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
);
530 if (s
->syncing
|| s
->expanding
|| s
->expanded
)
531 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
533 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
535 bi
->bi_bdev
= rdev
->bdev
;
536 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
537 __func__
, (unsigned long long)sh
->sector
,
539 atomic_inc(&sh
->count
);
540 bi
->bi_sector
= sh
->sector
+ rdev
->data_offset
;
541 bi
->bi_flags
= 1 << BIO_UPTODATE
;
545 bi
->bi_io_vec
= &sh
->dev
[i
].vec
;
546 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
547 bi
->bi_io_vec
[0].bv_offset
= 0;
548 bi
->bi_size
= STRIPE_SIZE
;
550 generic_make_request(bi
);
553 set_bit(STRIPE_DEGRADED
, &sh
->state
);
554 pr_debug("skip op %ld on disc %d for sector %llu\n",
555 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
556 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
557 set_bit(STRIPE_HANDLE
, &sh
->state
);
562 static struct dma_async_tx_descriptor
*
563 async_copy_data(int frombio
, struct bio
*bio
, struct page
*page
,
564 sector_t sector
, struct dma_async_tx_descriptor
*tx
)
567 struct page
*bio_page
;
570 struct async_submit_ctl submit
;
571 enum async_tx_flags flags
= 0;
573 if (bio
->bi_sector
>= sector
)
574 page_offset
= (signed)(bio
->bi_sector
- sector
) * 512;
576 page_offset
= (signed)(sector
- bio
->bi_sector
) * -512;
579 flags
|= ASYNC_TX_FENCE
;
580 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
582 bio_for_each_segment(bvl
, bio
, i
) {
583 int len
= bvl
->bv_len
;
587 if (page_offset
< 0) {
588 b_offset
= -page_offset
;
589 page_offset
+= b_offset
;
593 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
594 clen
= STRIPE_SIZE
- page_offset
;
599 b_offset
+= bvl
->bv_offset
;
600 bio_page
= bvl
->bv_page
;
602 tx
= async_memcpy(page
, bio_page
, page_offset
,
603 b_offset
, clen
, &submit
);
605 tx
= async_memcpy(bio_page
, page
, b_offset
,
606 page_offset
, clen
, &submit
);
608 /* chain the operations */
609 submit
.depend_tx
= tx
;
611 if (clen
< len
) /* hit end of page */
619 static void ops_complete_biofill(void *stripe_head_ref
)
621 struct stripe_head
*sh
= stripe_head_ref
;
622 struct bio
*return_bi
= NULL
;
623 raid5_conf_t
*conf
= sh
->raid_conf
;
626 pr_debug("%s: stripe %llu\n", __func__
,
627 (unsigned long long)sh
->sector
);
629 /* clear completed biofills */
630 spin_lock_irq(&conf
->device_lock
);
631 for (i
= sh
->disks
; i
--; ) {
632 struct r5dev
*dev
= &sh
->dev
[i
];
634 /* acknowledge completion of a biofill operation */
635 /* and check if we need to reply to a read request,
636 * new R5_Wantfill requests are held off until
637 * !STRIPE_BIOFILL_RUN
639 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
640 struct bio
*rbi
, *rbi2
;
645 while (rbi
&& rbi
->bi_sector
<
646 dev
->sector
+ STRIPE_SECTORS
) {
647 rbi2
= r5_next_bio(rbi
, dev
->sector
);
648 if (!raid5_dec_bi_phys_segments(rbi
)) {
649 rbi
->bi_next
= return_bi
;
656 spin_unlock_irq(&conf
->device_lock
);
657 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
659 return_io(return_bi
);
661 set_bit(STRIPE_HANDLE
, &sh
->state
);
665 static void ops_run_biofill(struct stripe_head
*sh
)
667 struct dma_async_tx_descriptor
*tx
= NULL
;
668 raid5_conf_t
*conf
= sh
->raid_conf
;
669 struct async_submit_ctl submit
;
672 pr_debug("%s: stripe %llu\n", __func__
,
673 (unsigned long long)sh
->sector
);
675 for (i
= sh
->disks
; i
--; ) {
676 struct r5dev
*dev
= &sh
->dev
[i
];
677 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
679 spin_lock_irq(&conf
->device_lock
);
680 dev
->read
= rbi
= dev
->toread
;
682 spin_unlock_irq(&conf
->device_lock
);
683 while (rbi
&& rbi
->bi_sector
<
684 dev
->sector
+ STRIPE_SECTORS
) {
685 tx
= async_copy_data(0, rbi
, dev
->page
,
687 rbi
= r5_next_bio(rbi
, dev
->sector
);
692 atomic_inc(&sh
->count
);
693 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
694 async_trigger_callback(&submit
);
697 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
704 tgt
= &sh
->dev
[target
];
705 set_bit(R5_UPTODATE
, &tgt
->flags
);
706 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
707 clear_bit(R5_Wantcompute
, &tgt
->flags
);
710 static void ops_complete_compute(void *stripe_head_ref
)
712 struct stripe_head
*sh
= stripe_head_ref
;
714 pr_debug("%s: stripe %llu\n", __func__
,
715 (unsigned long long)sh
->sector
);
717 /* mark the computed target(s) as uptodate */
718 mark_target_uptodate(sh
, sh
->ops
.target
);
719 mark_target_uptodate(sh
, sh
->ops
.target2
);
721 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
722 if (sh
->check_state
== check_state_compute_run
)
723 sh
->check_state
= check_state_compute_result
;
724 set_bit(STRIPE_HANDLE
, &sh
->state
);
728 /* return a pointer to the address conversion region of the scribble buffer */
729 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
730 struct raid5_percpu
*percpu
)
732 return percpu
->scribble
+ sizeof(struct page
*) * (sh
->disks
+ 2);
735 static struct dma_async_tx_descriptor
*
736 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
738 int disks
= sh
->disks
;
739 struct page
**xor_srcs
= percpu
->scribble
;
740 int target
= sh
->ops
.target
;
741 struct r5dev
*tgt
= &sh
->dev
[target
];
742 struct page
*xor_dest
= tgt
->page
;
744 struct dma_async_tx_descriptor
*tx
;
745 struct async_submit_ctl submit
;
748 pr_debug("%s: stripe %llu block: %d\n",
749 __func__
, (unsigned long long)sh
->sector
, target
);
750 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
752 for (i
= disks
; i
--; )
754 xor_srcs
[count
++] = sh
->dev
[i
].page
;
756 atomic_inc(&sh
->count
);
758 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
759 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
));
760 if (unlikely(count
== 1))
761 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
763 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
768 /* set_syndrome_sources - populate source buffers for gen_syndrome
769 * @srcs - (struct page *) array of size sh->disks
770 * @sh - stripe_head to parse
772 * Populates srcs in proper layout order for the stripe and returns the
773 * 'count' of sources to be used in a call to async_gen_syndrome. The P
774 * destination buffer is recorded in srcs[count] and the Q destination
775 * is recorded in srcs[count+1]].
777 static int set_syndrome_sources(struct page
**srcs
, struct stripe_head
*sh
)
779 int disks
= sh
->disks
;
780 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
781 int d0_idx
= raid6_d0(sh
);
785 for (i
= 0; i
< disks
; i
++)
791 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
793 srcs
[slot
] = sh
->dev
[i
].page
;
794 i
= raid6_next_disk(i
, disks
);
795 } while (i
!= d0_idx
);
797 return syndrome_disks
;
800 static struct dma_async_tx_descriptor
*
801 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
803 int disks
= sh
->disks
;
804 struct page
**blocks
= percpu
->scribble
;
806 int qd_idx
= sh
->qd_idx
;
807 struct dma_async_tx_descriptor
*tx
;
808 struct async_submit_ctl submit
;
814 if (sh
->ops
.target
< 0)
815 target
= sh
->ops
.target2
;
816 else if (sh
->ops
.target2
< 0)
817 target
= sh
->ops
.target
;
819 /* we should only have one valid target */
822 pr_debug("%s: stripe %llu block: %d\n",
823 __func__
, (unsigned long long)sh
->sector
, target
);
825 tgt
= &sh
->dev
[target
];
826 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
829 atomic_inc(&sh
->count
);
831 if (target
== qd_idx
) {
832 count
= set_syndrome_sources(blocks
, sh
);
833 blocks
[count
] = NULL
; /* regenerating p is not necessary */
834 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
835 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
836 ops_complete_compute
, sh
,
837 to_addr_conv(sh
, percpu
));
838 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
840 /* Compute any data- or p-drive using XOR */
842 for (i
= disks
; i
-- ; ) {
843 if (i
== target
|| i
== qd_idx
)
845 blocks
[count
++] = sh
->dev
[i
].page
;
848 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
849 NULL
, ops_complete_compute
, sh
,
850 to_addr_conv(sh
, percpu
));
851 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
857 static struct dma_async_tx_descriptor
*
858 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
860 int i
, count
, disks
= sh
->disks
;
861 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
862 int d0_idx
= raid6_d0(sh
);
863 int faila
= -1, failb
= -1;
864 int target
= sh
->ops
.target
;
865 int target2
= sh
->ops
.target2
;
866 struct r5dev
*tgt
= &sh
->dev
[target
];
867 struct r5dev
*tgt2
= &sh
->dev
[target2
];
868 struct dma_async_tx_descriptor
*tx
;
869 struct page
**blocks
= percpu
->scribble
;
870 struct async_submit_ctl submit
;
872 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
873 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
874 BUG_ON(target
< 0 || target2
< 0);
875 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
876 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
878 /* we need to open-code set_syndrome_sources to handle the
879 * slot number conversion for 'faila' and 'failb'
881 for (i
= 0; i
< disks
; i
++)
886 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
888 blocks
[slot
] = sh
->dev
[i
].page
;
894 i
= raid6_next_disk(i
, disks
);
895 } while (i
!= d0_idx
);
897 BUG_ON(faila
== failb
);
900 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
901 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
903 atomic_inc(&sh
->count
);
905 if (failb
== syndrome_disks
+1) {
906 /* Q disk is one of the missing disks */
907 if (faila
== syndrome_disks
) {
908 /* Missing P+Q, just recompute */
909 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
910 ops_complete_compute
, sh
,
911 to_addr_conv(sh
, percpu
));
912 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
913 STRIPE_SIZE
, &submit
);
917 int qd_idx
= sh
->qd_idx
;
919 /* Missing D+Q: recompute D from P, then recompute Q */
920 if (target
== qd_idx
)
921 data_target
= target2
;
923 data_target
= target
;
926 for (i
= disks
; i
-- ; ) {
927 if (i
== data_target
|| i
== qd_idx
)
929 blocks
[count
++] = sh
->dev
[i
].page
;
931 dest
= sh
->dev
[data_target
].page
;
932 init_async_submit(&submit
,
933 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
935 to_addr_conv(sh
, percpu
));
936 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
939 count
= set_syndrome_sources(blocks
, sh
);
940 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
941 ops_complete_compute
, sh
,
942 to_addr_conv(sh
, percpu
));
943 return async_gen_syndrome(blocks
, 0, count
+2,
944 STRIPE_SIZE
, &submit
);
947 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
948 ops_complete_compute
, sh
,
949 to_addr_conv(sh
, percpu
));
950 if (failb
== syndrome_disks
) {
951 /* We're missing D+P. */
952 return async_raid6_datap_recov(syndrome_disks
+2,
956 /* We're missing D+D. */
957 return async_raid6_2data_recov(syndrome_disks
+2,
958 STRIPE_SIZE
, faila
, failb
,
965 static void ops_complete_prexor(void *stripe_head_ref
)
967 struct stripe_head
*sh
= stripe_head_ref
;
969 pr_debug("%s: stripe %llu\n", __func__
,
970 (unsigned long long)sh
->sector
);
973 static struct dma_async_tx_descriptor
*
974 ops_run_prexor(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
975 struct dma_async_tx_descriptor
*tx
)
977 int disks
= sh
->disks
;
978 struct page
**xor_srcs
= percpu
->scribble
;
979 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
980 struct async_submit_ctl submit
;
982 /* existing parity data subtracted */
983 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
985 pr_debug("%s: stripe %llu\n", __func__
,
986 (unsigned long long)sh
->sector
);
988 for (i
= disks
; i
--; ) {
989 struct r5dev
*dev
= &sh
->dev
[i
];
990 /* Only process blocks that are known to be uptodate */
991 if (test_bit(R5_Wantdrain
, &dev
->flags
))
992 xor_srcs
[count
++] = dev
->page
;
995 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
996 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
));
997 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1002 static struct dma_async_tx_descriptor
*
1003 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1005 int disks
= sh
->disks
;
1008 pr_debug("%s: stripe %llu\n", __func__
,
1009 (unsigned long long)sh
->sector
);
1011 for (i
= disks
; i
--; ) {
1012 struct r5dev
*dev
= &sh
->dev
[i
];
1015 if (test_and_clear_bit(R5_Wantdrain
, &dev
->flags
)) {
1018 spin_lock_irq(&sh
->raid_conf
->device_lock
);
1019 chosen
= dev
->towrite
;
1020 dev
->towrite
= NULL
;
1021 BUG_ON(dev
->written
);
1022 wbi
= dev
->written
= chosen
;
1023 spin_unlock_irq(&sh
->raid_conf
->device_lock
);
1025 while (wbi
&& wbi
->bi_sector
<
1026 dev
->sector
+ STRIPE_SECTORS
) {
1027 if (wbi
->bi_rw
& REQ_FUA
)
1028 set_bit(R5_WantFUA
, &dev
->flags
);
1029 tx
= async_copy_data(1, wbi
, dev
->page
,
1031 wbi
= r5_next_bio(wbi
, dev
->sector
);
1039 static void ops_complete_reconstruct(void *stripe_head_ref
)
1041 struct stripe_head
*sh
= stripe_head_ref
;
1042 int disks
= sh
->disks
;
1043 int pd_idx
= sh
->pd_idx
;
1044 int qd_idx
= sh
->qd_idx
;
1048 pr_debug("%s: stripe %llu\n", __func__
,
1049 (unsigned long long)sh
->sector
);
1051 for (i
= disks
; i
--; )
1052 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1054 for (i
= disks
; i
--; ) {
1055 struct r5dev
*dev
= &sh
->dev
[i
];
1057 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1058 set_bit(R5_UPTODATE
, &dev
->flags
);
1060 set_bit(R5_WantFUA
, &dev
->flags
);
1064 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1065 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1066 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1067 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1069 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1070 sh
->reconstruct_state
= reconstruct_state_result
;
1073 set_bit(STRIPE_HANDLE
, &sh
->state
);
1078 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1079 struct dma_async_tx_descriptor
*tx
)
1081 int disks
= sh
->disks
;
1082 struct page
**xor_srcs
= percpu
->scribble
;
1083 struct async_submit_ctl submit
;
1084 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1085 struct page
*xor_dest
;
1087 unsigned long flags
;
1089 pr_debug("%s: stripe %llu\n", __func__
,
1090 (unsigned long long)sh
->sector
);
1092 /* check if prexor is active which means only process blocks
1093 * that are part of a read-modify-write (written)
1095 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1097 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1098 for (i
= disks
; i
--; ) {
1099 struct r5dev
*dev
= &sh
->dev
[i
];
1101 xor_srcs
[count
++] = dev
->page
;
1104 xor_dest
= sh
->dev
[pd_idx
].page
;
1105 for (i
= disks
; i
--; ) {
1106 struct r5dev
*dev
= &sh
->dev
[i
];
1108 xor_srcs
[count
++] = dev
->page
;
1112 /* 1/ if we prexor'd then the dest is reused as a source
1113 * 2/ if we did not prexor then we are redoing the parity
1114 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1115 * for the synchronous xor case
1117 flags
= ASYNC_TX_ACK
|
1118 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1120 atomic_inc(&sh
->count
);
1122 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, sh
,
1123 to_addr_conv(sh
, percpu
));
1124 if (unlikely(count
== 1))
1125 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1127 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1131 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1132 struct dma_async_tx_descriptor
*tx
)
1134 struct async_submit_ctl submit
;
1135 struct page
**blocks
= percpu
->scribble
;
1138 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1140 count
= set_syndrome_sources(blocks
, sh
);
1142 atomic_inc(&sh
->count
);
1144 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_reconstruct
,
1145 sh
, to_addr_conv(sh
, percpu
));
1146 async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1149 static void ops_complete_check(void *stripe_head_ref
)
1151 struct stripe_head
*sh
= stripe_head_ref
;
1153 pr_debug("%s: stripe %llu\n", __func__
,
1154 (unsigned long long)sh
->sector
);
1156 sh
->check_state
= check_state_check_result
;
1157 set_bit(STRIPE_HANDLE
, &sh
->state
);
1161 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1163 int disks
= sh
->disks
;
1164 int pd_idx
= sh
->pd_idx
;
1165 int qd_idx
= sh
->qd_idx
;
1166 struct page
*xor_dest
;
1167 struct page
**xor_srcs
= percpu
->scribble
;
1168 struct dma_async_tx_descriptor
*tx
;
1169 struct async_submit_ctl submit
;
1173 pr_debug("%s: stripe %llu\n", __func__
,
1174 (unsigned long long)sh
->sector
);
1177 xor_dest
= sh
->dev
[pd_idx
].page
;
1178 xor_srcs
[count
++] = xor_dest
;
1179 for (i
= disks
; i
--; ) {
1180 if (i
== pd_idx
|| i
== qd_idx
)
1182 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1185 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1186 to_addr_conv(sh
, percpu
));
1187 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1188 &sh
->ops
.zero_sum_result
, &submit
);
1190 atomic_inc(&sh
->count
);
1191 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1192 tx
= async_trigger_callback(&submit
);
1195 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1197 struct page
**srcs
= percpu
->scribble
;
1198 struct async_submit_ctl submit
;
1201 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1202 (unsigned long long)sh
->sector
, checkp
);
1204 count
= set_syndrome_sources(srcs
, sh
);
1208 atomic_inc(&sh
->count
);
1209 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1210 sh
, to_addr_conv(sh
, percpu
));
1211 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1212 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1215 static void __raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1217 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1218 struct dma_async_tx_descriptor
*tx
= NULL
;
1219 raid5_conf_t
*conf
= sh
->raid_conf
;
1220 int level
= conf
->level
;
1221 struct raid5_percpu
*percpu
;
1225 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1226 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1227 ops_run_biofill(sh
);
1231 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1233 tx
= ops_run_compute5(sh
, percpu
);
1235 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
1236 tx
= ops_run_compute6_1(sh
, percpu
);
1238 tx
= ops_run_compute6_2(sh
, percpu
);
1240 /* terminate the chain if reconstruct is not set to be run */
1241 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
1245 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
))
1246 tx
= ops_run_prexor(sh
, percpu
, tx
);
1248 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
1249 tx
= ops_run_biodrain(sh
, tx
);
1253 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
1255 ops_run_reconstruct5(sh
, percpu
, tx
);
1257 ops_run_reconstruct6(sh
, percpu
, tx
);
1260 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
1261 if (sh
->check_state
== check_state_run
)
1262 ops_run_check_p(sh
, percpu
);
1263 else if (sh
->check_state
== check_state_run_q
)
1264 ops_run_check_pq(sh
, percpu
, 0);
1265 else if (sh
->check_state
== check_state_run_pq
)
1266 ops_run_check_pq(sh
, percpu
, 1);
1272 for (i
= disks
; i
--; ) {
1273 struct r5dev
*dev
= &sh
->dev
[i
];
1274 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
1275 wake_up(&sh
->raid_conf
->wait_for_overlap
);
1280 #ifdef CONFIG_MULTICORE_RAID456
1281 static void async_run_ops(void *param
, async_cookie_t cookie
)
1283 struct stripe_head
*sh
= param
;
1284 unsigned long ops_request
= sh
->ops
.request
;
1286 clear_bit_unlock(STRIPE_OPS_REQ_PENDING
, &sh
->state
);
1287 wake_up(&sh
->ops
.wait_for_ops
);
1289 __raid_run_ops(sh
, ops_request
);
1293 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1295 /* since handle_stripe can be called outside of raid5d context
1296 * we need to ensure sh->ops.request is de-staged before another
1299 wait_event(sh
->ops
.wait_for_ops
,
1300 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING
, &sh
->state
));
1301 sh
->ops
.request
= ops_request
;
1303 atomic_inc(&sh
->count
);
1304 async_schedule(async_run_ops
, sh
);
1307 #define raid_run_ops __raid_run_ops
1310 static int grow_one_stripe(raid5_conf_t
*conf
)
1312 struct stripe_head
*sh
;
1313 sh
= kmem_cache_zalloc(conf
->slab_cache
, GFP_KERNEL
);
1317 sh
->raid_conf
= conf
;
1318 #ifdef CONFIG_MULTICORE_RAID456
1319 init_waitqueue_head(&sh
->ops
.wait_for_ops
);
1322 if (grow_buffers(sh
)) {
1324 kmem_cache_free(conf
->slab_cache
, sh
);
1327 /* we just created an active stripe so... */
1328 atomic_set(&sh
->count
, 1);
1329 atomic_inc(&conf
->active_stripes
);
1330 INIT_LIST_HEAD(&sh
->lru
);
1335 static int grow_stripes(raid5_conf_t
*conf
, int num
)
1337 struct kmem_cache
*sc
;
1338 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
1340 if (conf
->mddev
->gendisk
)
1341 sprintf(conf
->cache_name
[0],
1342 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
1344 sprintf(conf
->cache_name
[0],
1345 "raid%d-%p", conf
->level
, conf
->mddev
);
1346 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
1348 conf
->active_name
= 0;
1349 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
1350 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
1354 conf
->slab_cache
= sc
;
1355 conf
->pool_size
= devs
;
1357 if (!grow_one_stripe(conf
))
1363 * scribble_len - return the required size of the scribble region
1364 * @num - total number of disks in the array
1366 * The size must be enough to contain:
1367 * 1/ a struct page pointer for each device in the array +2
1368 * 2/ room to convert each entry in (1) to its corresponding dma
1369 * (dma_map_page()) or page (page_address()) address.
1371 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1372 * calculate over all devices (not just the data blocks), using zeros in place
1373 * of the P and Q blocks.
1375 static size_t scribble_len(int num
)
1379 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
1384 static int resize_stripes(raid5_conf_t
*conf
, int newsize
)
1386 /* Make all the stripes able to hold 'newsize' devices.
1387 * New slots in each stripe get 'page' set to a new page.
1389 * This happens in stages:
1390 * 1/ create a new kmem_cache and allocate the required number of
1392 * 2/ gather all the old stripe_heads and tranfer the pages across
1393 * to the new stripe_heads. This will have the side effect of
1394 * freezing the array as once all stripe_heads have been collected,
1395 * no IO will be possible. Old stripe heads are freed once their
1396 * pages have been transferred over, and the old kmem_cache is
1397 * freed when all stripes are done.
1398 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1399 * we simple return a failre status - no need to clean anything up.
1400 * 4/ allocate new pages for the new slots in the new stripe_heads.
1401 * If this fails, we don't bother trying the shrink the
1402 * stripe_heads down again, we just leave them as they are.
1403 * As each stripe_head is processed the new one is released into
1406 * Once step2 is started, we cannot afford to wait for a write,
1407 * so we use GFP_NOIO allocations.
1409 struct stripe_head
*osh
, *nsh
;
1410 LIST_HEAD(newstripes
);
1411 struct disk_info
*ndisks
;
1414 struct kmem_cache
*sc
;
1417 if (newsize
<= conf
->pool_size
)
1418 return 0; /* never bother to shrink */
1420 err
= md_allow_write(conf
->mddev
);
1425 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
1426 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
1431 for (i
= conf
->max_nr_stripes
; i
; i
--) {
1432 nsh
= kmem_cache_zalloc(sc
, GFP_KERNEL
);
1436 nsh
->raid_conf
= conf
;
1437 #ifdef CONFIG_MULTICORE_RAID456
1438 init_waitqueue_head(&nsh
->ops
.wait_for_ops
);
1441 list_add(&nsh
->lru
, &newstripes
);
1444 /* didn't get enough, give up */
1445 while (!list_empty(&newstripes
)) {
1446 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1447 list_del(&nsh
->lru
);
1448 kmem_cache_free(sc
, nsh
);
1450 kmem_cache_destroy(sc
);
1453 /* Step 2 - Must use GFP_NOIO now.
1454 * OK, we have enough stripes, start collecting inactive
1455 * stripes and copying them over
1457 list_for_each_entry(nsh
, &newstripes
, lru
) {
1458 spin_lock_irq(&conf
->device_lock
);
1459 wait_event_lock_irq(conf
->wait_for_stripe
,
1460 !list_empty(&conf
->inactive_list
),
1463 osh
= get_free_stripe(conf
);
1464 spin_unlock_irq(&conf
->device_lock
);
1465 atomic_set(&nsh
->count
, 1);
1466 for(i
=0; i
<conf
->pool_size
; i
++)
1467 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
1468 for( ; i
<newsize
; i
++)
1469 nsh
->dev
[i
].page
= NULL
;
1470 kmem_cache_free(conf
->slab_cache
, osh
);
1472 kmem_cache_destroy(conf
->slab_cache
);
1475 * At this point, we are holding all the stripes so the array
1476 * is completely stalled, so now is a good time to resize
1477 * conf->disks and the scribble region
1479 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
1481 for (i
=0; i
<conf
->raid_disks
; i
++)
1482 ndisks
[i
] = conf
->disks
[i
];
1484 conf
->disks
= ndisks
;
1489 conf
->scribble_len
= scribble_len(newsize
);
1490 for_each_present_cpu(cpu
) {
1491 struct raid5_percpu
*percpu
;
1494 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1495 scribble
= kmalloc(conf
->scribble_len
, GFP_NOIO
);
1498 kfree(percpu
->scribble
);
1499 percpu
->scribble
= scribble
;
1507 /* Step 4, return new stripes to service */
1508 while(!list_empty(&newstripes
)) {
1509 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1510 list_del_init(&nsh
->lru
);
1512 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
1513 if (nsh
->dev
[i
].page
== NULL
) {
1514 struct page
*p
= alloc_page(GFP_NOIO
);
1515 nsh
->dev
[i
].page
= p
;
1519 release_stripe(nsh
);
1521 /* critical section pass, GFP_NOIO no longer needed */
1523 conf
->slab_cache
= sc
;
1524 conf
->active_name
= 1-conf
->active_name
;
1525 conf
->pool_size
= newsize
;
1529 static int drop_one_stripe(raid5_conf_t
*conf
)
1531 struct stripe_head
*sh
;
1533 spin_lock_irq(&conf
->device_lock
);
1534 sh
= get_free_stripe(conf
);
1535 spin_unlock_irq(&conf
->device_lock
);
1538 BUG_ON(atomic_read(&sh
->count
));
1540 kmem_cache_free(conf
->slab_cache
, sh
);
1541 atomic_dec(&conf
->active_stripes
);
1545 static void shrink_stripes(raid5_conf_t
*conf
)
1547 while (drop_one_stripe(conf
))
1550 if (conf
->slab_cache
)
1551 kmem_cache_destroy(conf
->slab_cache
);
1552 conf
->slab_cache
= NULL
;
1555 static void raid5_end_read_request(struct bio
* bi
, int error
)
1557 struct stripe_head
*sh
= bi
->bi_private
;
1558 raid5_conf_t
*conf
= sh
->raid_conf
;
1559 int disks
= sh
->disks
, i
;
1560 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1561 char b
[BDEVNAME_SIZE
];
1565 for (i
=0 ; i
<disks
; i
++)
1566 if (bi
== &sh
->dev
[i
].req
)
1569 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1570 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1578 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1579 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
1580 rdev
= conf
->disks
[i
].rdev
;
1583 "md/raid:%s: read error corrected"
1584 " (%lu sectors at %llu on %s)\n",
1585 mdname(conf
->mddev
), STRIPE_SECTORS
,
1586 (unsigned long long)(sh
->sector
1587 + rdev
->data_offset
),
1588 bdevname(rdev
->bdev
, b
));
1589 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
1590 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1591 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1593 if (atomic_read(&conf
->disks
[i
].rdev
->read_errors
))
1594 atomic_set(&conf
->disks
[i
].rdev
->read_errors
, 0);
1596 const char *bdn
= bdevname(conf
->disks
[i
].rdev
->bdev
, b
);
1598 rdev
= conf
->disks
[i
].rdev
;
1600 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1601 atomic_inc(&rdev
->read_errors
);
1602 if (conf
->mddev
->degraded
>= conf
->max_degraded
)
1605 "md/raid:%s: read error not correctable "
1606 "(sector %llu on %s).\n",
1607 mdname(conf
->mddev
),
1608 (unsigned long long)(sh
->sector
1609 + rdev
->data_offset
),
1611 else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
))
1615 "md/raid:%s: read error NOT corrected!! "
1616 "(sector %llu on %s).\n",
1617 mdname(conf
->mddev
),
1618 (unsigned long long)(sh
->sector
1619 + rdev
->data_offset
),
1621 else if (atomic_read(&rdev
->read_errors
)
1622 > conf
->max_nr_stripes
)
1624 "md/raid:%s: Too many read errors, failing device %s.\n",
1625 mdname(conf
->mddev
), bdn
);
1629 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1631 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1632 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1633 md_error(conf
->mddev
, rdev
);
1636 rdev_dec_pending(conf
->disks
[i
].rdev
, conf
->mddev
);
1637 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1638 set_bit(STRIPE_HANDLE
, &sh
->state
);
1642 static void raid5_end_write_request(struct bio
*bi
, int error
)
1644 struct stripe_head
*sh
= bi
->bi_private
;
1645 raid5_conf_t
*conf
= sh
->raid_conf
;
1646 int disks
= sh
->disks
, i
;
1647 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1649 for (i
=0 ; i
<disks
; i
++)
1650 if (bi
== &sh
->dev
[i
].req
)
1653 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1654 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1662 md_error(conf
->mddev
, conf
->disks
[i
].rdev
);
1664 rdev_dec_pending(conf
->disks
[i
].rdev
, conf
->mddev
);
1666 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1667 set_bit(STRIPE_HANDLE
, &sh
->state
);
1672 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
);
1674 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
1676 struct r5dev
*dev
= &sh
->dev
[i
];
1678 bio_init(&dev
->req
);
1679 dev
->req
.bi_io_vec
= &dev
->vec
;
1681 dev
->req
.bi_max_vecs
++;
1682 dev
->vec
.bv_page
= dev
->page
;
1683 dev
->vec
.bv_len
= STRIPE_SIZE
;
1684 dev
->vec
.bv_offset
= 0;
1686 dev
->req
.bi_sector
= sh
->sector
;
1687 dev
->req
.bi_private
= sh
;
1690 dev
->sector
= compute_blocknr(sh
, i
, previous
);
1693 static void error(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1695 char b
[BDEVNAME_SIZE
];
1696 raid5_conf_t
*conf
= mddev
->private;
1697 pr_debug("raid456: error called\n");
1699 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1700 unsigned long flags
;
1701 spin_lock_irqsave(&conf
->device_lock
, flags
);
1703 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1705 * if recovery was running, make sure it aborts.
1707 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1709 set_bit(Blocked
, &rdev
->flags
);
1710 set_bit(Faulty
, &rdev
->flags
);
1711 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1713 "md/raid:%s: Disk failure on %s, disabling device.\n"
1714 "md/raid:%s: Operation continuing on %d devices.\n",
1716 bdevname(rdev
->bdev
, b
),
1718 conf
->raid_disks
- mddev
->degraded
);
1722 * Input: a 'big' sector number,
1723 * Output: index of the data and parity disk, and the sector # in them.
1725 static sector_t
raid5_compute_sector(raid5_conf_t
*conf
, sector_t r_sector
,
1726 int previous
, int *dd_idx
,
1727 struct stripe_head
*sh
)
1729 sector_t stripe
, stripe2
;
1730 sector_t chunk_number
;
1731 unsigned int chunk_offset
;
1734 sector_t new_sector
;
1735 int algorithm
= previous
? conf
->prev_algo
1737 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
1738 : conf
->chunk_sectors
;
1739 int raid_disks
= previous
? conf
->previous_raid_disks
1741 int data_disks
= raid_disks
- conf
->max_degraded
;
1743 /* First compute the information on this sector */
1746 * Compute the chunk number and the sector offset inside the chunk
1748 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
1749 chunk_number
= r_sector
;
1752 * Compute the stripe number
1754 stripe
= chunk_number
;
1755 *dd_idx
= sector_div(stripe
, data_disks
);
1758 * Select the parity disk based on the user selected algorithm.
1760 pd_idx
= qd_idx
= -1;
1761 switch(conf
->level
) {
1763 pd_idx
= data_disks
;
1766 switch (algorithm
) {
1767 case ALGORITHM_LEFT_ASYMMETRIC
:
1768 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
1769 if (*dd_idx
>= pd_idx
)
1772 case ALGORITHM_RIGHT_ASYMMETRIC
:
1773 pd_idx
= sector_div(stripe2
, raid_disks
);
1774 if (*dd_idx
>= pd_idx
)
1777 case ALGORITHM_LEFT_SYMMETRIC
:
1778 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
1779 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1781 case ALGORITHM_RIGHT_SYMMETRIC
:
1782 pd_idx
= sector_div(stripe2
, raid_disks
);
1783 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1785 case ALGORITHM_PARITY_0
:
1789 case ALGORITHM_PARITY_N
:
1790 pd_idx
= data_disks
;
1798 switch (algorithm
) {
1799 case ALGORITHM_LEFT_ASYMMETRIC
:
1800 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1801 qd_idx
= pd_idx
+ 1;
1802 if (pd_idx
== raid_disks
-1) {
1803 (*dd_idx
)++; /* Q D D D P */
1805 } else if (*dd_idx
>= pd_idx
)
1806 (*dd_idx
) += 2; /* D D P Q D */
1808 case ALGORITHM_RIGHT_ASYMMETRIC
:
1809 pd_idx
= sector_div(stripe2
, raid_disks
);
1810 qd_idx
= pd_idx
+ 1;
1811 if (pd_idx
== raid_disks
-1) {
1812 (*dd_idx
)++; /* Q D D D P */
1814 } else if (*dd_idx
>= pd_idx
)
1815 (*dd_idx
) += 2; /* D D P Q D */
1817 case ALGORITHM_LEFT_SYMMETRIC
:
1818 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1819 qd_idx
= (pd_idx
+ 1) % raid_disks
;
1820 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
1822 case ALGORITHM_RIGHT_SYMMETRIC
:
1823 pd_idx
= sector_div(stripe2
, raid_disks
);
1824 qd_idx
= (pd_idx
+ 1) % raid_disks
;
1825 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
1828 case ALGORITHM_PARITY_0
:
1833 case ALGORITHM_PARITY_N
:
1834 pd_idx
= data_disks
;
1835 qd_idx
= data_disks
+ 1;
1838 case ALGORITHM_ROTATING_ZERO_RESTART
:
1839 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1840 * of blocks for computing Q is different.
1842 pd_idx
= sector_div(stripe2
, raid_disks
);
1843 qd_idx
= pd_idx
+ 1;
1844 if (pd_idx
== raid_disks
-1) {
1845 (*dd_idx
)++; /* Q D D D P */
1847 } else if (*dd_idx
>= pd_idx
)
1848 (*dd_idx
) += 2; /* D D P Q D */
1852 case ALGORITHM_ROTATING_N_RESTART
:
1853 /* Same a left_asymmetric, by first stripe is
1854 * D D D P Q rather than
1858 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1859 qd_idx
= pd_idx
+ 1;
1860 if (pd_idx
== raid_disks
-1) {
1861 (*dd_idx
)++; /* Q D D D P */
1863 } else if (*dd_idx
>= pd_idx
)
1864 (*dd_idx
) += 2; /* D D P Q D */
1868 case ALGORITHM_ROTATING_N_CONTINUE
:
1869 /* Same as left_symmetric but Q is before P */
1870 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1871 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
1872 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1876 case ALGORITHM_LEFT_ASYMMETRIC_6
:
1877 /* RAID5 left_asymmetric, with Q on last device */
1878 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
1879 if (*dd_idx
>= pd_idx
)
1881 qd_idx
= raid_disks
- 1;
1884 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
1885 pd_idx
= sector_div(stripe2
, raid_disks
-1);
1886 if (*dd_idx
>= pd_idx
)
1888 qd_idx
= raid_disks
- 1;
1891 case ALGORITHM_LEFT_SYMMETRIC_6
:
1892 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
1893 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
1894 qd_idx
= raid_disks
- 1;
1897 case ALGORITHM_RIGHT_SYMMETRIC_6
:
1898 pd_idx
= sector_div(stripe2
, raid_disks
-1);
1899 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
1900 qd_idx
= raid_disks
- 1;
1903 case ALGORITHM_PARITY_0_6
:
1906 qd_idx
= raid_disks
- 1;
1916 sh
->pd_idx
= pd_idx
;
1917 sh
->qd_idx
= qd_idx
;
1918 sh
->ddf_layout
= ddf_layout
;
1921 * Finally, compute the new sector number
1923 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
1928 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
1930 raid5_conf_t
*conf
= sh
->raid_conf
;
1931 int raid_disks
= sh
->disks
;
1932 int data_disks
= raid_disks
- conf
->max_degraded
;
1933 sector_t new_sector
= sh
->sector
, check
;
1934 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
1935 : conf
->chunk_sectors
;
1936 int algorithm
= previous
? conf
->prev_algo
1940 sector_t chunk_number
;
1941 int dummy1
, dd_idx
= i
;
1943 struct stripe_head sh2
;
1946 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
1947 stripe
= new_sector
;
1949 if (i
== sh
->pd_idx
)
1951 switch(conf
->level
) {
1954 switch (algorithm
) {
1955 case ALGORITHM_LEFT_ASYMMETRIC
:
1956 case ALGORITHM_RIGHT_ASYMMETRIC
:
1960 case ALGORITHM_LEFT_SYMMETRIC
:
1961 case ALGORITHM_RIGHT_SYMMETRIC
:
1964 i
-= (sh
->pd_idx
+ 1);
1966 case ALGORITHM_PARITY_0
:
1969 case ALGORITHM_PARITY_N
:
1976 if (i
== sh
->qd_idx
)
1977 return 0; /* It is the Q disk */
1978 switch (algorithm
) {
1979 case ALGORITHM_LEFT_ASYMMETRIC
:
1980 case ALGORITHM_RIGHT_ASYMMETRIC
:
1981 case ALGORITHM_ROTATING_ZERO_RESTART
:
1982 case ALGORITHM_ROTATING_N_RESTART
:
1983 if (sh
->pd_idx
== raid_disks
-1)
1984 i
--; /* Q D D D P */
1985 else if (i
> sh
->pd_idx
)
1986 i
-= 2; /* D D P Q D */
1988 case ALGORITHM_LEFT_SYMMETRIC
:
1989 case ALGORITHM_RIGHT_SYMMETRIC
:
1990 if (sh
->pd_idx
== raid_disks
-1)
1991 i
--; /* Q D D D P */
1996 i
-= (sh
->pd_idx
+ 2);
1999 case ALGORITHM_PARITY_0
:
2002 case ALGORITHM_PARITY_N
:
2004 case ALGORITHM_ROTATING_N_CONTINUE
:
2005 /* Like left_symmetric, but P is before Q */
2006 if (sh
->pd_idx
== 0)
2007 i
--; /* P D D D Q */
2012 i
-= (sh
->pd_idx
+ 1);
2015 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2016 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2020 case ALGORITHM_LEFT_SYMMETRIC_6
:
2021 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2023 i
+= data_disks
+ 1;
2024 i
-= (sh
->pd_idx
+ 1);
2026 case ALGORITHM_PARITY_0_6
:
2035 chunk_number
= stripe
* data_disks
+ i
;
2036 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2038 check
= raid5_compute_sector(conf
, r_sector
,
2039 previous
, &dummy1
, &sh2
);
2040 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2041 || sh2
.qd_idx
!= sh
->qd_idx
) {
2042 printk(KERN_ERR
"md/raid:%s: compute_blocknr: map not correct\n",
2043 mdname(conf
->mddev
));
2051 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2052 int rcw
, int expand
)
2054 int i
, pd_idx
= sh
->pd_idx
, disks
= sh
->disks
;
2055 raid5_conf_t
*conf
= sh
->raid_conf
;
2056 int level
= conf
->level
;
2059 /* if we are not expanding this is a proper write request, and
2060 * there will be bios with new data to be drained into the
2064 sh
->reconstruct_state
= reconstruct_state_drain_run
;
2065 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2067 sh
->reconstruct_state
= reconstruct_state_run
;
2069 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2071 for (i
= disks
; i
--; ) {
2072 struct r5dev
*dev
= &sh
->dev
[i
];
2075 set_bit(R5_LOCKED
, &dev
->flags
);
2076 set_bit(R5_Wantdrain
, &dev
->flags
);
2078 clear_bit(R5_UPTODATE
, &dev
->flags
);
2082 if (s
->locked
+ conf
->max_degraded
== disks
)
2083 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2084 atomic_inc(&conf
->pending_full_writes
);
2087 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
2088 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
2090 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
2091 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
2092 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2093 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2095 for (i
= disks
; i
--; ) {
2096 struct r5dev
*dev
= &sh
->dev
[i
];
2101 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2102 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2103 set_bit(R5_Wantdrain
, &dev
->flags
);
2104 set_bit(R5_LOCKED
, &dev
->flags
);
2105 clear_bit(R5_UPTODATE
, &dev
->flags
);
2111 /* keep the parity disk(s) locked while asynchronous operations
2114 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
2115 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2119 int qd_idx
= sh
->qd_idx
;
2120 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
2122 set_bit(R5_LOCKED
, &dev
->flags
);
2123 clear_bit(R5_UPTODATE
, &dev
->flags
);
2127 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2128 __func__
, (unsigned long long)sh
->sector
,
2129 s
->locked
, s
->ops_request
);
2133 * Each stripe/dev can have one or more bion attached.
2134 * toread/towrite point to the first in a chain.
2135 * The bi_next chain must be in order.
2137 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
, int forwrite
)
2140 raid5_conf_t
*conf
= sh
->raid_conf
;
2143 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2144 (unsigned long long)bi
->bi_sector
,
2145 (unsigned long long)sh
->sector
);
2148 spin_lock_irq(&conf
->device_lock
);
2150 bip
= &sh
->dev
[dd_idx
].towrite
;
2151 if (*bip
== NULL
&& sh
->dev
[dd_idx
].written
== NULL
)
2154 bip
= &sh
->dev
[dd_idx
].toread
;
2155 while (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
) {
2156 if ((*bip
)->bi_sector
+ ((*bip
)->bi_size
>> 9) > bi
->bi_sector
)
2158 bip
= & (*bip
)->bi_next
;
2160 if (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
+ ((bi
->bi_size
)>>9))
2163 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
2167 bi
->bi_phys_segments
++;
2170 /* check if page is covered */
2171 sector_t sector
= sh
->dev
[dd_idx
].sector
;
2172 for (bi
=sh
->dev
[dd_idx
].towrite
;
2173 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
2174 bi
&& bi
->bi_sector
<= sector
;
2175 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
2176 if (bi
->bi_sector
+ (bi
->bi_size
>>9) >= sector
)
2177 sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
2179 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
2180 set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
);
2182 spin_unlock_irq(&conf
->device_lock
);
2184 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2185 (unsigned long long)(*bip
)->bi_sector
,
2186 (unsigned long long)sh
->sector
, dd_idx
);
2188 if (conf
->mddev
->bitmap
&& firstwrite
) {
2189 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
2191 sh
->bm_seq
= conf
->seq_flush
+1;
2192 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
2197 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
2198 spin_unlock_irq(&conf
->device_lock
);
2202 static void end_reshape(raid5_conf_t
*conf
);
2204 static void stripe_set_idx(sector_t stripe
, raid5_conf_t
*conf
, int previous
,
2205 struct stripe_head
*sh
)
2207 int sectors_per_chunk
=
2208 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
2210 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
2211 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
2213 raid5_compute_sector(conf
,
2214 stripe
* (disks
- conf
->max_degraded
)
2215 *sectors_per_chunk
+ chunk_offset
,
2221 handle_failed_stripe(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2222 struct stripe_head_state
*s
, int disks
,
2223 struct bio
**return_bi
)
2226 for (i
= disks
; i
--; ) {
2230 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2233 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2234 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
2235 atomic_inc(&rdev
->nr_pending
);
2240 if (!rdev_set_badblocks(
2244 md_error(conf
->mddev
, rdev
);
2245 rdev_dec_pending(rdev
, conf
->mddev
);
2248 spin_lock_irq(&conf
->device_lock
);
2249 /* fail all writes first */
2250 bi
= sh
->dev
[i
].towrite
;
2251 sh
->dev
[i
].towrite
= NULL
;
2257 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2258 wake_up(&conf
->wait_for_overlap
);
2260 while (bi
&& bi
->bi_sector
<
2261 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2262 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2263 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2264 if (!raid5_dec_bi_phys_segments(bi
)) {
2265 md_write_end(conf
->mddev
);
2266 bi
->bi_next
= *return_bi
;
2271 /* and fail all 'written' */
2272 bi
= sh
->dev
[i
].written
;
2273 sh
->dev
[i
].written
= NULL
;
2274 if (bi
) bitmap_end
= 1;
2275 while (bi
&& bi
->bi_sector
<
2276 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2277 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2278 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2279 if (!raid5_dec_bi_phys_segments(bi
)) {
2280 md_write_end(conf
->mddev
);
2281 bi
->bi_next
= *return_bi
;
2287 /* fail any reads if this device is non-operational and
2288 * the data has not reached the cache yet.
2290 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
2291 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
2292 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
2293 bi
= sh
->dev
[i
].toread
;
2294 sh
->dev
[i
].toread
= NULL
;
2295 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2296 wake_up(&conf
->wait_for_overlap
);
2297 if (bi
) s
->to_read
--;
2298 while (bi
&& bi
->bi_sector
<
2299 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2300 struct bio
*nextbi
=
2301 r5_next_bio(bi
, sh
->dev
[i
].sector
);
2302 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2303 if (!raid5_dec_bi_phys_segments(bi
)) {
2304 bi
->bi_next
= *return_bi
;
2310 spin_unlock_irq(&conf
->device_lock
);
2312 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2313 STRIPE_SECTORS
, 0, 0);
2314 /* If we were in the middle of a write the parity block might
2315 * still be locked - so just clear all R5_LOCKED flags
2317 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2320 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2321 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2322 md_wakeup_thread(conf
->mddev
->thread
);
2326 handle_failed_sync(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2327 struct stripe_head_state
*s
)
2332 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 0);
2333 clear_bit(STRIPE_SYNCING
, &sh
->state
);
2335 /* There is nothing more to do for sync/check/repair.
2336 * For recover we need to record a bad block on all
2337 * non-sync devices, or abort the recovery
2339 if (!test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
))
2341 /* During recovery devices cannot be removed, so locking and
2342 * refcounting of rdevs is not needed
2344 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2345 mdk_rdev_t
*rdev
= conf
->disks
[i
].rdev
;
2347 || test_bit(Faulty
, &rdev
->flags
)
2348 || test_bit(In_sync
, &rdev
->flags
))
2350 if (!rdev_set_badblocks(rdev
, sh
->sector
,
2355 conf
->recovery_disabled
= conf
->mddev
->recovery_disabled
;
2356 set_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
);
2360 /* fetch_block - checks the given member device to see if its data needs
2361 * to be read or computed to satisfy a request.
2363 * Returns 1 when no more member devices need to be checked, otherwise returns
2364 * 0 to tell the loop in handle_stripe_fill to continue
2366 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2367 int disk_idx
, int disks
)
2369 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
2370 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
2371 &sh
->dev
[s
->failed_num
[1]] };
2373 /* is the data in this block needed, and can we get it? */
2374 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2375 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2377 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2378 s
->syncing
|| s
->expanding
||
2379 (s
->failed
>= 1 && fdev
[0]->toread
) ||
2380 (s
->failed
>= 2 && fdev
[1]->toread
) ||
2381 (sh
->raid_conf
->level
<= 5 && s
->failed
&& fdev
[0]->towrite
&&
2382 !test_bit(R5_OVERWRITE
, &fdev
[0]->flags
)) ||
2383 (sh
->raid_conf
->level
== 6 && s
->failed
&& s
->to_write
))) {
2384 /* we would like to get this block, possibly by computing it,
2385 * otherwise read it if the backing disk is insync
2387 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
2388 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
2389 if ((s
->uptodate
== disks
- 1) &&
2390 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
2391 disk_idx
== s
->failed_num
[1]))) {
2392 /* have disk failed, and we're requested to fetch it;
2395 pr_debug("Computing stripe %llu block %d\n",
2396 (unsigned long long)sh
->sector
, disk_idx
);
2397 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2398 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2399 set_bit(R5_Wantcompute
, &dev
->flags
);
2400 sh
->ops
.target
= disk_idx
;
2401 sh
->ops
.target2
= -1; /* no 2nd target */
2403 /* Careful: from this point on 'uptodate' is in the eye
2404 * of raid_run_ops which services 'compute' operations
2405 * before writes. R5_Wantcompute flags a block that will
2406 * be R5_UPTODATE by the time it is needed for a
2407 * subsequent operation.
2411 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
2412 /* Computing 2-failure is *very* expensive; only
2413 * do it if failed >= 2
2416 for (other
= disks
; other
--; ) {
2417 if (other
== disk_idx
)
2419 if (!test_bit(R5_UPTODATE
,
2420 &sh
->dev
[other
].flags
))
2424 pr_debug("Computing stripe %llu blocks %d,%d\n",
2425 (unsigned long long)sh
->sector
,
2427 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2428 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2429 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
2430 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
2431 sh
->ops
.target
= disk_idx
;
2432 sh
->ops
.target2
= other
;
2436 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2437 set_bit(R5_LOCKED
, &dev
->flags
);
2438 set_bit(R5_Wantread
, &dev
->flags
);
2440 pr_debug("Reading block %d (sync=%d)\n",
2441 disk_idx
, s
->syncing
);
2449 * handle_stripe_fill - read or compute data to satisfy pending requests.
2451 static void handle_stripe_fill(struct stripe_head
*sh
,
2452 struct stripe_head_state
*s
,
2457 /* look for blocks to read/compute, skip this if a compute
2458 * is already in flight, or if the stripe contents are in the
2459 * midst of changing due to a write
2461 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2462 !sh
->reconstruct_state
)
2463 for (i
= disks
; i
--; )
2464 if (fetch_block(sh
, s
, i
, disks
))
2466 set_bit(STRIPE_HANDLE
, &sh
->state
);
2470 /* handle_stripe_clean_event
2471 * any written block on an uptodate or failed drive can be returned.
2472 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2473 * never LOCKED, so we don't need to test 'failed' directly.
2475 static void handle_stripe_clean_event(raid5_conf_t
*conf
,
2476 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
2481 for (i
= disks
; i
--; )
2482 if (sh
->dev
[i
].written
) {
2484 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2485 test_bit(R5_UPTODATE
, &dev
->flags
)) {
2486 /* We can return any write requests */
2487 struct bio
*wbi
, *wbi2
;
2489 pr_debug("Return write for disc %d\n", i
);
2490 spin_lock_irq(&conf
->device_lock
);
2492 dev
->written
= NULL
;
2493 while (wbi
&& wbi
->bi_sector
<
2494 dev
->sector
+ STRIPE_SECTORS
) {
2495 wbi2
= r5_next_bio(wbi
, dev
->sector
);
2496 if (!raid5_dec_bi_phys_segments(wbi
)) {
2497 md_write_end(conf
->mddev
);
2498 wbi
->bi_next
= *return_bi
;
2503 if (dev
->towrite
== NULL
)
2505 spin_unlock_irq(&conf
->device_lock
);
2507 bitmap_endwrite(conf
->mddev
->bitmap
,
2510 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
2515 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2516 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2517 md_wakeup_thread(conf
->mddev
->thread
);
2520 static void handle_stripe_dirtying(raid5_conf_t
*conf
,
2521 struct stripe_head
*sh
,
2522 struct stripe_head_state
*s
,
2525 int rmw
= 0, rcw
= 0, i
;
2526 if (conf
->max_degraded
== 2) {
2527 /* RAID6 requires 'rcw' in current implementation
2528 * Calculate the real rcw later - for now fake it
2529 * look like rcw is cheaper
2532 } else for (i
= disks
; i
--; ) {
2533 /* would I have to read this buffer for read_modify_write */
2534 struct r5dev
*dev
= &sh
->dev
[i
];
2535 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2536 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2537 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2538 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2539 if (test_bit(R5_Insync
, &dev
->flags
))
2542 rmw
+= 2*disks
; /* cannot read it */
2544 /* Would I have to read this buffer for reconstruct_write */
2545 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) && i
!= sh
->pd_idx
&&
2546 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2547 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2548 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2549 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2554 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2555 (unsigned long long)sh
->sector
, rmw
, rcw
);
2556 set_bit(STRIPE_HANDLE
, &sh
->state
);
2557 if (rmw
< rcw
&& rmw
> 0)
2558 /* prefer read-modify-write, but need to get some data */
2559 for (i
= disks
; i
--; ) {
2560 struct r5dev
*dev
= &sh
->dev
[i
];
2561 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2562 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2563 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2564 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2565 test_bit(R5_Insync
, &dev
->flags
)) {
2567 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2568 pr_debug("Read_old block "
2569 "%d for r-m-w\n", i
);
2570 set_bit(R5_LOCKED
, &dev
->flags
);
2571 set_bit(R5_Wantread
, &dev
->flags
);
2574 set_bit(STRIPE_DELAYED
, &sh
->state
);
2575 set_bit(STRIPE_HANDLE
, &sh
->state
);
2579 if (rcw
<= rmw
&& rcw
> 0) {
2580 /* want reconstruct write, but need to get some data */
2582 for (i
= disks
; i
--; ) {
2583 struct r5dev
*dev
= &sh
->dev
[i
];
2584 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2585 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
2586 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2587 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2588 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2590 if (!test_bit(R5_Insync
, &dev
->flags
))
2591 continue; /* it's a failed drive */
2593 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2594 pr_debug("Read_old block "
2595 "%d for Reconstruct\n", i
);
2596 set_bit(R5_LOCKED
, &dev
->flags
);
2597 set_bit(R5_Wantread
, &dev
->flags
);
2600 set_bit(STRIPE_DELAYED
, &sh
->state
);
2601 set_bit(STRIPE_HANDLE
, &sh
->state
);
2606 /* now if nothing is locked, and if we have enough data,
2607 * we can start a write request
2609 /* since handle_stripe can be called at any time we need to handle the
2610 * case where a compute block operation has been submitted and then a
2611 * subsequent call wants to start a write request. raid_run_ops only
2612 * handles the case where compute block and reconstruct are requested
2613 * simultaneously. If this is not the case then new writes need to be
2614 * held off until the compute completes.
2616 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
2617 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
2618 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
2619 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
2622 static void handle_parity_checks5(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2623 struct stripe_head_state
*s
, int disks
)
2625 struct r5dev
*dev
= NULL
;
2627 set_bit(STRIPE_HANDLE
, &sh
->state
);
2629 switch (sh
->check_state
) {
2630 case check_state_idle
:
2631 /* start a new check operation if there are no failures */
2632 if (s
->failed
== 0) {
2633 BUG_ON(s
->uptodate
!= disks
);
2634 sh
->check_state
= check_state_run
;
2635 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2636 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2640 dev
= &sh
->dev
[s
->failed_num
[0]];
2642 case check_state_compute_result
:
2643 sh
->check_state
= check_state_idle
;
2645 dev
= &sh
->dev
[sh
->pd_idx
];
2647 /* check that a write has not made the stripe insync */
2648 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2651 /* either failed parity check, or recovery is happening */
2652 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
2653 BUG_ON(s
->uptodate
!= disks
);
2655 set_bit(R5_LOCKED
, &dev
->flags
);
2657 set_bit(R5_Wantwrite
, &dev
->flags
);
2659 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2660 set_bit(STRIPE_INSYNC
, &sh
->state
);
2662 case check_state_run
:
2663 break; /* we will be called again upon completion */
2664 case check_state_check_result
:
2665 sh
->check_state
= check_state_idle
;
2667 /* if a failure occurred during the check operation, leave
2668 * STRIPE_INSYNC not set and let the stripe be handled again
2673 /* handle a successful check operation, if parity is correct
2674 * we are done. Otherwise update the mismatch count and repair
2675 * parity if !MD_RECOVERY_CHECK
2677 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
2678 /* parity is correct (on disc,
2679 * not in buffer any more)
2681 set_bit(STRIPE_INSYNC
, &sh
->state
);
2683 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2684 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2685 /* don't try to repair!! */
2686 set_bit(STRIPE_INSYNC
, &sh
->state
);
2688 sh
->check_state
= check_state_compute_run
;
2689 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2690 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2691 set_bit(R5_Wantcompute
,
2692 &sh
->dev
[sh
->pd_idx
].flags
);
2693 sh
->ops
.target
= sh
->pd_idx
;
2694 sh
->ops
.target2
= -1;
2699 case check_state_compute_run
:
2702 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
2703 __func__
, sh
->check_state
,
2704 (unsigned long long) sh
->sector
);
2710 static void handle_parity_checks6(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2711 struct stripe_head_state
*s
,
2714 int pd_idx
= sh
->pd_idx
;
2715 int qd_idx
= sh
->qd_idx
;
2718 set_bit(STRIPE_HANDLE
, &sh
->state
);
2720 BUG_ON(s
->failed
> 2);
2722 /* Want to check and possibly repair P and Q.
2723 * However there could be one 'failed' device, in which
2724 * case we can only check one of them, possibly using the
2725 * other to generate missing data
2728 switch (sh
->check_state
) {
2729 case check_state_idle
:
2730 /* start a new check operation if there are < 2 failures */
2731 if (s
->failed
== s
->q_failed
) {
2732 /* The only possible failed device holds Q, so it
2733 * makes sense to check P (If anything else were failed,
2734 * we would have used P to recreate it).
2736 sh
->check_state
= check_state_run
;
2738 if (!s
->q_failed
&& s
->failed
< 2) {
2739 /* Q is not failed, and we didn't use it to generate
2740 * anything, so it makes sense to check it
2742 if (sh
->check_state
== check_state_run
)
2743 sh
->check_state
= check_state_run_pq
;
2745 sh
->check_state
= check_state_run_q
;
2748 /* discard potentially stale zero_sum_result */
2749 sh
->ops
.zero_sum_result
= 0;
2751 if (sh
->check_state
== check_state_run
) {
2752 /* async_xor_zero_sum destroys the contents of P */
2753 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2756 if (sh
->check_state
>= check_state_run
&&
2757 sh
->check_state
<= check_state_run_pq
) {
2758 /* async_syndrome_zero_sum preserves P and Q, so
2759 * no need to mark them !uptodate here
2761 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2765 /* we have 2-disk failure */
2766 BUG_ON(s
->failed
!= 2);
2768 case check_state_compute_result
:
2769 sh
->check_state
= check_state_idle
;
2771 /* check that a write has not made the stripe insync */
2772 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2775 /* now write out any block on a failed drive,
2776 * or P or Q if they were recomputed
2778 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
2779 if (s
->failed
== 2) {
2780 dev
= &sh
->dev
[s
->failed_num
[1]];
2782 set_bit(R5_LOCKED
, &dev
->flags
);
2783 set_bit(R5_Wantwrite
, &dev
->flags
);
2785 if (s
->failed
>= 1) {
2786 dev
= &sh
->dev
[s
->failed_num
[0]];
2788 set_bit(R5_LOCKED
, &dev
->flags
);
2789 set_bit(R5_Wantwrite
, &dev
->flags
);
2791 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
2792 dev
= &sh
->dev
[pd_idx
];
2794 set_bit(R5_LOCKED
, &dev
->flags
);
2795 set_bit(R5_Wantwrite
, &dev
->flags
);
2797 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
2798 dev
= &sh
->dev
[qd_idx
];
2800 set_bit(R5_LOCKED
, &dev
->flags
);
2801 set_bit(R5_Wantwrite
, &dev
->flags
);
2803 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2805 set_bit(STRIPE_INSYNC
, &sh
->state
);
2807 case check_state_run
:
2808 case check_state_run_q
:
2809 case check_state_run_pq
:
2810 break; /* we will be called again upon completion */
2811 case check_state_check_result
:
2812 sh
->check_state
= check_state_idle
;
2814 /* handle a successful check operation, if parity is correct
2815 * we are done. Otherwise update the mismatch count and repair
2816 * parity if !MD_RECOVERY_CHECK
2818 if (sh
->ops
.zero_sum_result
== 0) {
2819 /* both parities are correct */
2821 set_bit(STRIPE_INSYNC
, &sh
->state
);
2823 /* in contrast to the raid5 case we can validate
2824 * parity, but still have a failure to write
2827 sh
->check_state
= check_state_compute_result
;
2828 /* Returning at this point means that we may go
2829 * off and bring p and/or q uptodate again so
2830 * we make sure to check zero_sum_result again
2831 * to verify if p or q need writeback
2835 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2836 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2837 /* don't try to repair!! */
2838 set_bit(STRIPE_INSYNC
, &sh
->state
);
2840 int *target
= &sh
->ops
.target
;
2842 sh
->ops
.target
= -1;
2843 sh
->ops
.target2
= -1;
2844 sh
->check_state
= check_state_compute_run
;
2845 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2846 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2847 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
2848 set_bit(R5_Wantcompute
,
2849 &sh
->dev
[pd_idx
].flags
);
2851 target
= &sh
->ops
.target2
;
2854 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
2855 set_bit(R5_Wantcompute
,
2856 &sh
->dev
[qd_idx
].flags
);
2863 case check_state_compute_run
:
2866 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
2867 __func__
, sh
->check_state
,
2868 (unsigned long long) sh
->sector
);
2873 static void handle_stripe_expansion(raid5_conf_t
*conf
, struct stripe_head
*sh
)
2877 /* We have read all the blocks in this stripe and now we need to
2878 * copy some of them into a target stripe for expand.
2880 struct dma_async_tx_descriptor
*tx
= NULL
;
2881 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
2882 for (i
= 0; i
< sh
->disks
; i
++)
2883 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
2885 struct stripe_head
*sh2
;
2886 struct async_submit_ctl submit
;
2888 sector_t bn
= compute_blocknr(sh
, i
, 1);
2889 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
2891 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
2893 /* so far only the early blocks of this stripe
2894 * have been requested. When later blocks
2895 * get requested, we will try again
2898 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
2899 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
2900 /* must have already done this block */
2901 release_stripe(sh2
);
2905 /* place all the copies on one channel */
2906 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
2907 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
2908 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
2911 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
2912 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
2913 for (j
= 0; j
< conf
->raid_disks
; j
++)
2914 if (j
!= sh2
->pd_idx
&&
2916 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
2918 if (j
== conf
->raid_disks
) {
2919 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
2920 set_bit(STRIPE_HANDLE
, &sh2
->state
);
2922 release_stripe(sh2
);
2925 /* done submitting copies, wait for them to complete */
2928 dma_wait_for_async_tx(tx
);
2934 * handle_stripe - do things to a stripe.
2936 * We lock the stripe and then examine the state of various bits
2937 * to see what needs to be done.
2939 * return some read request which now have data
2940 * return some write requests which are safely on disc
2941 * schedule a read on some buffers
2942 * schedule a write of some buffers
2943 * return confirmation of parity correctness
2945 * buffers are taken off read_list or write_list, and bh_cache buffers
2946 * get BH_Lock set before the stripe lock is released.
2950 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
2952 raid5_conf_t
*conf
= sh
->raid_conf
;
2953 int disks
= sh
->disks
;
2957 memset(s
, 0, sizeof(*s
));
2959 s
->syncing
= test_bit(STRIPE_SYNCING
, &sh
->state
);
2960 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
2961 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
2962 s
->failed_num
[0] = -1;
2963 s
->failed_num
[1] = -1;
2965 /* Now to look around and see what can be done */
2967 spin_lock_irq(&conf
->device_lock
);
2968 for (i
=disks
; i
--; ) {
2976 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2977 i
, dev
->flags
, dev
->toread
, dev
->towrite
, dev
->written
);
2978 /* maybe we can reply to a read
2980 * new wantfill requests are only permitted while
2981 * ops_complete_biofill is guaranteed to be inactive
2983 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
2984 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
2985 set_bit(R5_Wantfill
, &dev
->flags
);
2987 /* now count some things */
2988 if (test_bit(R5_LOCKED
, &dev
->flags
))
2990 if (test_bit(R5_UPTODATE
, &dev
->flags
))
2992 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
2994 BUG_ON(s
->compute
> 2);
2997 if (test_bit(R5_Wantfill
, &dev
->flags
))
2999 else if (dev
->toread
)
3003 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3008 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3010 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3011 &first_bad
, &bad_sectors
);
3012 if (s
->blocked_rdev
== NULL
3013 && (test_bit(Blocked
, &rdev
->flags
)
3016 set_bit(BlockedBadBlocks
,
3018 s
->blocked_rdev
= rdev
;
3019 atomic_inc(&rdev
->nr_pending
);
3022 clear_bit(R5_Insync
, &dev
->flags
);
3026 /* also not in-sync */
3027 if (!test_bit(WriteErrorSeen
, &rdev
->flags
)) {
3028 /* treat as in-sync, but with a read error
3029 * which we can now try to correct
3031 set_bit(R5_Insync
, &dev
->flags
);
3032 set_bit(R5_ReadError
, &dev
->flags
);
3034 } else if (test_bit(In_sync
, &rdev
->flags
))
3035 set_bit(R5_Insync
, &dev
->flags
);
3037 /* in sync if before recovery_offset */
3038 if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
3039 set_bit(R5_Insync
, &dev
->flags
);
3041 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3042 /* The ReadError flag will just be confusing now */
3043 clear_bit(R5_ReadError
, &dev
->flags
);
3044 clear_bit(R5_ReWrite
, &dev
->flags
);
3046 if (test_bit(R5_ReadError
, &dev
->flags
))
3047 clear_bit(R5_Insync
, &dev
->flags
);
3048 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3050 s
->failed_num
[s
->failed
] = i
;
3054 spin_unlock_irq(&conf
->device_lock
);
3058 static void handle_stripe(struct stripe_head
*sh
)
3060 struct stripe_head_state s
;
3061 raid5_conf_t
*conf
= sh
->raid_conf
;
3064 int disks
= sh
->disks
;
3065 struct r5dev
*pdev
, *qdev
;
3067 clear_bit(STRIPE_HANDLE
, &sh
->state
);
3068 if (test_and_set_bit(STRIPE_ACTIVE
, &sh
->state
)) {
3069 /* already being handled, ensure it gets handled
3070 * again when current action finishes */
3071 set_bit(STRIPE_HANDLE
, &sh
->state
);
3075 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
3076 set_bit(STRIPE_SYNCING
, &sh
->state
);
3077 clear_bit(STRIPE_INSYNC
, &sh
->state
);
3079 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3081 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3082 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3083 (unsigned long long)sh
->sector
, sh
->state
,
3084 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
3085 sh
->check_state
, sh
->reconstruct_state
);
3087 analyse_stripe(sh
, &s
);
3089 if (unlikely(s
.blocked_rdev
)) {
3090 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
3091 s
.to_write
|| s
.written
) {
3092 set_bit(STRIPE_HANDLE
, &sh
->state
);
3095 /* There is nothing for the blocked_rdev to block */
3096 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
3097 s
.blocked_rdev
= NULL
;
3100 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
3101 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
3102 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
3105 pr_debug("locked=%d uptodate=%d to_read=%d"
3106 " to_write=%d failed=%d failed_num=%d,%d\n",
3107 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
3108 s
.failed_num
[0], s
.failed_num
[1]);
3109 /* check if the array has lost more than max_degraded devices and,
3110 * if so, some requests might need to be failed.
3112 if (s
.failed
> conf
->max_degraded
&& s
.to_read
+s
.to_write
+s
.written
)
3113 handle_failed_stripe(conf
, sh
, &s
, disks
, &s
.return_bi
);
3114 if (s
.failed
> conf
->max_degraded
&& s
.syncing
)
3115 handle_failed_sync(conf
, sh
, &s
);
3118 * might be able to return some write requests if the parity blocks
3119 * are safe, or on a failed drive
3121 pdev
= &sh
->dev
[sh
->pd_idx
];
3122 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
3123 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
3124 qdev
= &sh
->dev
[sh
->qd_idx
];
3125 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
3126 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
3130 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
3131 && !test_bit(R5_LOCKED
, &pdev
->flags
)
3132 && test_bit(R5_UPTODATE
, &pdev
->flags
)))) &&
3133 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
3134 && !test_bit(R5_LOCKED
, &qdev
->flags
)
3135 && test_bit(R5_UPTODATE
, &qdev
->flags
)))))
3136 handle_stripe_clean_event(conf
, sh
, disks
, &s
.return_bi
);
3138 /* Now we might consider reading some blocks, either to check/generate
3139 * parity, or to satisfy requests
3140 * or to load a block that is being partially written.
3142 if (s
.to_read
|| s
.non_overwrite
3143 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
3144 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
)) || s
.expanding
)
3145 handle_stripe_fill(sh
, &s
, disks
);
3147 /* Now we check to see if any write operations have recently
3151 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
3153 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
3154 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
3155 sh
->reconstruct_state
= reconstruct_state_idle
;
3157 /* All the 'written' buffers and the parity block are ready to
3158 * be written back to disk
3160 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
));
3161 BUG_ON(sh
->qd_idx
>= 0 &&
3162 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
));
3163 for (i
= disks
; i
--; ) {
3164 struct r5dev
*dev
= &sh
->dev
[i
];
3165 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
3166 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3168 pr_debug("Writing block %d\n", i
);
3169 set_bit(R5_Wantwrite
, &dev
->flags
);
3172 if (!test_bit(R5_Insync
, &dev
->flags
) ||
3173 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3175 set_bit(STRIPE_INSYNC
, &sh
->state
);
3178 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3179 s
.dec_preread_active
= 1;
3182 /* Now to consider new write requests and what else, if anything
3183 * should be read. We do not handle new writes when:
3184 * 1/ A 'write' operation (copy+xor) is already in flight.
3185 * 2/ A 'check' operation is in flight, as it may clobber the parity
3188 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
3189 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
3191 /* maybe we need to check and possibly fix the parity for this stripe
3192 * Any reads will already have been scheduled, so we just see if enough
3193 * data is available. The parity check is held off while parity
3194 * dependent operations are in flight.
3196 if (sh
->check_state
||
3197 (s
.syncing
&& s
.locked
== 0 &&
3198 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
3199 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
3200 if (conf
->level
== 6)
3201 handle_parity_checks6(conf
, sh
, &s
, disks
);
3203 handle_parity_checks5(conf
, sh
, &s
, disks
);
3206 if (s
.syncing
&& s
.locked
== 0 && test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3207 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3208 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3211 /* If the failed drives are just a ReadError, then we might need
3212 * to progress the repair/check process
3214 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
3215 for (i
= 0; i
< s
.failed
; i
++) {
3216 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
3217 if (test_bit(R5_ReadError
, &dev
->flags
)
3218 && !test_bit(R5_LOCKED
, &dev
->flags
)
3219 && test_bit(R5_UPTODATE
, &dev
->flags
)
3221 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3222 set_bit(R5_Wantwrite
, &dev
->flags
);
3223 set_bit(R5_ReWrite
, &dev
->flags
);
3224 set_bit(R5_LOCKED
, &dev
->flags
);
3227 /* let's read it back */
3228 set_bit(R5_Wantread
, &dev
->flags
);
3229 set_bit(R5_LOCKED
, &dev
->flags
);
3236 /* Finish reconstruct operations initiated by the expansion process */
3237 if (sh
->reconstruct_state
== reconstruct_state_result
) {
3238 struct stripe_head
*sh_src
3239 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
3240 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
3241 /* sh cannot be written until sh_src has been read.
3242 * so arrange for sh to be delayed a little
3244 set_bit(STRIPE_DELAYED
, &sh
->state
);
3245 set_bit(STRIPE_HANDLE
, &sh
->state
);
3246 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
3248 atomic_inc(&conf
->preread_active_stripes
);
3249 release_stripe(sh_src
);
3253 release_stripe(sh_src
);
3255 sh
->reconstruct_state
= reconstruct_state_idle
;
3256 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3257 for (i
= conf
->raid_disks
; i
--; ) {
3258 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3259 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3264 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
3265 !sh
->reconstruct_state
) {
3266 /* Need to write out all blocks after computing parity */
3267 sh
->disks
= conf
->raid_disks
;
3268 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
3269 schedule_reconstruction(sh
, &s
, 1, 1);
3270 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
3271 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3272 atomic_dec(&conf
->reshape_stripes
);
3273 wake_up(&conf
->wait_for_overlap
);
3274 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3277 if (s
.expanding
&& s
.locked
== 0 &&
3278 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
3279 handle_stripe_expansion(conf
, sh
);
3282 /* wait for this device to become unblocked */
3283 if (unlikely(s
.blocked_rdev
))
3284 md_wait_for_blocked_rdev(s
.blocked_rdev
, conf
->mddev
);
3287 raid_run_ops(sh
, s
.ops_request
);
3292 if (s
.dec_preread_active
) {
3293 /* We delay this until after ops_run_io so that if make_request
3294 * is waiting on a flush, it won't continue until the writes
3295 * have actually been submitted.
3297 atomic_dec(&conf
->preread_active_stripes
);
3298 if (atomic_read(&conf
->preread_active_stripes
) <
3300 md_wakeup_thread(conf
->mddev
->thread
);
3303 return_io(s
.return_bi
);
3305 clear_bit(STRIPE_ACTIVE
, &sh
->state
);
3308 static void raid5_activate_delayed(raid5_conf_t
*conf
)
3310 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
3311 while (!list_empty(&conf
->delayed_list
)) {
3312 struct list_head
*l
= conf
->delayed_list
.next
;
3313 struct stripe_head
*sh
;
3314 sh
= list_entry(l
, struct stripe_head
, lru
);
3316 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3317 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3318 atomic_inc(&conf
->preread_active_stripes
);
3319 list_add_tail(&sh
->lru
, &conf
->hold_list
);
3324 static void activate_bit_delay(raid5_conf_t
*conf
)
3326 /* device_lock is held */
3327 struct list_head head
;
3328 list_add(&head
, &conf
->bitmap_list
);
3329 list_del_init(&conf
->bitmap_list
);
3330 while (!list_empty(&head
)) {
3331 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
3332 list_del_init(&sh
->lru
);
3333 atomic_inc(&sh
->count
);
3334 __release_stripe(conf
, sh
);
3338 int md_raid5_congested(mddev_t
*mddev
, int bits
)
3340 raid5_conf_t
*conf
= mddev
->private;
3342 /* No difference between reads and writes. Just check
3343 * how busy the stripe_cache is
3346 if (conf
->inactive_blocked
)
3350 if (list_empty_careful(&conf
->inactive_list
))
3355 EXPORT_SYMBOL_GPL(md_raid5_congested
);
3357 static int raid5_congested(void *data
, int bits
)
3359 mddev_t
*mddev
= data
;
3361 return mddev_congested(mddev
, bits
) ||
3362 md_raid5_congested(mddev
, bits
);
3365 /* We want read requests to align with chunks where possible,
3366 * but write requests don't need to.
3368 static int raid5_mergeable_bvec(struct request_queue
*q
,
3369 struct bvec_merge_data
*bvm
,
3370 struct bio_vec
*biovec
)
3372 mddev_t
*mddev
= q
->queuedata
;
3373 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
3375 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3376 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
3378 if ((bvm
->bi_rw
& 1) == WRITE
)
3379 return biovec
->bv_len
; /* always allow writes to be mergeable */
3381 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3382 chunk_sectors
= mddev
->new_chunk_sectors
;
3383 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
3384 if (max
< 0) max
= 0;
3385 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
3386 return biovec
->bv_len
;
3392 static int in_chunk_boundary(mddev_t
*mddev
, struct bio
*bio
)
3394 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
3395 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3396 unsigned int bio_sectors
= bio
->bi_size
>> 9;
3398 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3399 chunk_sectors
= mddev
->new_chunk_sectors
;
3400 return chunk_sectors
>=
3401 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
3405 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3406 * later sampled by raid5d.
3408 static void add_bio_to_retry(struct bio
*bi
,raid5_conf_t
*conf
)
3410 unsigned long flags
;
3412 spin_lock_irqsave(&conf
->device_lock
, flags
);
3414 bi
->bi_next
= conf
->retry_read_aligned_list
;
3415 conf
->retry_read_aligned_list
= bi
;
3417 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3418 md_wakeup_thread(conf
->mddev
->thread
);
3422 static struct bio
*remove_bio_from_retry(raid5_conf_t
*conf
)
3426 bi
= conf
->retry_read_aligned
;
3428 conf
->retry_read_aligned
= NULL
;
3431 bi
= conf
->retry_read_aligned_list
;
3433 conf
->retry_read_aligned_list
= bi
->bi_next
;
3436 * this sets the active strip count to 1 and the processed
3437 * strip count to zero (upper 8 bits)
3439 bi
->bi_phys_segments
= 1; /* biased count of active stripes */
3447 * The "raid5_align_endio" should check if the read succeeded and if it
3448 * did, call bio_endio on the original bio (having bio_put the new bio
3450 * If the read failed..
3452 static void raid5_align_endio(struct bio
*bi
, int error
)
3454 struct bio
* raid_bi
= bi
->bi_private
;
3457 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3462 rdev
= (void*)raid_bi
->bi_next
;
3463 raid_bi
->bi_next
= NULL
;
3464 mddev
= rdev
->mddev
;
3465 conf
= mddev
->private;
3467 rdev_dec_pending(rdev
, conf
->mddev
);
3469 if (!error
&& uptodate
) {
3470 bio_endio(raid_bi
, 0);
3471 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
3472 wake_up(&conf
->wait_for_stripe
);
3477 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3479 add_bio_to_retry(raid_bi
, conf
);
3482 static int bio_fits_rdev(struct bio
*bi
)
3484 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
3486 if ((bi
->bi_size
>>9) > queue_max_sectors(q
))
3488 blk_recount_segments(q
, bi
);
3489 if (bi
->bi_phys_segments
> queue_max_segments(q
))
3492 if (q
->merge_bvec_fn
)
3493 /* it's too hard to apply the merge_bvec_fn at this stage,
3502 static int chunk_aligned_read(mddev_t
*mddev
, struct bio
* raid_bio
)
3504 raid5_conf_t
*conf
= mddev
->private;
3506 struct bio
* align_bi
;
3509 if (!in_chunk_boundary(mddev
, raid_bio
)) {
3510 pr_debug("chunk_aligned_read : non aligned\n");
3514 * use bio_clone_mddev to make a copy of the bio
3516 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
3520 * set bi_end_io to a new function, and set bi_private to the
3523 align_bi
->bi_end_io
= raid5_align_endio
;
3524 align_bi
->bi_private
= raid_bio
;
3528 align_bi
->bi_sector
= raid5_compute_sector(conf
, raid_bio
->bi_sector
,
3533 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
3534 if (rdev
&& test_bit(In_sync
, &rdev
->flags
)) {
3538 atomic_inc(&rdev
->nr_pending
);
3540 raid_bio
->bi_next
= (void*)rdev
;
3541 align_bi
->bi_bdev
= rdev
->bdev
;
3542 align_bi
->bi_flags
&= ~(1 << BIO_SEG_VALID
);
3543 align_bi
->bi_sector
+= rdev
->data_offset
;
3545 if (!bio_fits_rdev(align_bi
) ||
3546 is_badblock(rdev
, align_bi
->bi_sector
, align_bi
->bi_size
>>9,
3547 &first_bad
, &bad_sectors
)) {
3548 /* too big in some way, or has a known bad block */
3550 rdev_dec_pending(rdev
, mddev
);
3554 spin_lock_irq(&conf
->device_lock
);
3555 wait_event_lock_irq(conf
->wait_for_stripe
,
3557 conf
->device_lock
, /* nothing */);
3558 atomic_inc(&conf
->active_aligned_reads
);
3559 spin_unlock_irq(&conf
->device_lock
);
3561 generic_make_request(align_bi
);
3570 /* __get_priority_stripe - get the next stripe to process
3572 * Full stripe writes are allowed to pass preread active stripes up until
3573 * the bypass_threshold is exceeded. In general the bypass_count
3574 * increments when the handle_list is handled before the hold_list; however, it
3575 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3576 * stripe with in flight i/o. The bypass_count will be reset when the
3577 * head of the hold_list has changed, i.e. the head was promoted to the
3580 static struct stripe_head
*__get_priority_stripe(raid5_conf_t
*conf
)
3582 struct stripe_head
*sh
;
3584 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3586 list_empty(&conf
->handle_list
) ? "empty" : "busy",
3587 list_empty(&conf
->hold_list
) ? "empty" : "busy",
3588 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
3590 if (!list_empty(&conf
->handle_list
)) {
3591 sh
= list_entry(conf
->handle_list
.next
, typeof(*sh
), lru
);
3593 if (list_empty(&conf
->hold_list
))
3594 conf
->bypass_count
= 0;
3595 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
3596 if (conf
->hold_list
.next
== conf
->last_hold
)
3597 conf
->bypass_count
++;
3599 conf
->last_hold
= conf
->hold_list
.next
;
3600 conf
->bypass_count
-= conf
->bypass_threshold
;
3601 if (conf
->bypass_count
< 0)
3602 conf
->bypass_count
= 0;
3605 } else if (!list_empty(&conf
->hold_list
) &&
3606 ((conf
->bypass_threshold
&&
3607 conf
->bypass_count
> conf
->bypass_threshold
) ||
3608 atomic_read(&conf
->pending_full_writes
) == 0)) {
3609 sh
= list_entry(conf
->hold_list
.next
,
3611 conf
->bypass_count
-= conf
->bypass_threshold
;
3612 if (conf
->bypass_count
< 0)
3613 conf
->bypass_count
= 0;
3617 list_del_init(&sh
->lru
);
3618 atomic_inc(&sh
->count
);
3619 BUG_ON(atomic_read(&sh
->count
) != 1);
3623 static int make_request(mddev_t
*mddev
, struct bio
* bi
)
3625 raid5_conf_t
*conf
= mddev
->private;
3627 sector_t new_sector
;
3628 sector_t logical_sector
, last_sector
;
3629 struct stripe_head
*sh
;
3630 const int rw
= bio_data_dir(bi
);
3634 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
3635 md_flush_request(mddev
, bi
);
3639 md_write_start(mddev
, bi
);
3642 mddev
->reshape_position
== MaxSector
&&
3643 chunk_aligned_read(mddev
,bi
))
3646 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
3647 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
3649 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
3651 plugged
= mddev_check_plugged(mddev
);
3652 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
3654 int disks
, data_disks
;
3659 disks
= conf
->raid_disks
;
3660 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
3661 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
3662 /* spinlock is needed as reshape_progress may be
3663 * 64bit on a 32bit platform, and so it might be
3664 * possible to see a half-updated value
3665 * Of course reshape_progress could change after
3666 * the lock is dropped, so once we get a reference
3667 * to the stripe that we think it is, we will have
3670 spin_lock_irq(&conf
->device_lock
);
3671 if (mddev
->delta_disks
< 0
3672 ? logical_sector
< conf
->reshape_progress
3673 : logical_sector
>= conf
->reshape_progress
) {
3674 disks
= conf
->previous_raid_disks
;
3677 if (mddev
->delta_disks
< 0
3678 ? logical_sector
< conf
->reshape_safe
3679 : logical_sector
>= conf
->reshape_safe
) {
3680 spin_unlock_irq(&conf
->device_lock
);
3685 spin_unlock_irq(&conf
->device_lock
);
3687 data_disks
= disks
- conf
->max_degraded
;
3689 new_sector
= raid5_compute_sector(conf
, logical_sector
,
3692 pr_debug("raid456: make_request, sector %llu logical %llu\n",
3693 (unsigned long long)new_sector
,
3694 (unsigned long long)logical_sector
);
3696 sh
= get_active_stripe(conf
, new_sector
, previous
,
3697 (bi
->bi_rw
&RWA_MASK
), 0);
3699 if (unlikely(previous
)) {
3700 /* expansion might have moved on while waiting for a
3701 * stripe, so we must do the range check again.
3702 * Expansion could still move past after this
3703 * test, but as we are holding a reference to
3704 * 'sh', we know that if that happens,
3705 * STRIPE_EXPANDING will get set and the expansion
3706 * won't proceed until we finish with the stripe.
3709 spin_lock_irq(&conf
->device_lock
);
3710 if (mddev
->delta_disks
< 0
3711 ? logical_sector
>= conf
->reshape_progress
3712 : logical_sector
< conf
->reshape_progress
)
3713 /* mismatch, need to try again */
3715 spin_unlock_irq(&conf
->device_lock
);
3724 logical_sector
>= mddev
->suspend_lo
&&
3725 logical_sector
< mddev
->suspend_hi
) {
3727 /* As the suspend_* range is controlled by
3728 * userspace, we want an interruptible
3731 flush_signals(current
);
3732 prepare_to_wait(&conf
->wait_for_overlap
,
3733 &w
, TASK_INTERRUPTIBLE
);
3734 if (logical_sector
>= mddev
->suspend_lo
&&
3735 logical_sector
< mddev
->suspend_hi
)
3740 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
3741 !add_stripe_bio(sh
, bi
, dd_idx
, rw
)) {
3742 /* Stripe is busy expanding or
3743 * add failed due to overlap. Flush everything
3746 md_wakeup_thread(mddev
->thread
);
3751 finish_wait(&conf
->wait_for_overlap
, &w
);
3752 set_bit(STRIPE_HANDLE
, &sh
->state
);
3753 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3754 if ((bi
->bi_rw
& REQ_SYNC
) &&
3755 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3756 atomic_inc(&conf
->preread_active_stripes
);
3759 /* cannot get stripe for read-ahead, just give-up */
3760 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3761 finish_wait(&conf
->wait_for_overlap
, &w
);
3767 md_wakeup_thread(mddev
->thread
);
3769 spin_lock_irq(&conf
->device_lock
);
3770 remaining
= raid5_dec_bi_phys_segments(bi
);
3771 spin_unlock_irq(&conf
->device_lock
);
3772 if (remaining
== 0) {
3775 md_write_end(mddev
);
3783 static sector_t
raid5_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
);
3785 static sector_t
reshape_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
)
3787 /* reshaping is quite different to recovery/resync so it is
3788 * handled quite separately ... here.
3790 * On each call to sync_request, we gather one chunk worth of
3791 * destination stripes and flag them as expanding.
3792 * Then we find all the source stripes and request reads.
3793 * As the reads complete, handle_stripe will copy the data
3794 * into the destination stripe and release that stripe.
3796 raid5_conf_t
*conf
= mddev
->private;
3797 struct stripe_head
*sh
;
3798 sector_t first_sector
, last_sector
;
3799 int raid_disks
= conf
->previous_raid_disks
;
3800 int data_disks
= raid_disks
- conf
->max_degraded
;
3801 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
3804 sector_t writepos
, readpos
, safepos
;
3805 sector_t stripe_addr
;
3806 int reshape_sectors
;
3807 struct list_head stripes
;
3809 if (sector_nr
== 0) {
3810 /* If restarting in the middle, skip the initial sectors */
3811 if (mddev
->delta_disks
< 0 &&
3812 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
3813 sector_nr
= raid5_size(mddev
, 0, 0)
3814 - conf
->reshape_progress
;
3815 } else if (mddev
->delta_disks
>= 0 &&
3816 conf
->reshape_progress
> 0)
3817 sector_nr
= conf
->reshape_progress
;
3818 sector_div(sector_nr
, new_data_disks
);
3820 mddev
->curr_resync_completed
= sector_nr
;
3821 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
3827 /* We need to process a full chunk at a time.
3828 * If old and new chunk sizes differ, we need to process the
3831 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
)
3832 reshape_sectors
= mddev
->new_chunk_sectors
;
3834 reshape_sectors
= mddev
->chunk_sectors
;
3836 /* we update the metadata when there is more than 3Meg
3837 * in the block range (that is rather arbitrary, should
3838 * probably be time based) or when the data about to be
3839 * copied would over-write the source of the data at
3840 * the front of the range.
3841 * i.e. one new_stripe along from reshape_progress new_maps
3842 * to after where reshape_safe old_maps to
3844 writepos
= conf
->reshape_progress
;
3845 sector_div(writepos
, new_data_disks
);
3846 readpos
= conf
->reshape_progress
;
3847 sector_div(readpos
, data_disks
);
3848 safepos
= conf
->reshape_safe
;
3849 sector_div(safepos
, data_disks
);
3850 if (mddev
->delta_disks
< 0) {
3851 writepos
-= min_t(sector_t
, reshape_sectors
, writepos
);
3852 readpos
+= reshape_sectors
;
3853 safepos
+= reshape_sectors
;
3855 writepos
+= reshape_sectors
;
3856 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
3857 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
3860 /* 'writepos' is the most advanced device address we might write.
3861 * 'readpos' is the least advanced device address we might read.
3862 * 'safepos' is the least address recorded in the metadata as having
3864 * If 'readpos' is behind 'writepos', then there is no way that we can
3865 * ensure safety in the face of a crash - that must be done by userspace
3866 * making a backup of the data. So in that case there is no particular
3867 * rush to update metadata.
3868 * Otherwise if 'safepos' is behind 'writepos', then we really need to
3869 * update the metadata to advance 'safepos' to match 'readpos' so that
3870 * we can be safe in the event of a crash.
3871 * So we insist on updating metadata if safepos is behind writepos and
3872 * readpos is beyond writepos.
3873 * In any case, update the metadata every 10 seconds.
3874 * Maybe that number should be configurable, but I'm not sure it is
3875 * worth it.... maybe it could be a multiple of safemode_delay???
3877 if ((mddev
->delta_disks
< 0
3878 ? (safepos
> writepos
&& readpos
< writepos
)
3879 : (safepos
< writepos
&& readpos
> writepos
)) ||
3880 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
3881 /* Cannot proceed until we've updated the superblock... */
3882 wait_event(conf
->wait_for_overlap
,
3883 atomic_read(&conf
->reshape_stripes
)==0);
3884 mddev
->reshape_position
= conf
->reshape_progress
;
3885 mddev
->curr_resync_completed
= sector_nr
;
3886 conf
->reshape_checkpoint
= jiffies
;
3887 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
3888 md_wakeup_thread(mddev
->thread
);
3889 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
3890 kthread_should_stop());
3891 spin_lock_irq(&conf
->device_lock
);
3892 conf
->reshape_safe
= mddev
->reshape_position
;
3893 spin_unlock_irq(&conf
->device_lock
);
3894 wake_up(&conf
->wait_for_overlap
);
3895 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
3898 if (mddev
->delta_disks
< 0) {
3899 BUG_ON(conf
->reshape_progress
== 0);
3900 stripe_addr
= writepos
;
3901 BUG_ON((mddev
->dev_sectors
&
3902 ~((sector_t
)reshape_sectors
- 1))
3903 - reshape_sectors
- stripe_addr
3906 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
3907 stripe_addr
= sector_nr
;
3909 INIT_LIST_HEAD(&stripes
);
3910 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
3912 int skipped_disk
= 0;
3913 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
3914 set_bit(STRIPE_EXPANDING
, &sh
->state
);
3915 atomic_inc(&conf
->reshape_stripes
);
3916 /* If any of this stripe is beyond the end of the old
3917 * array, then we need to zero those blocks
3919 for (j
=sh
->disks
; j
--;) {
3921 if (j
== sh
->pd_idx
)
3923 if (conf
->level
== 6 &&
3926 s
= compute_blocknr(sh
, j
, 0);
3927 if (s
< raid5_size(mddev
, 0, 0)) {
3931 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
3932 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
3933 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
3935 if (!skipped_disk
) {
3936 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3937 set_bit(STRIPE_HANDLE
, &sh
->state
);
3939 list_add(&sh
->lru
, &stripes
);
3941 spin_lock_irq(&conf
->device_lock
);
3942 if (mddev
->delta_disks
< 0)
3943 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
3945 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
3946 spin_unlock_irq(&conf
->device_lock
);
3947 /* Ok, those stripe are ready. We can start scheduling
3948 * reads on the source stripes.
3949 * The source stripes are determined by mapping the first and last
3950 * block on the destination stripes.
3953 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
3956 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
3957 * new_data_disks
- 1),
3959 if (last_sector
>= mddev
->dev_sectors
)
3960 last_sector
= mddev
->dev_sectors
- 1;
3961 while (first_sector
<= last_sector
) {
3962 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
3963 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3964 set_bit(STRIPE_HANDLE
, &sh
->state
);
3966 first_sector
+= STRIPE_SECTORS
;
3968 /* Now that the sources are clearly marked, we can release
3969 * the destination stripes
3971 while (!list_empty(&stripes
)) {
3972 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
3973 list_del_init(&sh
->lru
);
3976 /* If this takes us to the resync_max point where we have to pause,
3977 * then we need to write out the superblock.
3979 sector_nr
+= reshape_sectors
;
3980 if ((sector_nr
- mddev
->curr_resync_completed
) * 2
3981 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
3982 /* Cannot proceed until we've updated the superblock... */
3983 wait_event(conf
->wait_for_overlap
,
3984 atomic_read(&conf
->reshape_stripes
) == 0);
3985 mddev
->reshape_position
= conf
->reshape_progress
;
3986 mddev
->curr_resync_completed
= sector_nr
;
3987 conf
->reshape_checkpoint
= jiffies
;
3988 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
3989 md_wakeup_thread(mddev
->thread
);
3990 wait_event(mddev
->sb_wait
,
3991 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
3992 || kthread_should_stop());
3993 spin_lock_irq(&conf
->device_lock
);
3994 conf
->reshape_safe
= mddev
->reshape_position
;
3995 spin_unlock_irq(&conf
->device_lock
);
3996 wake_up(&conf
->wait_for_overlap
);
3997 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
3999 return reshape_sectors
;
4002 /* FIXME go_faster isn't used */
4003 static inline sector_t
sync_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
4005 raid5_conf_t
*conf
= mddev
->private;
4006 struct stripe_head
*sh
;
4007 sector_t max_sector
= mddev
->dev_sectors
;
4008 sector_t sync_blocks
;
4009 int still_degraded
= 0;
4012 if (sector_nr
>= max_sector
) {
4013 /* just being told to finish up .. nothing much to do */
4015 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
4020 if (mddev
->curr_resync
< max_sector
) /* aborted */
4021 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
4023 else /* completed sync */
4025 bitmap_close_sync(mddev
->bitmap
);
4030 /* Allow raid5_quiesce to complete */
4031 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
4033 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
4034 return reshape_request(mddev
, sector_nr
, skipped
);
4036 /* No need to check resync_max as we never do more than one
4037 * stripe, and as resync_max will always be on a chunk boundary,
4038 * if the check in md_do_sync didn't fire, there is no chance
4039 * of overstepping resync_max here
4042 /* if there is too many failed drives and we are trying
4043 * to resync, then assert that we are finished, because there is
4044 * nothing we can do.
4046 if (mddev
->degraded
>= conf
->max_degraded
&&
4047 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
4048 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
4052 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
4053 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
4054 !conf
->fullsync
&& sync_blocks
>= STRIPE_SECTORS
) {
4055 /* we can skip this block, and probably more */
4056 sync_blocks
/= STRIPE_SECTORS
;
4058 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
4062 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
4064 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
4066 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
4067 /* make sure we don't swamp the stripe cache if someone else
4068 * is trying to get access
4070 schedule_timeout_uninterruptible(1);
4072 /* Need to check if array will still be degraded after recovery/resync
4073 * We don't need to check the 'failed' flag as when that gets set,
4076 for (i
= 0; i
< conf
->raid_disks
; i
++)
4077 if (conf
->disks
[i
].rdev
== NULL
)
4080 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
4082 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
4087 return STRIPE_SECTORS
;
4090 static int retry_aligned_read(raid5_conf_t
*conf
, struct bio
*raid_bio
)
4092 /* We may not be able to submit a whole bio at once as there
4093 * may not be enough stripe_heads available.
4094 * We cannot pre-allocate enough stripe_heads as we may need
4095 * more than exist in the cache (if we allow ever large chunks).
4096 * So we do one stripe head at a time and record in
4097 * ->bi_hw_segments how many have been done.
4099 * We *know* that this entire raid_bio is in one chunk, so
4100 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4102 struct stripe_head
*sh
;
4104 sector_t sector
, logical_sector
, last_sector
;
4109 logical_sector
= raid_bio
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4110 sector
= raid5_compute_sector(conf
, logical_sector
,
4112 last_sector
= raid_bio
->bi_sector
+ (raid_bio
->bi_size
>>9);
4114 for (; logical_sector
< last_sector
;
4115 logical_sector
+= STRIPE_SECTORS
,
4116 sector
+= STRIPE_SECTORS
,
4119 if (scnt
< raid5_bi_hw_segments(raid_bio
))
4120 /* already done this stripe */
4123 sh
= get_active_stripe(conf
, sector
, 0, 1, 0);
4126 /* failed to get a stripe - must wait */
4127 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4128 conf
->retry_read_aligned
= raid_bio
;
4132 set_bit(R5_ReadError
, &sh
->dev
[dd_idx
].flags
);
4133 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0)) {
4135 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4136 conf
->retry_read_aligned
= raid_bio
;
4144 spin_lock_irq(&conf
->device_lock
);
4145 remaining
= raid5_dec_bi_phys_segments(raid_bio
);
4146 spin_unlock_irq(&conf
->device_lock
);
4148 bio_endio(raid_bio
, 0);
4149 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4150 wake_up(&conf
->wait_for_stripe
);
4156 * This is our raid5 kernel thread.
4158 * We scan the hash table for stripes which can be handled now.
4159 * During the scan, completed stripes are saved for us by the interrupt
4160 * handler, so that they will not have to wait for our next wakeup.
4162 static void raid5d(mddev_t
*mddev
)
4164 struct stripe_head
*sh
;
4165 raid5_conf_t
*conf
= mddev
->private;
4167 struct blk_plug plug
;
4169 pr_debug("+++ raid5d active\n");
4171 md_check_recovery(mddev
);
4173 blk_start_plug(&plug
);
4175 spin_lock_irq(&conf
->device_lock
);
4179 if (atomic_read(&mddev
->plug_cnt
) == 0 &&
4180 !list_empty(&conf
->bitmap_list
)) {
4181 /* Now is a good time to flush some bitmap updates */
4183 spin_unlock_irq(&conf
->device_lock
);
4184 bitmap_unplug(mddev
->bitmap
);
4185 spin_lock_irq(&conf
->device_lock
);
4186 conf
->seq_write
= conf
->seq_flush
;
4187 activate_bit_delay(conf
);
4189 if (atomic_read(&mddev
->plug_cnt
) == 0)
4190 raid5_activate_delayed(conf
);
4192 while ((bio
= remove_bio_from_retry(conf
))) {
4194 spin_unlock_irq(&conf
->device_lock
);
4195 ok
= retry_aligned_read(conf
, bio
);
4196 spin_lock_irq(&conf
->device_lock
);
4202 sh
= __get_priority_stripe(conf
);
4206 spin_unlock_irq(&conf
->device_lock
);
4213 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
))
4214 md_check_recovery(mddev
);
4216 spin_lock_irq(&conf
->device_lock
);
4218 pr_debug("%d stripes handled\n", handled
);
4220 spin_unlock_irq(&conf
->device_lock
);
4222 async_tx_issue_pending_all();
4223 blk_finish_plug(&plug
);
4225 pr_debug("--- raid5d inactive\n");
4229 raid5_show_stripe_cache_size(mddev_t
*mddev
, char *page
)
4231 raid5_conf_t
*conf
= mddev
->private;
4233 return sprintf(page
, "%d\n", conf
->max_nr_stripes
);
4239 raid5_set_cache_size(mddev_t
*mddev
, int size
)
4241 raid5_conf_t
*conf
= mddev
->private;
4244 if (size
<= 16 || size
> 32768)
4246 while (size
< conf
->max_nr_stripes
) {
4247 if (drop_one_stripe(conf
))
4248 conf
->max_nr_stripes
--;
4252 err
= md_allow_write(mddev
);
4255 while (size
> conf
->max_nr_stripes
) {
4256 if (grow_one_stripe(conf
))
4257 conf
->max_nr_stripes
++;
4262 EXPORT_SYMBOL(raid5_set_cache_size
);
4265 raid5_store_stripe_cache_size(mddev_t
*mddev
, const char *page
, size_t len
)
4267 raid5_conf_t
*conf
= mddev
->private;
4271 if (len
>= PAGE_SIZE
)
4276 if (strict_strtoul(page
, 10, &new))
4278 err
= raid5_set_cache_size(mddev
, new);
4284 static struct md_sysfs_entry
4285 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
4286 raid5_show_stripe_cache_size
,
4287 raid5_store_stripe_cache_size
);
4290 raid5_show_preread_threshold(mddev_t
*mddev
, char *page
)
4292 raid5_conf_t
*conf
= mddev
->private;
4294 return sprintf(page
, "%d\n", conf
->bypass_threshold
);
4300 raid5_store_preread_threshold(mddev_t
*mddev
, const char *page
, size_t len
)
4302 raid5_conf_t
*conf
= mddev
->private;
4304 if (len
>= PAGE_SIZE
)
4309 if (strict_strtoul(page
, 10, &new))
4311 if (new > conf
->max_nr_stripes
)
4313 conf
->bypass_threshold
= new;
4317 static struct md_sysfs_entry
4318 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
4320 raid5_show_preread_threshold
,
4321 raid5_store_preread_threshold
);
4324 stripe_cache_active_show(mddev_t
*mddev
, char *page
)
4326 raid5_conf_t
*conf
= mddev
->private;
4328 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
4333 static struct md_sysfs_entry
4334 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
4336 static struct attribute
*raid5_attrs
[] = {
4337 &raid5_stripecache_size
.attr
,
4338 &raid5_stripecache_active
.attr
,
4339 &raid5_preread_bypass_threshold
.attr
,
4342 static struct attribute_group raid5_attrs_group
= {
4344 .attrs
= raid5_attrs
,
4348 raid5_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
)
4350 raid5_conf_t
*conf
= mddev
->private;
4353 sectors
= mddev
->dev_sectors
;
4355 /* size is defined by the smallest of previous and new size */
4356 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
4358 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
4359 sectors
&= ~((sector_t
)mddev
->new_chunk_sectors
- 1);
4360 return sectors
* (raid_disks
- conf
->max_degraded
);
4363 static void raid5_free_percpu(raid5_conf_t
*conf
)
4365 struct raid5_percpu
*percpu
;
4372 for_each_possible_cpu(cpu
) {
4373 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4374 safe_put_page(percpu
->spare_page
);
4375 kfree(percpu
->scribble
);
4377 #ifdef CONFIG_HOTPLUG_CPU
4378 unregister_cpu_notifier(&conf
->cpu_notify
);
4382 free_percpu(conf
->percpu
);
4385 static void free_conf(raid5_conf_t
*conf
)
4387 shrink_stripes(conf
);
4388 raid5_free_percpu(conf
);
4390 kfree(conf
->stripe_hashtbl
);
4394 #ifdef CONFIG_HOTPLUG_CPU
4395 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
4398 raid5_conf_t
*conf
= container_of(nfb
, raid5_conf_t
, cpu_notify
);
4399 long cpu
= (long)hcpu
;
4400 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4403 case CPU_UP_PREPARE
:
4404 case CPU_UP_PREPARE_FROZEN
:
4405 if (conf
->level
== 6 && !percpu
->spare_page
)
4406 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
4407 if (!percpu
->scribble
)
4408 percpu
->scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
4410 if (!percpu
->scribble
||
4411 (conf
->level
== 6 && !percpu
->spare_page
)) {
4412 safe_put_page(percpu
->spare_page
);
4413 kfree(percpu
->scribble
);
4414 pr_err("%s: failed memory allocation for cpu%ld\n",
4416 return notifier_from_errno(-ENOMEM
);
4420 case CPU_DEAD_FROZEN
:
4421 safe_put_page(percpu
->spare_page
);
4422 kfree(percpu
->scribble
);
4423 percpu
->spare_page
= NULL
;
4424 percpu
->scribble
= NULL
;
4433 static int raid5_alloc_percpu(raid5_conf_t
*conf
)
4436 struct page
*spare_page
;
4437 struct raid5_percpu __percpu
*allcpus
;
4441 allcpus
= alloc_percpu(struct raid5_percpu
);
4444 conf
->percpu
= allcpus
;
4448 for_each_present_cpu(cpu
) {
4449 if (conf
->level
== 6) {
4450 spare_page
= alloc_page(GFP_KERNEL
);
4455 per_cpu_ptr(conf
->percpu
, cpu
)->spare_page
= spare_page
;
4457 scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
4462 per_cpu_ptr(conf
->percpu
, cpu
)->scribble
= scribble
;
4464 #ifdef CONFIG_HOTPLUG_CPU
4465 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
4466 conf
->cpu_notify
.priority
= 0;
4468 err
= register_cpu_notifier(&conf
->cpu_notify
);
4475 static raid5_conf_t
*setup_conf(mddev_t
*mddev
)
4478 int raid_disk
, memory
, max_disks
;
4480 struct disk_info
*disk
;
4482 if (mddev
->new_level
!= 5
4483 && mddev
->new_level
!= 4
4484 && mddev
->new_level
!= 6) {
4485 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4486 mdname(mddev
), mddev
->new_level
);
4487 return ERR_PTR(-EIO
);
4489 if ((mddev
->new_level
== 5
4490 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
4491 (mddev
->new_level
== 6
4492 && !algorithm_valid_raid6(mddev
->new_layout
))) {
4493 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
4494 mdname(mddev
), mddev
->new_layout
);
4495 return ERR_PTR(-EIO
);
4497 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
4498 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4499 mdname(mddev
), mddev
->raid_disks
);
4500 return ERR_PTR(-EINVAL
);
4503 if (!mddev
->new_chunk_sectors
||
4504 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
4505 !is_power_of_2(mddev
->new_chunk_sectors
)) {
4506 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
4507 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
4508 return ERR_PTR(-EINVAL
);
4511 conf
= kzalloc(sizeof(raid5_conf_t
), GFP_KERNEL
);
4514 spin_lock_init(&conf
->device_lock
);
4515 init_waitqueue_head(&conf
->wait_for_stripe
);
4516 init_waitqueue_head(&conf
->wait_for_overlap
);
4517 INIT_LIST_HEAD(&conf
->handle_list
);
4518 INIT_LIST_HEAD(&conf
->hold_list
);
4519 INIT_LIST_HEAD(&conf
->delayed_list
);
4520 INIT_LIST_HEAD(&conf
->bitmap_list
);
4521 INIT_LIST_HEAD(&conf
->inactive_list
);
4522 atomic_set(&conf
->active_stripes
, 0);
4523 atomic_set(&conf
->preread_active_stripes
, 0);
4524 atomic_set(&conf
->active_aligned_reads
, 0);
4525 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
4527 conf
->raid_disks
= mddev
->raid_disks
;
4528 if (mddev
->reshape_position
== MaxSector
)
4529 conf
->previous_raid_disks
= mddev
->raid_disks
;
4531 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4532 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
4533 conf
->scribble_len
= scribble_len(max_disks
);
4535 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
4540 conf
->mddev
= mddev
;
4542 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
4545 conf
->level
= mddev
->new_level
;
4546 if (raid5_alloc_percpu(conf
) != 0)
4549 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
4551 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
4552 raid_disk
= rdev
->raid_disk
;
4553 if (raid_disk
>= max_disks
4556 disk
= conf
->disks
+ raid_disk
;
4560 if (test_bit(In_sync
, &rdev
->flags
)) {
4561 char b
[BDEVNAME_SIZE
];
4562 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
4564 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
4565 } else if (rdev
->saved_raid_disk
!= raid_disk
)
4566 /* Cannot rely on bitmap to complete recovery */
4570 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
4571 conf
->level
= mddev
->new_level
;
4572 if (conf
->level
== 6)
4573 conf
->max_degraded
= 2;
4575 conf
->max_degraded
= 1;
4576 conf
->algorithm
= mddev
->new_layout
;
4577 conf
->max_nr_stripes
= NR_STRIPES
;
4578 conf
->reshape_progress
= mddev
->reshape_position
;
4579 if (conf
->reshape_progress
!= MaxSector
) {
4580 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
4581 conf
->prev_algo
= mddev
->layout
;
4584 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
4585 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
4586 if (grow_stripes(conf
, conf
->max_nr_stripes
)) {
4588 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4589 mdname(mddev
), memory
);
4592 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
4593 mdname(mddev
), memory
);
4595 conf
->thread
= md_register_thread(raid5d
, mddev
, NULL
);
4596 if (!conf
->thread
) {
4598 "md/raid:%s: couldn't allocate thread.\n",
4608 return ERR_PTR(-EIO
);
4610 return ERR_PTR(-ENOMEM
);
4614 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
4617 case ALGORITHM_PARITY_0
:
4618 if (raid_disk
< max_degraded
)
4621 case ALGORITHM_PARITY_N
:
4622 if (raid_disk
>= raid_disks
- max_degraded
)
4625 case ALGORITHM_PARITY_0_6
:
4626 if (raid_disk
== 0 ||
4627 raid_disk
== raid_disks
- 1)
4630 case ALGORITHM_LEFT_ASYMMETRIC_6
:
4631 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
4632 case ALGORITHM_LEFT_SYMMETRIC_6
:
4633 case ALGORITHM_RIGHT_SYMMETRIC_6
:
4634 if (raid_disk
== raid_disks
- 1)
4640 static int run(mddev_t
*mddev
)
4643 int working_disks
= 0;
4644 int dirty_parity_disks
= 0;
4646 sector_t reshape_offset
= 0;
4648 if (mddev
->recovery_cp
!= MaxSector
)
4649 printk(KERN_NOTICE
"md/raid:%s: not clean"
4650 " -- starting background reconstruction\n",
4652 if (mddev
->reshape_position
!= MaxSector
) {
4653 /* Check that we can continue the reshape.
4654 * Currently only disks can change, it must
4655 * increase, and we must be past the point where
4656 * a stripe over-writes itself
4658 sector_t here_new
, here_old
;
4660 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
4662 if (mddev
->new_level
!= mddev
->level
) {
4663 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
4664 "required - aborting.\n",
4668 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4669 /* reshape_position must be on a new-stripe boundary, and one
4670 * further up in new geometry must map after here in old
4673 here_new
= mddev
->reshape_position
;
4674 if (sector_div(here_new
, mddev
->new_chunk_sectors
*
4675 (mddev
->raid_disks
- max_degraded
))) {
4676 printk(KERN_ERR
"md/raid:%s: reshape_position not "
4677 "on a stripe boundary\n", mdname(mddev
));
4680 reshape_offset
= here_new
* mddev
->new_chunk_sectors
;
4681 /* here_new is the stripe we will write to */
4682 here_old
= mddev
->reshape_position
;
4683 sector_div(here_old
, mddev
->chunk_sectors
*
4684 (old_disks
-max_degraded
));
4685 /* here_old is the first stripe that we might need to read
4687 if (mddev
->delta_disks
== 0) {
4688 /* We cannot be sure it is safe to start an in-place
4689 * reshape. It is only safe if user-space if monitoring
4690 * and taking constant backups.
4691 * mdadm always starts a situation like this in
4692 * readonly mode so it can take control before
4693 * allowing any writes. So just check for that.
4695 if ((here_new
* mddev
->new_chunk_sectors
!=
4696 here_old
* mddev
->chunk_sectors
) ||
4698 printk(KERN_ERR
"md/raid:%s: in-place reshape must be started"
4699 " in read-only mode - aborting\n",
4703 } else if (mddev
->delta_disks
< 0
4704 ? (here_new
* mddev
->new_chunk_sectors
<=
4705 here_old
* mddev
->chunk_sectors
)
4706 : (here_new
* mddev
->new_chunk_sectors
>=
4707 here_old
* mddev
->chunk_sectors
)) {
4708 /* Reading from the same stripe as writing to - bad */
4709 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
4710 "auto-recovery - aborting.\n",
4714 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
4716 /* OK, we should be able to continue; */
4718 BUG_ON(mddev
->level
!= mddev
->new_level
);
4719 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
4720 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
4721 BUG_ON(mddev
->delta_disks
!= 0);
4724 if (mddev
->private == NULL
)
4725 conf
= setup_conf(mddev
);
4727 conf
= mddev
->private;
4730 return PTR_ERR(conf
);
4732 mddev
->thread
= conf
->thread
;
4733 conf
->thread
= NULL
;
4734 mddev
->private = conf
;
4737 * 0 for a fully functional array, 1 or 2 for a degraded array.
4739 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
4740 if (rdev
->raid_disk
< 0)
4742 if (test_bit(In_sync
, &rdev
->flags
)) {
4746 /* This disc is not fully in-sync. However if it
4747 * just stored parity (beyond the recovery_offset),
4748 * when we don't need to be concerned about the
4749 * array being dirty.
4750 * When reshape goes 'backwards', we never have
4751 * partially completed devices, so we only need
4752 * to worry about reshape going forwards.
4754 /* Hack because v0.91 doesn't store recovery_offset properly. */
4755 if (mddev
->major_version
== 0 &&
4756 mddev
->minor_version
> 90)
4757 rdev
->recovery_offset
= reshape_offset
;
4759 if (rdev
->recovery_offset
< reshape_offset
) {
4760 /* We need to check old and new layout */
4761 if (!only_parity(rdev
->raid_disk
,
4764 conf
->max_degraded
))
4767 if (!only_parity(rdev
->raid_disk
,
4769 conf
->previous_raid_disks
,
4770 conf
->max_degraded
))
4772 dirty_parity_disks
++;
4775 mddev
->degraded
= (max(conf
->raid_disks
, conf
->previous_raid_disks
)
4778 if (has_failed(conf
)) {
4779 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
4780 " (%d/%d failed)\n",
4781 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
4785 /* device size must be a multiple of chunk size */
4786 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
4787 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
4789 if (mddev
->degraded
> dirty_parity_disks
&&
4790 mddev
->recovery_cp
!= MaxSector
) {
4791 if (mddev
->ok_start_degraded
)
4793 "md/raid:%s: starting dirty degraded array"
4794 " - data corruption possible.\n",
4798 "md/raid:%s: cannot start dirty degraded array.\n",
4804 if (mddev
->degraded
== 0)
4805 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
4806 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
4807 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
4810 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
4811 " out of %d devices, algorithm %d\n",
4812 mdname(mddev
), conf
->level
,
4813 mddev
->raid_disks
- mddev
->degraded
,
4814 mddev
->raid_disks
, mddev
->new_layout
);
4816 print_raid5_conf(conf
);
4818 if (conf
->reshape_progress
!= MaxSector
) {
4819 conf
->reshape_safe
= conf
->reshape_progress
;
4820 atomic_set(&conf
->reshape_stripes
, 0);
4821 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4822 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4823 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4824 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4825 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4830 /* Ok, everything is just fine now */
4831 if (mddev
->to_remove
== &raid5_attrs_group
)
4832 mddev
->to_remove
= NULL
;
4833 else if (mddev
->kobj
.sd
&&
4834 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
4836 "raid5: failed to create sysfs attributes for %s\n",
4838 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
4842 /* read-ahead size must cover two whole stripes, which
4843 * is 2 * (datadisks) * chunksize where 'n' is the
4844 * number of raid devices
4846 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
4847 int stripe
= data_disks
*
4848 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4849 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4850 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4852 blk_queue_merge_bvec(mddev
->queue
, raid5_mergeable_bvec
);
4854 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
4855 mddev
->queue
->backing_dev_info
.congested_fn
= raid5_congested
;
4857 chunk_size
= mddev
->chunk_sectors
<< 9;
4858 blk_queue_io_min(mddev
->queue
, chunk_size
);
4859 blk_queue_io_opt(mddev
->queue
, chunk_size
*
4860 (conf
->raid_disks
- conf
->max_degraded
));
4862 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
4863 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
4864 rdev
->data_offset
<< 9);
4869 md_unregister_thread(mddev
->thread
);
4870 mddev
->thread
= NULL
;
4872 print_raid5_conf(conf
);
4875 mddev
->private = NULL
;
4876 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
4880 static int stop(mddev_t
*mddev
)
4882 raid5_conf_t
*conf
= mddev
->private;
4884 md_unregister_thread(mddev
->thread
);
4885 mddev
->thread
= NULL
;
4887 mddev
->queue
->backing_dev_info
.congested_fn
= NULL
;
4889 mddev
->private = NULL
;
4890 mddev
->to_remove
= &raid5_attrs_group
;
4895 static void print_sh(struct seq_file
*seq
, struct stripe_head
*sh
)
4899 seq_printf(seq
, "sh %llu, pd_idx %d, state %ld.\n",
4900 (unsigned long long)sh
->sector
, sh
->pd_idx
, sh
->state
);
4901 seq_printf(seq
, "sh %llu, count %d.\n",
4902 (unsigned long long)sh
->sector
, atomic_read(&sh
->count
));
4903 seq_printf(seq
, "sh %llu, ", (unsigned long long)sh
->sector
);
4904 for (i
= 0; i
< sh
->disks
; i
++) {
4905 seq_printf(seq
, "(cache%d: %p %ld) ",
4906 i
, sh
->dev
[i
].page
, sh
->dev
[i
].flags
);
4908 seq_printf(seq
, "\n");
4911 static void printall(struct seq_file
*seq
, raid5_conf_t
*conf
)
4913 struct stripe_head
*sh
;
4914 struct hlist_node
*hn
;
4917 spin_lock_irq(&conf
->device_lock
);
4918 for (i
= 0; i
< NR_HASH
; i
++) {
4919 hlist_for_each_entry(sh
, hn
, &conf
->stripe_hashtbl
[i
], hash
) {
4920 if (sh
->raid_conf
!= conf
)
4925 spin_unlock_irq(&conf
->device_lock
);
4929 static void status(struct seq_file
*seq
, mddev_t
*mddev
)
4931 raid5_conf_t
*conf
= mddev
->private;
4934 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
4935 mddev
->chunk_sectors
/ 2, mddev
->layout
);
4936 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
4937 for (i
= 0; i
< conf
->raid_disks
; i
++)
4938 seq_printf (seq
, "%s",
4939 conf
->disks
[i
].rdev
&&
4940 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
4941 seq_printf (seq
, "]");
4943 seq_printf (seq
, "\n");
4944 printall(seq
, conf
);
4948 static void print_raid5_conf (raid5_conf_t
*conf
)
4951 struct disk_info
*tmp
;
4953 printk(KERN_DEBUG
"RAID conf printout:\n");
4955 printk("(conf==NULL)\n");
4958 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
4960 conf
->raid_disks
- conf
->mddev
->degraded
);
4962 for (i
= 0; i
< conf
->raid_disks
; i
++) {
4963 char b
[BDEVNAME_SIZE
];
4964 tmp
= conf
->disks
+ i
;
4966 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
4967 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
4968 bdevname(tmp
->rdev
->bdev
, b
));
4972 static int raid5_spare_active(mddev_t
*mddev
)
4975 raid5_conf_t
*conf
= mddev
->private;
4976 struct disk_info
*tmp
;
4978 unsigned long flags
;
4980 for (i
= 0; i
< conf
->raid_disks
; i
++) {
4981 tmp
= conf
->disks
+ i
;
4983 && tmp
->rdev
->recovery_offset
== MaxSector
4984 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
4985 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
4987 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
4990 spin_lock_irqsave(&conf
->device_lock
, flags
);
4991 mddev
->degraded
-= count
;
4992 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
4993 print_raid5_conf(conf
);
4997 static int raid5_remove_disk(mddev_t
*mddev
, int number
)
4999 raid5_conf_t
*conf
= mddev
->private;
5002 struct disk_info
*p
= conf
->disks
+ number
;
5004 print_raid5_conf(conf
);
5007 if (number
>= conf
->raid_disks
&&
5008 conf
->reshape_progress
== MaxSector
)
5009 clear_bit(In_sync
, &rdev
->flags
);
5011 if (test_bit(In_sync
, &rdev
->flags
) ||
5012 atomic_read(&rdev
->nr_pending
)) {
5016 /* Only remove non-faulty devices if recovery
5019 if (!test_bit(Faulty
, &rdev
->flags
) &&
5020 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
5021 !has_failed(conf
) &&
5022 number
< conf
->raid_disks
) {
5028 if (atomic_read(&rdev
->nr_pending
)) {
5029 /* lost the race, try later */
5036 print_raid5_conf(conf
);
5040 static int raid5_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
5042 raid5_conf_t
*conf
= mddev
->private;
5045 struct disk_info
*p
;
5047 int last
= conf
->raid_disks
- 1;
5049 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
5052 if (has_failed(conf
))
5053 /* no point adding a device */
5056 if (rdev
->raid_disk
>= 0)
5057 first
= last
= rdev
->raid_disk
;
5060 * find the disk ... but prefer rdev->saved_raid_disk
5063 if (rdev
->saved_raid_disk
>= 0 &&
5064 rdev
->saved_raid_disk
>= first
&&
5065 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
5066 disk
= rdev
->saved_raid_disk
;
5069 for ( ; disk
<= last
; disk
++)
5070 if ((p
=conf
->disks
+ disk
)->rdev
== NULL
) {
5071 clear_bit(In_sync
, &rdev
->flags
);
5072 rdev
->raid_disk
= disk
;
5074 if (rdev
->saved_raid_disk
!= disk
)
5076 rcu_assign_pointer(p
->rdev
, rdev
);
5079 print_raid5_conf(conf
);
5083 static int raid5_resize(mddev_t
*mddev
, sector_t sectors
)
5085 /* no resync is happening, and there is enough space
5086 * on all devices, so we can resize.
5087 * We need to make sure resync covers any new space.
5088 * If the array is shrinking we should possibly wait until
5089 * any io in the removed space completes, but it hardly seems
5092 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
5093 md_set_array_sectors(mddev
, raid5_size(mddev
, sectors
,
5094 mddev
->raid_disks
));
5095 if (mddev
->array_sectors
>
5096 raid5_size(mddev
, sectors
, mddev
->raid_disks
))
5098 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5099 revalidate_disk(mddev
->gendisk
);
5100 if (sectors
> mddev
->dev_sectors
&&
5101 mddev
->recovery_cp
> mddev
->dev_sectors
) {
5102 mddev
->recovery_cp
= mddev
->dev_sectors
;
5103 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
5105 mddev
->dev_sectors
= sectors
;
5106 mddev
->resync_max_sectors
= sectors
;
5110 static int check_stripe_cache(mddev_t
*mddev
)
5112 /* Can only proceed if there are plenty of stripe_heads.
5113 * We need a minimum of one full stripe,, and for sensible progress
5114 * it is best to have about 4 times that.
5115 * If we require 4 times, then the default 256 4K stripe_heads will
5116 * allow for chunk sizes up to 256K, which is probably OK.
5117 * If the chunk size is greater, user-space should request more
5118 * stripe_heads first.
5120 raid5_conf_t
*conf
= mddev
->private;
5121 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5122 > conf
->max_nr_stripes
||
5123 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5124 > conf
->max_nr_stripes
) {
5125 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5127 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
5134 static int check_reshape(mddev_t
*mddev
)
5136 raid5_conf_t
*conf
= mddev
->private;
5138 if (mddev
->delta_disks
== 0 &&
5139 mddev
->new_layout
== mddev
->layout
&&
5140 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
5141 return 0; /* nothing to do */
5143 /* Cannot grow a bitmap yet */
5145 if (has_failed(conf
))
5147 if (mddev
->delta_disks
< 0) {
5148 /* We might be able to shrink, but the devices must
5149 * be made bigger first.
5150 * For raid6, 4 is the minimum size.
5151 * Otherwise 2 is the minimum
5154 if (mddev
->level
== 6)
5156 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
5160 if (!check_stripe_cache(mddev
))
5163 return resize_stripes(conf
, conf
->raid_disks
+ mddev
->delta_disks
);
5166 static int raid5_start_reshape(mddev_t
*mddev
)
5168 raid5_conf_t
*conf
= mddev
->private;
5171 unsigned long flags
;
5173 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
5176 if (!check_stripe_cache(mddev
))
5179 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
5180 if (!test_bit(In_sync
, &rdev
->flags
)
5181 && !test_bit(Faulty
, &rdev
->flags
))
5184 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
5185 /* Not enough devices even to make a degraded array
5190 /* Refuse to reduce size of the array. Any reductions in
5191 * array size must be through explicit setting of array_size
5194 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
5195 < mddev
->array_sectors
) {
5196 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
5197 "before number of disks\n", mdname(mddev
));
5201 atomic_set(&conf
->reshape_stripes
, 0);
5202 spin_lock_irq(&conf
->device_lock
);
5203 conf
->previous_raid_disks
= conf
->raid_disks
;
5204 conf
->raid_disks
+= mddev
->delta_disks
;
5205 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
5206 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
5207 conf
->prev_algo
= conf
->algorithm
;
5208 conf
->algorithm
= mddev
->new_layout
;
5209 if (mddev
->delta_disks
< 0)
5210 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
5212 conf
->reshape_progress
= 0;
5213 conf
->reshape_safe
= conf
->reshape_progress
;
5215 spin_unlock_irq(&conf
->device_lock
);
5217 /* Add some new drives, as many as will fit.
5218 * We know there are enough to make the newly sized array work.
5219 * Don't add devices if we are reducing the number of
5220 * devices in the array. This is because it is not possible
5221 * to correctly record the "partially reconstructed" state of
5222 * such devices during the reshape and confusion could result.
5224 if (mddev
->delta_disks
>= 0) {
5225 int added_devices
= 0;
5226 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
5227 if (rdev
->raid_disk
< 0 &&
5228 !test_bit(Faulty
, &rdev
->flags
)) {
5229 if (raid5_add_disk(mddev
, rdev
) == 0) {
5231 >= conf
->previous_raid_disks
) {
5232 set_bit(In_sync
, &rdev
->flags
);
5235 rdev
->recovery_offset
= 0;
5237 if (sysfs_link_rdev(mddev
, rdev
))
5238 /* Failure here is OK */;
5240 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
5241 && !test_bit(Faulty
, &rdev
->flags
)) {
5242 /* This is a spare that was manually added */
5243 set_bit(In_sync
, &rdev
->flags
);
5247 /* When a reshape changes the number of devices,
5248 * ->degraded is measured against the larger of the
5249 * pre and post number of devices.
5251 spin_lock_irqsave(&conf
->device_lock
, flags
);
5252 mddev
->degraded
+= (conf
->raid_disks
- conf
->previous_raid_disks
)
5254 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5256 mddev
->raid_disks
= conf
->raid_disks
;
5257 mddev
->reshape_position
= conf
->reshape_progress
;
5258 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5260 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5261 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5262 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5263 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5264 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5266 if (!mddev
->sync_thread
) {
5267 mddev
->recovery
= 0;
5268 spin_lock_irq(&conf
->device_lock
);
5269 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
5270 conf
->reshape_progress
= MaxSector
;
5271 spin_unlock_irq(&conf
->device_lock
);
5274 conf
->reshape_checkpoint
= jiffies
;
5275 md_wakeup_thread(mddev
->sync_thread
);
5276 md_new_event(mddev
);
5280 /* This is called from the reshape thread and should make any
5281 * changes needed in 'conf'
5283 static void end_reshape(raid5_conf_t
*conf
)
5286 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
5288 spin_lock_irq(&conf
->device_lock
);
5289 conf
->previous_raid_disks
= conf
->raid_disks
;
5290 conf
->reshape_progress
= MaxSector
;
5291 spin_unlock_irq(&conf
->device_lock
);
5292 wake_up(&conf
->wait_for_overlap
);
5294 /* read-ahead size must cover two whole stripes, which is
5295 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5297 if (conf
->mddev
->queue
) {
5298 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5299 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
5301 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
5302 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
5307 /* This is called from the raid5d thread with mddev_lock held.
5308 * It makes config changes to the device.
5310 static void raid5_finish_reshape(mddev_t
*mddev
)
5312 raid5_conf_t
*conf
= mddev
->private;
5314 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
5316 if (mddev
->delta_disks
> 0) {
5317 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5318 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5319 revalidate_disk(mddev
->gendisk
);
5322 mddev
->degraded
= conf
->raid_disks
;
5323 for (d
= 0; d
< conf
->raid_disks
; d
++)
5324 if (conf
->disks
[d
].rdev
&&
5326 &conf
->disks
[d
].rdev
->flags
))
5328 for (d
= conf
->raid_disks
;
5329 d
< conf
->raid_disks
- mddev
->delta_disks
;
5331 mdk_rdev_t
*rdev
= conf
->disks
[d
].rdev
;
5332 if (rdev
&& raid5_remove_disk(mddev
, d
) == 0) {
5333 sysfs_unlink_rdev(mddev
, rdev
);
5334 rdev
->raid_disk
= -1;
5338 mddev
->layout
= conf
->algorithm
;
5339 mddev
->chunk_sectors
= conf
->chunk_sectors
;
5340 mddev
->reshape_position
= MaxSector
;
5341 mddev
->delta_disks
= 0;
5345 static void raid5_quiesce(mddev_t
*mddev
, int state
)
5347 raid5_conf_t
*conf
= mddev
->private;
5350 case 2: /* resume for a suspend */
5351 wake_up(&conf
->wait_for_overlap
);
5354 case 1: /* stop all writes */
5355 spin_lock_irq(&conf
->device_lock
);
5356 /* '2' tells resync/reshape to pause so that all
5357 * active stripes can drain
5360 wait_event_lock_irq(conf
->wait_for_stripe
,
5361 atomic_read(&conf
->active_stripes
) == 0 &&
5362 atomic_read(&conf
->active_aligned_reads
) == 0,
5363 conf
->device_lock
, /* nothing */);
5365 spin_unlock_irq(&conf
->device_lock
);
5366 /* allow reshape to continue */
5367 wake_up(&conf
->wait_for_overlap
);
5370 case 0: /* re-enable writes */
5371 spin_lock_irq(&conf
->device_lock
);
5373 wake_up(&conf
->wait_for_stripe
);
5374 wake_up(&conf
->wait_for_overlap
);
5375 spin_unlock_irq(&conf
->device_lock
);
5381 static void *raid45_takeover_raid0(mddev_t
*mddev
, int level
)
5383 struct raid0_private_data
*raid0_priv
= mddev
->private;
5386 /* for raid0 takeover only one zone is supported */
5387 if (raid0_priv
->nr_strip_zones
> 1) {
5388 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5390 return ERR_PTR(-EINVAL
);
5393 sectors
= raid0_priv
->strip_zone
[0].zone_end
;
5394 sector_div(sectors
, raid0_priv
->strip_zone
[0].nb_dev
);
5395 mddev
->dev_sectors
= sectors
;
5396 mddev
->new_level
= level
;
5397 mddev
->new_layout
= ALGORITHM_PARITY_N
;
5398 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
5399 mddev
->raid_disks
+= 1;
5400 mddev
->delta_disks
= 1;
5401 /* make sure it will be not marked as dirty */
5402 mddev
->recovery_cp
= MaxSector
;
5404 return setup_conf(mddev
);
5408 static void *raid5_takeover_raid1(mddev_t
*mddev
)
5412 if (mddev
->raid_disks
!= 2 ||
5413 mddev
->degraded
> 1)
5414 return ERR_PTR(-EINVAL
);
5416 /* Should check if there are write-behind devices? */
5418 chunksect
= 64*2; /* 64K by default */
5420 /* The array must be an exact multiple of chunksize */
5421 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
5424 if ((chunksect
<<9) < STRIPE_SIZE
)
5425 /* array size does not allow a suitable chunk size */
5426 return ERR_PTR(-EINVAL
);
5428 mddev
->new_level
= 5;
5429 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5430 mddev
->new_chunk_sectors
= chunksect
;
5432 return setup_conf(mddev
);
5435 static void *raid5_takeover_raid6(mddev_t
*mddev
)
5439 switch (mddev
->layout
) {
5440 case ALGORITHM_LEFT_ASYMMETRIC_6
:
5441 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
5443 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
5444 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
5446 case ALGORITHM_LEFT_SYMMETRIC_6
:
5447 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5449 case ALGORITHM_RIGHT_SYMMETRIC_6
:
5450 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
5452 case ALGORITHM_PARITY_0_6
:
5453 new_layout
= ALGORITHM_PARITY_0
;
5455 case ALGORITHM_PARITY_N
:
5456 new_layout
= ALGORITHM_PARITY_N
;
5459 return ERR_PTR(-EINVAL
);
5461 mddev
->new_level
= 5;
5462 mddev
->new_layout
= new_layout
;
5463 mddev
->delta_disks
= -1;
5464 mddev
->raid_disks
-= 1;
5465 return setup_conf(mddev
);
5469 static int raid5_check_reshape(mddev_t
*mddev
)
5471 /* For a 2-drive array, the layout and chunk size can be changed
5472 * immediately as not restriping is needed.
5473 * For larger arrays we record the new value - after validation
5474 * to be used by a reshape pass.
5476 raid5_conf_t
*conf
= mddev
->private;
5477 int new_chunk
= mddev
->new_chunk_sectors
;
5479 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
5481 if (new_chunk
> 0) {
5482 if (!is_power_of_2(new_chunk
))
5484 if (new_chunk
< (PAGE_SIZE
>>9))
5486 if (mddev
->array_sectors
& (new_chunk
-1))
5487 /* not factor of array size */
5491 /* They look valid */
5493 if (mddev
->raid_disks
== 2) {
5494 /* can make the change immediately */
5495 if (mddev
->new_layout
>= 0) {
5496 conf
->algorithm
= mddev
->new_layout
;
5497 mddev
->layout
= mddev
->new_layout
;
5499 if (new_chunk
> 0) {
5500 conf
->chunk_sectors
= new_chunk
;
5501 mddev
->chunk_sectors
= new_chunk
;
5503 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5504 md_wakeup_thread(mddev
->thread
);
5506 return check_reshape(mddev
);
5509 static int raid6_check_reshape(mddev_t
*mddev
)
5511 int new_chunk
= mddev
->new_chunk_sectors
;
5513 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
5515 if (new_chunk
> 0) {
5516 if (!is_power_of_2(new_chunk
))
5518 if (new_chunk
< (PAGE_SIZE
>> 9))
5520 if (mddev
->array_sectors
& (new_chunk
-1))
5521 /* not factor of array size */
5525 /* They look valid */
5526 return check_reshape(mddev
);
5529 static void *raid5_takeover(mddev_t
*mddev
)
5531 /* raid5 can take over:
5532 * raid0 - if there is only one strip zone - make it a raid4 layout
5533 * raid1 - if there are two drives. We need to know the chunk size
5534 * raid4 - trivial - just use a raid4 layout.
5535 * raid6 - Providing it is a *_6 layout
5537 if (mddev
->level
== 0)
5538 return raid45_takeover_raid0(mddev
, 5);
5539 if (mddev
->level
== 1)
5540 return raid5_takeover_raid1(mddev
);
5541 if (mddev
->level
== 4) {
5542 mddev
->new_layout
= ALGORITHM_PARITY_N
;
5543 mddev
->new_level
= 5;
5544 return setup_conf(mddev
);
5546 if (mddev
->level
== 6)
5547 return raid5_takeover_raid6(mddev
);
5549 return ERR_PTR(-EINVAL
);
5552 static void *raid4_takeover(mddev_t
*mddev
)
5554 /* raid4 can take over:
5555 * raid0 - if there is only one strip zone
5556 * raid5 - if layout is right
5558 if (mddev
->level
== 0)
5559 return raid45_takeover_raid0(mddev
, 4);
5560 if (mddev
->level
== 5 &&
5561 mddev
->layout
== ALGORITHM_PARITY_N
) {
5562 mddev
->new_layout
= 0;
5563 mddev
->new_level
= 4;
5564 return setup_conf(mddev
);
5566 return ERR_PTR(-EINVAL
);
5569 static struct mdk_personality raid5_personality
;
5571 static void *raid6_takeover(mddev_t
*mddev
)
5573 /* Currently can only take over a raid5. We map the
5574 * personality to an equivalent raid6 personality
5575 * with the Q block at the end.
5579 if (mddev
->pers
!= &raid5_personality
)
5580 return ERR_PTR(-EINVAL
);
5581 if (mddev
->degraded
> 1)
5582 return ERR_PTR(-EINVAL
);
5583 if (mddev
->raid_disks
> 253)
5584 return ERR_PTR(-EINVAL
);
5585 if (mddev
->raid_disks
< 3)
5586 return ERR_PTR(-EINVAL
);
5588 switch (mddev
->layout
) {
5589 case ALGORITHM_LEFT_ASYMMETRIC
:
5590 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
5592 case ALGORITHM_RIGHT_ASYMMETRIC
:
5593 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
5595 case ALGORITHM_LEFT_SYMMETRIC
:
5596 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
5598 case ALGORITHM_RIGHT_SYMMETRIC
:
5599 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
5601 case ALGORITHM_PARITY_0
:
5602 new_layout
= ALGORITHM_PARITY_0_6
;
5604 case ALGORITHM_PARITY_N
:
5605 new_layout
= ALGORITHM_PARITY_N
;
5608 return ERR_PTR(-EINVAL
);
5610 mddev
->new_level
= 6;
5611 mddev
->new_layout
= new_layout
;
5612 mddev
->delta_disks
= 1;
5613 mddev
->raid_disks
+= 1;
5614 return setup_conf(mddev
);
5618 static struct mdk_personality raid6_personality
=
5622 .owner
= THIS_MODULE
,
5623 .make_request
= make_request
,
5627 .error_handler
= error
,
5628 .hot_add_disk
= raid5_add_disk
,
5629 .hot_remove_disk
= raid5_remove_disk
,
5630 .spare_active
= raid5_spare_active
,
5631 .sync_request
= sync_request
,
5632 .resize
= raid5_resize
,
5634 .check_reshape
= raid6_check_reshape
,
5635 .start_reshape
= raid5_start_reshape
,
5636 .finish_reshape
= raid5_finish_reshape
,
5637 .quiesce
= raid5_quiesce
,
5638 .takeover
= raid6_takeover
,
5640 static struct mdk_personality raid5_personality
=
5644 .owner
= THIS_MODULE
,
5645 .make_request
= make_request
,
5649 .error_handler
= error
,
5650 .hot_add_disk
= raid5_add_disk
,
5651 .hot_remove_disk
= raid5_remove_disk
,
5652 .spare_active
= raid5_spare_active
,
5653 .sync_request
= sync_request
,
5654 .resize
= raid5_resize
,
5656 .check_reshape
= raid5_check_reshape
,
5657 .start_reshape
= raid5_start_reshape
,
5658 .finish_reshape
= raid5_finish_reshape
,
5659 .quiesce
= raid5_quiesce
,
5660 .takeover
= raid5_takeover
,
5663 static struct mdk_personality raid4_personality
=
5667 .owner
= THIS_MODULE
,
5668 .make_request
= make_request
,
5672 .error_handler
= error
,
5673 .hot_add_disk
= raid5_add_disk
,
5674 .hot_remove_disk
= raid5_remove_disk
,
5675 .spare_active
= raid5_spare_active
,
5676 .sync_request
= sync_request
,
5677 .resize
= raid5_resize
,
5679 .check_reshape
= raid5_check_reshape
,
5680 .start_reshape
= raid5_start_reshape
,
5681 .finish_reshape
= raid5_finish_reshape
,
5682 .quiesce
= raid5_quiesce
,
5683 .takeover
= raid4_takeover
,
5686 static int __init
raid5_init(void)
5688 register_md_personality(&raid6_personality
);
5689 register_md_personality(&raid5_personality
);
5690 register_md_personality(&raid4_personality
);
5694 static void raid5_exit(void)
5696 unregister_md_personality(&raid6_personality
);
5697 unregister_md_personality(&raid5_personality
);
5698 unregister_md_personality(&raid4_personality
);
5701 module_init(raid5_init
);
5702 module_exit(raid5_exit
);
5703 MODULE_LICENSE("GPL");
5704 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
5705 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5706 MODULE_ALIAS("md-raid5");
5707 MODULE_ALIAS("md-raid4");
5708 MODULE_ALIAS("md-level-5");
5709 MODULE_ALIAS("md-level-4");
5710 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5711 MODULE_ALIAS("md-raid6");
5712 MODULE_ALIAS("md-level-6");
5714 /* This used to be two separate modules, they were: */
5715 MODULE_ALIAS("raid5");
5716 MODULE_ALIAS("raid6");