actually been submitted.
3376 atomic_dec(&conf
->preread_active_stripes
);
3377 if (atomic_read(&conf
->preread_active_stripes
) <
3379 md_wakeup_thread(conf
->mddev
->thread
);
3382 return_io(s
.return_bi
);
3384 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
3387 static void raid5_activate_delayed(raid5_conf_t
*conf
)
3389 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
3390 while (!list_empty(&conf
->delayed_list
)) {
3391 struct list_head
*l
= conf
->delayed_list
.next
;
3392 struct stripe_head
*sh
;
3393 sh
= list_entry(l
, struct stripe_head
, lru
);
3395 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3396 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3397 atomic_inc(&conf
->preread_active_stripes
);
3398 list_add_tail(&sh
->lru
, &conf
->hold_list
);
3403 static void activate_bit_delay(raid5_conf_t
*conf
)
3405 /* device_lock is held */
3406 struct list_head head
;
3407 list_add(&head
, &conf
->bitmap_list
);
3408 list_del_init(&conf
->bitmap_list
);
3409 while (!list_empty(&head
)) {
3410 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
3411 list_del_init(&sh
->lru
);
3412 atomic_inc(&sh
->count
);
3413 __release_stripe(conf
, sh
);
3417 int md_raid5_congested(mddev_t
*mddev
, int bits
)
3419 raid5_conf_t
*conf
= mddev
->private;
3421 /* No difference between reads and writes. Just check
3422 * how busy the stripe_cache is
3425 if (conf
->inactive_blocked
)
3429 if (list_empty_careful(&conf
->inactive_list
))
3434 EXPORT_SYMBOL_GPL(md_raid5_congested
);
3436 static int raid5_congested(void *data
, int bits
)
3438 mddev_t
*mddev
= data
;
3440 return mddev_congested(mddev
, bits
) ||
3441 md_raid5_congested(mddev
, bits
);
3444 /* We want read requests to align with chunks where possible,
3445 * but write requests don't need to.
3447 static int raid5_mergeable_bvec(struct request_queue
*q
,
3448 struct bvec_merge_data
*bvm
,
3449 struct bio_vec
*biovec
)
3451 mddev_t
*mddev
= q
->queuedata
;
3452 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
3454 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3455 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
3457 if ((bvm
->bi_rw
& 1) == WRITE
)
3458 return biovec
->bv_len
; /* always allow writes to be mergeable */
3460 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3461 chunk_sectors
= mddev
->new_chunk_sectors
;
3462 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
3463 if (max
< 0) max
= 0;
3464 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
3465 return biovec
->bv_len
;
3471 static int in_chunk_boundary(mddev_t
*mddev
, struct bio
*bio
)
3473 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
3474 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3475 unsigned int bio_sectors
= bio
->bi_size
>> 9;
3477 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3478 chunk_sectors
= mddev
->new_chunk_sectors
;
3479 return chunk_sectors
>=
3480 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
3484 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3485 * later sampled by raid5d.
3487 static void add_bio_to_retry(struct bio
*bi
,raid5_conf_t
*conf
)
3489 unsigned long flags
;
3491 spin_lock_irqsave(&conf
->device_lock
, flags
);
3493 bi
->bi_next
= conf
->retry_read_aligned_list
;
3494 conf
->retry_read_aligned_list
= bi
;
3496 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3497 md_wakeup_thread(conf
->mddev
->thread
);
3501 static struct bio
*remove_bio_from_retry(raid5_conf_t
*conf
)
3505 bi
= conf
->retry_read_aligned
;
3507 conf
->retry_read_aligned
= NULL
;
3510 bi
= conf
->retry_read_aligned_list
;
3512 conf
->retry_read_aligned_list
= bi
->bi_next
;
3515 * this sets the active strip count to 1 and the processed
3516 * strip count to zero (upper 8 bits)
3518 bi
->bi_phys_segments
= 1; /* biased count of active stripes */
3526 * The "raid5_align_endio" should check if the read succeeded and if it
3527 * did, call bio_endio on the original bio (having bio_put the new bio
3529 * If the read failed..
3531 static void raid5_align_endio(struct bio
*bi
, int error
)
3533 struct bio
* raid_bi
= bi
->bi_private
;
3536 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3541 rdev
= (void*)raid_bi
->bi_next
;
3542 raid_bi
->bi_next
= NULL
;
3543 mddev
= rdev
->mddev
;
3544 conf
= mddev
->private;
3546 rdev_dec_pending(rdev
, conf
->mddev
);
3548 if (!error
&& uptodate
) {
3549 bio_endio(raid_bi
, 0);
3550 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
3551 wake_up(&conf
->wait_for_stripe
);
3556 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3558 add_bio_to_retry(raid_bi
, conf
);
3561 static int bio_fits_rdev(struct bio
*bi
)
3563 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
3565 if ((bi
->bi_size
>>9) > queue_max_sectors(q
))
3567 blk_recount_segments(q
, bi
);
3568 if (bi
->bi_phys_segments
> queue_max_segments(q
))
3571 if (q
->merge_bvec_fn
)
3572 /* it's too hard to apply the merge_bvec_fn at this stage,
3581 static int chunk_aligned_read(mddev_t
*mddev
, struct bio
* raid_bio
)
3583 raid5_conf_t
*conf
= mddev
->private;
3585 struct bio
* align_bi
;
3588 if (!in_chunk_boundary(mddev
, raid_bio
)) {
3589 pr_debug("chunk_aligned_read : non aligned\n");
3593 * use bio_clone_mddev to make a copy of the bio
3595 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
3599 * set bi_end_io to a new function, and set bi_private to the
3602 align_bi
->bi_end_io
= raid5_align_endio
;
3603 align_bi
->bi_private
= raid_bio
;
3607 align_bi
->bi_sector
= raid5_compute_sector(conf
, raid_bio
->bi_sector
,
3612 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
3613 if (rdev
&& test_bit(In_sync
, &rdev
->flags
)) {
3617 atomic_inc(&rdev
->nr_pending
);
3619 raid_bio
->bi_next
= (void*)rdev
;
3620 align_bi
->bi_bdev
= rdev
->bdev
;
3621 align_bi
->bi_flags
&= ~(1 << BIO_SEG_VALID
);
3622 align_bi
->bi_sector
+= rdev
->data_offset
;
3624 if (!bio_fits_rdev(align_bi
) ||
3625 is_badblock(rdev
, align_bi
->bi_sector
, align_bi
->bi_size
>>9,
3626 &first_bad
, &bad_sectors
)) {
3627 /* too big in some way, or has a known bad block */
3629 rdev_dec_pending(rdev
, mddev
);
3633 spin_lock_irq(&conf
->device_lock
);
3634 wait_event_lock_irq(conf
->wait_for_stripe
,
3636 conf
->device_lock
, /* nothing */);
3637 atomic_inc(&conf
->active_aligned_reads
);
3638 spin_unlock_irq(&conf
->device_lock
);
3640 generic_make_request(align_bi
);
3649 /* __get_priority_stripe - get the next stripe to process
3651 * Full stripe writes are allowed to pass preread active stripes up until
3652 * the bypass_threshold is exceeded. In general the bypass_count
3653 * increments when the handle_list is handled before the hold_list; however, it
3654 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3655 * stripe with in flight i/o. The bypass_count will be reset when the
3656 * head of the hold_list has changed, i.e. the head was promoted to the
3659 static struct stripe_head
*__get_priority_stripe(raid5_conf_t
*conf
)
3661 struct stripe_head
*sh
;
3663 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3665 list_empty(&conf
->handle_list
) ? "empty" : "busy",
3666 list_empty(&conf
->hold_list
) ? "empty" : "busy",
3667 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
3669 if (!list_empty(&conf
->handle_list
)) {
3670 sh
= list_entry(conf
->handle_list
.next
, typeof(*sh
), lru
);
3672 if (list_empty(&conf
->hold_list
))
3673 conf
->bypass_count
= 0;
3674 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
3675 if (conf
->hold_list
.next
== conf
->last_hold
)
3676 conf
->bypass_count
++;
3678 conf
->last_hold
= conf
->hold_list
.next
;
3679 conf
->bypass_count
-= conf
->bypass_threshold
;
3680 if (conf
->bypass_count
< 0)
3681 conf
->bypass_count
= 0;
3684 } else if (!list_empty(&conf
->hold_list
) &&
3685 ((conf
->bypass_threshold
&&
3686 conf
->bypass_count
> conf
->bypass_threshold
) ||
3687 atomic_read(&conf
->pending_full_writes
) == 0)) {
3688 sh
= list_entry(conf
->hold_list
.next
,
3690 conf
->bypass_count
-= conf
->bypass_threshold
;
3691 if (conf
->bypass_count
< 0)
3692 conf
->bypass_count
= 0;
3696 list_del_init(&sh
->lru
);
3697 atomic_inc(&sh
->count
);
3698 BUG_ON(atomic_read(&sh
->count
) != 1);
3702 static int make_request(mddev_t
*mddev
, struct bio
* bi
)
3704 raid5_conf_t
*conf
= mddev
->private;
3706 sector_t new_sector
;
3707 sector_t logical_sector
, last_sector
;
3708 struct stripe_head
*sh
;
3709 const int rw
= bio_data_dir(bi
);
3713 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
3714 md_flush_request(mddev
, bi
);
3718 md_write_start(mddev
, bi
);
3721 mddev
->reshape_position
== MaxSector
&&
3722 chunk_aligned_read(mddev
,bi
))
3725 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
3726 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
3728 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
3730 plugged
= mddev_check_plugged(mddev
);
3731 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
3733 int disks
, data_disks
;
3738 disks
= conf
->raid_disks
;
3739 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
3740 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
3741 /* spinlock is needed as reshape_progress may be
3742 * 64bit on a 32bit platform, and so it might be
3743 * possible to see a half-updated value
3744 * Of course reshape_progress could change after
3745 * the lock is dropped, so once we get a reference
3746 * to the stripe that we think it is, we will have
3749 spin_lock_irq(&conf
->device_lock
);
3750 if (mddev
->delta_disks
< 0
3751 ? logical_sector
< conf
->reshape_progress
3752 : logical_sector
>= conf
->reshape_progress
) {
3753 disks
= conf
->previous_raid_disks
;
3756 if (mddev
->delta_disks
< 0
3757 ? logical_sector
< conf
->reshape_safe
3758 : logical_sector
>= conf
->reshape_safe
) {
3759 spin_unlock_irq(&conf
->device_lock
);
3764 spin_unlock_irq(&conf
->device_lock
);
3766 data_disks
= disks
- conf
->max_degraded
;
3768 new_sector
= raid5_compute_sector(conf
, logical_sector
,
3771 pr_debug("raid456: make_request, sector %llu logical %llu\n",
3772 (unsigned long long)new_sector
,
3773 (unsigned long long)logical_sector
);
3775 sh
= get_active_stripe(conf
, new_sector
, previous
,
3776 (bi
->bi_rw
&RWA_MASK
), 0);
3778 if (unlikely(previous
)) {
3779 /* expansion might have moved on while waiting for a
3780 * stripe, so we must do the range check again.
3781 * Expansion could still move past after this
3782 * test, but as we are holding a reference to
3783 * 'sh', we know that if that happens,
3784 * STRIPE_EXPANDING will get set and the expansion
3785 * won't proceed until we finish with the stripe.
3788 spin_lock_irq(&conf
->device_lock
);
3789 if (mddev
->delta_disks
< 0
3790 ? logical_sector
>= conf
->reshape_progress
3791 : logical_sector
< conf
->reshape_progress
)
3792 /* mismatch, need to try again */
3794 spin_unlock_irq(&conf
->device_lock
);
3803 logical_sector
>= mddev
->suspend_lo
&&
3804 logical_sector
< mddev
->suspend_hi
) {
3806 /* As the suspend_* range is controlled by
3807 * userspace, we want an interruptible
3810 flush_signals(current
);
3811 prepare_to_wait(&conf
->wait_for_overlap
,
3812 &w
, TASK_INTERRUPTIBLE
);
3813 if (logical_sector
>= mddev
->suspend_lo
&&
3814 logical_sector
< mddev
->suspend_hi
)
3819 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
3820 !add_stripe_bio(sh
, bi
, dd_idx
, rw
)) {
3821 /* Stripe is busy expanding or
3822 * add failed due to overlap. Flush everything
3825 md_wakeup_thread(mddev
->thread
);
3830 finish_wait(&conf
->wait_for_overlap
, &w
);
3831 set_bit(STRIPE_HANDLE
, &sh
->state
);
3832 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3833 if ((bi
->bi_rw
& REQ_SYNC
) &&
3834 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3835 atomic_inc(&conf
->preread_active_stripes
);
3838 /* cannot get stripe for read-ahead, just give-up */
3839 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3840 finish_wait(&conf
->wait_for_overlap
, &w
);
3846 md_wakeup_thread(mddev
->thread
);
3848 spin_lock_irq(&conf
->device_lock
);
3849 remaining
= raid5_dec_bi_phys_segments(bi
);
3850 spin_unlock_irq(&conf
->device_lock
);
3851 if (remaining
== 0) {
3854 md_write_end(mddev
);
3862 static sector_t
raid5_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
);
3864 static sector_t
reshape_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
)
3866 /* reshaping is quite different to recovery/resync so it is
3867 * handled quite separately ... here.
3869 * On each call to sync_request, we gather one chunk worth of
3870 * destination stripes and flag them as expanding.
3871 * Then we find all the source stripes and request reads.
3872 * As the reads complete, handle_stripe will copy the data
3873 * into the destination stripe and release that stripe.
3875 raid5_conf_t
*conf
= mddev
->private;
3876 struct stripe_head
*sh
;
3877 sector_t first_sector
, last_sector
;
3878 int raid_disks
= conf
->previous_raid_disks
;
3879 int data_disks
= raid_disks
- conf
->max_degraded
;
3880 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
3883 sector_t writepos
, readpos
, safepos
;
3884 sector_t stripe_addr
;
3885 int reshape_sectors
;
3886 struct list_head stripes
;
3888 if (sector_nr
== 0) {
3889 /* If restarting in the middle, skip the initial sectors */
3890 if (mddev
->delta_disks
< 0 &&
3891 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
3892 sector_nr
= raid5_size(mddev
, 0, 0)
3893 - conf
->reshape_progress
;
3894 } else if (mddev
->delta_disks
>= 0 &&
3895 conf
->reshape_progress
> 0)
3896 sector_nr
= conf
->reshape_progress
;
3897 sector_div(sector_nr
, new_data_disks
);
3899 mddev
->curr_resync_completed
= sector_nr
;
3900 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
3906 /* We need to process a full chunk at a time.
3907 * If old and new chunk sizes differ, we need to process the
3910 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
)
3911 reshape_sectors
= mddev
->new_chunk_sectors
;
3913 reshape_sectors
= mddev
->chunk_sectors
;
3915 /* we update the metadata when there is more than 3Meg
3916 * in the block range (that is rather arbitrary, should
3917 * probably be time based) or when the data about to be
3918 * copied would over-write the source of the data at
3919 * the front of the range.
3920 * i.e. one new_stripe along from reshape_progress new_maps
3921 * to after where reshape_safe old_maps to
3923 writepos
= conf
->reshape_progress
;
3924 sector_div(writepos
, new_data_disks
);
3925 readpos
= conf
->reshape_progress
;
3926 sector_div(readpos
, data_disks
);
3927 safepos
= conf
->reshape_safe
;
3928 sector_div(safepos
, data_disks
);
3929 if (mddev
->delta_disks
< 0) {
3930 writepos
-= min_t(sector_t
, reshape_sectors
, writepos
);
3931 readpos
+= reshape_sectors
;
3932 safepos
+= reshape_sectors
;
3934 writepos
+= reshape_sectors
;
3935 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
3936 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
3939 /* 'writepos' is the most advanced device address we might write.
3940 * 'readpos' is the least advanced device address we might read.
3941 * 'safepos' is the least address recorded in the metadata as having
3943 * If 'readpos' is behind 'writepos', then there is no way that we can
3944 * ensure safety in the face of a crash - that must be done by userspace
3945 * making a backup of the data. So in that case there is no particular
3946 * rush to update metadata.
3947 * Otherwise if 'safepos' is behind 'writepos', then we really need to
3948 * update the metadata to advance 'safepos' to match 'readpos' so that
3949 * we can be safe in the event of a crash.
3950 * So we insist on updating metadata if safepos is behind writepos and
3951 * readpos is beyond writepos.
3952 * In any case, update the metadata every 10 seconds.
3953 * Maybe that number should be configurable, but I'm not sure it is
3954 * worth it.... maybe it could be a multiple of safemode_delay???
3956 if ((mddev
->delta_disks
< 0
3957 ? (safepos
> writepos
&& readpos
< writepos
)
3958 : (safepos
< writepos
&& readpos
> writepos
)) ||
3959 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
3960 /* Cannot proceed until we've updated the superblock... */
3961 wait_event(conf
->wait_for_overlap
,
3962 atomic_read(&conf
->reshape_stripes
)==0);
3963 mddev
->reshape_position
= conf
->reshape_progress
;
3964 mddev
->curr_resync_completed
= sector_nr
;
3965 conf
->reshape_checkpoint
= jiffies
;
3966 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
3967 md_wakeup_thread(mddev
->thread
);
3968 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
3969 kthread_should_stop());
3970 spin_lock_irq(&conf
->device_lock
);
3971 conf
->reshape_safe
= mddev
->reshape_position
;
3972 spin_unlock_irq(&conf
->device_lock
);
3973 wake_up(&conf
->wait_for_overlap
);
3974 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
3977 if (mddev
->delta_disks
< 0) {
3978 BUG_ON(conf
->reshape_progress
== 0);
3979 stripe_addr
= writepos
;
3980 BUG_ON((mddev
->dev_sectors
&
3981 ~((sector_t
)reshape_sectors
- 1))
3982 - reshape_sectors
- stripe_addr
3985 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
3986 stripe_addr
= sector_nr
;
3988 INIT_LIST_HEAD(&stripes
);
3989 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
3991 int skipped_disk
= 0;
3992 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
3993 set_bit(STRIPE_EXPANDING
, &sh
->state
);
3994 atomic_inc(&conf
->reshape_stripes
);
3995 /* If any of this stripe is beyond the end of the old
3996 * array, then we need to zero those blocks
3998 for (j
=sh
->disks
; j
--;) {
4000 if (j
== sh
->pd_idx
)
4002 if (conf
->level
== 6 &&
4005 s
= compute_blocknr(sh
, j
, 0);
4006 if (s
< raid5_size(mddev
, 0, 0)) {
4010 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
4011 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
4012 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
4014 if (!skipped_disk
) {
4015 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4016 set_bit(STRIPE_HANDLE
, &sh
->state
);
4018 list_add(&sh
->lru
, &stripes
);
4020 spin_lock_irq(&conf
->device_lock
);
4021 if (mddev
->delta_disks
< 0)
4022 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
4024 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
4025 spin_unlock_irq(&conf
->device_lock
);
4026 /* Ok, those stripe are ready. We can start scheduling
4027 * reads on the source stripes.
4028 * The source stripes are determined by mapping the first and last
4029 * block on the destination stripes.
4032 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
4035 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
4036 * new_data_disks
- 1),
4038 if (last_sector
>= mddev
->dev_sectors
)
4039 last_sector
= mddev
->dev_sectors
- 1;
4040 while (first_sector
<= last_sector
) {
4041 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
4042 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
4043 set_bit(STRIPE_HANDLE
, &sh
->state
);
4045 first_sector
+= STRIPE_SECTORS
;
4047 /* Now that the sources are clearly marked, we can release
4048 * the destination stripes
4050 while (!list_empty(&stripes
)) {
4051 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
4052 list_del_init(&sh
->lru
);
4055 /* If this takes us to the resync_max point where we have to pause,
4056 * then we need to write out the superblock.
4058 sector_nr
+= reshape_sectors
;
4059 if ((sector_nr
- mddev
->curr_resync_completed
) * 2
4060 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
4061 /* Cannot proceed until we've updated the superblock... */
4062 wait_event(conf
->wait_for_overlap
,
4063 atomic_read(&conf
->reshape_stripes
) == 0);
4064 mddev
->reshape_position
= conf
->reshape_progress
;
4065 mddev
->curr_resync_completed
= sector_nr
;
4066 conf
->reshape_checkpoint
= jiffies
;
4067 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4068 md_wakeup_thread(mddev
->thread
);
4069 wait_event(mddev
->sb_wait
,
4070 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
4071 || kthread_should_stop());
4072 spin_lock_irq(&conf
->device_lock
);
4073 conf
->reshape_safe
= mddev
->reshape_position
;
4074 spin_unlock_irq(&conf
->device_lock
);
4075 wake_up(&conf
->wait_for_overlap
);
4076 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4078 return reshape_sectors
;
4081 /* FIXME go_faster isn't used */
4082 static inline sector_t
sync_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
4084 raid5_conf_t
*conf
= mddev
->private;
4085 struct stripe_head
*sh
;
4086 sector_t max_sector
= mddev
->dev_sectors
;
4087 sector_t sync_blocks
;
4088 int still_degraded
= 0;
4091 if (sector_nr
>= max_sector
) {
4092 /* just being told to finish up .. nothing much to do */
4094 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
4099 if (mddev
->curr_resync
< max_sector
) /* aborted */
4100 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
4102 else /* completed sync */
4104 bitmap_close_sync(mddev
->bitmap
);
4109 /* Allow raid5_quiesce to complete */
4110 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
4112 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
4113 return reshape_request(mddev
, sector_nr
, skipped
);
4115 /* No need to check resync_max as we never do more than one
4116 * stripe, and as resync_max will always be on a chunk boundary,
4117 * if the check in md_do_sync didn't fire, there is no chance
4118 * of overstepping resync_max here
4121 /* if there is too many failed drives and we are trying
4122 * to resync, then assert that we are finished, because there is
4123 * nothing we can do.
4125 if (mddev
->degraded
>= conf
->max_degraded
&&
4126 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
4127 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
4131 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
4132 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
4133 !conf
->fullsync
&& sync_blocks
>= STRIPE_SECTORS
) {
4134 /* we can skip this block, and probably more */
4135 sync_blocks
/= STRIPE_SECTORS
;
4137 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
4141 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
4143 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
4145 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
4146 /* make sure we don't swamp the stripe cache if someone else
4147 * is trying to get access
4149 schedule_timeout_uninterruptible(1);
4151 /* Need to check if array will still be degraded after recovery/resync
4152 * We don't need to check the 'failed' flag as when that gets set,
4155 for (i
= 0; i
< conf
->raid_disks
; i
++)
4156 if (conf
->disks
[i
].rdev
== NULL
)
4159 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
4161 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
4166 return STRIPE_SECTORS
;
4169 static int retry_aligned_read(raid5_conf_t
*conf
, struct bio
*raid_bio
)
4171 /* We may not be able to submit a whole bio at once as there
4172 * may not be enough stripe_heads available.
4173 * We cannot pre-allocate enough stripe_heads as we may need
4174 * more than exist in the cache (if we allow ever large chunks).
4175 * So we do one stripe head at a time and record in
4176 * ->bi_hw_segments how many have been done.
4178 * We *know* that this entire raid_bio is in one chunk, so
4179 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4181 struct stripe_head
*sh
;
4183 sector_t sector
, logical_sector
, last_sector
;
4188 logical_sector
= raid_bio
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4189 sector
= raid5_compute_sector(conf
, logical_sector
,
4191 last_sector
= raid_bio
->bi_sector
+ (raid_bio
->bi_size
>>9);
4193 for (; logical_sector
< last_sector
;
4194 logical_sector
+= STRIPE_SECTORS
,
4195 sector
+= STRIPE_SECTORS
,
4198 if (scnt
< raid5_bi_hw_segments(raid_bio
))
4199 /* already done this stripe */
4202 sh
= get_active_stripe(conf
, sector
, 0, 1, 0);
4205 /* failed to get a stripe - must wait */
4206 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4207 conf
->retry_read_aligned
= raid_bio
;
4211 set_bit(R5_ReadError
, &sh
->dev
[dd_idx
].flags
);
4212 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0)) {
4214 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4215 conf
->retry_read_aligned
= raid_bio
;
4223 spin_lock_irq(&conf
->device_lock
);
4224 remaining
= raid5_dec_bi_phys_segments(raid_bio
);
4225 spin_unlock_irq(&conf
->device_lock
);
4227 bio_endio(raid_bio
, 0);
4228 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4229 wake_up(&conf
->wait_for_stripe
);
4235 * This is our raid5 kernel thread.
4237 * We scan the hash table for stripes which can be handled now.
4238 * During the scan, completed stripes are saved for us by the interrupt
4239 * handler, so that they will not have to wait for our next wakeup.
4241 static void raid5d(mddev_t
*mddev
)
4243 struct stripe_head
*sh
;
4244 raid5_conf_t
*conf
= mddev
->private;
4246 struct blk_plug plug
;
4248 pr_debug("+++ raid5d active\n");
4250 md_check_recovery(mddev
);
4252 blk_start_plug(&plug
);
4254 spin_lock_irq(&conf
->device_lock
);
4258 if (atomic_read(&mddev
->plug_cnt
) == 0 &&
4259 !list_empty(&conf
->bitmap_list
)) {
4260 /* Now is a good time to flush some bitmap updates */
4262 spin_unlock_irq(&conf
->device_lock
);
4263 bitmap_unplug(mddev
->bitmap
);
4264 spin_lock_irq(&conf
->device_lock
);
4265 conf
->seq_write
= conf
->seq_flush
;
4266 activate_bit_delay(conf
);
4268 if (atomic_read(&mddev
->plug_cnt
) == 0)
4269 raid5_activate_delayed(conf
);
4271 while ((bio
= remove_bio_from_retry(conf
))) {
4273 spin_unlock_irq(&conf
->device_lock
);
4274 ok
= retry_aligned_read(conf
, bio
);
4275 spin_lock_irq(&conf
->device_lock
);
4281 sh
= __get_priority_stripe(conf
);
4285 spin_unlock_irq(&conf
->device_lock
);
4292 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
4293 md_check_recovery(mddev
);
4295 spin_lock_irq(&conf
->device_lock
);
4297 pr_debug("%d stripes handled\n", handled
);
4299 spin_unlock_irq(&conf
->device_lock
);
4301 async_tx_issue_pending_all();
4302 blk_finish_plug(&plug
);
4304 pr_debug("--- raid5d inactive\n");
4308 raid5_show_stripe_cache_size(mddev_t
*mddev
, char *page
)
4310 raid5_conf_t
*conf
= mddev
->private;
4312 return sprintf(page
, "%d\n", conf
->max_nr_stripes
);
4318 raid5_set_cache_size(mddev_t
*mddev
, int size
)
4320 raid5_conf_t
*conf
= mddev
->private;
4323 if (size
<= 16 || size
> 32768)
4325 while (size
< conf
->max_nr_stripes
) {
4326 if (drop_one_stripe(conf
))
4327 conf
->max_nr_stripes
--;
4331 err
= md_allow_write(mddev
);
4334 while (size
> conf
->max_nr_stripes
) {
4335 if (grow_one_stripe(conf
))
4336 conf
->max_nr_stripes
++;
4341 EXPORT_SYMBOL(raid5_set_cache_size
);
4344 raid5_store_stripe_cache_size(mddev_t
*mddev
, const char *page
, size_t len
)
4346 raid5_conf_t
*conf
= mddev
->private;
4350 if (len
>= PAGE_SIZE
)
4355 if (strict_strtoul(page
, 10, &new))
4357 err
= raid5_set_cache_size(mddev
, new);
4363 static struct md_sysfs_entry
4364 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
4365 raid5_show_stripe_cache_size
,
4366 raid5_store_stripe_cache_size
);
4369 raid5_show_preread_threshold(mddev_t
*mddev
, char *page
)
4371 raid5_conf_t
*conf
= mddev
->private;
4373 return sprintf(page
, "%d\n", conf
->bypass_threshold
);
4379 raid5_store_preread_threshold(mddev_t
*mddev
, const char *page
, size_t len
)
4381 raid5_conf_t
*conf
= mddev
->private;
4383 if (len
>= PAGE_SIZE
)
4388 if (strict_strtoul(page
, 10, &new))
4390 if (new > conf
->max_nr_stripes
)
4392 conf
->bypass_threshold
= new;
4396 static struct md_sysfs_entry
4397 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
4399 raid5_show_preread_threshold
,
4400 raid5_store_preread_threshold
);
4403 stripe_cache_active_show(mddev_t
*mddev
, char *page
)
4405 raid5_conf_t
*conf
= mddev
->private;
4407 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
4412 static struct md_sysfs_entry
4413 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
4415 static struct attribute
*raid5_attrs
[] = {
4416 &raid5_stripecache_size
.attr
,
4417 &raid5_stripecache_active
.attr
,
4418 &raid5_preread_bypass_threshold
.attr
,
4421 static struct attribute_group raid5_attrs_group
= {
4423 .attrs
= raid5_attrs
,
4427 raid5_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
)
4429 raid5_conf_t
*conf
= mddev
->private;
4432 sectors
= mddev
->dev_sectors
;
4434 /* size is defined by the smallest of previous and new size */
4435 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
4437 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
4438 sectors
&= ~((sector_t
)mddev
->new_chunk_sectors
- 1);
4439 return sectors
* (raid_disks
- conf
->max_degraded
);
4442 static void raid5_free_percpu(raid5_conf_t
*conf
)
4444 struct raid5_percpu
*percpu
;
4451 for_each_possible_cpu(cpu
) {
4452 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4453 safe_put_page(percpu
->spare_page
);
4454 kfree(percpu
->scribble
);
4456 #ifdef CONFIG_HOTPLUG_CPU
4457 unregister_cpu_notifier(&conf
->cpu_notify
);
4461 free_percpu(conf
->percpu
);
4464 static void free_conf(raid5_conf_t
*conf
)
4466 shrink_stripes(conf
);
4467 raid5_free_percpu(conf
);
4469 kfree(conf
->stripe_hashtbl
);
4473 #ifdef CONFIG_HOTPLUG_CPU
4474 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
4477 raid5_conf_t
*conf
= container_of(nfb
, raid5_conf_t
, cpu_notify
);
4478 long cpu
= (long)hcpu
;
4479 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4482 case CPU_UP_PREPARE
:
4483 case CPU_UP_PREPARE_FROZEN
:
4484 if (conf
->level
== 6 && !percpu
->spare_page
)
4485 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
4486 if (!percpu
->scribble
)
4487 percpu
->scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
4489 if (!percpu
->scribble
||
4490 (conf
->level
== 6 && !percpu
->spare_page
)) {
4491 safe_put_page(percpu
->spare_page
);
4492 kfree(percpu
->scribble
);
4493 pr_err("%s: failed memory allocation for cpu%ld\n",
4495 return notifier_from_errno(-ENOMEM
);
4499 case CPU_DEAD_FROZEN
:
4500 safe_put_page(percpu
->spare_page
);
4501 kfree(percpu
->scribble
);
4502 percpu
->spare_page
= NULL
;
4503 percpu
->scribble
= NULL
;
4512 static int raid5_alloc_percpu(raid5_conf_t
*conf
)
4515 struct page
*spare_page
;
4516 struct raid5_percpu __percpu
*allcpus
;
4520 allcpus
= alloc_percpu(struct raid5_percpu
);
4523 conf
->percpu
= allcpus
;
4527 for_each_present_cpu(cpu
) {
4528 if (conf
->level
== 6) {
4529 spare_page
= alloc_page(GFP_KERNEL
);
4534 per_cpu_ptr(conf
->percpu
, cpu
)->spare_page
= spare_page
;
4536 scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
4541 per_cpu_ptr(conf
->percpu
, cpu
)->scribble
= scribble
;
4543 #ifdef CONFIG_HOTPLUG_CPU
4544 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
4545 conf
->cpu_notify
.priority
= 0;
4547 err
= register_cpu_notifier(&conf
->cpu_notify
);
4554 static raid5_conf_t
*setup_conf(mddev_t
*mddev
)
4557 int raid_disk
, memory
, max_disks
;
4559 struct disk_info
*disk
;
4561 if (mddev
->new_level
!= 5
4562 && mddev
->new_level
!= 4
4563 && mddev
->new_level
!= 6) {
4564 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4565 mdname(mddev
), mddev
->new_level
);
4566 return ERR_PTR(-EIO
);
4568 if ((mddev
->new_level
== 5
4569 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
4570 (mddev
->new_level
== 6
4571 && !algorithm_valid_raid6(mddev
->new_layout
))) {
4572 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
4573 mdname(mddev
), mddev
->new_layout
);
4574 return ERR_PTR(-EIO
);
4576 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
4577 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4578 mdname(mddev
), mddev
->raid_disks
);
4579 return ERR_PTR(-EINVAL
);
4582 if (!mddev
->new_chunk_sectors
||
4583 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
4584 !is_power_of_2(mddev
->new_chunk_sectors
)) {
4585 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
4586 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
4587 return ERR_PTR(-EINVAL
);
4590 conf
= kzalloc(sizeof(raid5_conf_t
), GFP_KERNEL
);
4593 spin_lock_init(&conf
->device_lock
);
4594 init_waitqueue_head(&conf
->wait_for_stripe
);
4595 init_waitqueue_head(&conf
->wait_for_overlap
);
4596 INIT_LIST_HEAD(&conf
->handle_list
);
4597 INIT_LIST_HEAD(&conf
->hold_list
);
4598 INIT_LIST_HEAD(&conf
->delayed_list
);
4599 INIT_LIST_HEAD(&conf
->bitmap_list
);
4600 INIT_LIST_HEAD(&conf
->inactive_list
);
4601 atomic_set(&conf
->active_stripes
, 0);
4602 atomic_set(&conf
->preread_active_stripes
, 0);
4603 atomic_set(&conf
->active_aligned_reads
, 0);
4604 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
4606 conf
->raid_disks
= mddev
->raid_disks
;
4607 if (mddev
->reshape_position
== MaxSector
)
4608 conf
->previous_raid_disks
= mddev
->raid_disks
;
4610 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4611 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
4612 conf
->scribble_len
= scribble_len(max_disks
);
4614 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
4619 conf
->mddev
= mddev
;
4621 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
4624 conf
->level
= mddev
->new_level
;
4625 if (raid5_alloc_percpu(conf
) != 0)
4628 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
4630 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
4631 raid_disk
= rdev
->raid_disk
;
4632 if (raid_disk
>= max_disks
4635 disk
= conf
->disks
+ raid_disk
;
4639 if (test_bit(In_sync
, &rdev
->flags
)) {
4640 char b
[BDEVNAME_SIZE
];
4641 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
4643 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
4644 } else if (rdev
->saved_raid_disk
!= raid_disk
)
4645 /* Cannot rely on bitmap to complete recovery */
4649 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
4650 conf
->level
= mddev
->new_level
;
4651 if (conf
->level
== 6)
4652 conf
->max_degraded
= 2;
4654 conf
->max_degraded
= 1;
4655 conf
->algorithm
= mddev
->new_layout
;
4656 conf
->max_nr_stripes
= NR_STRIPES
;
4657 conf
->reshape_progress
= mddev
->reshape_position
;
4658 if (conf
->reshape_progress
!= MaxSector
) {
4659 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
4660 conf
->prev_algo
= mddev
->layout
;
4663 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
4664 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
4665 if (grow_stripes(conf
, conf
->max_nr_stripes
)) {
4667 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4668 mdname(mddev
), memory
);
4671 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
4672 mdname(mddev
), memory
);
4674 conf
->thread
= md_register_thread(raid5d
, mddev
, NULL
);
4675 if (!conf
->thread
) {
4677 "md/raid:%s: couldn't allocate thread.\n",
4687 return ERR_PTR(-EIO
);
4689 return ERR_PTR(-ENOMEM
);
4693 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
4696 case ALGORITHM_PARITY_0
:
4697 if (raid_disk
< max_degraded
)
4700 case ALGORITHM_PARITY_N
:
4701 if (raid_disk
>= raid_disks
- max_degraded
)
4704 case ALGORITHM_PARITY_0_6
:
4705 if (raid_disk
== 0 ||
4706 raid_disk
== raid_disks
- 1)
4709 case ALGORITHM_LEFT_ASYMMETRIC_6
:
4710 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
4711 case ALGORITHM_LEFT_SYMMETRIC_6
:
4712 case ALGORITHM_RIGHT_SYMMETRIC_6
:
4713 if (raid_disk
== raid_disks
- 1)
4719 static int run(mddev_t
*mddev
)
4722 int working_disks
= 0;
4723 int dirty_parity_disks
= 0;
4725 sector_t reshape_offset
= 0;
4727 if (mddev
->recovery_cp
!= MaxSector
)
4728 printk(KERN_NOTICE
"md/raid:%s: not clean"
4729 " -- starting background reconstruction\n",
4731 if (mddev
->reshape_position
!= MaxSector
) {
4732 /* Check that we can continue the reshape.
4733 * Currently only disks can change, it must
4734 * increase, and we must be past the point where
4735 * a stripe over-writes itself
4737 sector_t here_new
, here_old
;
4739 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
4741 if (mddev
->new_level
!= mddev
->level
) {
4742 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
4743 "required - aborting.\n",
4747 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4748 /* reshape_position must be on a new-stripe boundary, and one
4749 * further up in new geometry must map after here in old
4752 here_new
= mddev
->reshape_position
;
4753 if (sector_div(here_new
, mddev
->new_chunk_sectors
*
4754 (mddev
->raid_disks
- max_degraded
))) {
4755 printk(KERN_ERR
"md/raid:%s: reshape_position not "
4756 "on a stripe boundary\n", mdname(mddev
));
4759 reshape_offset
= here_new
* mddev
->new_chunk_sectors
;
4760 /* here_new is the stripe we will write to */
4761 here_old
= mddev
->reshape_position
;
4762 sector_div(here_old
, mddev
->chunk_sectors
*
4763 (old_disks
-max_degraded
));
4764 /* here_old is the first stripe that we might need to read
4766 if (mddev
->delta_disks
== 0) {
4767 /* We cannot be sure it is safe to start an in-place
4768 * reshape. It is only safe if user-space if monitoring
4769 * and taking constant backups.
4770 * mdadm always starts a situation like this in
4771 * readonly mode so it can take control before
4772 * allowing any writes. So just check for that.
4774 if ((here_new
* mddev
->new_chunk_sectors
!=
4775 here_old
* mddev
->chunk_sectors
) ||
4777 printk(KERN_ERR
"md/raid:%s: in-place reshape must be started"
4778 " in read-only mode - aborting\n",
4782 } else if (mddev
->delta_disks
< 0
4783 ? (here_new
* mddev
->new_chunk_sectors
<=
4784 here_old
* mddev
->chunk_sectors
)
4785 : (here_new
* mddev
->new_chunk_sectors
>=
4786 here_old
* mddev
->chunk_sectors
)) {
4787 /* Reading from the same stripe as writing to - bad */
4788 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
4789 "auto-recovery - aborting.\n",
4793 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
4795 /* OK, we should be able to continue; */
4797 BUG_ON(mddev
->level
!= mddev
->new_level
);
4798 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
4799 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
4800 BUG_ON(mddev
->delta_disks
!= 0);
4803 if (mddev
->private == NULL
)
4804 conf
= setup_conf(mddev
);
4806 conf
= mddev
->private;
4809 return PTR_ERR(conf
);
4811 mddev
->thread
= conf
->thread
;
4812 conf
->thread
= NULL
;
4813 mddev
->private = conf
;
4816 * 0 for a fully functional array, 1 or 2 for a degraded array.
4818 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
4819 if (rdev
->raid_disk
< 0)
4821 if (test_bit(In_sync
, &rdev
->flags
)) {
4825 /* This disc is not fully in-sync. However if it
4826 * just stored parity (beyond the recovery_offset),
4827 * when we don't need to be concerned about the
4828 * array being dirty.
4829 * When reshape goes 'backwards', we never have
4830 * partially completed devices, so we only need
4831 * to worry about reshape going forwards.
4833 /* Hack because v0.91 doesn't store recovery_offset properly. */
4834 if (mddev
->major_version
== 0 &&
4835 mddev
->minor_version
> 90)
4836 rdev
->recovery_offset
= reshape_offset
;
4838 if (rdev
->recovery_offset
< reshape_offset
) {
4839 /* We need to check old and new layout */
4840 if (!only_parity(rdev
->raid_disk
,
4843 conf
->max_degraded
))
4846 if (!only_parity(rdev
->raid_disk
,
4848 conf
->previous_raid_disks
,
4849 conf
->max_degraded
))
4851 dirty_parity_disks
++;
4854 mddev
->degraded
= (max(conf
->raid_disks
, conf
->previous_raid_disks
)
4857 if (has_failed(conf
)) {
4858 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
4859 " (%d/%d failed)\n",
4860 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
4864 /* device size must be a multiple of chunk size */
4865 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
4866 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
4868 if (mddev
->degraded
> dirty_parity_disks
&&
4869 mddev
->recovery_cp
!= MaxSector
) {
4870 if (mddev
->ok_start_degraded
)
4872 "md/raid:%s: starting dirty degraded array"
4873 " - data corruption possible.\n",
4877 "md/raid:%s: cannot start dirty degraded array.\n",
4883 if (mddev
->degraded
== 0)
4884 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
4885 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
4886 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
4889 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
4890 " out of %d devices, algorithm %d\n",
4891 mdname(mddev
), conf
->level
,
4892 mddev
->raid_disks
- mddev
->degraded
,
4893 mddev
->raid_disks
, mddev
->new_layout
);
4895 print_raid5_conf(conf
);
4897 if (conf
->reshape_progress
!= MaxSector
) {
4898 conf
->reshape_safe
= conf
->reshape_progress
;
4899 atomic_set(&conf
->reshape_stripes
, 0);
4900 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4901 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4902 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4903 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4904 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4909 /* Ok, everything is just fine now */
4910 if (mddev
->to_remove
== &raid5_attrs_group
)
4911 mddev
->to_remove
= NULL
;
4912 else if (mddev
->kobj
.sd
&&
4913 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
4915 "raid5: failed to create sysfs attributes for %s\n",
4917 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
4921 /* read-ahead size must cover two whole stripes, which
4922 * is 2 * (datadisks) * chunksize where 'n' is the
4923 * number of raid devices
4925 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
4926 int stripe
= data_disks
*
4927 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4928 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4929 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4931 blk_queue_merge_bvec(mddev
->queue
, raid5_mergeable_bvec
);
4933 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
4934 mddev
->queue
->backing_dev_info
.congested_fn
= raid5_congested
;
4936 chunk_size
= mddev
->chunk_sectors
<< 9;
4937 blk_queue_io_min(mddev
->queue
, chunk_size
);
4938 blk_queue_io_opt(mddev
->queue
, chunk_size
*
4939 (conf
->raid_disks
- conf
->max_degraded
));
4941 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
4942 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
4943 rdev
->data_offset
<< 9);
4948 md_unregister_thread(&mddev
->thread
);
4950 print_raid5_conf(conf
);
4953 mddev
->private = NULL
;
4954 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
4958 static int stop(mddev_t
*mddev
)
4960 raid5_conf_t
*conf
= mddev
->private;
4962 md_unregister_thread(&mddev
->thread
);
4964 mddev
->queue
->backing_dev_info
.congested_fn
= NULL
;
4966 mddev
->private = NULL
;
4967 mddev
->to_remove
= &raid5_attrs_group
;
4972 static void print_sh(struct seq_file
*seq
, struct stripe_head
*sh
)
4976 seq_printf(seq
, "sh %llu, pd_idx %d, state %ld.\n",
4977 (unsigned long long)sh
->sector
, sh
->pd_idx
, sh
->state
);
4978 seq_printf(seq
, "sh %llu, count %d.\n",
4979 (unsigned long long)sh
->sector
, atomic_read(&sh
->count
));
4980 seq_printf(seq
, "sh %llu, ", (unsigned long long)sh
->sector
);
4981 for (i
= 0; i
< sh
->disks
; i
++) {
4982 seq_printf(seq
, "(cache%d: %p %ld) ",
4983 i
, sh
->dev
[i
].page
, sh
->dev
[i
].flags
);
4985 seq_printf(seq
, "\n");
4988 static void printall(struct seq_file
*seq
, raid5_conf_t
*conf
)
4990 struct stripe_head
*sh
;
4991 struct hlist_node
*hn
;
4994 spin_lock_irq(&conf
->device_lock
);
4995 for (i
= 0; i
< NR_HASH
; i
++) {
4996 hlist_for_each_entry(sh
, hn
, &conf
->stripe_hashtbl
[i
], hash
) {
4997 if (sh
->raid_conf
!= conf
)
5002 spin_unlock_irq(&conf
->device_lock
);
5006 static void status(struct seq_file
*seq
, mddev_t
*mddev
)
5008 raid5_conf_t
*conf
= mddev
->private;
5011 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
5012 mddev
->chunk_sectors
/ 2, mddev
->layout
);
5013 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
5014 for (i
= 0; i
< conf
->raid_disks
; i
++)
5015 seq_printf (seq
, "%s",
5016 conf
->disks
[i
].rdev
&&
5017 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
5018 seq_printf (seq
, "]");
5020 seq_printf (seq
, "\n");
5021 printall(seq
, conf
);
5025 static void print_raid5_conf (raid5_conf_t
*conf
)
5028 struct disk_info
*tmp
;
5030 printk(KERN_DEBUG
"RAID conf printout:\n");
5032 printk("(conf==NULL)\n");
5035 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
5037 conf
->raid_disks
- conf
->mddev
->degraded
);
5039 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5040 char b
[BDEVNAME_SIZE
];
5041 tmp
= conf
->disks
+ i
;
5043 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
5044 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
5045 bdevname(tmp
->rdev
->bdev
, b
));
5049 static int raid5_spare_active(mddev_t
*mddev
)
5052 raid5_conf_t
*conf
= mddev
->private;
5053 struct disk_info
*tmp
;
5055 unsigned long flags
;
5057 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5058 tmp
= conf
->disks
+ i
;
5060 && tmp
->rdev
->recovery_offset
== MaxSector
5061 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
5062 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
5064 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
5067 spin_lock_irqsave(&conf
->device_lock
, flags
);
5068 mddev
->degraded
-= count
;
5069 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5070 print_raid5_conf(conf
);
5074 static int raid5_remove_disk(mddev_t
*mddev
, int number
)
5076 raid5_conf_t
*conf
= mddev
->private;
5079 struct disk_info
*p
= conf
->disks
+ number
;
5081 print_raid5_conf(conf
);
5084 if (number
>= conf
->raid_disks
&&
5085 conf
->reshape_progress
== MaxSector
)
5086 clear_bit(In_sync
, &rdev
->flags
);
5088 if (test_bit(In_sync
, &rdev
->flags
) ||
5089 atomic_read(&rdev
->nr_pending
)) {
5093 /* Only remove non-faulty devices if recovery
5096 if (!test_bit(Faulty
, &rdev
->flags
) &&
5097 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
5098 !has_failed(conf
) &&
5099 number
< conf
->raid_disks
) {
5105 if (atomic_read(&rdev
->nr_pending
)) {
5106 /* lost the race, try later */
5113 print_raid5_conf(conf
);
5117 static int raid5_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
5119 raid5_conf_t
*conf
= mddev
->private;
5122 struct disk_info
*p
;
5124 int last
= conf
->raid_disks
- 1;
5126 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
5129 if (has_failed(conf
))
5130 /* no point adding a device */
5133 if (rdev
->raid_disk
>= 0)
5134 first
= last
= rdev
->raid_disk
;
5137 * find the disk ... but prefer rdev->saved_raid_disk
5140 if (rdev
->saved_raid_disk
>= 0 &&
5141 rdev
->saved_raid_disk
>= first
&&
5142 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
5143 disk
= rdev
->saved_raid_disk
;
5146 for ( ; disk
<= last
; disk
++)
5147 if ((p
=conf
->disks
+ disk
)->rdev
== NULL
) {
5148 clear_bit(In_sync
, &rdev
->flags
);
5149 rdev
->raid_disk
= disk
;
5151 if (rdev
->saved_raid_disk
!= disk
)
5153 rcu_assign_pointer(p
->rdev
, rdev
);
5156 print_raid5_conf(conf
);
5160 static int raid5_resize(mddev_t
*mddev
, sector_t sectors
)
5162 /* no resync is happening, and there is enough space
5163 * on all devices, so we can resize.
5164 * We need to make sure resync covers any new space.
5165 * If the array is shrinking we should possibly wait until
5166 * any io in the removed space completes, but it hardly seems
5169 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
5170 md_set_array_sectors(mddev
, raid5_size(mddev
, sectors
,
5171 mddev
->raid_disks
));
5172 if (mddev
->array_sectors
>
5173 raid5_size(mddev
, sectors
, mddev
->raid_disks
))
5175 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5176 revalidate_disk(mddev
->gendisk
);
5177 if (sectors
> mddev
->dev_sectors
&&
5178 mddev
->recovery_cp
> mddev
->dev_sectors
) {
5179 mddev
->recovery_cp
= mddev
->dev_sectors
;
5180 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
5182 mddev
->dev_sectors
= sectors
;
5183 mddev
->resync_max_sectors
= sectors
;
5187 static int check_stripe_cache(mddev_t
*mddev
)
5189 /* Can only proceed if there are plenty of stripe_heads.
5190 * We need a minimum of one full stripe,, and for sensible progress
5191 * it is best to have about 4 times that.
5192 * If we require 4 times, then the default 256 4K stripe_heads will
5193 * allow for chunk sizes up to 256K, which is probably OK.
5194 * If the chunk size is greater, user-space should request more
5195 * stripe_heads first.
5197 raid5_conf_t
*conf
= mddev
->private;
5198 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5199 > conf
->max_nr_stripes
||
5200 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5201 > conf
->max_nr_stripes
) {
5202 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5204 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
5211 static int check_reshape(mddev_t
*mddev
)
5213 raid5_conf_t
*conf
= mddev
->private;
5215 if (mddev
->delta_disks
== 0 &&
5216 mddev
->new_layout
== mddev
->layout
&&
5217 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
5218 return 0; /* nothing to do */
5220 /* Cannot grow a bitmap yet */
5222 if (has_failed(conf
))
5224 if (mddev
->delta_disks
< 0) {
5225 /* We might be able to shrink, but the devices must
5226 * be made bigger first.
5227 * For raid6, 4 is the minimum size.
5228 * Otherwise 2 is the minimum
5231 if (mddev
->level
== 6)
5233 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
5237 if (!check_stripe_cache(mddev
))
5240 return resize_stripes(conf
, conf
->raid_disks
+ mddev
->delta_disks
);
5243 static int raid5_start_reshape(mddev_t
*mddev
)
5245 raid5_conf_t
*conf
= mddev
->private;
5248 unsigned long flags
;
5250 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
5253 if (!check_stripe_cache(mddev
))
5256 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
5257 if (!test_bit(In_sync
, &rdev
->flags
)
5258 && !test_bit(Faulty
, &rdev
->flags
))
5261 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
5262 /* Not enough devices even to make a degraded array
5267 /* Refuse to reduce size of the array. Any reductions in
5268 * array size must be through explicit setting of array_size
5271 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
5272 < mddev
->array_sectors
) {
5273 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
5274 "before number of disks\n", mdname(mddev
));
5278 atomic_set(&conf
->reshape_stripes
, 0);
5279 spin_lock_irq(&conf
->device_lock
);
5280 conf
->previous_raid_disks
= conf
->raid_disks
;
5281 conf
->raid_disks
+= mddev
->delta_disks
;
5282 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
5283 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
5284 conf
->prev_algo
= conf
->algorithm
;
5285 conf
->algorithm
= mddev
->new_layout
;
5286 if (mddev
->delta_disks
< 0)
5287 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
5289 conf
->reshape_progress
= 0;
5290 conf
->reshape_safe
= conf
->reshape_progress
;
5292 spin_unlock_irq(&conf
->device_lock
);
5294 /* Add some new drives, as many as will fit.
5295 * We know there are enough to make the newly sized array work.
5296 * Don't add devices if we are reducing the number of
5297 * devices in the array. This is because it is not possible
5298 * to correctly record the "partially reconstructed" state of
5299 * such devices during the reshape and confusion could result.
5301 if (mddev
->delta_disks
>= 0) {
5302 int added_devices
= 0;
5303 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
5304 if (rdev
->raid_disk
< 0 &&
5305 !test_bit(Faulty
, &rdev
->flags
)) {
5306 if (raid5_add_disk(mddev
, rdev
) == 0) {
5308 >= conf
->previous_raid_disks
) {
5309 set_bit(In_sync
, &rdev
->flags
);
5312 rdev
->recovery_offset
= 0;
5314 if (sysfs_link_rdev(mddev
, rdev
))
5315 /* Failure here is OK */;
5317 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
5318 && !test_bit(Faulty
, &rdev
->flags
)) {
5319 /* This is a spare that was manually added */
5320 set_bit(In_sync
, &rdev
->flags
);
5324 /* When a reshape changes the number of devices,
5325 * ->degraded is measured against the larger of the
5326 * pre and post number of devices.
5328 spin_lock_irqsave(&conf
->device_lock
, flags
);
5329 mddev
->degraded
+= (conf
->raid_disks
- conf
->previous_raid_disks
)
5331 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5333 mddev
->raid_disks
= conf
->raid_disks
;
5334 mddev
->reshape_position
= conf
->reshape_progress
;
5335 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5337 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5338 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5339 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5340 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5341 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5343 if (!mddev
->sync_thread
) {
5344 mddev
->recovery
= 0;
5345 spin_lock_irq(&conf
->device_lock
);
5346 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
5347 conf
->reshape_progress
= MaxSector
;
5348 spin_unlock_irq(&conf
->device_lock
);
5351 conf
->reshape_checkpoint
= jiffies
;
5352 md_wakeup_thread(mddev
->sync_thread
);
5353 md_new_event(mddev
);
5357 /* This is called from the reshape thread and should make any
5358 * changes needed in 'conf'
5360 static void end_reshape(raid5_conf_t
*conf
)
5363 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
5365 spin_lock_irq(&conf
->device_lock
);
5366 conf
->previous_raid_disks
= conf
->raid_disks
;
5367 conf
->reshape_progress
= MaxSector
;
5368 spin_unlock_irq(&conf
->device_lock
);
5369 wake_up(&conf
->wait_for_overlap
);
5371 /* read-ahead size must cover two whole stripes, which is
5372 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5374 if (conf
->mddev
->queue
) {
5375 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5376 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
5378 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
5379 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
5384 /* This is called from the raid5d thread with mddev_lock held.
5385 * It makes config changes to the device.
5387 static void raid5_finish_reshape(mddev_t
*mddev
)
5389 raid5_conf_t
*conf
= mddev
->private;
5391 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
5393 if (mddev
->delta_disks
> 0) {
5394 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5395 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5396 revalidate_disk(mddev
->gendisk
);
5399 mddev
->degraded
= conf
->raid_disks
;
5400 for (d
= 0; d
< conf
->raid_disks
; d
++)
5401 if (conf
->disks
[d
].rdev
&&
5403 &conf
->disks
[d
].rdev
->flags
))
5405 for (d
= conf
->raid_disks
;
5406 d
< conf
->raid_disks
- mddev
->delta_disks
;
5408 mdk_rdev_t
*rdev
= conf
->disks
[d
].rdev
;
5409 if (rdev
&& raid5_remove_disk(mddev
, d
) == 0) {
5410 sysfs_unlink_rdev(mddev
, rdev
);
5411 rdev
->raid_disk
= -1;
5415 mddev
->layout
= conf
->algorithm
;
5416 mddev
->chunk_sectors
= conf
->chunk_sectors
;
5417 mddev
->reshape_position
= MaxSector
;
5418 mddev
->delta_disks
= 0;
5422 static void raid5_quiesce(mddev_t
*mddev
, int state
)
5424 raid5_conf_t
*conf
= mddev
->private;
5427 case 2: /* resume for a suspend */
5428 wake_up(&conf
->wait_for_overlap
);
5431 case 1: /* stop all writes */
5432 spin_lock_irq(&conf
->device_lock
);
5433 /* '2' tells resync/reshape to pause so that all
5434 * active stripes can drain
5437 wait_event_lock_irq(conf
->wait_for_stripe
,
5438 atomic_read(&conf
->active_stripes
) == 0 &&
5439 atomic_read(&conf
->active_aligned_reads
) == 0,
5440 conf
->device_lock
, /* nothing */);
5442 spin_unlock_irq(&conf
->device_lock
);
5443 /* allow reshape to continue */
5444 wake_up(&conf
->wait_for_overlap
);
5447 case 0: /* re-enable writes */
5448 spin_lock_irq(&conf
->device_lock
);
5450 wake_up(&conf
->wait_for_stripe
);
5451 wake_up(&conf
->wait_for_overlap
);
5452 spin_unlock_irq(&conf
->device_lock
);
5458 static void *raid45_takeover_raid0(mddev_t
*mddev
, int level
)
5460 struct raid0_private_data
*raid0_priv
= mddev
->private;
5463 /* for raid0 takeover only one zone is supported */
5464 if (raid0_priv
->nr_strip_zones
> 1) {
5465 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5467 return ERR_PTR(-EINVAL
);
5470 sectors
= raid0_priv
->strip_zone
[0].zone_end
;
5471 sector_div(sectors
, raid0_priv
->strip_zone
[0].nb_dev
);
5472 mddev
->dev_sectors
= sectors
;
5473 mddev
->new_level
= level
;
5474 mddev
->new_layout
= ALGORITHM_PARITY_N
;
5475 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
5476 mddev
->raid_disks
+= 1;
5477 mddev
->delta_disks
= 1;
5478 /* make sure it will be not marked as dirty */
5479 mddev
->recovery_cp
= MaxSector
;
5481 return setup_conf(mddev
);
5485 static void *raid5_takeover_raid1(mddev_t
*mddev
)
5489 if (mddev
->raid_disks
!= 2 ||
5490 mddev
->degraded
> 1)
5491 return ERR_PTR(-EINVAL
);
5493 /* Should check if there are write-behind devices? */
5495 chunksect
= 64*2; /* 64K by default */
5497 /* The array must be an exact multiple of chunksize */
5498 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
5501 if ((chunksect
<<9) < STRIPE_SIZE
)
5502 /* array size does not allow a suitable chunk size */
5503 return ERR_PTR(-EINVAL
);
5505 mddev
->new_level
= 5;
5506 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5507 mddev
->new_chunk_sectors
= chunksect
;
5509 return setup_conf(mddev
);
5512 static void *raid5_takeover_raid6(mddev_t
*mddev
)
5516 switch (mddev
->layout
) {
5517 case ALGORITHM_LEFT_ASYMMETRIC_6
:
5518 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
5520 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
5521 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
5523 case ALGORITHM_LEFT_SYMMETRIC_6
:
5524 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5526 case ALGORITHM_RIGHT_SYMMETRIC_6
:
5527 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
5529 case ALGORITHM_PARITY_0_6
:
5530 new_layout
= ALGORITHM_PARITY_0
;
5532 case ALGORITHM_PARITY_N
:
5533 new_layout
= ALGORITHM_PARITY_N
;
5536 return ERR_PTR(-EINVAL
);
5538 mddev
->new_level
= 5;
5539 mddev
->new_layout
= new_layout
;
5540 mddev
->delta_disks
= -1;
5541 mddev
->raid_disks
-= 1;
5542 return setup_conf(mddev
);
5546 static int raid5_check_reshape(mddev_t
*mddev
)
5548 /* For a 2-drive array, the layout and chunk size can be changed
5549 * immediately as not restriping is needed.
5550 * For larger arrays we record the new value - after validation
5551 * to be used by a reshape pass.
5553 raid5_conf_t
*conf
= mddev
->private;
5554 int new_chunk
= mddev
->new_chunk_sectors
;
5556 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
5558 if (new_chunk
> 0) {
5559 if (!is_power_of_2(new_chunk
))
5561 if (new_chunk
< (PAGE_SIZE
>>9))
5563 if (mddev
->array_sectors
& (new_chunk
-1))
5564 /* not factor of array size */
5568 /* They look valid */
5570 if (mddev
->raid_disks
== 2) {
5571 /* can make the change immediately */
5572 if (mddev
->new_layout
>= 0) {
5573 conf
->algorithm
= mddev
->new_layout
;
5574 mddev
->layout
= mddev
->new_layout
;
5576 if (new_chunk
> 0) {
5577 conf
->chunk_sectors
= new_chunk
;
5578 mddev
->chunk_sectors
= new_chunk
;
5580 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5581 md_wakeup_thread(mddev
->thread
);
5583 return check_reshape(mddev
);
5586 static int raid6_check_reshape(mddev_t
*mddev
)
5588 int new_chunk
= mddev
->new_chunk_sectors
;
5590 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
5592 if (new_chunk
> 0) {
5593 if (!is_power_of_2(new_chunk
))
5595 if (new_chunk
< (PAGE_SIZE
>> 9))
5597 if (mddev
->array_sectors
& (new_chunk
-1))
5598 /* not factor of array size */
5602 /* They look valid */
5603 return check_reshape(mddev
);
5606 static void *raid5_takeover(mddev_t
*mddev
)
5608 /* raid5 can take over:
5609 * raid0 - if there is only one strip zone - make it a raid4 layout
5610 * raid1 - if there are two drives. We need to know the chunk size
5611 * raid4 - trivial - just use a raid4 layout.
5612 * raid6 - Providing it is a *_6 layout
5614 if (mddev
->level
== 0)
5615 return raid45_takeover_raid0(mddev
, 5);
5616 if (mddev
->level
== 1)
5617 return raid5_takeover_raid1(mddev
);
5618 if (mddev
->level
== 4) {
5619 mddev
->new_layout
= ALGORITHM_PARITY_N
;
5620 mddev
->new_level
= 5;
5621 return setup_conf(mddev
);
5623 if (mddev
->level
== 6)
5624 return raid5_takeover_raid6(mddev
);
5626 return ERR_PTR(-EINVAL
);
5629 static void *raid4_takeover(mddev_t
*mddev
)
5631 /* raid4 can take over:
5632 * raid0 - if there is only one strip zone
5633 * raid5 - if layout is right
5635 if (mddev
->level
== 0)
5636 return raid45_takeover_raid0(mddev
, 4);
5637 if (mddev
->level
== 5 &&
5638 mddev
->layout
== ALGORITHM_PARITY_N
) {
5639 mddev
->new_layout
= 0;
5640 mddev
->new_level
= 4;
5641 return setup_conf(mddev
);
5643 return ERR_PTR(-EINVAL
);
5646 static struct mdk_personality raid5_personality
;
5648 static void *raid6_takeover(mddev_t
*mddev
)
5650 /* Currently can only take over a raid5. We map the
5651 * personality to an equivalent raid6 personality
5652 * with the Q block at the end.
5656 if (mddev
->pers
!= &raid5_personality
)
5657 return ERR_PTR(-EINVAL
);
5658 if (mddev
->degraded
> 1)
5659 return ERR_PTR(-EINVAL
);
5660 if (mddev
->raid_disks
> 253)
5661 return ERR_PTR(-EINVAL
);
5662 if (mddev
->raid_disks
< 3)
5663 return ERR_PTR(-EINVAL
);
5665 switch (mddev
->layout
) {
5666 case ALGORITHM_LEFT_ASYMMETRIC
:
5667 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
5669 case ALGORITHM_RIGHT_ASYMMETRIC
:
5670 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
5672 case ALGORITHM_LEFT_SYMMETRIC
:
5673 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
5675 case ALGORITHM_RIGHT_SYMMETRIC
:
5676 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
5678 case ALGORITHM_PARITY_0
:
5679 new_layout
= ALGORITHM_PARITY_0_6
;
5681 case ALGORITHM_PARITY_N
:
5682 new_layout
= ALGORITHM_PARITY_N
;
5685 return ERR_PTR(-EINVAL
);
5687 mddev
->new_level
= 6;
5688 mddev
->new_layout
= new_layout
;
5689 mddev
->delta_disks
= 1;
5690 mddev
->raid_disks
+= 1;
5691 return setup_conf(mddev
);
5695 static struct mdk_personality raid6_personality
=
5699 .owner
= THIS_MODULE
,
5700 .make_request
= make_request
,
5704 .error_handler
= error
,
5705 .hot_add_disk
= raid5_add_disk
,
5706 .hot_remove_disk
= raid5_remove_disk
,
5707 .spare_active
= raid5_spare_active
,
5708 .sync_request
= sync_request
,
5709 .resize
= raid5_resize
,
5711 .check_reshape
= raid6_check_reshape
,
5712 .start_reshape
= raid5_start_reshape
,
5713 .finish_reshape
= raid5_finish_reshape
,
5714 .quiesce
= raid5_quiesce
,
5715 .takeover
= raid6_takeover
,
5717 static struct mdk_personality raid5_personality
=
5721 .owner
= THIS_MODULE
,
5722 .make_request
= make_request
,
5726 .error_handler
= error
,
5727 .hot_add_disk
= raid5_add_disk
,
5728 .hot_remove_disk
= raid5_remove_disk
,
5729 .spare_active
= raid5_spare_active
,
5730 .sync_request
= sync_request
,
5731 .resize
= raid5_resize
,
5733 .check_reshape
= raid5_check_reshape
,
5734 .start_reshape
= raid5_start_reshape
,
5735 .finish_reshape
= raid5_finish_reshape
,
5736 .quiesce
= raid5_quiesce
,
5737 .takeover
= raid5_takeover
,
5740 static struct mdk_personality raid4_personality
=
5744 .owner
= THIS_MODULE
,
5745 .make_request
= make_request
,
5749 .error_handler
= error
,
5750 .hot_add_disk
= raid5_add_disk
,
5751 .hot_remove_disk
= raid5_remove_disk
,
5752 .spare_active
= raid5_spare_active
,
5753 .sync_request
= sync_request
,
5754 .resize
= raid5_resize
,
5756 .check_reshape
= raid5_check_reshape
,
5757 .start_reshape
= raid5_start_reshape
,
5758 .finish_reshape
= raid5_finish_reshape
,
5759 .quiesce
= raid5_quiesce
,
5760 .takeover
= raid4_takeover
,
5763 static int __init
raid5_init(void)
5765 register_md_personality(&raid6_personality
);
5766 register_md_personality(&raid5_personality
);
5767 register_md_personality(&raid4_personality
);
5771 static void raid5_exit(void)
5773 unregister_md_personality(&raid6_personality
);
5774 unregister_md_personality(&raid5_personality
);
5775 unregister_md_personality(&raid4_personality
);
5778 module_init(raid5_init
);
5779 module_exit(raid5_exit
);
5780 MODULE_LICENSE("GPL");
5781 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
5782 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5783 MODULE_ALIAS("md-raid5");
5784 MODULE_ALIAS("md-raid4");
5785 MODULE_ALIAS("md-level-5");
5786 MODULE_ALIAS("md-level-4");
5787 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5788 MODULE_ALIAS("md-raid6");
5789 MODULE_ALIAS("md-level-6");
5791 /* This used to be two separate modules, they were: */
5792 MODULE_ALIAS("raid5");
5793 MODULE_ALIAS("raid6");