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>
63 #define NR_STRIPES 256
64 #define STRIPE_SIZE PAGE_SIZE
65 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
66 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
67 #define IO_THRESHOLD 1
68 #define BYPASS_THRESHOLD 1
69 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
70 #define HASH_MASK (NR_HASH - 1)
72 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
74 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
75 * order without overlap. There may be several bio's per stripe+device, and
76 * a bio could span several devices.
77 * When walking this list for a particular stripe+device, we must never proceed
78 * beyond a bio that extends past this device, as the next bio might no longer
80 * This macro is used to determine the 'next' bio in the list, given the sector
81 * of the current stripe+device
83 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
85 * The following can be used to debug the driver
87 #define RAID5_PARANOIA 1
88 #if RAID5_PARANOIA && defined(CONFIG_SMP)
89 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
91 # define CHECK_DEVLOCK()
99 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
102 * We maintain a biased count of active stripes in the bottom 16 bits of
103 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
105 static inline int raid5_bi_phys_segments(struct bio
*bio
)
107 return bio
->bi_phys_segments
& 0xffff;
110 static inline int raid5_bi_hw_segments(struct bio
*bio
)
112 return (bio
->bi_phys_segments
>> 16) & 0xffff;
115 static inline int raid5_dec_bi_phys_segments(struct bio
*bio
)
117 --bio
->bi_phys_segments
;
118 return raid5_bi_phys_segments(bio
);
121 static inline int raid5_dec_bi_hw_segments(struct bio
*bio
)
123 unsigned short val
= raid5_bi_hw_segments(bio
);
126 bio
->bi_phys_segments
= (val
<< 16) | raid5_bi_phys_segments(bio
);
130 static inline void raid5_set_bi_hw_segments(struct bio
*bio
, unsigned int cnt
)
132 bio
->bi_phys_segments
= raid5_bi_phys_segments(bio
) | (cnt
<< 16);
135 /* Find first data disk in a raid6 stripe */
136 static inline int raid6_d0(struct stripe_head
*sh
)
139 /* ddf always start from first device */
141 /* md starts just after Q block */
142 if (sh
->qd_idx
== sh
->disks
- 1)
145 return sh
->qd_idx
+ 1;
147 static inline int raid6_next_disk(int disk
, int raid_disks
)
150 return (disk
< raid_disks
) ? disk
: 0;
153 /* When walking through the disks in a raid5, starting at raid6_d0,
154 * We need to map each disk to a 'slot', where the data disks are slot
155 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
156 * is raid_disks-1. This help does that mapping.
158 static int raid6_idx_to_slot(int idx
, struct stripe_head
*sh
,
159 int *count
, int syndrome_disks
)
165 if (idx
== sh
->pd_idx
)
166 return syndrome_disks
;
167 if (idx
== sh
->qd_idx
)
168 return syndrome_disks
+ 1;
174 static void return_io(struct bio
*return_bi
)
176 struct bio
*bi
= return_bi
;
179 return_bi
= bi
->bi_next
;
187 static void print_raid5_conf (raid5_conf_t
*conf
);
189 static int stripe_operations_active(struct stripe_head
*sh
)
191 return sh
->check_state
|| sh
->reconstruct_state
||
192 test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
) ||
193 test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
196 static void __release_stripe(raid5_conf_t
*conf
, struct stripe_head
*sh
)
198 if (atomic_dec_and_test(&sh
->count
)) {
199 BUG_ON(!list_empty(&sh
->lru
));
200 BUG_ON(atomic_read(&conf
->active_stripes
)==0);
201 if (test_bit(STRIPE_HANDLE
, &sh
->state
)) {
202 if (test_bit(STRIPE_DELAYED
, &sh
->state
))
203 list_add_tail(&sh
->lru
, &conf
->delayed_list
);
204 else if (test_bit(STRIPE_BIT_DELAY
, &sh
->state
) &&
205 sh
->bm_seq
- conf
->seq_write
> 0)
206 list_add_tail(&sh
->lru
, &conf
->bitmap_list
);
208 clear_bit(STRIPE_BIT_DELAY
, &sh
->state
);
209 list_add_tail(&sh
->lru
, &conf
->handle_list
);
211 md_wakeup_thread(conf
->mddev
->thread
);
213 BUG_ON(stripe_operations_active(sh
));
214 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
215 atomic_dec(&conf
->preread_active_stripes
);
216 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
)
217 md_wakeup_thread(conf
->mddev
->thread
);
219 atomic_dec(&conf
->active_stripes
);
220 if (!test_bit(STRIPE_EXPANDING
, &sh
->state
)) {
221 list_add_tail(&sh
->lru
, &conf
->inactive_list
);
222 wake_up(&conf
->wait_for_stripe
);
223 if (conf
->retry_read_aligned
)
224 md_wakeup_thread(conf
->mddev
->thread
);
230 static void release_stripe(struct stripe_head
*sh
)
232 raid5_conf_t
*conf
= sh
->raid_conf
;
235 spin_lock_irqsave(&conf
->device_lock
, flags
);
236 __release_stripe(conf
, sh
);
237 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
240 static inline void remove_hash(struct stripe_head
*sh
)
242 pr_debug("remove_hash(), stripe %llu\n",
243 (unsigned long long)sh
->sector
);
245 hlist_del_init(&sh
->hash
);
248 static inline void insert_hash(raid5_conf_t
*conf
, struct stripe_head
*sh
)
250 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
252 pr_debug("insert_hash(), stripe %llu\n",
253 (unsigned long long)sh
->sector
);
256 hlist_add_head(&sh
->hash
, hp
);
260 /* find an idle stripe, make sure it is unhashed, and return it. */
261 static struct stripe_head
*get_free_stripe(raid5_conf_t
*conf
)
263 struct stripe_head
*sh
= NULL
;
264 struct list_head
*first
;
267 if (list_empty(&conf
->inactive_list
))
269 first
= conf
->inactive_list
.next
;
270 sh
= list_entry(first
, struct stripe_head
, lru
);
271 list_del_init(first
);
273 atomic_inc(&conf
->active_stripes
);
278 static void shrink_buffers(struct stripe_head
*sh
)
282 int num
= sh
->raid_conf
->pool_size
;
284 for (i
= 0; i
< num
; i
++) {
288 sh
->dev
[i
].page
= NULL
;
293 static int grow_buffers(struct stripe_head
*sh
)
296 int num
= sh
->raid_conf
->pool_size
;
298 for (i
= 0; i
< num
; i
++) {
301 if (!(page
= alloc_page(GFP_KERNEL
))) {
304 sh
->dev
[i
].page
= page
;
309 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
);
310 static void stripe_set_idx(sector_t stripe
, raid5_conf_t
*conf
, int previous
,
311 struct stripe_head
*sh
);
313 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int previous
)
315 raid5_conf_t
*conf
= sh
->raid_conf
;
318 BUG_ON(atomic_read(&sh
->count
) != 0);
319 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
320 BUG_ON(stripe_operations_active(sh
));
323 pr_debug("init_stripe called, stripe %llu\n",
324 (unsigned long long)sh
->sector
);
328 sh
->generation
= conf
->generation
- previous
;
329 sh
->disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
331 stripe_set_idx(sector
, conf
, previous
, sh
);
335 for (i
= sh
->disks
; i
--; ) {
336 struct r5dev
*dev
= &sh
->dev
[i
];
338 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
339 test_bit(R5_LOCKED
, &dev
->flags
)) {
340 printk(KERN_ERR
"sector=%llx i=%d %p %p %p %p %d\n",
341 (unsigned long long)sh
->sector
, i
, dev
->toread
,
342 dev
->read
, dev
->towrite
, dev
->written
,
343 test_bit(R5_LOCKED
, &dev
->flags
));
347 raid5_build_block(sh
, i
, previous
);
349 insert_hash(conf
, sh
);
352 static struct stripe_head
*__find_stripe(raid5_conf_t
*conf
, sector_t sector
,
355 struct stripe_head
*sh
;
356 struct hlist_node
*hn
;
359 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
360 hlist_for_each_entry(sh
, hn
, stripe_hash(conf
, sector
), hash
)
361 if (sh
->sector
== sector
&& sh
->generation
== generation
)
363 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
368 * Need to check if array has failed when deciding whether to:
370 * - remove non-faulty devices
373 * This determination is simple when no reshape is happening.
374 * However if there is a reshape, we need to carefully check
375 * both the before and after sections.
376 * This is because some failed devices may only affect one
377 * of the two sections, and some non-in_sync devices may
378 * be insync in the section most affected by failed devices.
380 static int has_failed(raid5_conf_t
*conf
)
384 if (conf
->mddev
->reshape_position
== MaxSector
)
385 return conf
->mddev
->degraded
> conf
->max_degraded
;
389 for (i
= 0; i
< conf
->previous_raid_disks
; i
++) {
390 mdk_rdev_t
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
391 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
393 else if (test_bit(In_sync
, &rdev
->flags
))
396 /* not in-sync or faulty.
397 * If the reshape increases the number of devices,
398 * this is being recovered by the reshape, so
399 * this 'previous' section is not in_sync.
400 * If the number of devices is being reduced however,
401 * the device can only be part of the array if
402 * we are reverting a reshape, so this section will
405 if (conf
->raid_disks
>= conf
->previous_raid_disks
)
409 if (degraded
> conf
->max_degraded
)
413 for (i
= 0; i
< conf
->raid_disks
; i
++) {
414 mdk_rdev_t
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
415 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
417 else if (test_bit(In_sync
, &rdev
->flags
))
420 /* not in-sync or faulty.
421 * If reshape increases the number of devices, this
422 * section has already been recovered, else it
423 * almost certainly hasn't.
425 if (conf
->raid_disks
<= conf
->previous_raid_disks
)
429 if (degraded
> conf
->max_degraded
)
434 static struct stripe_head
*
435 get_active_stripe(raid5_conf_t
*conf
, sector_t sector
,
436 int previous
, int noblock
, int noquiesce
)
438 struct stripe_head
*sh
;
440 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
442 spin_lock_irq(&conf
->device_lock
);
445 wait_event_lock_irq(conf
->wait_for_stripe
,
446 conf
->quiesce
== 0 || noquiesce
,
447 conf
->device_lock
, /* nothing */);
448 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
450 if (!conf
->inactive_blocked
)
451 sh
= get_free_stripe(conf
);
452 if (noblock
&& sh
== NULL
)
455 conf
->inactive_blocked
= 1;
456 wait_event_lock_irq(conf
->wait_for_stripe
,
457 !list_empty(&conf
->inactive_list
) &&
458 (atomic_read(&conf
->active_stripes
)
459 < (conf
->max_nr_stripes
*3/4)
460 || !conf
->inactive_blocked
),
463 conf
->inactive_blocked
= 0;
465 init_stripe(sh
, sector
, previous
);
467 if (atomic_read(&sh
->count
)) {
468 BUG_ON(!list_empty(&sh
->lru
)
469 && !test_bit(STRIPE_EXPANDING
, &sh
->state
));
471 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
472 atomic_inc(&conf
->active_stripes
);
473 if (list_empty(&sh
->lru
) &&
474 !test_bit(STRIPE_EXPANDING
, &sh
->state
))
476 list_del_init(&sh
->lru
);
479 } while (sh
== NULL
);
482 atomic_inc(&sh
->count
);
484 spin_unlock_irq(&conf
->device_lock
);
489 raid5_end_read_request(struct bio
*bi
, int error
);
491 raid5_end_write_request(struct bio
*bi
, int error
);
493 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
495 raid5_conf_t
*conf
= sh
->raid_conf
;
496 int i
, disks
= sh
->disks
;
500 for (i
= disks
; i
--; ) {
504 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
505 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
509 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
514 bi
= &sh
->dev
[i
].req
;
518 bi
->bi_end_io
= raid5_end_write_request
;
520 bi
->bi_end_io
= raid5_end_read_request
;
523 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
524 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
527 atomic_inc(&rdev
->nr_pending
);
531 if (s
->syncing
|| s
->expanding
|| s
->expanded
)
532 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
534 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
536 bi
->bi_bdev
= rdev
->bdev
;
537 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
538 __func__
, (unsigned long long)sh
->sector
,
540 atomic_inc(&sh
->count
);
541 bi
->bi_sector
= sh
->sector
+ rdev
->data_offset
;
542 bi
->bi_flags
= 1 << BIO_UPTODATE
;
546 bi
->bi_io_vec
= &sh
->dev
[i
].vec
;
547 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
548 bi
->bi_io_vec
[0].bv_offset
= 0;
549 bi
->bi_size
= STRIPE_SIZE
;
552 test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
))
553 atomic_add(STRIPE_SECTORS
,
554 &rdev
->corrected_errors
);
555 generic_make_request(bi
);
558 set_bit(STRIPE_DEGRADED
, &sh
->state
);
559 pr_debug("skip op %ld on disc %d for sector %llu\n",
560 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
561 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
562 set_bit(STRIPE_HANDLE
, &sh
->state
);
567 static struct dma_async_tx_descriptor
*
568 async_copy_data(int frombio
, struct bio
*bio
, struct page
*page
,
569 sector_t sector
, struct dma_async_tx_descriptor
*tx
)
572 struct page
*bio_page
;
575 struct async_submit_ctl submit
;
576 enum async_tx_flags flags
= 0;
578 if (bio
->bi_sector
>= sector
)
579 page_offset
= (signed)(bio
->bi_sector
- sector
) * 512;
581 page_offset
= (signed)(sector
- bio
->bi_sector
) * -512;
584 flags
|= ASYNC_TX_FENCE
;
585 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
587 bio_for_each_segment(bvl
, bio
, i
) {
588 int len
= bvl
->bv_len
;
592 if (page_offset
< 0) {
593 b_offset
= -page_offset
;
594 page_offset
+= b_offset
;
598 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
599 clen
= STRIPE_SIZE
- page_offset
;
604 b_offset
+= bvl
->bv_offset
;
605 bio_page
= bvl
->bv_page
;
607 tx
= async_memcpy(page
, bio_page
, page_offset
,
608 b_offset
, clen
, &submit
);
610 tx
= async_memcpy(bio_page
, page
, b_offset
,
611 page_offset
, clen
, &submit
);
613 /* chain the operations */
614 submit
.depend_tx
= tx
;
616 if (clen
< len
) /* hit end of page */
624 static void ops_complete_biofill(void *stripe_head_ref
)
626 struct stripe_head
*sh
= stripe_head_ref
;
627 struct bio
*return_bi
= NULL
;
628 raid5_conf_t
*conf
= sh
->raid_conf
;
631 pr_debug("%s: stripe %llu\n", __func__
,
632 (unsigned long long)sh
->sector
);
634 /* clear completed biofills */
635 spin_lock_irq(&conf
->device_lock
);
636 for (i
= sh
->disks
; i
--; ) {
637 struct r5dev
*dev
= &sh
->dev
[i
];
639 /* acknowledge completion of a biofill operation */
640 /* and check if we need to reply to a read request,
641 * new R5_Wantfill requests are held off until
642 * !STRIPE_BIOFILL_RUN
644 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
645 struct bio
*rbi
, *rbi2
;
650 while (rbi
&& rbi
->bi_sector
<
651 dev
->sector
+ STRIPE_SECTORS
) {
652 rbi2
= r5_next_bio(rbi
, dev
->sector
);
653 if (!raid5_dec_bi_phys_segments(rbi
)) {
654 rbi
->bi_next
= return_bi
;
661 spin_unlock_irq(&conf
->device_lock
);
662 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
664 return_io(return_bi
);
666 set_bit(STRIPE_HANDLE
, &sh
->state
);
670 static void ops_run_biofill(struct stripe_head
*sh
)
672 struct dma_async_tx_descriptor
*tx
= NULL
;
673 raid5_conf_t
*conf
= sh
->raid_conf
;
674 struct async_submit_ctl submit
;
677 pr_debug("%s: stripe %llu\n", __func__
,
678 (unsigned long long)sh
->sector
);
680 for (i
= sh
->disks
; i
--; ) {
681 struct r5dev
*dev
= &sh
->dev
[i
];
682 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
684 spin_lock_irq(&conf
->device_lock
);
685 dev
->read
= rbi
= dev
->toread
;
687 spin_unlock_irq(&conf
->device_lock
);
688 while (rbi
&& rbi
->bi_sector
<
689 dev
->sector
+ STRIPE_SECTORS
) {
690 tx
= async_copy_data(0, rbi
, dev
->page
,
692 rbi
= r5_next_bio(rbi
, dev
->sector
);
697 atomic_inc(&sh
->count
);
698 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
699 async_trigger_callback(&submit
);
702 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
709 tgt
= &sh
->dev
[target
];
710 set_bit(R5_UPTODATE
, &tgt
->flags
);
711 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
712 clear_bit(R5_Wantcompute
, &tgt
->flags
);
715 static void ops_complete_compute(void *stripe_head_ref
)
717 struct stripe_head
*sh
= stripe_head_ref
;
719 pr_debug("%s: stripe %llu\n", __func__
,
720 (unsigned long long)sh
->sector
);
722 /* mark the computed target(s) as uptodate */
723 mark_target_uptodate(sh
, sh
->ops
.target
);
724 mark_target_uptodate(sh
, sh
->ops
.target2
);
726 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
727 if (sh
->check_state
== check_state_compute_run
)
728 sh
->check_state
= check_state_compute_result
;
729 set_bit(STRIPE_HANDLE
, &sh
->state
);
733 /* return a pointer to the address conversion region of the scribble buffer */
734 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
735 struct raid5_percpu
*percpu
)
737 return percpu
->scribble
+ sizeof(struct page
*) * (sh
->disks
+ 2);
740 static struct dma_async_tx_descriptor
*
741 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
743 int disks
= sh
->disks
;
744 struct page
**xor_srcs
= percpu
->scribble
;
745 int target
= sh
->ops
.target
;
746 struct r5dev
*tgt
= &sh
->dev
[target
];
747 struct page
*xor_dest
= tgt
->page
;
749 struct dma_async_tx_descriptor
*tx
;
750 struct async_submit_ctl submit
;
753 pr_debug("%s: stripe %llu block: %d\n",
754 __func__
, (unsigned long long)sh
->sector
, target
);
755 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
757 for (i
= disks
; i
--; )
759 xor_srcs
[count
++] = sh
->dev
[i
].page
;
761 atomic_inc(&sh
->count
);
763 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
764 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
));
765 if (unlikely(count
== 1))
766 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
768 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
773 /* set_syndrome_sources - populate source buffers for gen_syndrome
774 * @srcs - (struct page *) array of size sh->disks
775 * @sh - stripe_head to parse
777 * Populates srcs in proper layout order for the stripe and returns the
778 * 'count' of sources to be used in a call to async_gen_syndrome. The P
779 * destination buffer is recorded in srcs[count] and the Q destination
780 * is recorded in srcs[count+1]].
782 static int set_syndrome_sources(struct page
**srcs
, struct stripe_head
*sh
)
784 int disks
= sh
->disks
;
785 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
786 int d0_idx
= raid6_d0(sh
);
790 for (i
= 0; i
< disks
; i
++)
796 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
798 srcs
[slot
] = sh
->dev
[i
].page
;
799 i
= raid6_next_disk(i
, disks
);
800 } while (i
!= d0_idx
);
802 return syndrome_disks
;
805 static struct dma_async_tx_descriptor
*
806 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
808 int disks
= sh
->disks
;
809 struct page
**blocks
= percpu
->scribble
;
811 int qd_idx
= sh
->qd_idx
;
812 struct dma_async_tx_descriptor
*tx
;
813 struct async_submit_ctl submit
;
819 if (sh
->ops
.target
< 0)
820 target
= sh
->ops
.target2
;
821 else if (sh
->ops
.target2
< 0)
822 target
= sh
->ops
.target
;
824 /* we should only have one valid target */
827 pr_debug("%s: stripe %llu block: %d\n",
828 __func__
, (unsigned long long)sh
->sector
, target
);
830 tgt
= &sh
->dev
[target
];
831 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
834 atomic_inc(&sh
->count
);
836 if (target
== qd_idx
) {
837 count
= set_syndrome_sources(blocks
, sh
);
838 blocks
[count
] = NULL
; /* regenerating p is not necessary */
839 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
840 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
841 ops_complete_compute
, sh
,
842 to_addr_conv(sh
, percpu
));
843 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
845 /* Compute any data- or p-drive using XOR */
847 for (i
= disks
; i
-- ; ) {
848 if (i
== target
|| i
== qd_idx
)
850 blocks
[count
++] = sh
->dev
[i
].page
;
853 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
854 NULL
, ops_complete_compute
, sh
,
855 to_addr_conv(sh
, percpu
));
856 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
862 static struct dma_async_tx_descriptor
*
863 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
865 int i
, count
, disks
= sh
->disks
;
866 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
867 int d0_idx
= raid6_d0(sh
);
868 int faila
= -1, failb
= -1;
869 int target
= sh
->ops
.target
;
870 int target2
= sh
->ops
.target2
;
871 struct r5dev
*tgt
= &sh
->dev
[target
];
872 struct r5dev
*tgt2
= &sh
->dev
[target2
];
873 struct dma_async_tx_descriptor
*tx
;
874 struct page
**blocks
= percpu
->scribble
;
875 struct async_submit_ctl submit
;
877 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
878 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
879 BUG_ON(target
< 0 || target2
< 0);
880 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
881 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
883 /* we need to open-code set_syndrome_sources to handle the
884 * slot number conversion for 'faila' and 'failb'
886 for (i
= 0; i
< disks
; i
++)
891 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
893 blocks
[slot
] = sh
->dev
[i
].page
;
899 i
= raid6_next_disk(i
, disks
);
900 } while (i
!= d0_idx
);
902 BUG_ON(faila
== failb
);
905 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
906 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
908 atomic_inc(&sh
->count
);
910 if (failb
== syndrome_disks
+1) {
911 /* Q disk is one of the missing disks */
912 if (faila
== syndrome_disks
) {
913 /* Missing P+Q, just recompute */
914 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
915 ops_complete_compute
, sh
,
916 to_addr_conv(sh
, percpu
));
917 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
918 STRIPE_SIZE
, &submit
);
922 int qd_idx
= sh
->qd_idx
;
924 /* Missing D+Q: recompute D from P, then recompute Q */
925 if (target
== qd_idx
)
926 data_target
= target2
;
928 data_target
= target
;
931 for (i
= disks
; i
-- ; ) {
932 if (i
== data_target
|| i
== qd_idx
)
934 blocks
[count
++] = sh
->dev
[i
].page
;
936 dest
= sh
->dev
[data_target
].page
;
937 init_async_submit(&submit
,
938 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
940 to_addr_conv(sh
, percpu
));
941 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
944 count
= set_syndrome_sources(blocks
, sh
);
945 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
946 ops_complete_compute
, sh
,
947 to_addr_conv(sh
, percpu
));
948 return async_gen_syndrome(blocks
, 0, count
+2,
949 STRIPE_SIZE
, &submit
);
952 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
953 ops_complete_compute
, sh
,
954 to_addr_conv(sh
, percpu
));
955 if (failb
== syndrome_disks
) {
956 /* We're missing D+P. */
957 return async_raid6_datap_recov(syndrome_disks
+2,
961 /* We're missing D+D. */
962 return async_raid6_2data_recov(syndrome_disks
+2,
963 STRIPE_SIZE
, faila
, failb
,
970 static void ops_complete_prexor(void *stripe_head_ref
)
972 struct stripe_head
*sh
= stripe_head_ref
;
974 pr_debug("%s: stripe %llu\n", __func__
,
975 (unsigned long long)sh
->sector
);
978 static struct dma_async_tx_descriptor
*
979 ops_run_prexor(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
980 struct dma_async_tx_descriptor
*tx
)
982 int disks
= sh
->disks
;
983 struct page
**xor_srcs
= percpu
->scribble
;
984 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
985 struct async_submit_ctl submit
;
987 /* existing parity data subtracted */
988 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
990 pr_debug("%s: stripe %llu\n", __func__
,
991 (unsigned long long)sh
->sector
);
993 for (i
= disks
; i
--; ) {
994 struct r5dev
*dev
= &sh
->dev
[i
];
995 /* Only process blocks that are known to be uptodate */
996 if (test_bit(R5_Wantdrain
, &dev
->flags
))
997 xor_srcs
[count
++] = dev
->page
;
1000 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1001 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
));
1002 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1007 static struct dma_async_tx_descriptor
*
1008 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1010 int disks
= sh
->disks
;
1013 pr_debug("%s: stripe %llu\n", __func__
,
1014 (unsigned long long)sh
->sector
);
1016 for (i
= disks
; i
--; ) {
1017 struct r5dev
*dev
= &sh
->dev
[i
];
1020 if (test_and_clear_bit(R5_Wantdrain
, &dev
->flags
)) {
1023 spin_lock_irq(&sh
->raid_conf
->device_lock
);
1024 chosen
= dev
->towrite
;
1025 dev
->towrite
= NULL
;
1026 BUG_ON(dev
->written
);
1027 wbi
= dev
->written
= chosen
;
1028 spin_unlock_irq(&sh
->raid_conf
->device_lock
);
1030 while (wbi
&& wbi
->bi_sector
<
1031 dev
->sector
+ STRIPE_SECTORS
) {
1032 if (wbi
->bi_rw
& REQ_FUA
)
1033 set_bit(R5_WantFUA
, &dev
->flags
);
1034 tx
= async_copy_data(1, wbi
, dev
->page
,
1036 wbi
= r5_next_bio(wbi
, dev
->sector
);
1044 static void ops_complete_reconstruct(void *stripe_head_ref
)
1046 struct stripe_head
*sh
= stripe_head_ref
;
1047 int disks
= sh
->disks
;
1048 int pd_idx
= sh
->pd_idx
;
1049 int qd_idx
= sh
->qd_idx
;
1053 pr_debug("%s: stripe %llu\n", __func__
,
1054 (unsigned long long)sh
->sector
);
1056 for (i
= disks
; i
--; )
1057 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1059 for (i
= disks
; i
--; ) {
1060 struct r5dev
*dev
= &sh
->dev
[i
];
1062 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1063 set_bit(R5_UPTODATE
, &dev
->flags
);
1065 set_bit(R5_WantFUA
, &dev
->flags
);
1069 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1070 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1071 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1072 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1074 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1075 sh
->reconstruct_state
= reconstruct_state_result
;
1078 set_bit(STRIPE_HANDLE
, &sh
->state
);
1083 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1084 struct dma_async_tx_descriptor
*tx
)
1086 int disks
= sh
->disks
;
1087 struct page
**xor_srcs
= percpu
->scribble
;
1088 struct async_submit_ctl submit
;
1089 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1090 struct page
*xor_dest
;
1092 unsigned long flags
;
1094 pr_debug("%s: stripe %llu\n", __func__
,
1095 (unsigned long long)sh
->sector
);
1097 /* check if prexor is active which means only process blocks
1098 * that are part of a read-modify-write (written)
1100 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1102 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1103 for (i
= disks
; i
--; ) {
1104 struct r5dev
*dev
= &sh
->dev
[i
];
1106 xor_srcs
[count
++] = dev
->page
;
1109 xor_dest
= sh
->dev
[pd_idx
].page
;
1110 for (i
= disks
; i
--; ) {
1111 struct r5dev
*dev
= &sh
->dev
[i
];
1113 xor_srcs
[count
++] = dev
->page
;
1117 /* 1/ if we prexor'd then the dest is reused as a source
1118 * 2/ if we did not prexor then we are redoing the parity
1119 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1120 * for the synchronous xor case
1122 flags
= ASYNC_TX_ACK
|
1123 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1125 atomic_inc(&sh
->count
);
1127 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, sh
,
1128 to_addr_conv(sh
, percpu
));
1129 if (unlikely(count
== 1))
1130 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1132 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1136 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1137 struct dma_async_tx_descriptor
*tx
)
1139 struct async_submit_ctl submit
;
1140 struct page
**blocks
= percpu
->scribble
;
1143 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1145 count
= set_syndrome_sources(blocks
, sh
);
1147 atomic_inc(&sh
->count
);
1149 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_reconstruct
,
1150 sh
, to_addr_conv(sh
, percpu
));
1151 async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1154 static void ops_complete_check(void *stripe_head_ref
)
1156 struct stripe_head
*sh
= stripe_head_ref
;
1158 pr_debug("%s: stripe %llu\n", __func__
,
1159 (unsigned long long)sh
->sector
);
1161 sh
->check_state
= check_state_check_result
;
1162 set_bit(STRIPE_HANDLE
, &sh
->state
);
1166 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1168 int disks
= sh
->disks
;
1169 int pd_idx
= sh
->pd_idx
;
1170 int qd_idx
= sh
->qd_idx
;
1171 struct page
*xor_dest
;
1172 struct page
**xor_srcs
= percpu
->scribble
;
1173 struct dma_async_tx_descriptor
*tx
;
1174 struct async_submit_ctl submit
;
1178 pr_debug("%s: stripe %llu\n", __func__
,
1179 (unsigned long long)sh
->sector
);
1182 xor_dest
= sh
->dev
[pd_idx
].page
;
1183 xor_srcs
[count
++] = xor_dest
;
1184 for (i
= disks
; i
--; ) {
1185 if (i
== pd_idx
|| i
== qd_idx
)
1187 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1190 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1191 to_addr_conv(sh
, percpu
));
1192 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1193 &sh
->ops
.zero_sum_result
, &submit
);
1195 atomic_inc(&sh
->count
);
1196 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1197 tx
= async_trigger_callback(&submit
);
1200 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1202 struct page
**srcs
= percpu
->scribble
;
1203 struct async_submit_ctl submit
;
1206 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1207 (unsigned long long)sh
->sector
, checkp
);
1209 count
= set_syndrome_sources(srcs
, sh
);
1213 atomic_inc(&sh
->count
);
1214 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1215 sh
, to_addr_conv(sh
, percpu
));
1216 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1217 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1220 static void __raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1222 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1223 struct dma_async_tx_descriptor
*tx
= NULL
;
1224 raid5_conf_t
*conf
= sh
->raid_conf
;
1225 int level
= conf
->level
;
1226 struct raid5_percpu
*percpu
;
1230 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1231 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1232 ops_run_biofill(sh
);
1236 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1238 tx
= ops_run_compute5(sh
, percpu
);
1240 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
1241 tx
= ops_run_compute6_1(sh
, percpu
);
1243 tx
= ops_run_compute6_2(sh
, percpu
);
1245 /* terminate the chain if reconstruct is not set to be run */
1246 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
1250 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
))
1251 tx
= ops_run_prexor(sh
, percpu
, tx
);
1253 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
1254 tx
= ops_run_biodrain(sh
, tx
);
1258 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
1260 ops_run_reconstruct5(sh
, percpu
, tx
);
1262 ops_run_reconstruct6(sh
, percpu
, tx
);
1265 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
1266 if (sh
->check_state
== check_state_run
)
1267 ops_run_check_p(sh
, percpu
);
1268 else if (sh
->check_state
== check_state_run_q
)
1269 ops_run_check_pq(sh
, percpu
, 0);
1270 else if (sh
->check_state
== check_state_run_pq
)
1271 ops_run_check_pq(sh
, percpu
, 1);
1277 for (i
= disks
; i
--; ) {
1278 struct r5dev
*dev
= &sh
->dev
[i
];
1279 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
1280 wake_up(&sh
->raid_conf
->wait_for_overlap
);
1285 #ifdef CONFIG_MULTICORE_RAID456
1286 static void async_run_ops(void *param
, async_cookie_t cookie
)
1288 struct stripe_head
*sh
= param
;
1289 unsigned long ops_request
= sh
->ops
.request
;
1291 clear_bit_unlock(STRIPE_OPS_REQ_PENDING
, &sh
->state
);
1292 wake_up(&sh
->ops
.wait_for_ops
);
1294 __raid_run_ops(sh
, ops_request
);
1298 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1300 /* since handle_stripe can be called outside of raid5d context
1301 * we need to ensure sh->ops.request is de-staged before another
1304 wait_event(sh
->ops
.wait_for_ops
,
1305 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING
, &sh
->state
));
1306 sh
->ops
.request
= ops_request
;
1308 atomic_inc(&sh
->count
);
1309 async_schedule(async_run_ops
, sh
);
1312 #define raid_run_ops __raid_run_ops
1315 static int grow_one_stripe(raid5_conf_t
*conf
)
1317 struct stripe_head
*sh
;
1318 sh
= kmem_cache_zalloc(conf
->slab_cache
, GFP_KERNEL
);
1322 sh
->raid_conf
= conf
;
1323 #ifdef CONFIG_MULTICORE_RAID456
1324 init_waitqueue_head(&sh
->ops
.wait_for_ops
);
1327 if (grow_buffers(sh
)) {
1329 kmem_cache_free(conf
->slab_cache
, sh
);
1332 /* we just created an active stripe so... */
1333 atomic_set(&sh
->count
, 1);
1334 atomic_inc(&conf
->active_stripes
);
1335 INIT_LIST_HEAD(&sh
->lru
);
1340 static int grow_stripes(raid5_conf_t
*conf
, int num
)
1342 struct kmem_cache
*sc
;
1343 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
1345 if (conf
->mddev
->gendisk
)
1346 sprintf(conf
->cache_name
[0],
1347 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
1349 sprintf(conf
->cache_name
[0],
1350 "raid%d-%p", conf
->level
, conf
->mddev
);
1351 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
1353 conf
->active_name
= 0;
1354 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
1355 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
1359 conf
->slab_cache
= sc
;
1360 conf
->pool_size
= devs
;
1362 if (!grow_one_stripe(conf
))
1368 * scribble_len - return the required size of the scribble region
1369 * @num - total number of disks in the array
1371 * The size must be enough to contain:
1372 * 1/ a struct page pointer for each device in the array +2
1373 * 2/ room to convert each entry in (1) to its corresponding dma
1374 * (dma_map_page()) or page (page_address()) address.
1376 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1377 * calculate over all devices (not just the data blocks), using zeros in place
1378 * of the P and Q blocks.
1380 static size_t scribble_len(int num
)
1384 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
1389 static int resize_stripes(raid5_conf_t
*conf
, int newsize
)
1391 /* Make all the stripes able to hold 'newsize' devices.
1392 * New slots in each stripe get 'page' set to a new page.
1394 * This happens in stages:
1395 * 1/ create a new kmem_cache and allocate the required number of
1397 * 2/ gather all the old stripe_heads and tranfer the pages across
1398 * to the new stripe_heads. This will have the side effect of
1399 * freezing the array as once all stripe_heads have been collected,
1400 * no IO will be possible. Old stripe heads are freed once their
1401 * pages have been transferred over, and the old kmem_cache is
1402 * freed when all stripes are done.
1403 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1404 * we simple return a failre status - no need to clean anything up.
1405 * 4/ allocate new pages for the new slots in the new stripe_heads.
1406 * If this fails, we don't bother trying the shrink the
1407 * stripe_heads down again, we just leave them as they are.
1408 * As each stripe_head is processed the new one is released into
1411 * Once step2 is started, we cannot afford to wait for a write,
1412 * so we use GFP_NOIO allocations.
1414 struct stripe_head
*osh
, *nsh
;
1415 LIST_HEAD(newstripes
);
1416 struct disk_info
*ndisks
;
1419 struct kmem_cache
*sc
;
1422 if (newsize
<= conf
->pool_size
)
1423 return 0; /* never bother to shrink */
1425 err
= md_allow_write(conf
->mddev
);
1430 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
1431 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
1436 for (i
= conf
->max_nr_stripes
; i
; i
--) {
1437 nsh
= kmem_cache_zalloc(sc
, GFP_KERNEL
);
1441 nsh
->raid_conf
= conf
;
1442 #ifdef CONFIG_MULTICORE_RAID456
1443 init_waitqueue_head(&nsh
->ops
.wait_for_ops
);
1446 list_add(&nsh
->lru
, &newstripes
);
1449 /* didn't get enough, give up */
1450 while (!list_empty(&newstripes
)) {
1451 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1452 list_del(&nsh
->lru
);
1453 kmem_cache_free(sc
, nsh
);
1455 kmem_cache_destroy(sc
);
1458 /* Step 2 - Must use GFP_NOIO now.
1459 * OK, we have enough stripes, start collecting inactive
1460 * stripes and copying them over
1462 list_for_each_entry(nsh
, &newstripes
, lru
) {
1463 spin_lock_irq(&conf
->device_lock
);
1464 wait_event_lock_irq(conf
->wait_for_stripe
,
1465 !list_empty(&conf
->inactive_list
),
1468 osh
= get_free_stripe(conf
);
1469 spin_unlock_irq(&conf
->device_lock
);
1470 atomic_set(&nsh
->count
, 1);
1471 for(i
=0; i
<conf
->pool_size
; i
++)
1472 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
1473 for( ; i
<newsize
; i
++)
1474 nsh
->dev
[i
].page
= NULL
;
1475 kmem_cache_free(conf
->slab_cache
, osh
);
1477 kmem_cache_destroy(conf
->slab_cache
);
1480 * At this point, we are holding all the stripes so the array
1481 * is completely stalled, so now is a good time to resize
1482 * conf->disks and the scribble region
1484 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
1486 for (i
=0; i
<conf
->raid_disks
; i
++)
1487 ndisks
[i
] = conf
->disks
[i
];
1489 conf
->disks
= ndisks
;
1494 conf
->scribble_len
= scribble_len(newsize
);
1495 for_each_present_cpu(cpu
) {
1496 struct raid5_percpu
*percpu
;
1499 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1500 scribble
= kmalloc(conf
->scribble_len
, GFP_NOIO
);
1503 kfree(percpu
->scribble
);
1504 percpu
->scribble
= scribble
;
1512 /* Step 4, return new stripes to service */
1513 while(!list_empty(&newstripes
)) {
1514 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1515 list_del_init(&nsh
->lru
);
1517 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
1518 if (nsh
->dev
[i
].page
== NULL
) {
1519 struct page
*p
= alloc_page(GFP_NOIO
);
1520 nsh
->dev
[i
].page
= p
;
1524 release_stripe(nsh
);
1526 /* critical section pass, GFP_NOIO no longer needed */
1528 conf
->slab_cache
= sc
;
1529 conf
->active_name
= 1-conf
->active_name
;
1530 conf
->pool_size
= newsize
;
1534 static int drop_one_stripe(raid5_conf_t
*conf
)
1536 struct stripe_head
*sh
;
1538 spin_lock_irq(&conf
->device_lock
);
1539 sh
= get_free_stripe(conf
);
1540 spin_unlock_irq(&conf
->device_lock
);
1543 BUG_ON(atomic_read(&sh
->count
));
1545 kmem_cache_free(conf
->slab_cache
, sh
);
1546 atomic_dec(&conf
->active_stripes
);
1550 static void shrink_stripes(raid5_conf_t
*conf
)
1552 while (drop_one_stripe(conf
))
1555 if (conf
->slab_cache
)
1556 kmem_cache_destroy(conf
->slab_cache
);
1557 conf
->slab_cache
= NULL
;
1560 static void raid5_end_read_request(struct bio
* bi
, int error
)
1562 struct stripe_head
*sh
= bi
->bi_private
;
1563 raid5_conf_t
*conf
= sh
->raid_conf
;
1564 int disks
= sh
->disks
, i
;
1565 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1566 char b
[BDEVNAME_SIZE
];
1570 for (i
=0 ; i
<disks
; i
++)
1571 if (bi
== &sh
->dev
[i
].req
)
1574 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1575 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1583 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1584 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
1585 rdev
= conf
->disks
[i
].rdev
;
1586 printk_rl(KERN_INFO
"md/raid:%s: read error corrected"
1587 " (%lu sectors at %llu on %s)\n",
1588 mdname(conf
->mddev
), STRIPE_SECTORS
,
1589 (unsigned long long)(sh
->sector
1590 + rdev
->data_offset
),
1591 bdevname(rdev
->bdev
, b
));
1592 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1593 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1595 if (atomic_read(&conf
->disks
[i
].rdev
->read_errors
))
1596 atomic_set(&conf
->disks
[i
].rdev
->read_errors
, 0);
1598 const char *bdn
= bdevname(conf
->disks
[i
].rdev
->bdev
, b
);
1600 rdev
= conf
->disks
[i
].rdev
;
1602 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1603 atomic_inc(&rdev
->read_errors
);
1604 if (conf
->mddev
->degraded
>= conf
->max_degraded
)
1605 printk_rl(KERN_WARNING
1606 "md/raid:%s: read error not correctable "
1607 "(sector %llu on %s).\n",
1608 mdname(conf
->mddev
),
1609 (unsigned long long)(sh
->sector
1610 + rdev
->data_offset
),
1612 else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
))
1614 printk_rl(KERN_WARNING
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(Faulty
, &rdev
->flags
);
1710 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1712 "md/raid:%s: Disk failure on %s, disabling device.\n"
1713 "md/raid:%s: Operation continuing on %d devices.\n",
1715 bdevname(rdev
->bdev
, b
),
1717 conf
->raid_disks
- mddev
->degraded
);
1721 * Input: a 'big' sector number,
1722 * Output: index of the data and parity disk, and the sector # in them.
1724 static sector_t
raid5_compute_sector(raid5_conf_t
*conf
, sector_t r_sector
,
1725 int previous
, int *dd_idx
,
1726 struct stripe_head
*sh
)
1728 sector_t stripe
, stripe2
;
1729 sector_t chunk_number
;
1730 unsigned int chunk_offset
;
1733 sector_t new_sector
;
1734 int algorithm
= previous
? conf
->prev_algo
1736 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
1737 : conf
->chunk_sectors
;
1738 int raid_disks
= previous
? conf
->previous_raid_disks
1740 int data_disks
= raid_disks
- conf
->max_degraded
;
1742 /* First compute the information on this sector */
1745 * Compute the chunk number and the sector offset inside the chunk
1747 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
1748 chunk_number
= r_sector
;
1751 * Compute the stripe number
1753 stripe
= chunk_number
;
1754 *dd_idx
= sector_div(stripe
, data_disks
);
1757 * Select the parity disk based on the user selected algorithm.
1759 pd_idx
= qd_idx
= -1;
1760 switch(conf
->level
) {
1762 pd_idx
= data_disks
;
1765 switch (algorithm
) {
1766 case ALGORITHM_LEFT_ASYMMETRIC
:
1767 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
1768 if (*dd_idx
>= pd_idx
)
1771 case ALGORITHM_RIGHT_ASYMMETRIC
:
1772 pd_idx
= sector_div(stripe2
, raid_disks
);
1773 if (*dd_idx
>= pd_idx
)
1776 case ALGORITHM_LEFT_SYMMETRIC
:
1777 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
1778 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1780 case ALGORITHM_RIGHT_SYMMETRIC
:
1781 pd_idx
= sector_div(stripe2
, raid_disks
);
1782 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1784 case ALGORITHM_PARITY_0
:
1788 case ALGORITHM_PARITY_N
:
1789 pd_idx
= data_disks
;
1797 switch (algorithm
) {
1798 case ALGORITHM_LEFT_ASYMMETRIC
:
1799 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1800 qd_idx
= pd_idx
+ 1;
1801 if (pd_idx
== raid_disks
-1) {
1802 (*dd_idx
)++; /* Q D D D P */
1804 } else if (*dd_idx
>= pd_idx
)
1805 (*dd_idx
) += 2; /* D D P Q D */
1807 case ALGORITHM_RIGHT_ASYMMETRIC
:
1808 pd_idx
= sector_div(stripe2
, raid_disks
);
1809 qd_idx
= pd_idx
+ 1;
1810 if (pd_idx
== raid_disks
-1) {
1811 (*dd_idx
)++; /* Q D D D P */
1813 } else if (*dd_idx
>= pd_idx
)
1814 (*dd_idx
) += 2; /* D D P Q D */
1816 case ALGORITHM_LEFT_SYMMETRIC
:
1817 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1818 qd_idx
= (pd_idx
+ 1) % raid_disks
;
1819 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
1821 case ALGORITHM_RIGHT_SYMMETRIC
:
1822 pd_idx
= sector_div(stripe2
, raid_disks
);
1823 qd_idx
= (pd_idx
+ 1) % raid_disks
;
1824 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
1827 case ALGORITHM_PARITY_0
:
1832 case ALGORITHM_PARITY_N
:
1833 pd_idx
= data_disks
;
1834 qd_idx
= data_disks
+ 1;
1837 case ALGORITHM_ROTATING_ZERO_RESTART
:
1838 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1839 * of blocks for computing Q is different.
1841 pd_idx
= sector_div(stripe2
, raid_disks
);
1842 qd_idx
= pd_idx
+ 1;
1843 if (pd_idx
== raid_disks
-1) {
1844 (*dd_idx
)++; /* Q D D D P */
1846 } else if (*dd_idx
>= pd_idx
)
1847 (*dd_idx
) += 2; /* D D P Q D */
1851 case ALGORITHM_ROTATING_N_RESTART
:
1852 /* Same a left_asymmetric, by first stripe is
1853 * D D D P Q rather than
1857 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1858 qd_idx
= pd_idx
+ 1;
1859 if (pd_idx
== raid_disks
-1) {
1860 (*dd_idx
)++; /* Q D D D P */
1862 } else if (*dd_idx
>= pd_idx
)
1863 (*dd_idx
) += 2; /* D D P Q D */
1867 case ALGORITHM_ROTATING_N_CONTINUE
:
1868 /* Same as left_symmetric but Q is before P */
1869 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1870 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
1871 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1875 case ALGORITHM_LEFT_ASYMMETRIC_6
:
1876 /* RAID5 left_asymmetric, with Q on last device */
1877 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
1878 if (*dd_idx
>= pd_idx
)
1880 qd_idx
= raid_disks
- 1;
1883 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
1884 pd_idx
= sector_div(stripe2
, raid_disks
-1);
1885 if (*dd_idx
>= pd_idx
)
1887 qd_idx
= raid_disks
- 1;
1890 case ALGORITHM_LEFT_SYMMETRIC_6
:
1891 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
1892 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
1893 qd_idx
= raid_disks
- 1;
1896 case ALGORITHM_RIGHT_SYMMETRIC_6
:
1897 pd_idx
= sector_div(stripe2
, raid_disks
-1);
1898 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
1899 qd_idx
= raid_disks
- 1;
1902 case ALGORITHM_PARITY_0_6
:
1905 qd_idx
= raid_disks
- 1;
1915 sh
->pd_idx
= pd_idx
;
1916 sh
->qd_idx
= qd_idx
;
1917 sh
->ddf_layout
= ddf_layout
;
1920 * Finally, compute the new sector number
1922 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
1927 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
1929 raid5_conf_t
*conf
= sh
->raid_conf
;
1930 int raid_disks
= sh
->disks
;
1931 int data_disks
= raid_disks
- conf
->max_degraded
;
1932 sector_t new_sector
= sh
->sector
, check
;
1933 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
1934 : conf
->chunk_sectors
;
1935 int algorithm
= previous
? conf
->prev_algo
1939 sector_t chunk_number
;
1940 int dummy1
, dd_idx
= i
;
1942 struct stripe_head sh2
;
1945 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
1946 stripe
= new_sector
;
1948 if (i
== sh
->pd_idx
)
1950 switch(conf
->level
) {
1953 switch (algorithm
) {
1954 case ALGORITHM_LEFT_ASYMMETRIC
:
1955 case ALGORITHM_RIGHT_ASYMMETRIC
:
1959 case ALGORITHM_LEFT_SYMMETRIC
:
1960 case ALGORITHM_RIGHT_SYMMETRIC
:
1963 i
-= (sh
->pd_idx
+ 1);
1965 case ALGORITHM_PARITY_0
:
1968 case ALGORITHM_PARITY_N
:
1975 if (i
== sh
->qd_idx
)
1976 return 0; /* It is the Q disk */
1977 switch (algorithm
) {
1978 case ALGORITHM_LEFT_ASYMMETRIC
:
1979 case ALGORITHM_RIGHT_ASYMMETRIC
:
1980 case ALGORITHM_ROTATING_ZERO_RESTART
:
1981 case ALGORITHM_ROTATING_N_RESTART
:
1982 if (sh
->pd_idx
== raid_disks
-1)
1983 i
--; /* Q D D D P */
1984 else if (i
> sh
->pd_idx
)
1985 i
-= 2; /* D D P Q D */
1987 case ALGORITHM_LEFT_SYMMETRIC
:
1988 case ALGORITHM_RIGHT_SYMMETRIC
:
1989 if (sh
->pd_idx
== raid_disks
-1)
1990 i
--; /* Q D D D P */
1995 i
-= (sh
->pd_idx
+ 2);
1998 case ALGORITHM_PARITY_0
:
2001 case ALGORITHM_PARITY_N
:
2003 case ALGORITHM_ROTATING_N_CONTINUE
:
2004 /* Like left_symmetric, but P is before Q */
2005 if (sh
->pd_idx
== 0)
2006 i
--; /* P D D D Q */
2011 i
-= (sh
->pd_idx
+ 1);
2014 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2015 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2019 case ALGORITHM_LEFT_SYMMETRIC_6
:
2020 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2022 i
+= data_disks
+ 1;
2023 i
-= (sh
->pd_idx
+ 1);
2025 case ALGORITHM_PARITY_0_6
:
2034 chunk_number
= stripe
* data_disks
+ i
;
2035 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2037 check
= raid5_compute_sector(conf
, r_sector
,
2038 previous
, &dummy1
, &sh2
);
2039 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2040 || sh2
.qd_idx
!= sh
->qd_idx
) {
2041 printk(KERN_ERR
"md/raid:%s: compute_blocknr: map not correct\n",
2042 mdname(conf
->mddev
));
2050 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2051 int rcw
, int expand
)
2053 int i
, pd_idx
= sh
->pd_idx
, disks
= sh
->disks
;
2054 raid5_conf_t
*conf
= sh
->raid_conf
;
2055 int level
= conf
->level
;
2058 /* if we are not expanding this is a proper write request, and
2059 * there will be bios with new data to be drained into the
2063 sh
->reconstruct_state
= reconstruct_state_drain_run
;
2064 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2066 sh
->reconstruct_state
= reconstruct_state_run
;
2068 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2070 for (i
= disks
; i
--; ) {
2071 struct r5dev
*dev
= &sh
->dev
[i
];
2074 set_bit(R5_LOCKED
, &dev
->flags
);
2075 set_bit(R5_Wantdrain
, &dev
->flags
);
2077 clear_bit(R5_UPTODATE
, &dev
->flags
);
2081 if (s
->locked
+ conf
->max_degraded
== disks
)
2082 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2083 atomic_inc(&conf
->pending_full_writes
);
2086 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
2087 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
2089 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
2090 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
2091 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2092 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2094 for (i
= disks
; i
--; ) {
2095 struct r5dev
*dev
= &sh
->dev
[i
];
2100 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2101 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2102 set_bit(R5_Wantdrain
, &dev
->flags
);
2103 set_bit(R5_LOCKED
, &dev
->flags
);
2104 clear_bit(R5_UPTODATE
, &dev
->flags
);
2110 /* keep the parity disk(s) locked while asynchronous operations
2113 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
2114 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2118 int qd_idx
= sh
->qd_idx
;
2119 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
2121 set_bit(R5_LOCKED
, &dev
->flags
);
2122 clear_bit(R5_UPTODATE
, &dev
->flags
);
2126 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2127 __func__
, (unsigned long long)sh
->sector
,
2128 s
->locked
, s
->ops_request
);
2132 * Each stripe/dev can have one or more bion attached.
2133 * toread/towrite point to the first in a chain.
2134 * The bi_next chain must be in order.
2136 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
, int forwrite
)
2139 raid5_conf_t
*conf
= sh
->raid_conf
;
2142 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2143 (unsigned long long)bi
->bi_sector
,
2144 (unsigned long long)sh
->sector
);
2147 spin_lock_irq(&conf
->device_lock
);
2149 bip
= &sh
->dev
[dd_idx
].towrite
;
2150 if (*bip
== NULL
&& sh
->dev
[dd_idx
].written
== NULL
)
2153 bip
= &sh
->dev
[dd_idx
].toread
;
2154 while (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
) {
2155 if ((*bip
)->bi_sector
+ ((*bip
)->bi_size
>> 9) > bi
->bi_sector
)
2157 bip
= & (*bip
)->bi_next
;
2159 if (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
+ ((bi
->bi_size
)>>9))
2162 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
2166 bi
->bi_phys_segments
++;
2169 /* check if page is covered */
2170 sector_t sector
= sh
->dev
[dd_idx
].sector
;
2171 for (bi
=sh
->dev
[dd_idx
].towrite
;
2172 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
2173 bi
&& bi
->bi_sector
<= sector
;
2174 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
2175 if (bi
->bi_sector
+ (bi
->bi_size
>>9) >= sector
)
2176 sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
2178 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
2179 set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
);
2181 spin_unlock_irq(&conf
->device_lock
);
2183 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2184 (unsigned long long)(*bip
)->bi_sector
,
2185 (unsigned long long)sh
->sector
, dd_idx
);
2187 if (conf
->mddev
->bitmap
&& firstwrite
) {
2188 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
2190 sh
->bm_seq
= conf
->seq_flush
+1;
2191 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
2196 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
2197 spin_unlock_irq(&conf
->device_lock
);
2201 static void end_reshape(raid5_conf_t
*conf
);
2203 static void stripe_set_idx(sector_t stripe
, raid5_conf_t
*conf
, int previous
,
2204 struct stripe_head
*sh
)
2206 int sectors_per_chunk
=
2207 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
2209 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
2210 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
2212 raid5_compute_sector(conf
,
2213 stripe
* (disks
- conf
->max_degraded
)
2214 *sectors_per_chunk
+ chunk_offset
,
2220 handle_failed_stripe(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2221 struct stripe_head_state
*s
, int disks
,
2222 struct bio
**return_bi
)
2225 for (i
= disks
; i
--; ) {
2229 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2232 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2233 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
2234 /* multiple read failures in one stripe */
2235 md_error(conf
->mddev
, rdev
);
2238 spin_lock_irq(&conf
->device_lock
);
2239 /* fail all writes first */
2240 bi
= sh
->dev
[i
].towrite
;
2241 sh
->dev
[i
].towrite
= NULL
;
2247 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2248 wake_up(&conf
->wait_for_overlap
);
2250 while (bi
&& bi
->bi_sector
<
2251 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2252 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2253 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2254 if (!raid5_dec_bi_phys_segments(bi
)) {
2255 md_write_end(conf
->mddev
);
2256 bi
->bi_next
= *return_bi
;
2261 /* and fail all 'written' */
2262 bi
= sh
->dev
[i
].written
;
2263 sh
->dev
[i
].written
= NULL
;
2264 if (bi
) bitmap_end
= 1;
2265 while (bi
&& bi
->bi_sector
<
2266 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2267 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2268 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2269 if (!raid5_dec_bi_phys_segments(bi
)) {
2270 md_write_end(conf
->mddev
);
2271 bi
->bi_next
= *return_bi
;
2277 /* fail any reads if this device is non-operational and
2278 * the data has not reached the cache yet.
2280 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
2281 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
2282 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
2283 bi
= sh
->dev
[i
].toread
;
2284 sh
->dev
[i
].toread
= NULL
;
2285 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2286 wake_up(&conf
->wait_for_overlap
);
2287 if (bi
) s
->to_read
--;
2288 while (bi
&& bi
->bi_sector
<
2289 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2290 struct bio
*nextbi
=
2291 r5_next_bio(bi
, sh
->dev
[i
].sector
);
2292 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2293 if (!raid5_dec_bi_phys_segments(bi
)) {
2294 bi
->bi_next
= *return_bi
;
2300 spin_unlock_irq(&conf
->device_lock
);
2302 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2303 STRIPE_SECTORS
, 0, 0);
2306 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2307 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2308 md_wakeup_thread(conf
->mddev
->thread
);
2311 /* fetch_block - checks the given member device to see if its data needs
2312 * to be read or computed to satisfy a request.
2314 * Returns 1 when no more member devices need to be checked, otherwise returns
2315 * 0 to tell the loop in handle_stripe_fill to continue
2317 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2318 int disk_idx
, int disks
)
2320 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
2321 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
2322 &sh
->dev
[s
->failed_num
[1]] };
2324 /* is the data in this block needed, and can we get it? */
2325 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2326 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2328 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2329 s
->syncing
|| s
->expanding
||
2330 (s
->failed
>= 1 && fdev
[0]->toread
) ||
2331 (s
->failed
>= 2 && fdev
[1]->toread
) ||
2332 (sh
->raid_conf
->level
<= 5 && s
->failed
&& fdev
[0]->towrite
&&
2333 !test_bit(R5_OVERWRITE
, &fdev
[0]->flags
)) ||
2334 (sh
->raid_conf
->level
== 6 && s
->failed
&& s
->to_write
))) {
2335 /* we would like to get this block, possibly by computing it,
2336 * otherwise read it if the backing disk is insync
2338 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
2339 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
2340 if ((s
->uptodate
== disks
- 1) &&
2341 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
2342 disk_idx
== s
->failed_num
[1]))) {
2343 /* have disk failed, and we're requested to fetch it;
2346 pr_debug("Computing stripe %llu block %d\n",
2347 (unsigned long long)sh
->sector
, disk_idx
);
2348 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2349 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2350 set_bit(R5_Wantcompute
, &dev
->flags
);
2351 sh
->ops
.target
= disk_idx
;
2352 sh
->ops
.target2
= -1; /* no 2nd target */
2354 /* Careful: from this point on 'uptodate' is in the eye
2355 * of raid_run_ops which services 'compute' operations
2356 * before writes. R5_Wantcompute flags a block that will
2357 * be R5_UPTODATE by the time it is needed for a
2358 * subsequent operation.
2362 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
2363 /* Computing 2-failure is *very* expensive; only
2364 * do it if failed >= 2
2367 for (other
= disks
; other
--; ) {
2368 if (other
== disk_idx
)
2370 if (!test_bit(R5_UPTODATE
,
2371 &sh
->dev
[other
].flags
))
2375 pr_debug("Computing stripe %llu blocks %d,%d\n",
2376 (unsigned long long)sh
->sector
,
2378 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2379 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2380 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
2381 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
2382 sh
->ops
.target
= disk_idx
;
2383 sh
->ops
.target2
= other
;
2387 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2388 set_bit(R5_LOCKED
, &dev
->flags
);
2389 set_bit(R5_Wantread
, &dev
->flags
);
2391 pr_debug("Reading block %d (sync=%d)\n",
2392 disk_idx
, s
->syncing
);
2400 * handle_stripe_fill - read or compute data to satisfy pending requests.
2402 static void handle_stripe_fill(struct stripe_head
*sh
,
2403 struct stripe_head_state
*s
,
2408 /* look for blocks to read/compute, skip this if a compute
2409 * is already in flight, or if the stripe contents are in the
2410 * midst of changing due to a write
2412 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2413 !sh
->reconstruct_state
)
2414 for (i
= disks
; i
--; )
2415 if (fetch_block(sh
, s
, i
, disks
))
2417 set_bit(STRIPE_HANDLE
, &sh
->state
);
2421 /* handle_stripe_clean_event
2422 * any written block on an uptodate or failed drive can be returned.
2423 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2424 * never LOCKED, so we don't need to test 'failed' directly.
2426 static void handle_stripe_clean_event(raid5_conf_t
*conf
,
2427 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
2432 for (i
= disks
; i
--; )
2433 if (sh
->dev
[i
].written
) {
2435 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2436 test_bit(R5_UPTODATE
, &dev
->flags
)) {
2437 /* We can return any write requests */
2438 struct bio
*wbi
, *wbi2
;
2440 pr_debug("Return write for disc %d\n", i
);
2441 spin_lock_irq(&conf
->device_lock
);
2443 dev
->written
= NULL
;
2444 while (wbi
&& wbi
->bi_sector
<
2445 dev
->sector
+ STRIPE_SECTORS
) {
2446 wbi2
= r5_next_bio(wbi
, dev
->sector
);
2447 if (!raid5_dec_bi_phys_segments(wbi
)) {
2448 md_write_end(conf
->mddev
);
2449 wbi
->bi_next
= *return_bi
;
2454 if (dev
->towrite
== NULL
)
2456 spin_unlock_irq(&conf
->device_lock
);
2458 bitmap_endwrite(conf
->mddev
->bitmap
,
2461 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
2466 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2467 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2468 md_wakeup_thread(conf
->mddev
->thread
);
2471 static void handle_stripe_dirtying(raid5_conf_t
*conf
,
2472 struct stripe_head
*sh
,
2473 struct stripe_head_state
*s
,
2476 int rmw
= 0, rcw
= 0, i
;
2477 if (conf
->max_degraded
== 2) {
2478 /* RAID6 requires 'rcw' in current implementation
2479 * Calculate the real rcw later - for now fake it
2480 * look like rcw is cheaper
2483 } else for (i
= disks
; i
--; ) {
2484 /* would I have to read this buffer for read_modify_write */
2485 struct r5dev
*dev
= &sh
->dev
[i
];
2486 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2487 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2488 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2489 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2490 if (test_bit(R5_Insync
, &dev
->flags
))
2493 rmw
+= 2*disks
; /* cannot read it */
2495 /* Would I have to read this buffer for reconstruct_write */
2496 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) && i
!= sh
->pd_idx
&&
2497 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2498 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2499 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2500 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2505 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2506 (unsigned long long)sh
->sector
, rmw
, rcw
);
2507 set_bit(STRIPE_HANDLE
, &sh
->state
);
2508 if (rmw
< rcw
&& rmw
> 0)
2509 /* prefer read-modify-write, but need to get some data */
2510 for (i
= disks
; i
--; ) {
2511 struct r5dev
*dev
= &sh
->dev
[i
];
2512 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2513 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2514 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2515 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2516 test_bit(R5_Insync
, &dev
->flags
)) {
2518 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2519 pr_debug("Read_old block "
2520 "%d for r-m-w\n", i
);
2521 set_bit(R5_LOCKED
, &dev
->flags
);
2522 set_bit(R5_Wantread
, &dev
->flags
);
2525 set_bit(STRIPE_DELAYED
, &sh
->state
);
2526 set_bit(STRIPE_HANDLE
, &sh
->state
);
2530 if (rcw
<= rmw
&& rcw
> 0) {
2531 /* want reconstruct write, but need to get some data */
2533 for (i
= disks
; i
--; ) {
2534 struct r5dev
*dev
= &sh
->dev
[i
];
2535 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2536 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
2537 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2538 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2539 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2541 if (!test_bit(R5_Insync
, &dev
->flags
))
2542 continue; /* it's a failed drive */
2544 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2545 pr_debug("Read_old block "
2546 "%d for Reconstruct\n", i
);
2547 set_bit(R5_LOCKED
, &dev
->flags
);
2548 set_bit(R5_Wantread
, &dev
->flags
);
2551 set_bit(STRIPE_DELAYED
, &sh
->state
);
2552 set_bit(STRIPE_HANDLE
, &sh
->state
);
2557 /* now if nothing is locked, and if we have enough data,
2558 * we can start a write request
2560 /* since handle_stripe can be called at any time we need to handle the
2561 * case where a compute block operation has been submitted and then a
2562 * subsequent call wants to start a write request. raid_run_ops only
2563 * handles the case where compute block and reconstruct are requested
2564 * simultaneously. If this is not the case then new writes need to be
2565 * held off until the compute completes.
2567 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
2568 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
2569 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
2570 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
2573 static void handle_parity_checks5(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2574 struct stripe_head_state
*s
, int disks
)
2576 struct r5dev
*dev
= NULL
;
2578 set_bit(STRIPE_HANDLE
, &sh
->state
);
2580 switch (sh
->check_state
) {
2581 case check_state_idle
:
2582 /* start a new check operation if there are no failures */
2583 if (s
->failed
== 0) {
2584 BUG_ON(s
->uptodate
!= disks
);
2585 sh
->check_state
= check_state_run
;
2586 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2587 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2591 dev
= &sh
->dev
[s
->failed_num
[0]];
2593 case check_state_compute_result
:
2594 sh
->check_state
= check_state_idle
;
2596 dev
= &sh
->dev
[sh
->pd_idx
];
2598 /* check that a write has not made the stripe insync */
2599 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2602 /* either failed parity check, or recovery is happening */
2603 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
2604 BUG_ON(s
->uptodate
!= disks
);
2606 set_bit(R5_LOCKED
, &dev
->flags
);
2608 set_bit(R5_Wantwrite
, &dev
->flags
);
2610 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2611 set_bit(STRIPE_INSYNC
, &sh
->state
);
2613 case check_state_run
:
2614 break; /* we will be called again upon completion */
2615 case check_state_check_result
:
2616 sh
->check_state
= check_state_idle
;
2618 /* if a failure occurred during the check operation, leave
2619 * STRIPE_INSYNC not set and let the stripe be handled again
2624 /* handle a successful check operation, if parity is correct
2625 * we are done. Otherwise update the mismatch count and repair
2626 * parity if !MD_RECOVERY_CHECK
2628 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
2629 /* parity is correct (on disc,
2630 * not in buffer any more)
2632 set_bit(STRIPE_INSYNC
, &sh
->state
);
2634 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2635 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2636 /* don't try to repair!! */
2637 set_bit(STRIPE_INSYNC
, &sh
->state
);
2639 sh
->check_state
= check_state_compute_run
;
2640 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2641 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2642 set_bit(R5_Wantcompute
,
2643 &sh
->dev
[sh
->pd_idx
].flags
);
2644 sh
->ops
.target
= sh
->pd_idx
;
2645 sh
->ops
.target2
= -1;
2650 case check_state_compute_run
:
2653 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
2654 __func__
, sh
->check_state
,
2655 (unsigned long long) sh
->sector
);
2661 static void handle_parity_checks6(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2662 struct stripe_head_state
*s
,
2665 int pd_idx
= sh
->pd_idx
;
2666 int qd_idx
= sh
->qd_idx
;
2669 set_bit(STRIPE_HANDLE
, &sh
->state
);
2671 BUG_ON(s
->failed
> 2);
2673 /* Want to check and possibly repair P and Q.
2674 * However there could be one 'failed' device, in which
2675 * case we can only check one of them, possibly using the
2676 * other to generate missing data
2679 switch (sh
->check_state
) {
2680 case check_state_idle
:
2681 /* start a new check operation if there are < 2 failures */
2682 if (s
->failed
== s
->q_failed
) {
2683 /* The only possible failed device holds Q, so it
2684 * makes sense to check P (If anything else were failed,
2685 * we would have used P to recreate it).
2687 sh
->check_state
= check_state_run
;
2689 if (!s
->q_failed
&& s
->failed
< 2) {
2690 /* Q is not failed, and we didn't use it to generate
2691 * anything, so it makes sense to check it
2693 if (sh
->check_state
== check_state_run
)
2694 sh
->check_state
= check_state_run_pq
;
2696 sh
->check_state
= check_state_run_q
;
2699 /* discard potentially stale zero_sum_result */
2700 sh
->ops
.zero_sum_result
= 0;
2702 if (sh
->check_state
== check_state_run
) {
2703 /* async_xor_zero_sum destroys the contents of P */
2704 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2707 if (sh
->check_state
>= check_state_run
&&
2708 sh
->check_state
<= check_state_run_pq
) {
2709 /* async_syndrome_zero_sum preserves P and Q, so
2710 * no need to mark them !uptodate here
2712 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2716 /* we have 2-disk failure */
2717 BUG_ON(s
->failed
!= 2);
2719 case check_state_compute_result
:
2720 sh
->check_state
= check_state_idle
;
2722 /* check that a write has not made the stripe insync */
2723 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2726 /* now write out any block on a failed drive,
2727 * or P or Q if they were recomputed
2729 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
2730 if (s
->failed
== 2) {
2731 dev
= &sh
->dev
[s
->failed_num
[1]];
2733 set_bit(R5_LOCKED
, &dev
->flags
);
2734 set_bit(R5_Wantwrite
, &dev
->flags
);
2736 if (s
->failed
>= 1) {
2737 dev
= &sh
->dev
[s
->failed_num
[0]];
2739 set_bit(R5_LOCKED
, &dev
->flags
);
2740 set_bit(R5_Wantwrite
, &dev
->flags
);
2742 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
2743 dev
= &sh
->dev
[pd_idx
];
2745 set_bit(R5_LOCKED
, &dev
->flags
);
2746 set_bit(R5_Wantwrite
, &dev
->flags
);
2748 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
2749 dev
= &sh
->dev
[qd_idx
];
2751 set_bit(R5_LOCKED
, &dev
->flags
);
2752 set_bit(R5_Wantwrite
, &dev
->flags
);
2754 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2756 set_bit(STRIPE_INSYNC
, &sh
->state
);
2758 case check_state_run
:
2759 case check_state_run_q
:
2760 case check_state_run_pq
:
2761 break; /* we will be called again upon completion */
2762 case check_state_check_result
:
2763 sh
->check_state
= check_state_idle
;
2765 /* handle a successful check operation, if parity is correct
2766 * we are done. Otherwise update the mismatch count and repair
2767 * parity if !MD_RECOVERY_CHECK
2769 if (sh
->ops
.zero_sum_result
== 0) {
2770 /* both parities are correct */
2772 set_bit(STRIPE_INSYNC
, &sh
->state
);
2774 /* in contrast to the raid5 case we can validate
2775 * parity, but still have a failure to write
2778 sh
->check_state
= check_state_compute_result
;
2779 /* Returning at this point means that we may go
2780 * off and bring p and/or q uptodate again so
2781 * we make sure to check zero_sum_result again
2782 * to verify if p or q need writeback
2786 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2787 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2788 /* don't try to repair!! */
2789 set_bit(STRIPE_INSYNC
, &sh
->state
);
2791 int *target
= &sh
->ops
.target
;
2793 sh
->ops
.target
= -1;
2794 sh
->ops
.target2
= -1;
2795 sh
->check_state
= check_state_compute_run
;
2796 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2797 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2798 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
2799 set_bit(R5_Wantcompute
,
2800 &sh
->dev
[pd_idx
].flags
);
2802 target
= &sh
->ops
.target2
;
2805 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
2806 set_bit(R5_Wantcompute
,
2807 &sh
->dev
[qd_idx
].flags
);
2814 case check_state_compute_run
:
2817 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
2818 __func__
, sh
->check_state
,
2819 (unsigned long long) sh
->sector
);
2824 static void handle_stripe_expansion(raid5_conf_t
*conf
, struct stripe_head
*sh
)
2828 /* We have read all the blocks in this stripe and now we need to
2829 * copy some of them into a target stripe for expand.
2831 struct dma_async_tx_descriptor
*tx
= NULL
;
2832 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
2833 for (i
= 0; i
< sh
->disks
; i
++)
2834 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
2836 struct stripe_head
*sh2
;
2837 struct async_submit_ctl submit
;
2839 sector_t bn
= compute_blocknr(sh
, i
, 1);
2840 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
2842 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
2844 /* so far only the early blocks of this stripe
2845 * have been requested. When later blocks
2846 * get requested, we will try again
2849 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
2850 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
2851 /* must have already done this block */
2852 release_stripe(sh2
);
2856 /* place all the copies on one channel */
2857 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
2858 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
2859 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
2862 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
2863 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
2864 for (j
= 0; j
< conf
->raid_disks
; j
++)
2865 if (j
!= sh2
->pd_idx
&&
2867 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
2869 if (j
== conf
->raid_disks
) {
2870 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
2871 set_bit(STRIPE_HANDLE
, &sh2
->state
);
2873 release_stripe(sh2
);
2876 /* done submitting copies, wait for them to complete */
2879 dma_wait_for_async_tx(tx
);
2885 * handle_stripe - do things to a stripe.
2887 * We lock the stripe and then examine the state of various bits
2888 * to see what needs to be done.
2890 * return some read request which now have data
2891 * return some write requests which are safely on disc
2892 * schedule a read on some buffers
2893 * schedule a write of some buffers
2894 * return confirmation of parity correctness
2896 * buffers are taken off read_list or write_list, and bh_cache buffers
2897 * get BH_Lock set before the stripe lock is released.
2901 static int handle_stripe5(struct stripe_head
*sh
, struct stripe_head_state
*s
)
2903 raid5_conf_t
*conf
= sh
->raid_conf
;
2904 int disks
= sh
->disks
, i
;
2907 /* Now to look around and see what can be done */
2909 spin_lock_irq(&conf
->device_lock
);
2910 for (i
=disks
; i
--; ) {
2915 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2916 "written %p\n", i
, dev
->flags
, dev
->toread
, dev
->read
,
2917 dev
->towrite
, dev
->written
);
2919 /* maybe we can request a biofill operation
2921 * new wantfill requests are only permitted while
2922 * ops_complete_biofill is guaranteed to be inactive
2924 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
2925 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
2926 set_bit(R5_Wantfill
, &dev
->flags
);
2928 /* now count some things */
2929 if (test_bit(R5_LOCKED
, &dev
->flags
))
2931 if (test_bit(R5_UPTODATE
, &dev
->flags
))
2933 if (test_bit(R5_Wantcompute
, &dev
->flags
))
2936 if (test_bit(R5_Wantfill
, &dev
->flags
))
2938 else if (dev
->toread
)
2942 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
2947 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2948 if (s
->blocked_rdev
== NULL
&&
2949 rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
2950 s
->blocked_rdev
= rdev
;
2951 atomic_inc(&rdev
->nr_pending
);
2953 clear_bit(R5_Insync
, &dev
->flags
);
2956 else if (test_bit(In_sync
, &rdev
->flags
))
2957 set_bit(R5_Insync
, &dev
->flags
);
2959 /* could be in-sync depending on recovery/reshape status */
2960 if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
2961 set_bit(R5_Insync
, &dev
->flags
);
2963 if (!test_bit(R5_Insync
, &dev
->flags
)) {
2964 /* The ReadError flag will just be confusing now */
2965 clear_bit(R5_ReadError
, &dev
->flags
);
2966 clear_bit(R5_ReWrite
, &dev
->flags
);
2968 if (test_bit(R5_ReadError
, &dev
->flags
))
2969 clear_bit(R5_Insync
, &dev
->flags
);
2970 if (!test_bit(R5_Insync
, &dev
->flags
)) {
2972 s
->failed_num
[0] = i
;
2975 spin_unlock_irq(&conf
->device_lock
);
2978 if (unlikely(s
->blocked_rdev
)) {
2979 if (s
->syncing
|| s
->expanding
|| s
->expanded
||
2980 s
->to_write
|| s
->written
) {
2981 set_bit(STRIPE_HANDLE
, &sh
->state
);
2984 /* There is nothing for the blocked_rdev to block */
2985 rdev_dec_pending(s
->blocked_rdev
, conf
->mddev
);
2986 s
->blocked_rdev
= NULL
;
2989 if (s
->to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
2990 set_bit(STRIPE_OP_BIOFILL
, &s
->ops_request
);
2991 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
2994 pr_debug("locked=%d uptodate=%d to_read=%d"
2995 " to_write=%d failed=%d failed_num=%d\n",
2996 s
->locked
, s
->uptodate
, s
->to_read
, s
->to_write
,
2997 s
->failed
, s
->failed_num
[0]);
2998 /* check if the array has lost two devices and, if so, some requests might
3001 if (s
->failed
> 1 && s
->to_read
+s
->to_write
+s
->written
)
3002 handle_failed_stripe(conf
, sh
, s
, disks
, &s
->return_bi
);
3003 if (s
->failed
> 1 && s
->syncing
) {
3004 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,0);
3005 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3009 /* might be able to return some write requests if the parity block
3010 * is safe, or on a failed drive
3012 dev
= &sh
->dev
[sh
->pd_idx
];
3014 ((test_bit(R5_Insync
, &dev
->flags
) &&
3015 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3016 test_bit(R5_UPTODATE
, &dev
->flags
)) ||
3017 (s
->failed
== 1 && s
->failed_num
[0] == sh
->pd_idx
)))
3018 handle_stripe_clean_event(conf
, sh
, disks
, &s
->return_bi
);
3020 /* Now we might consider reading some blocks, either to check/generate
3021 * parity, or to satisfy requests
3022 * or to load a block that is being partially written.
3024 if (s
->to_read
|| s
->non_overwrite
||
3025 (s
->syncing
&& (s
->uptodate
+ s
->compute
< disks
)) || s
->expanding
)
3026 handle_stripe_fill(sh
, s
, disks
);
3031 static int handle_stripe6(struct stripe_head
*sh
, struct stripe_head_state
*s
)
3033 raid5_conf_t
*conf
= sh
->raid_conf
;
3034 int disks
= sh
->disks
;
3035 int i
, pd_idx
= sh
->pd_idx
, qd_idx
= sh
->qd_idx
;
3036 struct r5dev
*dev
, *pdev
, *qdev
;
3038 /* Now to look around and see what can be done */
3041 spin_lock_irq(&conf
->device_lock
);
3042 for (i
=disks
; i
--; ) {
3046 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3047 i
, dev
->flags
, dev
->toread
, dev
->towrite
, dev
->written
);
3048 /* maybe we can reply to a read
3050 * new wantfill requests are only permitted while
3051 * ops_complete_biofill is guaranteed to be inactive
3053 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
3054 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
3055 set_bit(R5_Wantfill
, &dev
->flags
);
3057 /* now count some things */
3058 if (test_bit(R5_LOCKED
, &dev
->flags
))
3060 if (test_bit(R5_UPTODATE
, &dev
->flags
))
3062 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
3064 BUG_ON(s
->compute
> 2);
3067 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
3069 } else if (dev
->toread
)
3073 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3078 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3079 if (s
->blocked_rdev
== NULL
&&
3080 rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
3081 s
->blocked_rdev
= rdev
;
3082 atomic_inc(&rdev
->nr_pending
);
3084 clear_bit(R5_Insync
, &dev
->flags
);
3087 else if (test_bit(In_sync
, &rdev
->flags
))
3088 set_bit(R5_Insync
, &dev
->flags
);
3090 /* in sync if before recovery_offset */
3091 if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
3092 set_bit(R5_Insync
, &dev
->flags
);
3094 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3095 /* The ReadError flag will just be confusing now */
3096 clear_bit(R5_ReadError
, &dev
->flags
);
3097 clear_bit(R5_ReWrite
, &dev
->flags
);
3099 if (test_bit(R5_ReadError
, &dev
->flags
))
3100 clear_bit(R5_Insync
, &dev
->flags
);
3101 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3103 s
->failed_num
[s
->failed
] = i
;
3107 spin_unlock_irq(&conf
->device_lock
);
3110 if (unlikely(s
->blocked_rdev
)) {
3111 if (s
->syncing
|| s
->expanding
|| s
->expanded
||
3112 s
->to_write
|| s
->written
) {
3113 set_bit(STRIPE_HANDLE
, &sh
->state
);
3116 /* There is nothing for the blocked_rdev to block */
3117 rdev_dec_pending(s
->blocked_rdev
, conf
->mddev
);
3118 s
->blocked_rdev
= NULL
;
3121 if (s
->to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
3122 set_bit(STRIPE_OP_BIOFILL
, &s
->ops_request
);
3123 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
3126 pr_debug("locked=%d uptodate=%d to_read=%d"
3127 " to_write=%d failed=%d failed_num=%d,%d\n",
3128 s
->locked
, s
->uptodate
, s
->to_read
, s
->to_write
, s
->failed
,
3129 s
->failed_num
[0], s
->failed_num
[1]);
3130 /* check if the array has lost >2 devices and, if so, some requests
3131 * might need to be failed
3133 if (s
->failed
> 2 && s
->to_read
+s
->to_write
+s
->written
)
3134 handle_failed_stripe(conf
, sh
, s
, disks
, &s
->return_bi
);
3135 if (s
->failed
> 2 && s
->syncing
) {
3136 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,0);
3137 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3142 * might be able to return some write requests if the parity blocks
3143 * are safe, or on a failed drive
3145 pdev
= &sh
->dev
[pd_idx
];
3146 s
->p_failed
= (s
->failed
>= 1 && s
->failed_num
[0] == pd_idx
)
3147 || (s
->failed
>= 2 && s
->failed_num
[1] == pd_idx
);
3148 qdev
= &sh
->dev
[qd_idx
];
3149 s
->q_failed
= (s
->failed
>= 1 && s
->failed_num
[0] == qd_idx
)
3150 || (s
->failed
>= 2 && s
->failed_num
[1] == qd_idx
);
3153 (s
->p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
3154 && !test_bit(R5_LOCKED
, &pdev
->flags
)
3155 && test_bit(R5_UPTODATE
, &pdev
->flags
)))) &&
3156 (s
->q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
3157 && !test_bit(R5_LOCKED
, &qdev
->flags
)
3158 && test_bit(R5_UPTODATE
, &qdev
->flags
)))))
3159 handle_stripe_clean_event(conf
, sh
, disks
, &s
->return_bi
);
3161 /* Now we might consider reading some blocks, either to check/generate
3162 * parity, or to satisfy requests
3163 * or to load a block that is being partially written.
3165 if (s
->to_read
|| s
->non_overwrite
|| (s
->to_write
&& s
->failed
) ||
3166 (s
->syncing
&& (s
->uptodate
+ s
->compute
< disks
)) || s
->expanding
)
3167 handle_stripe_fill(sh
, s
, disks
);
3172 static void handle_stripe(struct stripe_head
*sh
)
3174 struct stripe_head_state s
;
3178 int disks
= sh
->disks
;
3179 raid5_conf_t
*conf
= sh
->raid_conf
;
3181 clear_bit(STRIPE_HANDLE
, &sh
->state
);
3182 if (test_and_set_bit(STRIPE_ACTIVE
, &sh
->state
)) {
3183 /* already being handled, ensure it gets handled
3184 * again when current action finishes */
3185 set_bit(STRIPE_HANDLE
, &sh
->state
);
3189 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
3190 set_bit(STRIPE_SYNCING
, &sh
->state
);
3191 clear_bit(STRIPE_INSYNC
, &sh
->state
);
3193 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3195 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3196 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3197 (unsigned long long)sh
->sector
, sh
->state
,
3198 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
3199 sh
->check_state
, sh
->reconstruct_state
);
3200 memset(&s
, 0, sizeof(s
));
3202 s
.syncing
= test_bit(STRIPE_SYNCING
, &sh
->state
);
3203 s
.expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3204 s
.expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3205 s
.failed_num
[0] = -1;
3206 s
.failed_num
[1] = -1;
3208 if (conf
->level
== 6)
3209 done
= handle_stripe6(sh
, &s
);
3211 done
= handle_stripe5(sh
, &s
);
3216 /* Now we check to see if any write operations have recently
3220 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
3222 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
3223 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
3224 sh
->reconstruct_state
= reconstruct_state_idle
;
3226 /* All the 'written' buffers and the parity block are ready to
3227 * be written back to disk
3229 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
));
3230 BUG_ON(sh
->qd_idx
>= 0 &&
3231 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
));
3232 for (i
= disks
; i
--; ) {
3233 struct r5dev
*dev
= &sh
->dev
[i
];
3234 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
3235 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3237 pr_debug("Writing block %d\n", i
);
3238 set_bit(R5_Wantwrite
, &dev
->flags
);
3241 if (!test_bit(R5_Insync
, &dev
->flags
) ||
3242 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3244 set_bit(STRIPE_INSYNC
, &sh
->state
);
3247 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3248 s
.dec_preread_active
= 1;
3251 /* Now to consider new write requests and what else, if anything
3252 * should be read. We do not handle new writes when:
3253 * 1/ A 'write' operation (copy+xor) is already in flight.
3254 * 2/ A 'check' operation is in flight, as it may clobber the parity
3257 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
3258 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
3260 /* maybe we need to check and possibly fix the parity for this stripe
3261 * Any reads will already have been scheduled, so we just see if enough
3262 * data is available. The parity check is held off while parity
3263 * dependent operations are in flight.
3265 if (sh
->check_state
||
3266 (s
.syncing
&& s
.locked
== 0 &&
3267 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
3268 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
3269 if (conf
->level
== 6)
3270 handle_parity_checks6(conf
, sh
, &s
, disks
);
3272 handle_parity_checks5(conf
, sh
, &s
, disks
);
3275 if (s
.syncing
&& s
.locked
== 0 && test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3276 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3277 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3280 /* If the failed drives are just a ReadError, then we might need
3281 * to progress the repair/check process
3283 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
3284 for (i
= 0; i
< s
.failed
; i
++) {
3285 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
3286 if (test_bit(R5_ReadError
, &dev
->flags
)
3287 && !test_bit(R5_LOCKED
, &dev
->flags
)
3288 && test_bit(R5_UPTODATE
, &dev
->flags
)
3290 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3291 set_bit(R5_Wantwrite
, &dev
->flags
);
3292 set_bit(R5_ReWrite
, &dev
->flags
);
3293 set_bit(R5_LOCKED
, &dev
->flags
);
3296 /* let's read it back */
3297 set_bit(R5_Wantread
, &dev
->flags
);
3298 set_bit(R5_LOCKED
, &dev
->flags
);
3305 /* Finish reconstruct operations initiated by the expansion process */
3306 if (sh
->reconstruct_state
== reconstruct_state_result
) {
3307 struct stripe_head
*sh_src
3308 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
3309 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
3310 /* sh cannot be written until sh_src has been read.
3311 * so arrange for sh to be delayed a little
3313 set_bit(STRIPE_DELAYED
, &sh
->state
);
3314 set_bit(STRIPE_HANDLE
, &sh
->state
);
3315 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
3317 atomic_inc(&conf
->preread_active_stripes
);
3318 release_stripe(sh_src
);
3322 release_stripe(sh_src
);
3324 sh
->reconstruct_state
= reconstruct_state_idle
;
3325 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3326 for (i
= conf
->raid_disks
; i
--; ) {
3327 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3328 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3333 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
3334 !sh
->reconstruct_state
) {
3335 /* Need to write out all blocks after computing parity */
3336 sh
->disks
= conf
->raid_disks
;
3337 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
3338 schedule_reconstruction(sh
, &s
, 1, 1);
3339 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
3340 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3341 atomic_dec(&conf
->reshape_stripes
);
3342 wake_up(&conf
->wait_for_overlap
);
3343 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3346 if (s
.expanding
&& s
.locked
== 0 &&
3347 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
3348 handle_stripe_expansion(conf
, sh
);
3351 /* wait for this device to become unblocked */
3352 if (unlikely(s
.blocked_rdev
))
3353 md_wait_for_blocked_rdev(s
.blocked_rdev
, conf
->mddev
);
3356 raid_run_ops(sh
, s
.ops_request
);
3361 if (s
.dec_preread_active
) {
3362 /* We delay this until after ops_run_io so that if make_request
3363 * is waiting on a flush, it won't continue until the writes
3364 * have actually been submitted.
3366 atomic_dec(&conf
->preread_active_stripes
);
3367 if (atomic_read(&conf
->preread_active_stripes
) <
3369 md_wakeup_thread(conf
->mddev
->thread
);
3372 return_io(s
.return_bi
);
3374 clear_bit(STRIPE_ACTIVE
, &sh
->state
);
3377 static void raid5_activate_delayed(raid5_conf_t
*conf
)
3379 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
3380 while (!list_empty(&conf
->delayed_list
)) {
3381 struct list_head
*l
= conf
->delayed_list
.next
;
3382 struct stripe_head
*sh
;
3383 sh
= list_entry(l
, struct stripe_head
, lru
);
3385 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3386 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3387 atomic_inc(&conf
->preread_active_stripes
);
3388 list_add_tail(&sh
->lru
, &conf
->hold_list
);
3393 static void activate_bit_delay(raid5_conf_t
*conf
)
3395 /* device_lock is held */
3396 struct list_head head
;
3397 list_add(&head
, &conf
->bitmap_list
);
3398 list_del_init(&conf
->bitmap_list
);
3399 while (!list_empty(&head
)) {
3400 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
3401 list_del_init(&sh
->lru
);
3402 atomic_inc(&sh
->count
);
3403 __release_stripe(conf
, sh
);
3407 int md_raid5_congested(mddev_t
*mddev
, int bits
)
3409 raid5_conf_t
*conf
= mddev
->private;
3411 /* No difference between reads and writes. Just check
3412 * how busy the stripe_cache is
3415 if (conf
->inactive_blocked
)
3419 if (list_empty_careful(&conf
->inactive_list
))
3424 EXPORT_SYMBOL_GPL(md_raid5_congested
);
3426 static int raid5_congested(void *data
, int bits
)
3428 mddev_t
*mddev
= data
;
3430 return mddev_congested(mddev
, bits
) ||
3431 md_raid5_congested(mddev
, bits
);
3434 /* We want read requests to align with chunks where possible,
3435 * but write requests don't need to.
3437 static int raid5_mergeable_bvec(struct request_queue
*q
,
3438 struct bvec_merge_data
*bvm
,
3439 struct bio_vec
*biovec
)
3441 mddev_t
*mddev
= q
->queuedata
;
3442 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
3444 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3445 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
3447 if ((bvm
->bi_rw
& 1) == WRITE
)
3448 return biovec
->bv_len
; /* always allow writes to be mergeable */
3450 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3451 chunk_sectors
= mddev
->new_chunk_sectors
;
3452 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
3453 if (max
< 0) max
= 0;
3454 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
3455 return biovec
->bv_len
;
3461 static int in_chunk_boundary(mddev_t
*mddev
, struct bio
*bio
)
3463 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
3464 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3465 unsigned int bio_sectors
= bio
->bi_size
>> 9;
3467 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3468 chunk_sectors
= mddev
->new_chunk_sectors
;
3469 return chunk_sectors
>=
3470 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
3474 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3475 * later sampled by raid5d.
3477 static void add_bio_to_retry(struct bio
*bi
,raid5_conf_t
*conf
)
3479 unsigned long flags
;
3481 spin_lock_irqsave(&conf
->device_lock
, flags
);
3483 bi
->bi_next
= conf
->retry_read_aligned_list
;
3484 conf
->retry_read_aligned_list
= bi
;
3486 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3487 md_wakeup_thread(conf
->mddev
->thread
);
3491 static struct bio
*remove_bio_from_retry(raid5_conf_t
*conf
)
3495 bi
= conf
->retry_read_aligned
;
3497 conf
->retry_read_aligned
= NULL
;
3500 bi
= conf
->retry_read_aligned_list
;
3502 conf
->retry_read_aligned_list
= bi
->bi_next
;
3505 * this sets the active strip count to 1 and the processed
3506 * strip count to zero (upper 8 bits)
3508 bi
->bi_phys_segments
= 1; /* biased count of active stripes */
3516 * The "raid5_align_endio" should check if the read succeeded and if it
3517 * did, call bio_endio on the original bio (having bio_put the new bio
3519 * If the read failed..
3521 static void raid5_align_endio(struct bio
*bi
, int error
)
3523 struct bio
* raid_bi
= bi
->bi_private
;
3526 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3531 rdev
= (void*)raid_bi
->bi_next
;
3532 raid_bi
->bi_next
= NULL
;
3533 mddev
= rdev
->mddev
;
3534 conf
= mddev
->private;
3536 rdev_dec_pending(rdev
, conf
->mddev
);
3538 if (!error
&& uptodate
) {
3539 bio_endio(raid_bi
, 0);
3540 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
3541 wake_up(&conf
->wait_for_stripe
);
3546 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3548 add_bio_to_retry(raid_bi
, conf
);
3551 static int bio_fits_rdev(struct bio
*bi
)
3553 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
3555 if ((bi
->bi_size
>>9) > queue_max_sectors(q
))
3557 blk_recount_segments(q
, bi
);
3558 if (bi
->bi_phys_segments
> queue_max_segments(q
))
3561 if (q
->merge_bvec_fn
)
3562 /* it's too hard to apply the merge_bvec_fn at this stage,
3571 static int chunk_aligned_read(mddev_t
*mddev
, struct bio
* raid_bio
)
3573 raid5_conf_t
*conf
= mddev
->private;
3575 struct bio
* align_bi
;
3578 if (!in_chunk_boundary(mddev
, raid_bio
)) {
3579 pr_debug("chunk_aligned_read : non aligned\n");
3583 * use bio_clone_mddev to make a copy of the bio
3585 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
3589 * set bi_end_io to a new function, and set bi_private to the
3592 align_bi
->bi_end_io
= raid5_align_endio
;
3593 align_bi
->bi_private
= raid_bio
;
3597 align_bi
->bi_sector
= raid5_compute_sector(conf
, raid_bio
->bi_sector
,
3602 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
3603 if (rdev
&& test_bit(In_sync
, &rdev
->flags
)) {
3604 atomic_inc(&rdev
->nr_pending
);
3606 raid_bio
->bi_next
= (void*)rdev
;
3607 align_bi
->bi_bdev
= rdev
->bdev
;
3608 align_bi
->bi_flags
&= ~(1 << BIO_SEG_VALID
);
3609 align_bi
->bi_sector
+= rdev
->data_offset
;
3611 if (!bio_fits_rdev(align_bi
)) {
3612 /* too big in some way */
3614 rdev_dec_pending(rdev
, mddev
);
3618 spin_lock_irq(&conf
->device_lock
);
3619 wait_event_lock_irq(conf
->wait_for_stripe
,
3621 conf
->device_lock
, /* nothing */);
3622 atomic_inc(&conf
->active_aligned_reads
);
3623 spin_unlock_irq(&conf
->device_lock
);
3625 generic_make_request(align_bi
);
3634 /* __get_priority_stripe - get the next stripe to process
3636 * Full stripe writes are allowed to pass preread active stripes up until
3637 * the bypass_threshold is exceeded. In general the bypass_count
3638 * increments when the handle_list is handled before the hold_list; however, it
3639 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3640 * stripe with in flight i/o. The bypass_count will be reset when the
3641 * head of the hold_list has changed, i.e. the head was promoted to the
3644 static struct stripe_head
*__get_priority_stripe(raid5_conf_t
*conf
)
3646 struct stripe_head
*sh
;
3648 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3650 list_empty(&conf
->handle_list
) ? "empty" : "busy",
3651 list_empty(&conf
->hold_list
) ? "empty" : "busy",
3652 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
3654 if (!list_empty(&conf
->handle_list
)) {
3655 sh
= list_entry(conf
->handle_list
.next
, typeof(*sh
), lru
);
3657 if (list_empty(&conf
->hold_list
))
3658 conf
->bypass_count
= 0;
3659 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
3660 if (conf
->hold_list
.next
== conf
->last_hold
)
3661 conf
->bypass_count
++;
3663 conf
->last_hold
= conf
->hold_list
.next
;
3664 conf
->bypass_count
-= conf
->bypass_threshold
;
3665 if (conf
->bypass_count
< 0)
3666 conf
->bypass_count
= 0;
3669 } else if (!list_empty(&conf
->hold_list
) &&
3670 ((conf
->bypass_threshold
&&
3671 conf
->bypass_count
> conf
->bypass_threshold
) ||
3672 atomic_read(&conf
->pending_full_writes
) == 0)) {
3673 sh
= list_entry(conf
->hold_list
.next
,
3675 conf
->bypass_count
-= conf
->bypass_threshold
;
3676 if (conf
->bypass_count
< 0)
3677 conf
->bypass_count
= 0;
3681 list_del_init(&sh
->lru
);
3682 atomic_inc(&sh
->count
);
3683 BUG_ON(atomic_read(&sh
->count
) != 1);
3687 static int make_request(mddev_t
*mddev
, struct bio
* bi
)
3689 raid5_conf_t
*conf
= mddev
->private;
3691 sector_t new_sector
;
3692 sector_t logical_sector
, last_sector
;
3693 struct stripe_head
*sh
;
3694 const int rw
= bio_data_dir(bi
);
3698 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
3699 md_flush_request(mddev
, bi
);
3703 md_write_start(mddev
, bi
);
3706 mddev
->reshape_position
== MaxSector
&&
3707 chunk_aligned_read(mddev
,bi
))
3710 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
3711 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
3713 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
3715 plugged
= mddev_check_plugged(mddev
);
3716 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
3718 int disks
, data_disks
;
3723 disks
= conf
->raid_disks
;
3724 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
3725 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
3726 /* spinlock is needed as reshape_progress may be
3727 * 64bit on a 32bit platform, and so it might be
3728 * possible to see a half-updated value
3729 * Of course reshape_progress could change after
3730 * the lock is dropped, so once we get a reference
3731 * to the stripe that we think it is, we will have
3734 spin_lock_irq(&conf
->device_lock
);
3735 if (mddev
->delta_disks
< 0
3736 ? logical_sector
< conf
->reshape_progress
3737 : logical_sector
>= conf
->reshape_progress
) {
3738 disks
= conf
->previous_raid_disks
;
3741 if (mddev
->delta_disks
< 0
3742 ? logical_sector
< conf
->reshape_safe
3743 : logical_sector
>= conf
->reshape_safe
) {
3744 spin_unlock_irq(&conf
->device_lock
);
3749 spin_unlock_irq(&conf
->device_lock
);
3751 data_disks
= disks
- conf
->max_degraded
;
3753 new_sector
= raid5_compute_sector(conf
, logical_sector
,
3756 pr_debug("raid456: make_request, sector %llu logical %llu\n",
3757 (unsigned long long)new_sector
,
3758 (unsigned long long)logical_sector
);
3760 sh
= get_active_stripe(conf
, new_sector
, previous
,
3761 (bi
->bi_rw
&RWA_MASK
), 0);
3763 if (unlikely(previous
)) {
3764 /* expansion might have moved on while waiting for a
3765 * stripe, so we must do the range check again.
3766 * Expansion could still move past after this
3767 * test, but as we are holding a reference to
3768 * 'sh', we know that if that happens,
3769 * STRIPE_EXPANDING will get set and the expansion
3770 * won't proceed until we finish with the stripe.
3773 spin_lock_irq(&conf
->device_lock
);
3774 if (mddev
->delta_disks
< 0
3775 ? logical_sector
>= conf
->reshape_progress
3776 : logical_sector
< conf
->reshape_progress
)
3777 /* mismatch, need to try again */
3779 spin_unlock_irq(&conf
->device_lock
);
3788 logical_sector
>= mddev
->suspend_lo
&&
3789 logical_sector
< mddev
->suspend_hi
) {
3791 /* As the suspend_* range is controlled by
3792 * userspace, we want an interruptible
3795 flush_signals(current
);
3796 prepare_to_wait(&conf
->wait_for_overlap
,
3797 &w
, TASK_INTERRUPTIBLE
);
3798 if (logical_sector
>= mddev
->suspend_lo
&&
3799 logical_sector
< mddev
->suspend_hi
)
3804 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
3805 !add_stripe_bio(sh
, bi
, dd_idx
, rw
)) {
3806 /* Stripe is busy expanding or
3807 * add failed due to overlap. Flush everything
3810 md_wakeup_thread(mddev
->thread
);
3815 finish_wait(&conf
->wait_for_overlap
, &w
);
3816 set_bit(STRIPE_HANDLE
, &sh
->state
);
3817 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3818 if ((bi
->bi_rw
& REQ_SYNC
) &&
3819 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3820 atomic_inc(&conf
->preread_active_stripes
);
3823 /* cannot get stripe for read-ahead, just give-up */
3824 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3825 finish_wait(&conf
->wait_for_overlap
, &w
);
3831 md_wakeup_thread(mddev
->thread
);
3833 spin_lock_irq(&conf
->device_lock
);
3834 remaining
= raid5_dec_bi_phys_segments(bi
);
3835 spin_unlock_irq(&conf
->device_lock
);
3836 if (remaining
== 0) {
3839 md_write_end(mddev
);
3847 static sector_t
raid5_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
);
3849 static sector_t
reshape_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
)
3851 /* reshaping is quite different to recovery/resync so it is
3852 * handled quite separately ... here.
3854 * On each call to sync_request, we gather one chunk worth of
3855 * destination stripes and flag them as expanding.
3856 * Then we find all the source stripes and request reads.
3857 * As the reads complete, handle_stripe will copy the data
3858 * into the destination stripe and release that stripe.
3860 raid5_conf_t
*conf
= mddev
->private;
3861 struct stripe_head
*sh
;
3862 sector_t first_sector
, last_sector
;
3863 int raid_disks
= conf
->previous_raid_disks
;
3864 int data_disks
= raid_disks
- conf
->max_degraded
;
3865 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
3868 sector_t writepos
, readpos
, safepos
;
3869 sector_t stripe_addr
;
3870 int reshape_sectors
;
3871 struct list_head stripes
;
3873 if (sector_nr
== 0) {
3874 /* If restarting in the middle, skip the initial sectors */
3875 if (mddev
->delta_disks
< 0 &&
3876 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
3877 sector_nr
= raid5_size(mddev
, 0, 0)
3878 - conf
->reshape_progress
;
3879 } else if (mddev
->delta_disks
>= 0 &&
3880 conf
->reshape_progress
> 0)
3881 sector_nr
= conf
->reshape_progress
;
3882 sector_div(sector_nr
, new_data_disks
);
3884 mddev
->curr_resync_completed
= sector_nr
;
3885 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
3891 /* We need to process a full chunk at a time.
3892 * If old and new chunk sizes differ, we need to process the
3895 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
)
3896 reshape_sectors
= mddev
->new_chunk_sectors
;
3898 reshape_sectors
= mddev
->chunk_sectors
;
3900 /* we update the metadata when there is more than 3Meg
3901 * in the block range (that is rather arbitrary, should
3902 * probably be time based) or when the data about to be
3903 * copied would over-write the source of the data at
3904 * the front of the range.
3905 * i.e. one new_stripe along from reshape_progress new_maps
3906 * to after where reshape_safe old_maps to
3908 writepos
= conf
->reshape_progress
;
3909 sector_div(writepos
, new_data_disks
);
3910 readpos
= conf
->reshape_progress
;
3911 sector_div(readpos
, data_disks
);
3912 safepos
= conf
->reshape_safe
;
3913 sector_div(safepos
, data_disks
);
3914 if (mddev
->delta_disks
< 0) {
3915 writepos
-= min_t(sector_t
, reshape_sectors
, writepos
);
3916 readpos
+= reshape_sectors
;
3917 safepos
+= reshape_sectors
;
3919 writepos
+= reshape_sectors
;
3920 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
3921 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
3924 /* 'writepos' is the most advanced device address we might write.
3925 * 'readpos' is the least advanced device address we might read.
3926 * 'safepos' is the least address recorded in the metadata as having
3928 * If 'readpos' is behind 'writepos', then there is no way that we can
3929 * ensure safety in the face of a crash - that must be done by userspace
3930 * making a backup of the data. So in that case there is no particular
3931 * rush to update metadata.
3932 * Otherwise if 'safepos' is behind 'writepos', then we really need to
3933 * update the metadata to advance 'safepos' to match 'readpos' so that
3934 * we can be safe in the event of a crash.
3935 * So we insist on updating metadata if safepos is behind writepos and
3936 * readpos is beyond writepos.
3937 * In any case, update the metadata every 10 seconds.
3938 * Maybe that number should be configurable, but I'm not sure it is
3939 * worth it.... maybe it could be a multiple of safemode_delay???
3941 if ((mddev
->delta_disks
< 0
3942 ? (safepos
> writepos
&& readpos
< writepos
)
3943 : (safepos
< writepos
&& readpos
> writepos
)) ||
3944 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
3945 /* Cannot proceed until we've updated the superblock... */
3946 wait_event(conf
->wait_for_overlap
,
3947 atomic_read(&conf
->reshape_stripes
)==0);
3948 mddev
->reshape_position
= conf
->reshape_progress
;
3949 mddev
->curr_resync_completed
= sector_nr
;
3950 conf
->reshape_checkpoint
= jiffies
;
3951 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
3952 md_wakeup_thread(mddev
->thread
);
3953 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
3954 kthread_should_stop());
3955 spin_lock_irq(&conf
->device_lock
);
3956 conf
->reshape_safe
= mddev
->reshape_position
;
3957 spin_unlock_irq(&conf
->device_lock
);
3958 wake_up(&conf
->wait_for_overlap
);
3959 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
3962 if (mddev
->delta_disks
< 0) {
3963 BUG_ON(conf
->reshape_progress
== 0);
3964 stripe_addr
= writepos
;
3965 BUG_ON((mddev
->dev_sectors
&
3966 ~((sector_t
)reshape_sectors
- 1))
3967 - reshape_sectors
- stripe_addr
3970 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
3971 stripe_addr
= sector_nr
;
3973 INIT_LIST_HEAD(&stripes
);
3974 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
3976 int skipped_disk
= 0;
3977 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
3978 set_bit(STRIPE_EXPANDING
, &sh
->state
);
3979 atomic_inc(&conf
->reshape_stripes
);
3980 /* If any of this stripe is beyond the end of the old
3981 * array, then we need to zero those blocks
3983 for (j
=sh
->disks
; j
--;) {
3985 if (j
== sh
->pd_idx
)
3987 if (conf
->level
== 6 &&
3990 s
= compute_blocknr(sh
, j
, 0);
3991 if (s
< raid5_size(mddev
, 0, 0)) {
3995 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
3996 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
3997 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
3999 if (!skipped_disk
) {
4000 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4001 set_bit(STRIPE_HANDLE
, &sh
->state
);
4003 list_add(&sh
->lru
, &stripes
);
4005 spin_lock_irq(&conf
->device_lock
);
4006 if (mddev
->delta_disks
< 0)
4007 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
4009 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
4010 spin_unlock_irq(&conf
->device_lock
);
4011 /* Ok, those stripe are ready. We can start scheduling
4012 * reads on the source stripes.
4013 * The source stripes are determined by mapping the first and last
4014 * block on the destination stripes.
4017 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
4020 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
4021 * new_data_disks
- 1),
4023 if (last_sector
>= mddev
->dev_sectors
)
4024 last_sector
= mddev
->dev_sectors
- 1;
4025 while (first_sector
<= last_sector
) {
4026 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
4027 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
4028 set_bit(STRIPE_HANDLE
, &sh
->state
);
4030 first_sector
+= STRIPE_SECTORS
;
4032 /* Now that the sources are clearly marked, we can release
4033 * the destination stripes
4035 while (!list_empty(&stripes
)) {
4036 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
4037 list_del_init(&sh
->lru
);
4040 /* If this takes us to the resync_max point where we have to pause,
4041 * then we need to write out the superblock.
4043 sector_nr
+= reshape_sectors
;
4044 if ((sector_nr
- mddev
->curr_resync_completed
) * 2
4045 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
4046 /* Cannot proceed until we've updated the superblock... */
4047 wait_event(conf
->wait_for_overlap
,
4048 atomic_read(&conf
->reshape_stripes
) == 0);
4049 mddev
->reshape_position
= conf
->reshape_progress
;
4050 mddev
->curr_resync_completed
= sector_nr
;
4051 conf
->reshape_checkpoint
= jiffies
;
4052 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4053 md_wakeup_thread(mddev
->thread
);
4054 wait_event(mddev
->sb_wait
,
4055 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
4056 || kthread_should_stop());
4057 spin_lock_irq(&conf
->device_lock
);
4058 conf
->reshape_safe
= mddev
->reshape_position
;
4059 spin_unlock_irq(&conf
->device_lock
);
4060 wake_up(&conf
->wait_for_overlap
);
4061 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4063 return reshape_sectors
;
4066 /* FIXME go_faster isn't used */
4067 static inline sector_t
sync_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
4069 raid5_conf_t
*conf
= mddev
->private;
4070 struct stripe_head
*sh
;
4071 sector_t max_sector
= mddev
->dev_sectors
;
4072 sector_t sync_blocks
;
4073 int still_degraded
= 0;
4076 if (sector_nr
>= max_sector
) {
4077 /* just being told to finish up .. nothing much to do */
4079 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
4084 if (mddev
->curr_resync
< max_sector
) /* aborted */
4085 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
4087 else /* completed sync */
4089 bitmap_close_sync(mddev
->bitmap
);
4094 /* Allow raid5_quiesce to complete */
4095 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
4097 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
4098 return reshape_request(mddev
, sector_nr
, skipped
);
4100 /* No need to check resync_max as we never do more than one
4101 * stripe, and as resync_max will always be on a chunk boundary,
4102 * if the check in md_do_sync didn't fire, there is no chance
4103 * of overstepping resync_max here
4106 /* if there is too many failed drives and we are trying
4107 * to resync, then assert that we are finished, because there is
4108 * nothing we can do.
4110 if (mddev
->degraded
>= conf
->max_degraded
&&
4111 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
4112 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
4116 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
4117 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
4118 !conf
->fullsync
&& sync_blocks
>= STRIPE_SECTORS
) {
4119 /* we can skip this block, and probably more */
4120 sync_blocks
/= STRIPE_SECTORS
;
4122 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
4126 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
4128 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
4130 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
4131 /* make sure we don't swamp the stripe cache if someone else
4132 * is trying to get access
4134 schedule_timeout_uninterruptible(1);
4136 /* Need to check if array will still be degraded after recovery/resync
4137 * We don't need to check the 'failed' flag as when that gets set,
4140 for (i
= 0; i
< conf
->raid_disks
; i
++)
4141 if (conf
->disks
[i
].rdev
== NULL
)
4144 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
4146 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
4151 return STRIPE_SECTORS
;
4154 static int retry_aligned_read(raid5_conf_t
*conf
, struct bio
*raid_bio
)
4156 /* We may not be able to submit a whole bio at once as there
4157 * may not be enough stripe_heads available.
4158 * We cannot pre-allocate enough stripe_heads as we may need
4159 * more than exist in the cache (if we allow ever large chunks).
4160 * So we do one stripe head at a time and record in
4161 * ->bi_hw_segments how many have been done.
4163 * We *know* that this entire raid_bio is in one chunk, so
4164 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4166 struct stripe_head
*sh
;
4168 sector_t sector
, logical_sector
, last_sector
;
4173 logical_sector
= raid_bio
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4174 sector
= raid5_compute_sector(conf
, logical_sector
,
4176 last_sector
= raid_bio
->bi_sector
+ (raid_bio
->bi_size
>>9);
4178 for (; logical_sector
< last_sector
;
4179 logical_sector
+= STRIPE_SECTORS
,
4180 sector
+= STRIPE_SECTORS
,
4183 if (scnt
< raid5_bi_hw_segments(raid_bio
))
4184 /* already done this stripe */
4187 sh
= get_active_stripe(conf
, sector
, 0, 1, 0);
4190 /* failed to get a stripe - must wait */
4191 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4192 conf
->retry_read_aligned
= raid_bio
;
4196 set_bit(R5_ReadError
, &sh
->dev
[dd_idx
].flags
);
4197 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0)) {
4199 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4200 conf
->retry_read_aligned
= raid_bio
;
4208 spin_lock_irq(&conf
->device_lock
);
4209 remaining
= raid5_dec_bi_phys_segments(raid_bio
);
4210 spin_unlock_irq(&conf
->device_lock
);
4212 bio_endio(raid_bio
, 0);
4213 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4214 wake_up(&conf
->wait_for_stripe
);
4220 * This is our raid5 kernel thread.
4222 * We scan the hash table for stripes which can be handled now.
4223 * During the scan, completed stripes are saved for us by the interrupt
4224 * handler, so that they will not have to wait for our next wakeup.
4226 static void raid5d(mddev_t
*mddev
)
4228 struct stripe_head
*sh
;
4229 raid5_conf_t
*conf
= mddev
->private;
4231 struct blk_plug plug
;
4233 pr_debug("+++ raid5d active\n");
4235 md_check_recovery(mddev
);
4237 blk_start_plug(&plug
);
4239 spin_lock_irq(&conf
->device_lock
);
4243 if (atomic_read(&mddev
->plug_cnt
) == 0 &&
4244 !list_empty(&conf
->bitmap_list
)) {
4245 /* Now is a good time to flush some bitmap updates */
4247 spin_unlock_irq(&conf
->device_lock
);
4248 bitmap_unplug(mddev
->bitmap
);
4249 spin_lock_irq(&conf
->device_lock
);
4250 conf
->seq_write
= conf
->seq_flush
;
4251 activate_bit_delay(conf
);
4253 if (atomic_read(&mddev
->plug_cnt
) == 0)
4254 raid5_activate_delayed(conf
);
4256 while ((bio
= remove_bio_from_retry(conf
))) {
4258 spin_unlock_irq(&conf
->device_lock
);
4259 ok
= retry_aligned_read(conf
, bio
);
4260 spin_lock_irq(&conf
->device_lock
);
4266 sh
= __get_priority_stripe(conf
);
4270 spin_unlock_irq(&conf
->device_lock
);
4277 spin_lock_irq(&conf
->device_lock
);
4279 pr_debug("%d stripes handled\n", handled
);
4281 spin_unlock_irq(&conf
->device_lock
);
4283 async_tx_issue_pending_all();
4284 blk_finish_plug(&plug
);
4286 pr_debug("--- raid5d inactive\n");
4290 raid5_show_stripe_cache_size(mddev_t
*mddev
, char *page
)
4292 raid5_conf_t
*conf
= mddev
->private;
4294 return sprintf(page
, "%d\n", conf
->max_nr_stripes
);
4300 raid5_set_cache_size(mddev_t
*mddev
, int size
)
4302 raid5_conf_t
*conf
= mddev
->private;
4305 if (size
<= 16 || size
> 32768)
4307 while (size
< conf
->max_nr_stripes
) {
4308 if (drop_one_stripe(conf
))
4309 conf
->max_nr_stripes
--;
4313 err
= md_allow_write(mddev
);
4316 while (size
> conf
->max_nr_stripes
) {
4317 if (grow_one_stripe(conf
))
4318 conf
->max_nr_stripes
++;
4323 EXPORT_SYMBOL(raid5_set_cache_size
);
4326 raid5_store_stripe_cache_size(mddev_t
*mddev
, const char *page
, size_t len
)
4328 raid5_conf_t
*conf
= mddev
->private;
4332 if (len
>= PAGE_SIZE
)
4337 if (strict_strtoul(page
, 10, &new))
4339 err
= raid5_set_cache_size(mddev
, new);
4345 static struct md_sysfs_entry
4346 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
4347 raid5_show_stripe_cache_size
,
4348 raid5_store_stripe_cache_size
);
4351 raid5_show_preread_threshold(mddev_t
*mddev
, char *page
)
4353 raid5_conf_t
*conf
= mddev
->private;
4355 return sprintf(page
, "%d\n", conf
->bypass_threshold
);
4361 raid5_store_preread_threshold(mddev_t
*mddev
, const char *page
, size_t len
)
4363 raid5_conf_t
*conf
= mddev
->private;
4365 if (len
>= PAGE_SIZE
)
4370 if (strict_strtoul(page
, 10, &new))
4372 if (new > conf
->max_nr_stripes
)
4374 conf
->bypass_threshold
= new;
4378 static struct md_sysfs_entry
4379 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
4381 raid5_show_preread_threshold
,
4382 raid5_store_preread_threshold
);
4385 stripe_cache_active_show(mddev_t
*mddev
, char *page
)
4387 raid5_conf_t
*conf
= mddev
->private;
4389 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
4394 static struct md_sysfs_entry
4395 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
4397 static struct attribute
*raid5_attrs
[] = {
4398 &raid5_stripecache_size
.attr
,
4399 &raid5_stripecache_active
.attr
,
4400 &raid5_preread_bypass_threshold
.attr
,
4403 static struct attribute_group raid5_attrs_group
= {
4405 .attrs
= raid5_attrs
,
4409 raid5_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
)
4411 raid5_conf_t
*conf
= mddev
->private;
4414 sectors
= mddev
->dev_sectors
;
4416 /* size is defined by the smallest of previous and new size */
4417 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
4419 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
4420 sectors
&= ~((sector_t
)mddev
->new_chunk_sectors
- 1);
4421 return sectors
* (raid_disks
- conf
->max_degraded
);
4424 static void raid5_free_percpu(raid5_conf_t
*conf
)
4426 struct raid5_percpu
*percpu
;
4433 for_each_possible_cpu(cpu
) {
4434 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4435 safe_put_page(percpu
->spare_page
);
4436 kfree(percpu
->scribble
);
4438 #ifdef CONFIG_HOTPLUG_CPU
4439 unregister_cpu_notifier(&conf
->cpu_notify
);
4443 free_percpu(conf
->percpu
);
4446 static void free_conf(raid5_conf_t
*conf
)
4448 shrink_stripes(conf
);
4449 raid5_free_percpu(conf
);
4451 kfree(conf
->stripe_hashtbl
);
4455 #ifdef CONFIG_HOTPLUG_CPU
4456 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
4459 raid5_conf_t
*conf
= container_of(nfb
, raid5_conf_t
, cpu_notify
);
4460 long cpu
= (long)hcpu
;
4461 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4464 case CPU_UP_PREPARE
:
4465 case CPU_UP_PREPARE_FROZEN
:
4466 if (conf
->level
== 6 && !percpu
->spare_page
)
4467 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
4468 if (!percpu
->scribble
)
4469 percpu
->scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
4471 if (!percpu
->scribble
||
4472 (conf
->level
== 6 && !percpu
->spare_page
)) {
4473 safe_put_page(percpu
->spare_page
);
4474 kfree(percpu
->scribble
);
4475 pr_err("%s: failed memory allocation for cpu%ld\n",
4477 return notifier_from_errno(-ENOMEM
);
4481 case CPU_DEAD_FROZEN
:
4482 safe_put_page(percpu
->spare_page
);
4483 kfree(percpu
->scribble
);
4484 percpu
->spare_page
= NULL
;
4485 percpu
->scribble
= NULL
;
4494 static int raid5_alloc_percpu(raid5_conf_t
*conf
)
4497 struct page
*spare_page
;
4498 struct raid5_percpu __percpu
*allcpus
;
4502 allcpus
= alloc_percpu(struct raid5_percpu
);
4505 conf
->percpu
= allcpus
;
4509 for_each_present_cpu(cpu
) {
4510 if (conf
->level
== 6) {
4511 spare_page
= alloc_page(GFP_KERNEL
);
4516 per_cpu_ptr(conf
->percpu
, cpu
)->spare_page
= spare_page
;
4518 scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
4523 per_cpu_ptr(conf
->percpu
, cpu
)->scribble
= scribble
;
4525 #ifdef CONFIG_HOTPLUG_CPU
4526 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
4527 conf
->cpu_notify
.priority
= 0;
4529 err
= register_cpu_notifier(&conf
->cpu_notify
);
4536 static raid5_conf_t
*setup_conf(mddev_t
*mddev
)
4539 int raid_disk
, memory
, max_disks
;
4541 struct disk_info
*disk
;
4543 if (mddev
->new_level
!= 5
4544 && mddev
->new_level
!= 4
4545 && mddev
->new_level
!= 6) {
4546 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4547 mdname(mddev
), mddev
->new_level
);
4548 return ERR_PTR(-EIO
);
4550 if ((mddev
->new_level
== 5
4551 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
4552 (mddev
->new_level
== 6
4553 && !algorithm_valid_raid6(mddev
->new_layout
))) {
4554 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
4555 mdname(mddev
), mddev
->new_layout
);
4556 return ERR_PTR(-EIO
);
4558 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
4559 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4560 mdname(mddev
), mddev
->raid_disks
);
4561 return ERR_PTR(-EINVAL
);
4564 if (!mddev
->new_chunk_sectors
||
4565 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
4566 !is_power_of_2(mddev
->new_chunk_sectors
)) {
4567 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
4568 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
4569 return ERR_PTR(-EINVAL
);
4572 conf
= kzalloc(sizeof(raid5_conf_t
), GFP_KERNEL
);
4575 spin_lock_init(&conf
->device_lock
);
4576 init_waitqueue_head(&conf
->wait_for_stripe
);
4577 init_waitqueue_head(&conf
->wait_for_overlap
);
4578 INIT_LIST_HEAD(&conf
->handle_list
);
4579 INIT_LIST_HEAD(&conf
->hold_list
);
4580 INIT_LIST_HEAD(&conf
->delayed_list
);
4581 INIT_LIST_HEAD(&conf
->bitmap_list
);
4582 INIT_LIST_HEAD(&conf
->inactive_list
);
4583 atomic_set(&conf
->active_stripes
, 0);
4584 atomic_set(&conf
->preread_active_stripes
, 0);
4585 atomic_set(&conf
->active_aligned_reads
, 0);
4586 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
4588 conf
->raid_disks
= mddev
->raid_disks
;
4589 if (mddev
->reshape_position
== MaxSector
)
4590 conf
->previous_raid_disks
= mddev
->raid_disks
;
4592 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4593 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
4594 conf
->scribble_len
= scribble_len(max_disks
);
4596 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
4601 conf
->mddev
= mddev
;
4603 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
4606 conf
->level
= mddev
->new_level
;
4607 if (raid5_alloc_percpu(conf
) != 0)
4610 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
4612 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
4613 raid_disk
= rdev
->raid_disk
;
4614 if (raid_disk
>= max_disks
4617 disk
= conf
->disks
+ raid_disk
;
4621 if (test_bit(In_sync
, &rdev
->flags
)) {
4622 char b
[BDEVNAME_SIZE
];
4623 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
4625 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
4626 } else if (rdev
->saved_raid_disk
!= raid_disk
)
4627 /* Cannot rely on bitmap to complete recovery */
4631 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
4632 conf
->level
= mddev
->new_level
;
4633 if (conf
->level
== 6)
4634 conf
->max_degraded
= 2;
4636 conf
->max_degraded
= 1;
4637 conf
->algorithm
= mddev
->new_layout
;
4638 conf
->max_nr_stripes
= NR_STRIPES
;
4639 conf
->reshape_progress
= mddev
->reshape_position
;
4640 if (conf
->reshape_progress
!= MaxSector
) {
4641 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
4642 conf
->prev_algo
= mddev
->layout
;
4645 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
4646 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
4647 if (grow_stripes(conf
, conf
->max_nr_stripes
)) {
4649 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4650 mdname(mddev
), memory
);
4653 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
4654 mdname(mddev
), memory
);
4656 conf
->thread
= md_register_thread(raid5d
, mddev
, NULL
);
4657 if (!conf
->thread
) {
4659 "md/raid:%s: couldn't allocate thread.\n",
4669 return ERR_PTR(-EIO
);
4671 return ERR_PTR(-ENOMEM
);
4675 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
4678 case ALGORITHM_PARITY_0
:
4679 if (raid_disk
< max_degraded
)
4682 case ALGORITHM_PARITY_N
:
4683 if (raid_disk
>= raid_disks
- max_degraded
)
4686 case ALGORITHM_PARITY_0_6
:
4687 if (raid_disk
== 0 ||
4688 raid_disk
== raid_disks
- 1)
4691 case ALGORITHM_LEFT_ASYMMETRIC_6
:
4692 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
4693 case ALGORITHM_LEFT_SYMMETRIC_6
:
4694 case ALGORITHM_RIGHT_SYMMETRIC_6
:
4695 if (raid_disk
== raid_disks
- 1)
4701 static int run(mddev_t
*mddev
)
4704 int working_disks
= 0;
4705 int dirty_parity_disks
= 0;
4707 sector_t reshape_offset
= 0;
4709 if (mddev
->recovery_cp
!= MaxSector
)
4710 printk(KERN_NOTICE
"md/raid:%s: not clean"
4711 " -- starting background reconstruction\n",
4713 if (mddev
->reshape_position
!= MaxSector
) {
4714 /* Check that we can continue the reshape.
4715 * Currently only disks can change, it must
4716 * increase, and we must be past the point where
4717 * a stripe over-writes itself
4719 sector_t here_new
, here_old
;
4721 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
4723 if (mddev
->new_level
!= mddev
->level
) {
4724 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
4725 "required - aborting.\n",
4729 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4730 /* reshape_position must be on a new-stripe boundary, and one
4731 * further up in new geometry must map after here in old
4734 here_new
= mddev
->reshape_position
;
4735 if (sector_div(here_new
, mddev
->new_chunk_sectors
*
4736 (mddev
->raid_disks
- max_degraded
))) {
4737 printk(KERN_ERR
"md/raid:%s: reshape_position not "
4738 "on a stripe boundary\n", mdname(mddev
));
4741 reshape_offset
= here_new
* mddev
->new_chunk_sectors
;
4742 /* here_new is the stripe we will write to */
4743 here_old
= mddev
->reshape_position
;
4744 sector_div(here_old
, mddev
->chunk_sectors
*
4745 (old_disks
-max_degraded
));
4746 /* here_old is the first stripe that we might need to read
4748 if (mddev
->delta_disks
== 0) {
4749 /* We cannot be sure it is safe to start an in-place
4750 * reshape. It is only safe if user-space if monitoring
4751 * and taking constant backups.
4752 * mdadm always starts a situation like this in
4753 * readonly mode so it can take control before
4754 * allowing any writes. So just check for that.
4756 if ((here_new
* mddev
->new_chunk_sectors
!=
4757 here_old
* mddev
->chunk_sectors
) ||
4759 printk(KERN_ERR
"md/raid:%s: in-place reshape must be started"
4760 " in read-only mode - aborting\n",
4764 } else if (mddev
->delta_disks
< 0
4765 ? (here_new
* mddev
->new_chunk_sectors
<=
4766 here_old
* mddev
->chunk_sectors
)
4767 : (here_new
* mddev
->new_chunk_sectors
>=
4768 here_old
* mddev
->chunk_sectors
)) {
4769 /* Reading from the same stripe as writing to - bad */
4770 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
4771 "auto-recovery - aborting.\n",
4775 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
4777 /* OK, we should be able to continue; */
4779 BUG_ON(mddev
->level
!= mddev
->new_level
);
4780 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
4781 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
4782 BUG_ON(mddev
->delta_disks
!= 0);
4785 if (mddev
->private == NULL
)
4786 conf
= setup_conf(mddev
);
4788 conf
= mddev
->private;
4791 return PTR_ERR(conf
);
4793 mddev
->thread
= conf
->thread
;
4794 conf
->thread
= NULL
;
4795 mddev
->private = conf
;
4798 * 0 for a fully functional array, 1 or 2 for a degraded array.
4800 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
4801 if (rdev
->raid_disk
< 0)
4803 if (test_bit(In_sync
, &rdev
->flags
)) {
4807 /* This disc is not fully in-sync. However if it
4808 * just stored parity (beyond the recovery_offset),
4809 * when we don't need to be concerned about the
4810 * array being dirty.
4811 * When reshape goes 'backwards', we never have
4812 * partially completed devices, so we only need
4813 * to worry about reshape going forwards.
4815 /* Hack because v0.91 doesn't store recovery_offset properly. */
4816 if (mddev
->major_version
== 0 &&
4817 mddev
->minor_version
> 90)
4818 rdev
->recovery_offset
= reshape_offset
;
4820 if (rdev
->recovery_offset
< reshape_offset
) {
4821 /* We need to check old and new layout */
4822 if (!only_parity(rdev
->raid_disk
,
4825 conf
->max_degraded
))
4828 if (!only_parity(rdev
->raid_disk
,
4830 conf
->previous_raid_disks
,
4831 conf
->max_degraded
))
4833 dirty_parity_disks
++;
4836 mddev
->degraded
= (max(conf
->raid_disks
, conf
->previous_raid_disks
)
4839 if (has_failed(conf
)) {
4840 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
4841 " (%d/%d failed)\n",
4842 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
4846 /* device size must be a multiple of chunk size */
4847 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
4848 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
4850 if (mddev
->degraded
> dirty_parity_disks
&&
4851 mddev
->recovery_cp
!= MaxSector
) {
4852 if (mddev
->ok_start_degraded
)
4854 "md/raid:%s: starting dirty degraded array"
4855 " - data corruption possible.\n",
4859 "md/raid:%s: cannot start dirty degraded array.\n",
4865 if (mddev
->degraded
== 0)
4866 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
4867 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
4868 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
4871 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
4872 " out of %d devices, algorithm %d\n",
4873 mdname(mddev
), conf
->level
,
4874 mddev
->raid_disks
- mddev
->degraded
,
4875 mddev
->raid_disks
, mddev
->new_layout
);
4877 print_raid5_conf(conf
);
4879 if (conf
->reshape_progress
!= MaxSector
) {
4880 conf
->reshape_safe
= conf
->reshape_progress
;
4881 atomic_set(&conf
->reshape_stripes
, 0);
4882 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4883 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4884 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4885 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4886 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4891 /* Ok, everything is just fine now */
4892 if (mddev
->to_remove
== &raid5_attrs_group
)
4893 mddev
->to_remove
= NULL
;
4894 else if (mddev
->kobj
.sd
&&
4895 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
4897 "raid5: failed to create sysfs attributes for %s\n",
4899 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
4903 /* read-ahead size must cover two whole stripes, which
4904 * is 2 * (datadisks) * chunksize where 'n' is the
4905 * number of raid devices
4907 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
4908 int stripe
= data_disks
*
4909 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
4910 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4911 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4913 blk_queue_merge_bvec(mddev
->queue
, raid5_mergeable_bvec
);
4915 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
4916 mddev
->queue
->backing_dev_info
.congested_fn
= raid5_congested
;
4918 chunk_size
= mddev
->chunk_sectors
<< 9;
4919 blk_queue_io_min(mddev
->queue
, chunk_size
);
4920 blk_queue_io_opt(mddev
->queue
, chunk_size
*
4921 (conf
->raid_disks
- conf
->max_degraded
));
4923 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
4924 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
4925 rdev
->data_offset
<< 9);
4930 md_unregister_thread(mddev
->thread
);
4931 mddev
->thread
= NULL
;
4933 print_raid5_conf(conf
);
4936 mddev
->private = NULL
;
4937 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
4941 static int stop(mddev_t
*mddev
)
4943 raid5_conf_t
*conf
= mddev
->private;
4945 md_unregister_thread(mddev
->thread
);
4946 mddev
->thread
= NULL
;
4948 mddev
->queue
->backing_dev_info
.congested_fn
= NULL
;
4950 mddev
->private = NULL
;
4951 mddev
->to_remove
= &raid5_attrs_group
;
4956 static void print_sh(struct seq_file
*seq
, struct stripe_head
*sh
)
4960 seq_printf(seq
, "sh %llu, pd_idx %d, state %ld.\n",
4961 (unsigned long long)sh
->sector
, sh
->pd_idx
, sh
->state
);
4962 seq_printf(seq
, "sh %llu, count %d.\n",
4963 (unsigned long long)sh
->sector
, atomic_read(&sh
->count
));
4964 seq_printf(seq
, "sh %llu, ", (unsigned long long)sh
->sector
);
4965 for (i
= 0; i
< sh
->disks
; i
++) {
4966 seq_printf(seq
, "(cache%d: %p %ld) ",
4967 i
, sh
->dev
[i
].page
, sh
->dev
[i
].flags
);
4969 seq_printf(seq
, "\n");
4972 static void printall(struct seq_file
*seq
, raid5_conf_t
*conf
)
4974 struct stripe_head
*sh
;
4975 struct hlist_node
*hn
;
4978 spin_lock_irq(&conf
->device_lock
);
4979 for (i
= 0; i
< NR_HASH
; i
++) {
4980 hlist_for_each_entry(sh
, hn
, &conf
->stripe_hashtbl
[i
], hash
) {
4981 if (sh
->raid_conf
!= conf
)
4986 spin_unlock_irq(&conf
->device_lock
);
4990 static void status(struct seq_file
*seq
, mddev_t
*mddev
)
4992 raid5_conf_t
*conf
= mddev
->private;
4995 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
4996 mddev
->chunk_sectors
/ 2, mddev
->layout
);
4997 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
4998 for (i
= 0; i
< conf
->raid_disks
; i
++)
4999 seq_printf (seq
, "%s",
5000 conf
->disks
[i
].rdev
&&
5001 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
5002 seq_printf (seq
, "]");
5004 seq_printf (seq
, "\n");
5005 printall(seq
, conf
);
5009 static void print_raid5_conf (raid5_conf_t
*conf
)
5012 struct disk_info
*tmp
;
5014 printk(KERN_DEBUG
"RAID conf printout:\n");
5016 printk("(conf==NULL)\n");
5019 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
5021 conf
->raid_disks
- conf
->mddev
->degraded
);
5023 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5024 char b
[BDEVNAME_SIZE
];
5025 tmp
= conf
->disks
+ i
;
5027 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
5028 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
5029 bdevname(tmp
->rdev
->bdev
, b
));
5033 static int raid5_spare_active(mddev_t
*mddev
)
5036 raid5_conf_t
*conf
= mddev
->private;
5037 struct disk_info
*tmp
;
5039 unsigned long flags
;
5041 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5042 tmp
= conf
->disks
+ i
;
5044 && tmp
->rdev
->recovery_offset
== MaxSector
5045 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
5046 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
5048 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
5051 spin_lock_irqsave(&conf
->device_lock
, flags
);
5052 mddev
->degraded
-= count
;
5053 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5054 print_raid5_conf(conf
);
5058 static int raid5_remove_disk(mddev_t
*mddev
, int number
)
5060 raid5_conf_t
*conf
= mddev
->private;
5063 struct disk_info
*p
= conf
->disks
+ number
;
5065 print_raid5_conf(conf
);
5068 if (number
>= conf
->raid_disks
&&
5069 conf
->reshape_progress
== MaxSector
)
5070 clear_bit(In_sync
, &rdev
->flags
);
5072 if (test_bit(In_sync
, &rdev
->flags
) ||
5073 atomic_read(&rdev
->nr_pending
)) {
5077 /* Only remove non-faulty devices if recovery
5080 if (!test_bit(Faulty
, &rdev
->flags
) &&
5081 !has_failed(conf
) &&
5082 number
< conf
->raid_disks
) {
5088 if (atomic_read(&rdev
->nr_pending
)) {
5089 /* lost the race, try later */
5096 print_raid5_conf(conf
);
5100 static int raid5_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
5102 raid5_conf_t
*conf
= mddev
->private;
5105 struct disk_info
*p
;
5107 int last
= conf
->raid_disks
- 1;
5109 if (has_failed(conf
))
5110 /* no point adding a device */
5113 if (rdev
->raid_disk
>= 0)
5114 first
= last
= rdev
->raid_disk
;
5117 * find the disk ... but prefer rdev->saved_raid_disk
5120 if (rdev
->saved_raid_disk
>= 0 &&
5121 rdev
->saved_raid_disk
>= first
&&
5122 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
5123 disk
= rdev
->saved_raid_disk
;
5126 for ( ; disk
<= last
; disk
++)
5127 if ((p
=conf
->disks
+ disk
)->rdev
== NULL
) {
5128 clear_bit(In_sync
, &rdev
->flags
);
5129 rdev
->raid_disk
= disk
;
5131 if (rdev
->saved_raid_disk
!= disk
)
5133 rcu_assign_pointer(p
->rdev
, rdev
);
5136 print_raid5_conf(conf
);
5140 static int raid5_resize(mddev_t
*mddev
, sector_t sectors
)
5142 /* no resync is happening, and there is enough space
5143 * on all devices, so we can resize.
5144 * We need to make sure resync covers any new space.
5145 * If the array is shrinking we should possibly wait until
5146 * any io in the removed space completes, but it hardly seems
5149 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
5150 md_set_array_sectors(mddev
, raid5_size(mddev
, sectors
,
5151 mddev
->raid_disks
));
5152 if (mddev
->array_sectors
>
5153 raid5_size(mddev
, sectors
, mddev
->raid_disks
))
5155 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5156 revalidate_disk(mddev
->gendisk
);
5157 if (sectors
> mddev
->dev_sectors
&&
5158 mddev
->recovery_cp
> mddev
->dev_sectors
) {
5159 mddev
->recovery_cp
= mddev
->dev_sectors
;
5160 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
5162 mddev
->dev_sectors
= sectors
;
5163 mddev
->resync_max_sectors
= sectors
;
5167 static int check_stripe_cache(mddev_t
*mddev
)
5169 /* Can only proceed if there are plenty of stripe_heads.
5170 * We need a minimum of one full stripe,, and for sensible progress
5171 * it is best to have about 4 times that.
5172 * If we require 4 times, then the default 256 4K stripe_heads will
5173 * allow for chunk sizes up to 256K, which is probably OK.
5174 * If the chunk size is greater, user-space should request more
5175 * stripe_heads first.
5177 raid5_conf_t
*conf
= mddev
->private;
5178 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5179 > conf
->max_nr_stripes
||
5180 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5181 > conf
->max_nr_stripes
) {
5182 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5184 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
5191 static int check_reshape(mddev_t
*mddev
)
5193 raid5_conf_t
*conf
= mddev
->private;
5195 if (mddev
->delta_disks
== 0 &&
5196 mddev
->new_layout
== mddev
->layout
&&
5197 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
5198 return 0; /* nothing to do */
5200 /* Cannot grow a bitmap yet */
5202 if (has_failed(conf
))
5204 if (mddev
->delta_disks
< 0) {
5205 /* We might be able to shrink, but the devices must
5206 * be made bigger first.
5207 * For raid6, 4 is the minimum size.
5208 * Otherwise 2 is the minimum
5211 if (mddev
->level
== 6)
5213 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
5217 if (!check_stripe_cache(mddev
))
5220 return resize_stripes(conf
, conf
->raid_disks
+ mddev
->delta_disks
);
5223 static int raid5_start_reshape(mddev_t
*mddev
)
5225 raid5_conf_t
*conf
= mddev
->private;
5228 unsigned long flags
;
5230 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
5233 if (!check_stripe_cache(mddev
))
5236 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
5237 if (!test_bit(In_sync
, &rdev
->flags
)
5238 && !test_bit(Faulty
, &rdev
->flags
))
5241 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
5242 /* Not enough devices even to make a degraded array
5247 /* Refuse to reduce size of the array. Any reductions in
5248 * array size must be through explicit setting of array_size
5251 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
5252 < mddev
->array_sectors
) {
5253 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
5254 "before number of disks\n", mdname(mddev
));
5258 atomic_set(&conf
->reshape_stripes
, 0);
5259 spin_lock_irq(&conf
->device_lock
);
5260 conf
->previous_raid_disks
= conf
->raid_disks
;
5261 conf
->raid_disks
+= mddev
->delta_disks
;
5262 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
5263 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
5264 conf
->prev_algo
= conf
->algorithm
;
5265 conf
->algorithm
= mddev
->new_layout
;
5266 if (mddev
->delta_disks
< 0)
5267 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
5269 conf
->reshape_progress
= 0;
5270 conf
->reshape_safe
= conf
->reshape_progress
;
5272 spin_unlock_irq(&conf
->device_lock
);
5274 /* Add some new drives, as many as will fit.
5275 * We know there are enough to make the newly sized array work.
5276 * Don't add devices if we are reducing the number of
5277 * devices in the array. This is because it is not possible
5278 * to correctly record the "partially reconstructed" state of
5279 * such devices during the reshape and confusion could result.
5281 if (mddev
->delta_disks
>= 0) {
5282 int added_devices
= 0;
5283 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
5284 if (rdev
->raid_disk
< 0 &&
5285 !test_bit(Faulty
, &rdev
->flags
)) {
5286 if (raid5_add_disk(mddev
, rdev
) == 0) {
5289 >= conf
->previous_raid_disks
) {
5290 set_bit(In_sync
, &rdev
->flags
);
5293 rdev
->recovery_offset
= 0;
5294 sprintf(nm
, "rd%d", rdev
->raid_disk
);
5295 if (sysfs_create_link(&mddev
->kobj
,
5297 /* Failure here is OK */;
5299 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
5300 && !test_bit(Faulty
, &rdev
->flags
)) {
5301 /* This is a spare that was manually added */
5302 set_bit(In_sync
, &rdev
->flags
);
5306 /* When a reshape changes the number of devices,
5307 * ->degraded is measured against the larger of the
5308 * pre and post number of devices.
5310 spin_lock_irqsave(&conf
->device_lock
, flags
);
5311 mddev
->degraded
+= (conf
->raid_disks
- conf
->previous_raid_disks
)
5313 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5315 mddev
->raid_disks
= conf
->raid_disks
;
5316 mddev
->reshape_position
= conf
->reshape_progress
;
5317 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5319 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5320 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5321 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5322 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5323 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5325 if (!mddev
->sync_thread
) {
5326 mddev
->recovery
= 0;
5327 spin_lock_irq(&conf
->device_lock
);
5328 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
5329 conf
->reshape_progress
= MaxSector
;
5330 spin_unlock_irq(&conf
->device_lock
);
5333 conf
->reshape_checkpoint
= jiffies
;
5334 md_wakeup_thread(mddev
->sync_thread
);
5335 md_new_event(mddev
);
5339 /* This is called from the reshape thread and should make any
5340 * changes needed in 'conf'
5342 static void end_reshape(raid5_conf_t
*conf
)
5345 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
5347 spin_lock_irq(&conf
->device_lock
);
5348 conf
->previous_raid_disks
= conf
->raid_disks
;
5349 conf
->reshape_progress
= MaxSector
;
5350 spin_unlock_irq(&conf
->device_lock
);
5351 wake_up(&conf
->wait_for_overlap
);
5353 /* read-ahead size must cover two whole stripes, which is
5354 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5356 if (conf
->mddev
->queue
) {
5357 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5358 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
5360 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
5361 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
5366 /* This is called from the raid5d thread with mddev_lock held.
5367 * It makes config changes to the device.
5369 static void raid5_finish_reshape(mddev_t
*mddev
)
5371 raid5_conf_t
*conf
= mddev
->private;
5373 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
5375 if (mddev
->delta_disks
> 0) {
5376 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5377 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5378 revalidate_disk(mddev
->gendisk
);
5381 mddev
->degraded
= conf
->raid_disks
;
5382 for (d
= 0; d
< conf
->raid_disks
; d
++)
5383 if (conf
->disks
[d
].rdev
&&
5385 &conf
->disks
[d
].rdev
->flags
))
5387 for (d
= conf
->raid_disks
;
5388 d
< conf
->raid_disks
- mddev
->delta_disks
;
5390 mdk_rdev_t
*rdev
= conf
->disks
[d
].rdev
;
5391 if (rdev
&& raid5_remove_disk(mddev
, d
) == 0) {
5393 sprintf(nm
, "rd%d", rdev
->raid_disk
);
5394 sysfs_remove_link(&mddev
->kobj
, nm
);
5395 rdev
->raid_disk
= -1;
5399 mddev
->layout
= conf
->algorithm
;
5400 mddev
->chunk_sectors
= conf
->chunk_sectors
;
5401 mddev
->reshape_position
= MaxSector
;
5402 mddev
->delta_disks
= 0;
5406 static void raid5_quiesce(mddev_t
*mddev
, int state
)
5408 raid5_conf_t
*conf
= mddev
->private;
5411 case 2: /* resume for a suspend */
5412 wake_up(&conf
->wait_for_overlap
);
5415 case 1: /* stop all writes */
5416 spin_lock_irq(&conf
->device_lock
);
5417 /* '2' tells resync/reshape to pause so that all
5418 * active stripes can drain
5421 wait_event_lock_irq(conf
->wait_for_stripe
,
5422 atomic_read(&conf
->active_stripes
) == 0 &&
5423 atomic_read(&conf
->active_aligned_reads
) == 0,
5424 conf
->device_lock
, /* nothing */);
5426 spin_unlock_irq(&conf
->device_lock
);
5427 /* allow reshape to continue */
5428 wake_up(&conf
->wait_for_overlap
);
5431 case 0: /* re-enable writes */
5432 spin_lock_irq(&conf
->device_lock
);
5434 wake_up(&conf
->wait_for_stripe
);
5435 wake_up(&conf
->wait_for_overlap
);
5436 spin_unlock_irq(&conf
->device_lock
);
5442 static void *raid45_takeover_raid0(mddev_t
*mddev
, int level
)
5444 struct raid0_private_data
*raid0_priv
= mddev
->private;
5447 /* for raid0 takeover only one zone is supported */
5448 if (raid0_priv
->nr_strip_zones
> 1) {
5449 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5451 return ERR_PTR(-EINVAL
);
5454 sectors
= raid0_priv
->strip_zone
[0].zone_end
;
5455 sector_div(sectors
, raid0_priv
->strip_zone
[0].nb_dev
);
5456 mddev
->dev_sectors
= sectors
;
5457 mddev
->new_level
= level
;
5458 mddev
->new_layout
= ALGORITHM_PARITY_N
;
5459 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
5460 mddev
->raid_disks
+= 1;
5461 mddev
->delta_disks
= 1;
5462 /* make sure it will be not marked as dirty */
5463 mddev
->recovery_cp
= MaxSector
;
5465 return setup_conf(mddev
);
5469 static void *raid5_takeover_raid1(mddev_t
*mddev
)
5473 if (mddev
->raid_disks
!= 2 ||
5474 mddev
->degraded
> 1)
5475 return ERR_PTR(-EINVAL
);
5477 /* Should check if there are write-behind devices? */
5479 chunksect
= 64*2; /* 64K by default */
5481 /* The array must be an exact multiple of chunksize */
5482 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
5485 if ((chunksect
<<9) < STRIPE_SIZE
)
5486 /* array size does not allow a suitable chunk size */
5487 return ERR_PTR(-EINVAL
);
5489 mddev
->new_level
= 5;
5490 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5491 mddev
->new_chunk_sectors
= chunksect
;
5493 return setup_conf(mddev
);
5496 static void *raid5_takeover_raid6(mddev_t
*mddev
)
5500 switch (mddev
->layout
) {
5501 case ALGORITHM_LEFT_ASYMMETRIC_6
:
5502 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
5504 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
5505 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
5507 case ALGORITHM_LEFT_SYMMETRIC_6
:
5508 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5510 case ALGORITHM_RIGHT_SYMMETRIC_6
:
5511 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
5513 case ALGORITHM_PARITY_0_6
:
5514 new_layout
= ALGORITHM_PARITY_0
;
5516 case ALGORITHM_PARITY_N
:
5517 new_layout
= ALGORITHM_PARITY_N
;
5520 return ERR_PTR(-EINVAL
);
5522 mddev
->new_level
= 5;
5523 mddev
->new_layout
= new_layout
;
5524 mddev
->delta_disks
= -1;
5525 mddev
->raid_disks
-= 1;
5526 return setup_conf(mddev
);
5530 static int raid5_check_reshape(mddev_t
*mddev
)
5532 /* For a 2-drive array, the layout and chunk size can be changed
5533 * immediately as not restriping is needed.
5534 * For larger arrays we record the new value - after validation
5535 * to be used by a reshape pass.
5537 raid5_conf_t
*conf
= mddev
->private;
5538 int new_chunk
= mddev
->new_chunk_sectors
;
5540 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
5542 if (new_chunk
> 0) {
5543 if (!is_power_of_2(new_chunk
))
5545 if (new_chunk
< (PAGE_SIZE
>>9))
5547 if (mddev
->array_sectors
& (new_chunk
-1))
5548 /* not factor of array size */
5552 /* They look valid */
5554 if (mddev
->raid_disks
== 2) {
5555 /* can make the change immediately */
5556 if (mddev
->new_layout
>= 0) {
5557 conf
->algorithm
= mddev
->new_layout
;
5558 mddev
->layout
= mddev
->new_layout
;
5560 if (new_chunk
> 0) {
5561 conf
->chunk_sectors
= new_chunk
;
5562 mddev
->chunk_sectors
= new_chunk
;
5564 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5565 md_wakeup_thread(mddev
->thread
);
5567 return check_reshape(mddev
);
5570 static int raid6_check_reshape(mddev_t
*mddev
)
5572 int new_chunk
= mddev
->new_chunk_sectors
;
5574 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
5576 if (new_chunk
> 0) {
5577 if (!is_power_of_2(new_chunk
))
5579 if (new_chunk
< (PAGE_SIZE
>> 9))
5581 if (mddev
->array_sectors
& (new_chunk
-1))
5582 /* not factor of array size */
5586 /* They look valid */
5587 return check_reshape(mddev
);
5590 static void *raid5_takeover(mddev_t
*mddev
)
5592 /* raid5 can take over:
5593 * raid0 - if there is only one strip zone - make it a raid4 layout
5594 * raid1 - if there are two drives. We need to know the chunk size
5595 * raid4 - trivial - just use a raid4 layout.
5596 * raid6 - Providing it is a *_6 layout
5598 if (mddev
->level
== 0)
5599 return raid45_takeover_raid0(mddev
, 5);
5600 if (mddev
->level
== 1)
5601 return raid5_takeover_raid1(mddev
);
5602 if (mddev
->level
== 4) {
5603 mddev
->new_layout
= ALGORITHM_PARITY_N
;
5604 mddev
->new_level
= 5;
5605 return setup_conf(mddev
);
5607 if (mddev
->level
== 6)
5608 return raid5_takeover_raid6(mddev
);
5610 return ERR_PTR(-EINVAL
);
5613 static void *raid4_takeover(mddev_t
*mddev
)
5615 /* raid4 can take over:
5616 * raid0 - if there is only one strip zone
5617 * raid5 - if layout is right
5619 if (mddev
->level
== 0)
5620 return raid45_takeover_raid0(mddev
, 4);
5621 if (mddev
->level
== 5 &&
5622 mddev
->layout
== ALGORITHM_PARITY_N
) {
5623 mddev
->new_layout
= 0;
5624 mddev
->new_level
= 4;
5625 return setup_conf(mddev
);
5627 return ERR_PTR(-EINVAL
);
5630 static struct mdk_personality raid5_personality
;
5632 static void *raid6_takeover(mddev_t
*mddev
)
5634 /* Currently can only take over a raid5. We map the
5635 * personality to an equivalent raid6 personality
5636 * with the Q block at the end.
5640 if (mddev
->pers
!= &raid5_personality
)
5641 return ERR_PTR(-EINVAL
);
5642 if (mddev
->degraded
> 1)
5643 return ERR_PTR(-EINVAL
);
5644 if (mddev
->raid_disks
> 253)
5645 return ERR_PTR(-EINVAL
);
5646 if (mddev
->raid_disks
< 3)
5647 return ERR_PTR(-EINVAL
);
5649 switch (mddev
->layout
) {
5650 case ALGORITHM_LEFT_ASYMMETRIC
:
5651 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
5653 case ALGORITHM_RIGHT_ASYMMETRIC
:
5654 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
5656 case ALGORITHM_LEFT_SYMMETRIC
:
5657 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
5659 case ALGORITHM_RIGHT_SYMMETRIC
:
5660 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
5662 case ALGORITHM_PARITY_0
:
5663 new_layout
= ALGORITHM_PARITY_0_6
;
5665 case ALGORITHM_PARITY_N
:
5666 new_layout
= ALGORITHM_PARITY_N
;
5669 return ERR_PTR(-EINVAL
);
5671 mddev
->new_level
= 6;
5672 mddev
->new_layout
= new_layout
;
5673 mddev
->delta_disks
= 1;
5674 mddev
->raid_disks
+= 1;
5675 return setup_conf(mddev
);
5679 static struct mdk_personality raid6_personality
=
5683 .owner
= THIS_MODULE
,
5684 .make_request
= make_request
,
5688 .error_handler
= error
,
5689 .hot_add_disk
= raid5_add_disk
,
5690 .hot_remove_disk
= raid5_remove_disk
,
5691 .spare_active
= raid5_spare_active
,
5692 .sync_request
= sync_request
,
5693 .resize
= raid5_resize
,
5695 .check_reshape
= raid6_check_reshape
,
5696 .start_reshape
= raid5_start_reshape
,
5697 .finish_reshape
= raid5_finish_reshape
,
5698 .quiesce
= raid5_quiesce
,
5699 .takeover
= raid6_takeover
,
5701 static struct mdk_personality raid5_personality
=
5705 .owner
= THIS_MODULE
,
5706 .make_request
= make_request
,
5710 .error_handler
= error
,
5711 .hot_add_disk
= raid5_add_disk
,
5712 .hot_remove_disk
= raid5_remove_disk
,
5713 .spare_active
= raid5_spare_active
,
5714 .sync_request
= sync_request
,
5715 .resize
= raid5_resize
,
5717 .check_reshape
= raid5_check_reshape
,
5718 .start_reshape
= raid5_start_reshape
,
5719 .finish_reshape
= raid5_finish_reshape
,
5720 .quiesce
= raid5_quiesce
,
5721 .takeover
= raid5_takeover
,
5724 static struct mdk_personality raid4_personality
=
5728 .owner
= THIS_MODULE
,
5729 .make_request
= make_request
,
5733 .error_handler
= error
,
5734 .hot_add_disk
= raid5_add_disk
,
5735 .hot_remove_disk
= raid5_remove_disk
,
5736 .spare_active
= raid5_spare_active
,
5737 .sync_request
= sync_request
,
5738 .resize
= raid5_resize
,
5740 .check_reshape
= raid5_check_reshape
,
5741 .start_reshape
= raid5_start_reshape
,
5742 .finish_reshape
= raid5_finish_reshape
,
5743 .quiesce
= raid5_quiesce
,
5744 .takeover
= raid4_takeover
,
5747 static int __init
raid5_init(void)
5749 register_md_personality(&raid6_personality
);
5750 register_md_personality(&raid5_personality
);
5751 register_md_personality(&raid4_personality
);
5755 static void raid5_exit(void)
5757 unregister_md_personality(&raid6_personality
);
5758 unregister_md_personality(&raid5_personality
);
5759 unregister_md_personality(&raid4_personality
);
5762 module_init(raid5_init
);
5763 module_exit(raid5_exit
);
5764 MODULE_LICENSE("GPL");
5765 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
5766 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5767 MODULE_ALIAS("md-raid5");
5768 MODULE_ALIAS("md-raid4");
5769 MODULE_ALIAS("md-level-5");
5770 MODULE_ALIAS("md-level-4");
5771 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5772 MODULE_ALIAS("md-raid6");
5773 MODULE_ALIAS("md-level-6");
5775 /* This used to be two separate modules, they were: */
5776 MODULE_ALIAS("raid5");
5777 MODULE_ALIAS("raid6");