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
= ~0;
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_block5 - 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_fill5 to continue
2317 static int fetch_block5(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
*failed_dev
= &sh
->dev
[s
->failed_num
[0]];
2323 /* is the data in this block needed, and can we get it? */
2324 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2325 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2327 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2328 s
->syncing
|| s
->expanding
||
2329 (s
->failed
&& failed_dev
->toread
) ||
2330 (s
->failed
&& failed_dev
->towrite
&&
2331 !test_bit(R5_OVERWRITE
, &failed_dev
->flags
)))))) {
2332 /* We would like to get this block, possibly by computing it,
2333 * otherwise read it if the backing disk is insync
2335 if ((s
->uptodate
== disks
- 1) &&
2336 (s
->failed
&& disk_idx
== s
->failed_num
[0])) {
2337 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2338 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2339 set_bit(R5_Wantcompute
, &dev
->flags
);
2340 sh
->ops
.target
= disk_idx
;
2341 sh
->ops
.target2
= -1;
2343 /* Careful: from this point on 'uptodate' is in the eye
2344 * of raid_run_ops which services 'compute' operations
2345 * before writes. R5_Wantcompute flags a block that will
2346 * be R5_UPTODATE by the time it is needed for a
2347 * subsequent operation.
2350 return 1; /* uptodate + compute == disks */
2351 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2352 set_bit(R5_LOCKED
, &dev
->flags
);
2353 set_bit(R5_Wantread
, &dev
->flags
);
2355 pr_debug("Reading block %d (sync=%d)\n", disk_idx
,
2364 * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2366 static void handle_stripe_fill5(struct stripe_head
*sh
,
2367 struct stripe_head_state
*s
, int disks
)
2371 /* look for blocks to read/compute, skip this if a compute
2372 * is already in flight, or if the stripe contents are in the
2373 * midst of changing due to a write
2375 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2376 !sh
->reconstruct_state
)
2377 for (i
= disks
; i
--; )
2378 if (fetch_block5(sh
, s
, i
, disks
))
2380 set_bit(STRIPE_HANDLE
, &sh
->state
);
2383 /* fetch_block6 - checks the given member device to see if its data needs
2384 * to be read or computed to satisfy a request.
2386 * Returns 1 when no more member devices need to be checked, otherwise returns
2387 * 0 to tell the loop in handle_stripe_fill6 to continue
2389 static int fetch_block6(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2390 int disk_idx
, int disks
)
2392 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
2393 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
2394 &sh
->dev
[s
->failed_num
[1]] };
2396 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2397 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2399 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2400 s
->syncing
|| s
->expanding
||
2401 (s
->failed
>= 1 && fdev
[0]->toread
) ||
2402 (s
->failed
>= 2 && fdev
[1]->toread
) ||
2403 (s
->failed
&& s
->to_write
)) {
2404 /* we would like to get this block, possibly by computing it,
2405 * otherwise read it if the backing disk is insync
2407 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
2408 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
2409 if ((s
->uptodate
== disks
- 1) &&
2410 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
2411 disk_idx
== s
->failed_num
[1]))) {
2412 /* have disk failed, and we're requested to fetch it;
2415 pr_debug("Computing stripe %llu block %d\n",
2416 (unsigned long long)sh
->sector
, disk_idx
);
2417 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2418 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2419 set_bit(R5_Wantcompute
, &dev
->flags
);
2420 sh
->ops
.target
= disk_idx
;
2421 sh
->ops
.target2
= -1; /* no 2nd target */
2425 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
2426 /* Computing 2-failure is *very* expensive; only
2427 * do it if failed >= 2
2430 for (other
= disks
; other
--; ) {
2431 if (other
== disk_idx
)
2433 if (!test_bit(R5_UPTODATE
,
2434 &sh
->dev
[other
].flags
))
2438 pr_debug("Computing stripe %llu blocks %d,%d\n",
2439 (unsigned long long)sh
->sector
,
2441 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2442 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2443 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
2444 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
2445 sh
->ops
.target
= disk_idx
;
2446 sh
->ops
.target2
= other
;
2450 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2451 set_bit(R5_LOCKED
, &dev
->flags
);
2452 set_bit(R5_Wantread
, &dev
->flags
);
2454 pr_debug("Reading block %d (sync=%d)\n",
2455 disk_idx
, s
->syncing
);
2463 * handle_stripe_fill6 - read or compute data to satisfy pending requests.
2465 static void handle_stripe_fill6(struct stripe_head
*sh
,
2466 struct stripe_head_state
*s
,
2471 /* look for blocks to read/compute, skip this if a compute
2472 * is already in flight, or if the stripe contents are in the
2473 * midst of changing due to a write
2475 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2476 !sh
->reconstruct_state
)
2477 for (i
= disks
; i
--; )
2478 if (fetch_block6(sh
, s
, i
, disks
))
2480 set_bit(STRIPE_HANDLE
, &sh
->state
);
2484 /* handle_stripe_clean_event
2485 * any written block on an uptodate or failed drive can be returned.
2486 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2487 * never LOCKED, so we don't need to test 'failed' directly.
2489 static void handle_stripe_clean_event(raid5_conf_t
*conf
,
2490 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
2495 for (i
= disks
; i
--; )
2496 if (sh
->dev
[i
].written
) {
2498 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2499 test_bit(R5_UPTODATE
, &dev
->flags
)) {
2500 /* We can return any write requests */
2501 struct bio
*wbi
, *wbi2
;
2503 pr_debug("Return write for disc %d\n", i
);
2504 spin_lock_irq(&conf
->device_lock
);
2506 dev
->written
= NULL
;
2507 while (wbi
&& wbi
->bi_sector
<
2508 dev
->sector
+ STRIPE_SECTORS
) {
2509 wbi2
= r5_next_bio(wbi
, dev
->sector
);
2510 if (!raid5_dec_bi_phys_segments(wbi
)) {
2511 md_write_end(conf
->mddev
);
2512 wbi
->bi_next
= *return_bi
;
2517 if (dev
->towrite
== NULL
)
2519 spin_unlock_irq(&conf
->device_lock
);
2521 bitmap_endwrite(conf
->mddev
->bitmap
,
2524 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
2529 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2530 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2531 md_wakeup_thread(conf
->mddev
->thread
);
2534 static void handle_stripe_dirtying5(raid5_conf_t
*conf
,
2535 struct stripe_head
*sh
, struct stripe_head_state
*s
, int disks
)
2537 int rmw
= 0, rcw
= 0, i
;
2538 for (i
= disks
; i
--; ) {
2539 /* would I have to read this buffer for read_modify_write */
2540 struct r5dev
*dev
= &sh
->dev
[i
];
2541 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2542 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2543 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2544 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2545 if (test_bit(R5_Insync
, &dev
->flags
))
2548 rmw
+= 2*disks
; /* cannot read it */
2550 /* Would I have to read this buffer for reconstruct_write */
2551 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) && i
!= sh
->pd_idx
&&
2552 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2553 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2554 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2555 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2560 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2561 (unsigned long long)sh
->sector
, rmw
, rcw
);
2562 set_bit(STRIPE_HANDLE
, &sh
->state
);
2563 if (rmw
< rcw
&& rmw
> 0)
2564 /* prefer read-modify-write, but need to get some data */
2565 for (i
= disks
; i
--; ) {
2566 struct r5dev
*dev
= &sh
->dev
[i
];
2567 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2568 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2569 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2570 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2571 test_bit(R5_Insync
, &dev
->flags
)) {
2573 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2574 pr_debug("Read_old block "
2575 "%d for r-m-w\n", i
);
2576 set_bit(R5_LOCKED
, &dev
->flags
);
2577 set_bit(R5_Wantread
, &dev
->flags
);
2580 set_bit(STRIPE_DELAYED
, &sh
->state
);
2581 set_bit(STRIPE_HANDLE
, &sh
->state
);
2585 if (rcw
<= rmw
&& rcw
> 0)
2586 /* want reconstruct write, but need to get some data */
2587 for (i
= disks
; i
--; ) {
2588 struct r5dev
*dev
= &sh
->dev
[i
];
2589 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2591 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2592 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2593 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2594 test_bit(R5_Insync
, &dev
->flags
)) {
2596 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2597 pr_debug("Read_old block "
2598 "%d for Reconstruct\n", i
);
2599 set_bit(R5_LOCKED
, &dev
->flags
);
2600 set_bit(R5_Wantread
, &dev
->flags
);
2603 set_bit(STRIPE_DELAYED
, &sh
->state
);
2604 set_bit(STRIPE_HANDLE
, &sh
->state
);
2608 /* now if nothing is locked, and if we have enough data,
2609 * we can start a write request
2611 /* since handle_stripe can be called at any time we need to handle the
2612 * case where a compute block operation has been submitted and then a
2613 * subsequent call wants to start a write request. raid_run_ops only
2614 * handles the case where compute block and reconstruct are requested
2615 * simultaneously. If this is not the case then new writes need to be
2616 * held off until the compute completes.
2618 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
2619 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
2620 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
2621 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
2624 static void handle_stripe_dirtying6(raid5_conf_t
*conf
,
2625 struct stripe_head
*sh
, struct stripe_head_state
*s
,
2628 int rcw
= 0, pd_idx
= sh
->pd_idx
, i
;
2629 int qd_idx
= sh
->qd_idx
;
2631 set_bit(STRIPE_HANDLE
, &sh
->state
);
2632 for (i
= disks
; i
--; ) {
2633 struct r5dev
*dev
= &sh
->dev
[i
];
2634 /* check if we haven't enough data */
2635 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2636 i
!= pd_idx
&& i
!= qd_idx
&&
2637 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2638 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2639 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2641 if (!test_bit(R5_Insync
, &dev
->flags
))
2642 continue; /* it's a failed drive */
2645 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2646 pr_debug("Read_old stripe %llu "
2647 "block %d for Reconstruct\n",
2648 (unsigned long long)sh
->sector
, i
);
2649 set_bit(R5_LOCKED
, &dev
->flags
);
2650 set_bit(R5_Wantread
, &dev
->flags
);
2653 pr_debug("Request delayed stripe %llu "
2654 "block %d for Reconstruct\n",
2655 (unsigned long long)sh
->sector
, i
);
2656 set_bit(STRIPE_DELAYED
, &sh
->state
);
2657 set_bit(STRIPE_HANDLE
, &sh
->state
);
2661 /* now if nothing is locked, and if we have enough data, we can start a
2664 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
2665 s
->locked
== 0 && rcw
== 0 &&
2666 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)) {
2667 schedule_reconstruction(sh
, s
, 1, 0);
2671 static void handle_parity_checks5(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2672 struct stripe_head_state
*s
, int disks
)
2674 struct r5dev
*dev
= NULL
;
2676 set_bit(STRIPE_HANDLE
, &sh
->state
);
2678 switch (sh
->check_state
) {
2679 case check_state_idle
:
2680 /* start a new check operation if there are no failures */
2681 if (s
->failed
== 0) {
2682 BUG_ON(s
->uptodate
!= disks
);
2683 sh
->check_state
= check_state_run
;
2684 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2685 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2689 dev
= &sh
->dev
[s
->failed_num
[0]];
2691 case check_state_compute_result
:
2692 sh
->check_state
= check_state_idle
;
2694 dev
= &sh
->dev
[sh
->pd_idx
];
2696 /* check that a write has not made the stripe insync */
2697 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2700 /* either failed parity check, or recovery is happening */
2701 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
2702 BUG_ON(s
->uptodate
!= disks
);
2704 set_bit(R5_LOCKED
, &dev
->flags
);
2706 set_bit(R5_Wantwrite
, &dev
->flags
);
2708 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2709 set_bit(STRIPE_INSYNC
, &sh
->state
);
2711 case check_state_run
:
2712 break; /* we will be called again upon completion */
2713 case check_state_check_result
:
2714 sh
->check_state
= check_state_idle
;
2716 /* if a failure occurred during the check operation, leave
2717 * STRIPE_INSYNC not set and let the stripe be handled again
2722 /* handle a successful check operation, if parity is correct
2723 * we are done. Otherwise update the mismatch count and repair
2724 * parity if !MD_RECOVERY_CHECK
2726 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
2727 /* parity is correct (on disc,
2728 * not in buffer any more)
2730 set_bit(STRIPE_INSYNC
, &sh
->state
);
2732 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2733 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2734 /* don't try to repair!! */
2735 set_bit(STRIPE_INSYNC
, &sh
->state
);
2737 sh
->check_state
= check_state_compute_run
;
2738 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2739 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2740 set_bit(R5_Wantcompute
,
2741 &sh
->dev
[sh
->pd_idx
].flags
);
2742 sh
->ops
.target
= sh
->pd_idx
;
2743 sh
->ops
.target2
= -1;
2748 case check_state_compute_run
:
2751 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
2752 __func__
, sh
->check_state
,
2753 (unsigned long long) sh
->sector
);
2759 static void handle_parity_checks6(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2760 struct stripe_head_state
*s
,
2763 int pd_idx
= sh
->pd_idx
;
2764 int qd_idx
= sh
->qd_idx
;
2767 set_bit(STRIPE_HANDLE
, &sh
->state
);
2769 BUG_ON(s
->failed
> 2);
2771 /* Want to check and possibly repair P and Q.
2772 * However there could be one 'failed' device, in which
2773 * case we can only check one of them, possibly using the
2774 * other to generate missing data
2777 switch (sh
->check_state
) {
2778 case check_state_idle
:
2779 /* start a new check operation if there are < 2 failures */
2780 if (s
->failed
== s
->q_failed
) {
2781 /* The only possible failed device holds Q, so it
2782 * makes sense to check P (If anything else were failed,
2783 * we would have used P to recreate it).
2785 sh
->check_state
= check_state_run
;
2787 if (!s
->q_failed
&& s
->failed
< 2) {
2788 /* Q is not failed, and we didn't use it to generate
2789 * anything, so it makes sense to check it
2791 if (sh
->check_state
== check_state_run
)
2792 sh
->check_state
= check_state_run_pq
;
2794 sh
->check_state
= check_state_run_q
;
2797 /* discard potentially stale zero_sum_result */
2798 sh
->ops
.zero_sum_result
= 0;
2800 if (sh
->check_state
== check_state_run
) {
2801 /* async_xor_zero_sum destroys the contents of P */
2802 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2805 if (sh
->check_state
>= check_state_run
&&
2806 sh
->check_state
<= check_state_run_pq
) {
2807 /* async_syndrome_zero_sum preserves P and Q, so
2808 * no need to mark them !uptodate here
2810 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2814 /* we have 2-disk failure */
2815 BUG_ON(s
->failed
!= 2);
2817 case check_state_compute_result
:
2818 sh
->check_state
= check_state_idle
;
2820 /* check that a write has not made the stripe insync */
2821 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2824 /* now write out any block on a failed drive,
2825 * or P or Q if they were recomputed
2827 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
2828 if (s
->failed
== 2) {
2829 dev
= &sh
->dev
[s
->failed_num
[1]];
2831 set_bit(R5_LOCKED
, &dev
->flags
);
2832 set_bit(R5_Wantwrite
, &dev
->flags
);
2834 if (s
->failed
>= 1) {
2835 dev
= &sh
->dev
[s
->failed_num
[0]];
2837 set_bit(R5_LOCKED
, &dev
->flags
);
2838 set_bit(R5_Wantwrite
, &dev
->flags
);
2840 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
2841 dev
= &sh
->dev
[pd_idx
];
2843 set_bit(R5_LOCKED
, &dev
->flags
);
2844 set_bit(R5_Wantwrite
, &dev
->flags
);
2846 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
2847 dev
= &sh
->dev
[qd_idx
];
2849 set_bit(R5_LOCKED
, &dev
->flags
);
2850 set_bit(R5_Wantwrite
, &dev
->flags
);
2852 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2854 set_bit(STRIPE_INSYNC
, &sh
->state
);
2856 case check_state_run
:
2857 case check_state_run_q
:
2858 case check_state_run_pq
:
2859 break; /* we will be called again upon completion */
2860 case check_state_check_result
:
2861 sh
->check_state
= check_state_idle
;
2863 /* handle a successful check operation, if parity is correct
2864 * we are done. Otherwise update the mismatch count and repair
2865 * parity if !MD_RECOVERY_CHECK
2867 if (sh
->ops
.zero_sum_result
== 0) {
2868 /* both parities are correct */
2870 set_bit(STRIPE_INSYNC
, &sh
->state
);
2872 /* in contrast to the raid5 case we can validate
2873 * parity, but still have a failure to write
2876 sh
->check_state
= check_state_compute_result
;
2877 /* Returning at this point means that we may go
2878 * off and bring p and/or q uptodate again so
2879 * we make sure to check zero_sum_result again
2880 * to verify if p or q need writeback
2884 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2885 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2886 /* don't try to repair!! */
2887 set_bit(STRIPE_INSYNC
, &sh
->state
);
2889 int *target
= &sh
->ops
.target
;
2891 sh
->ops
.target
= -1;
2892 sh
->ops
.target2
= -1;
2893 sh
->check_state
= check_state_compute_run
;
2894 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2895 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2896 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
2897 set_bit(R5_Wantcompute
,
2898 &sh
->dev
[pd_idx
].flags
);
2900 target
= &sh
->ops
.target2
;
2903 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
2904 set_bit(R5_Wantcompute
,
2905 &sh
->dev
[qd_idx
].flags
);
2912 case check_state_compute_run
:
2915 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
2916 __func__
, sh
->check_state
,
2917 (unsigned long long) sh
->sector
);
2922 static void handle_stripe_expansion(raid5_conf_t
*conf
, struct stripe_head
*sh
)
2926 /* We have read all the blocks in this stripe and now we need to
2927 * copy some of them into a target stripe for expand.
2929 struct dma_async_tx_descriptor
*tx
= NULL
;
2930 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
2931 for (i
= 0; i
< sh
->disks
; i
++)
2932 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
2934 struct stripe_head
*sh2
;
2935 struct async_submit_ctl submit
;
2937 sector_t bn
= compute_blocknr(sh
, i
, 1);
2938 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
2940 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
2942 /* so far only the early blocks of this stripe
2943 * have been requested. When later blocks
2944 * get requested, we will try again
2947 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
2948 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
2949 /* must have already done this block */
2950 release_stripe(sh2
);
2954 /* place all the copies on one channel */
2955 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
2956 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
2957 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
2960 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
2961 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
2962 for (j
= 0; j
< conf
->raid_disks
; j
++)
2963 if (j
!= sh2
->pd_idx
&&
2965 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
2967 if (j
== conf
->raid_disks
) {
2968 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
2969 set_bit(STRIPE_HANDLE
, &sh2
->state
);
2971 release_stripe(sh2
);
2974 /* done submitting copies, wait for them to complete */
2977 dma_wait_for_async_tx(tx
);
2983 * handle_stripe - do things to a stripe.
2985 * We lock the stripe and then examine the state of various bits
2986 * to see what needs to be done.
2988 * return some read request which now have data
2989 * return some write requests which are safely on disc
2990 * schedule a read on some buffers
2991 * schedule a write of some buffers
2992 * return confirmation of parity correctness
2994 * buffers are taken off read_list or write_list, and bh_cache buffers
2995 * get BH_Lock set before the stripe lock is released.
2999 static int handle_stripe5(struct stripe_head
*sh
, struct stripe_head_state
*s
)
3001 raid5_conf_t
*conf
= sh
->raid_conf
;
3002 int disks
= sh
->disks
, i
;
3006 /* Now to look around and see what can be done */
3008 spin_lock_irq(&conf
->device_lock
);
3009 for (i
=disks
; i
--; ) {
3014 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
3015 "written %p\n", i
, dev
->flags
, dev
->toread
, dev
->read
,
3016 dev
->towrite
, dev
->written
);
3018 /* maybe we can request a biofill operation
3020 * new wantfill requests are only permitted while
3021 * ops_complete_biofill is guaranteed to be inactive
3023 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
3024 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
3025 set_bit(R5_Wantfill
, &dev
->flags
);
3027 /* now count some things */
3028 if (test_bit(R5_LOCKED
, &dev
->flags
))
3030 if (test_bit(R5_UPTODATE
, &dev
->flags
))
3032 if (test_bit(R5_Wantcompute
, &dev
->flags
))
3035 if (test_bit(R5_Wantfill
, &dev
->flags
))
3037 else if (dev
->toread
)
3041 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3046 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3047 if (s
->blocked_rdev
== NULL
&&
3048 rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
3049 s
->blocked_rdev
= rdev
;
3050 atomic_inc(&rdev
->nr_pending
);
3052 clear_bit(R5_Insync
, &dev
->flags
);
3055 else if (test_bit(In_sync
, &rdev
->flags
))
3056 set_bit(R5_Insync
, &dev
->flags
);
3058 /* could be in-sync depending on recovery/reshape status */
3059 if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
3060 set_bit(R5_Insync
, &dev
->flags
);
3062 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3063 /* The ReadError flag will just be confusing now */
3064 clear_bit(R5_ReadError
, &dev
->flags
);
3065 clear_bit(R5_ReWrite
, &dev
->flags
);
3067 if (test_bit(R5_ReadError
, &dev
->flags
))
3068 clear_bit(R5_Insync
, &dev
->flags
);
3069 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3071 s
->failed_num
[0] = i
;
3074 spin_unlock_irq(&conf
->device_lock
);
3077 if (unlikely(s
->blocked_rdev
)) {
3078 if (s
->syncing
|| s
->expanding
|| s
->expanded
||
3079 s
->to_write
|| s
->written
) {
3080 set_bit(STRIPE_HANDLE
, &sh
->state
);
3083 /* There is nothing for the blocked_rdev to block */
3084 rdev_dec_pending(s
->blocked_rdev
, conf
->mddev
);
3085 s
->blocked_rdev
= NULL
;
3088 if (s
->to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
3089 set_bit(STRIPE_OP_BIOFILL
, &s
->ops_request
);
3090 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
3093 pr_debug("locked=%d uptodate=%d to_read=%d"
3094 " to_write=%d failed=%d failed_num=%d\n",
3095 s
->locked
, s
->uptodate
, s
->to_read
, s
->to_write
,
3096 s
->failed
, s
->failed_num
[0]);
3097 /* check if the array has lost two devices and, if so, some requests might
3100 if (s
->failed
> 1 && s
->to_read
+s
->to_write
+s
->written
)
3101 handle_failed_stripe(conf
, sh
, s
, disks
, &s
->return_bi
);
3102 if (s
->failed
> 1 && s
->syncing
) {
3103 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,0);
3104 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3108 /* might be able to return some write requests if the parity block
3109 * is safe, or on a failed drive
3111 dev
= &sh
->dev
[sh
->pd_idx
];
3113 ((test_bit(R5_Insync
, &dev
->flags
) &&
3114 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3115 test_bit(R5_UPTODATE
, &dev
->flags
)) ||
3116 (s
->failed
== 1 && s
->failed_num
[0] == sh
->pd_idx
)))
3117 handle_stripe_clean_event(conf
, sh
, disks
, &s
->return_bi
);
3119 /* Now we might consider reading some blocks, either to check/generate
3120 * parity, or to satisfy requests
3121 * or to load a block that is being partially written.
3123 if (s
->to_read
|| s
->non_overwrite
||
3124 (s
->syncing
&& (s
->uptodate
+ s
->compute
< disks
)) || s
->expanding
)
3125 handle_stripe_fill5(sh
, s
, disks
);
3127 /* Now we check to see if any write operations have recently
3131 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
3133 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
3134 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
3135 sh
->reconstruct_state
= reconstruct_state_idle
;
3137 /* All the 'written' buffers and the parity block are ready to
3138 * be written back to disk
3140 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
));
3141 for (i
= disks
; i
--; ) {
3143 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
3144 (i
== sh
->pd_idx
|| dev
->written
)) {
3145 pr_debug("Writing block %d\n", i
);
3146 set_bit(R5_Wantwrite
, &dev
->flags
);
3149 if (!test_bit(R5_Insync
, &dev
->flags
) ||
3150 (i
== sh
->pd_idx
&& s
->failed
== 0))
3151 set_bit(STRIPE_INSYNC
, &sh
->state
);
3154 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3155 s
->dec_preread_active
= 1;
3158 /* Now to consider new write requests and what else, if anything
3159 * should be read. We do not handle new writes when:
3160 * 1/ A 'write' operation (copy+xor) is already in flight.
3161 * 2/ A 'check' operation is in flight, as it may clobber the parity
3164 if (s
->to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
3165 handle_stripe_dirtying5(conf
, sh
, s
, disks
);
3167 /* maybe we need to check and possibly fix the parity for this stripe
3168 * Any reads will already have been scheduled, so we just see if enough
3169 * data is available. The parity check is held off while parity
3170 * dependent operations are in flight.
3172 if (sh
->check_state
||
3173 (s
->syncing
&& s
->locked
== 0 &&
3174 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
3175 !test_bit(STRIPE_INSYNC
, &sh
->state
)))
3176 handle_parity_checks5(conf
, sh
, s
, disks
);
3180 static int handle_stripe6(struct stripe_head
*sh
, struct stripe_head_state
*s
)
3182 raid5_conf_t
*conf
= sh
->raid_conf
;
3183 int disks
= sh
->disks
;
3184 int i
, pd_idx
= sh
->pd_idx
, qd_idx
= sh
->qd_idx
;
3185 struct r5dev
*dev
, *pdev
, *qdev
;
3187 /* Now to look around and see what can be done */
3190 spin_lock_irq(&conf
->device_lock
);
3191 for (i
=disks
; i
--; ) {
3195 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3196 i
, dev
->flags
, dev
->toread
, dev
->towrite
, dev
->written
);
3197 /* maybe we can reply to a read
3199 * new wantfill requests are only permitted while
3200 * ops_complete_biofill is guaranteed to be inactive
3202 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
3203 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
3204 set_bit(R5_Wantfill
, &dev
->flags
);
3206 /* now count some things */
3207 if (test_bit(R5_LOCKED
, &dev
->flags
))
3209 if (test_bit(R5_UPTODATE
, &dev
->flags
))
3211 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
3213 BUG_ON(s
->compute
> 2);
3216 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
3218 } else if (dev
->toread
)
3222 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3227 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3228 if (s
->blocked_rdev
== NULL
&&
3229 rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
3230 s
->blocked_rdev
= rdev
;
3231 atomic_inc(&rdev
->nr_pending
);
3233 clear_bit(R5_Insync
, &dev
->flags
);
3236 else if (test_bit(In_sync
, &rdev
->flags
))
3237 set_bit(R5_Insync
, &dev
->flags
);
3239 /* in sync if before recovery_offset */
3240 if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
3241 set_bit(R5_Insync
, &dev
->flags
);
3243 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3244 /* The ReadError flag will just be confusing now */
3245 clear_bit(R5_ReadError
, &dev
->flags
);
3246 clear_bit(R5_ReWrite
, &dev
->flags
);
3248 if (test_bit(R5_ReadError
, &dev
->flags
))
3249 clear_bit(R5_Insync
, &dev
->flags
);
3250 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3252 s
->failed_num
[s
->failed
] = i
;
3256 spin_unlock_irq(&conf
->device_lock
);
3259 if (unlikely(s
->blocked_rdev
)) {
3260 if (s
->syncing
|| s
->expanding
|| s
->expanded
||
3261 s
->to_write
|| s
->written
) {
3262 set_bit(STRIPE_HANDLE
, &sh
->state
);
3265 /* There is nothing for the blocked_rdev to block */
3266 rdev_dec_pending(s
->blocked_rdev
, conf
->mddev
);
3267 s
->blocked_rdev
= NULL
;
3270 if (s
->to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
3271 set_bit(STRIPE_OP_BIOFILL
, &s
->ops_request
);
3272 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
3275 pr_debug("locked=%d uptodate=%d to_read=%d"
3276 " to_write=%d failed=%d failed_num=%d,%d\n",
3277 s
->locked
, s
->uptodate
, s
->to_read
, s
->to_write
, s
->failed
,
3278 s
->failed_num
[0], s
->failed_num
[1]);
3279 /* check if the array has lost >2 devices and, if so, some requests
3280 * might need to be failed
3282 if (s
->failed
> 2 && s
->to_read
+s
->to_write
+s
->written
)
3283 handle_failed_stripe(conf
, sh
, s
, disks
, &s
->return_bi
);
3284 if (s
->failed
> 2 && s
->syncing
) {
3285 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,0);
3286 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3291 * might be able to return some write requests if the parity blocks
3292 * are safe, or on a failed drive
3294 pdev
= &sh
->dev
[pd_idx
];
3295 s
->p_failed
= (s
->failed
>= 1 && s
->failed_num
[0] == pd_idx
)
3296 || (s
->failed
>= 2 && s
->failed_num
[1] == pd_idx
);
3297 qdev
= &sh
->dev
[qd_idx
];
3298 s
->q_failed
= (s
->failed
>= 1 && s
->failed_num
[0] == qd_idx
)
3299 || (s
->failed
>= 2 && s
->failed_num
[1] == qd_idx
);
3302 (s
->p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
3303 && !test_bit(R5_LOCKED
, &pdev
->flags
)
3304 && test_bit(R5_UPTODATE
, &pdev
->flags
)))) &&
3305 (s
->q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
3306 && !test_bit(R5_LOCKED
, &qdev
->flags
)
3307 && test_bit(R5_UPTODATE
, &qdev
->flags
)))))
3308 handle_stripe_clean_event(conf
, sh
, disks
, &s
->return_bi
);
3310 /* Now we might consider reading some blocks, either to check/generate
3311 * parity, or to satisfy requests
3312 * or to load a block that is being partially written.
3314 if (s
->to_read
|| s
->non_overwrite
|| (s
->to_write
&& s
->failed
) ||
3315 (s
->syncing
&& (s
->uptodate
+ s
->compute
< disks
)) || s
->expanding
)
3316 handle_stripe_fill6(sh
, s
, disks
);
3318 /* Now we check to see if any write operations have recently
3321 if (sh
->reconstruct_state
== reconstruct_state_drain_result
) {
3323 sh
->reconstruct_state
= reconstruct_state_idle
;
3324 /* All the 'written' buffers and the parity blocks are ready to
3325 * be written back to disk
3327 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
));
3328 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
));
3329 for (i
= disks
; i
--; ) {
3331 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
3332 (i
== sh
->pd_idx
|| i
== qd_idx
||
3334 pr_debug("Writing block %d\n", i
);
3335 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
3336 set_bit(R5_Wantwrite
, &dev
->flags
);
3337 if (!test_bit(R5_Insync
, &dev
->flags
) ||
3338 ((i
== sh
->pd_idx
|| i
== qd_idx
) &&
3340 set_bit(STRIPE_INSYNC
, &sh
->state
);
3343 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3344 s
->dec_preread_active
= 1;
3347 /* Now to consider new write requests and what else, if anything
3348 * should be read. We do not handle new writes when:
3349 * 1/ A 'write' operation (copy+gen_syndrome) is already in flight.
3350 * 2/ A 'check' operation is in flight, as it may clobber the parity
3353 if (s
->to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
3354 handle_stripe_dirtying6(conf
, sh
, s
, disks
);
3356 /* maybe we need to check and possibly fix the parity for this stripe
3357 * Any reads will already have been scheduled, so we just see if enough
3358 * data is available. The parity check is held off while parity
3359 * dependent operations are in flight.
3361 if (sh
->check_state
||
3362 (s
->syncing
&& s
->locked
== 0 &&
3363 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
3364 !test_bit(STRIPE_INSYNC
, &sh
->state
)))
3365 handle_parity_checks6(conf
, sh
, s
, disks
);
3369 static void handle_stripe(struct stripe_head
*sh
)
3371 struct stripe_head_state s
;
3374 raid5_conf_t
*conf
= sh
->raid_conf
;
3376 clear_bit(STRIPE_HANDLE
, &sh
->state
);
3377 if (test_and_set_bit(STRIPE_ACTIVE
, &sh
->state
)) {
3378 /* already being handled, ensure it gets handled
3379 * again when current action finishes */
3380 set_bit(STRIPE_HANDLE
, &sh
->state
);
3384 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
3385 set_bit(STRIPE_SYNCING
, &sh
->state
);
3386 clear_bit(STRIPE_INSYNC
, &sh
->state
);
3388 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3390 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3391 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3392 (unsigned long long)sh
->sector
, sh
->state
,
3393 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
3394 sh
->check_state
, sh
->reconstruct_state
);
3395 memset(&s
, 0, sizeof(s
));
3397 s
.syncing
= test_bit(STRIPE_SYNCING
, &sh
->state
);
3398 s
.expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3399 s
.expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3401 if (conf
->level
== 6)
3402 done
= handle_stripe6(sh
, &s
);
3404 done
= handle_stripe5(sh
, &s
);
3410 if (s
.syncing
&& s
.locked
== 0 && test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3411 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3412 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3415 /* If the failed drives are just a ReadError, then we might need
3416 * to progress the repair/check process
3418 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
3419 for (i
= 0; i
< s
.failed
; i
++) {
3420 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
3421 if (test_bit(R5_ReadError
, &dev
->flags
)
3422 && !test_bit(R5_LOCKED
, &dev
->flags
)
3423 && test_bit(R5_UPTODATE
, &dev
->flags
)
3425 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3426 set_bit(R5_Wantwrite
, &dev
->flags
);
3427 set_bit(R5_ReWrite
, &dev
->flags
);
3428 set_bit(R5_LOCKED
, &dev
->flags
);
3431 /* let's read it back */
3432 set_bit(R5_Wantread
, &dev
->flags
);
3433 set_bit(R5_LOCKED
, &dev
->flags
);
3440 /* Finish reconstruct operations initiated by the expansion process */
3441 if (sh
->reconstruct_state
== reconstruct_state_result
) {
3442 struct stripe_head
*sh_src
3443 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
3444 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
3445 /* sh cannot be written until sh_src has been read.
3446 * so arrange for sh to be delayed a little
3448 set_bit(STRIPE_DELAYED
, &sh
->state
);
3449 set_bit(STRIPE_HANDLE
, &sh
->state
);
3450 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
3452 atomic_inc(&conf
->preread_active_stripes
);
3453 release_stripe(sh_src
);
3457 release_stripe(sh_src
);
3459 sh
->reconstruct_state
= reconstruct_state_idle
;
3460 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3461 for (i
= conf
->raid_disks
; i
--; ) {
3462 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3463 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3468 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
3469 !sh
->reconstruct_state
) {
3470 /* Need to write out all blocks after computing parity */
3471 sh
->disks
= conf
->raid_disks
;
3472 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
3473 schedule_reconstruction(sh
, &s
, 1, 1);
3474 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
3475 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3476 atomic_dec(&conf
->reshape_stripes
);
3477 wake_up(&conf
->wait_for_overlap
);
3478 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3481 if (s
.expanding
&& s
.locked
== 0 &&
3482 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
3483 handle_stripe_expansion(conf
, sh
);
3486 /* wait for this device to become unblocked */
3487 if (unlikely(s
.blocked_rdev
))
3488 md_wait_for_blocked_rdev(s
.blocked_rdev
, conf
->mddev
);
3491 raid_run_ops(sh
, s
.ops_request
);
3496 if (s
.dec_preread_active
) {
3497 /* We delay this until after ops_run_io so that if make_request
3498 * is waiting on a flush, it won't continue until the writes
3499 * have actually been submitted.
3501 atomic_dec(&conf
->preread_active_stripes
);
3502 if (atomic_read(&conf
->preread_active_stripes
) <
3504 md_wakeup_thread(conf
->mddev
->thread
);
3507 return_io(s
.return_bi
);
3509 clear_bit(STRIPE_ACTIVE
, &sh
->state
);
3512 static void raid5_activate_delayed(raid5_conf_t
*conf
)
3514 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
3515 while (!list_empty(&conf
->delayed_list
)) {
3516 struct list_head
*l
= conf
->delayed_list
.next
;
3517 struct stripe_head
*sh
;
3518 sh
= list_entry(l
, struct stripe_head
, lru
);
3520 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3521 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3522 atomic_inc(&conf
->preread_active_stripes
);
3523 list_add_tail(&sh
->lru
, &conf
->hold_list
);
3528 static void activate_bit_delay(raid5_conf_t
*conf
)
3530 /* device_lock is held */
3531 struct list_head head
;
3532 list_add(&head
, &conf
->bitmap_list
);
3533 list_del_init(&conf
->bitmap_list
);
3534 while (!list_empty(&head
)) {
3535 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
3536 list_del_init(&sh
->lru
);
3537 atomic_inc(&sh
->count
);
3538 __release_stripe(conf
, sh
);
3542 int md_raid5_congested(mddev_t
*mddev
, int bits
)
3544 raid5_conf_t
*conf
= mddev
->private;
3546 /* No difference between reads and writes. Just check
3547 * how busy the stripe_cache is
3550 if (conf
->inactive_blocked
)
3554 if (list_empty_careful(&conf
->inactive_list
))
3559 EXPORT_SYMBOL_GPL(md_raid5_congested
);
3561 static int raid5_congested(void *data
, int bits
)
3563 mddev_t
*mddev
= data
;
3565 return mddev_congested(mddev
, bits
) ||
3566 md_raid5_congested(mddev
, bits
);
3569 /* We want read requests to align with chunks where possible,
3570 * but write requests don't need to.
3572 static int raid5_mergeable_bvec(struct request_queue
*q
,
3573 struct bvec_merge_data
*bvm
,
3574 struct bio_vec
*biovec
)
3576 mddev_t
*mddev
= q
->queuedata
;
3577 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
3579 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3580 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
3582 if ((bvm
->bi_rw
& 1) == WRITE
)
3583 return biovec
->bv_len
; /* always allow writes to be mergeable */
3585 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3586 chunk_sectors
= mddev
->new_chunk_sectors
;
3587 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
3588 if (max
< 0) max
= 0;
3589 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
3590 return biovec
->bv_len
;
3596 static int in_chunk_boundary(mddev_t
*mddev
, struct bio
*bio
)
3598 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
3599 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3600 unsigned int bio_sectors
= bio
->bi_size
>> 9;
3602 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3603 chunk_sectors
= mddev
->new_chunk_sectors
;
3604 return chunk_sectors
>=
3605 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
3609 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3610 * later sampled by raid5d.
3612 static void add_bio_to_retry(struct bio
*bi
,raid5_conf_t
*conf
)
3614 unsigned long flags
;
3616 spin_lock_irqsave(&conf
->device_lock
, flags
);
3618 bi
->bi_next
= conf
->retry_read_aligned_list
;
3619 conf
->retry_read_aligned_list
= bi
;
3621 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3622 md_wakeup_thread(conf
->mddev
->thread
);
3626 static struct bio
*remove_bio_from_retry(raid5_conf_t
*conf
)
3630 bi
= conf
->retry_read_aligned
;
3632 conf
->retry_read_aligned
= NULL
;
3635 bi
= conf
->retry_read_aligned_list
;
3637 conf
->retry_read_aligned_list
= bi
->bi_next
;
3640 * this sets the active strip count to 1 and the processed
3641 * strip count to zero (upper 8 bits)
3643 bi
->bi_phys_segments
= 1; /* biased count of active stripes */
3651 * The "raid5_align_endio" should check if the read succeeded and if it
3652 * did, call bio_endio on the original bio (having bio_put the new bio
3654 * If the read failed..
3656 static void raid5_align_endio(struct bio
*bi
, int error
)
3658 struct bio
* raid_bi
= bi
->bi_private
;
3661 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3666 rdev
= (void*)raid_bi
->bi_next
;
3667 raid_bi
->bi_next
= NULL
;
3668 mddev
= rdev
->mddev
;
3669 conf
= mddev
->private;
3671 rdev_dec_pending(rdev
, conf
->mddev
);
3673 if (!error
&& uptodate
) {
3674 bio_endio(raid_bi
, 0);
3675 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
3676 wake_up(&conf
->wait_for_stripe
);
3681 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3683 add_bio_to_retry(raid_bi
, conf
);
3686 static int bio_fits_rdev(struct bio
*bi
)
3688 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
3690 if ((bi
->bi_size
>>9) > queue_max_sectors(q
))
3692 blk_recount_segments(q
, bi
);
3693 if (bi
->bi_phys_segments
> queue_max_segments(q
))
3696 if (q
->merge_bvec_fn
)
3697 /* it's too hard to apply the merge_bvec_fn at this stage,
3706 static int chunk_aligned_read(mddev_t
*mddev
, struct bio
* raid_bio
)
3708 raid5_conf_t
*conf
= mddev
->private;
3710 struct bio
* align_bi
;
3713 if (!in_chunk_boundary(mddev
, raid_bio
)) {
3714 pr_debug("chunk_aligned_read : non aligned\n");
3718 * use bio_clone_mddev to make a copy of the bio
3720 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
3724 * set bi_end_io to a new function, and set bi_private to the
3727 align_bi
->bi_end_io
= raid5_align_endio
;
3728 align_bi
->bi_private
= raid_bio
;
3732 align_bi
->bi_sector
= raid5_compute_sector(conf
, raid_bio
->bi_sector
,
3737 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
3738 if (rdev
&& test_bit(In_sync
, &rdev
->flags
)) {
3739 atomic_inc(&rdev
->nr_pending
);
3741 raid_bio
->bi_next
= (void*)rdev
;
3742 align_bi
->bi_bdev
= rdev
->bdev
;
3743 align_bi
->bi_flags
&= ~(1 << BIO_SEG_VALID
);
3744 align_bi
->bi_sector
+= rdev
->data_offset
;
3746 if (!bio_fits_rdev(align_bi
)) {
3747 /* too big in some way */
3749 rdev_dec_pending(rdev
, mddev
);
3753 spin_lock_irq(&conf
->device_lock
);
3754 wait_event_lock_irq(conf
->wait_for_stripe
,
3756 conf
->device_lock
, /* nothing */);
3757 atomic_inc(&conf
->active_aligned_reads
);
3758 spin_unlock_irq(&conf
->device_lock
);
3760 generic_make_request(align_bi
);
3769 /* __get_priority_stripe - get the next stripe to process
3771 * Full stripe writes are allowed to pass preread active stripes up until
3772 * the bypass_threshold is exceeded. In general the bypass_count
3773 * increments when the handle_list is handled before the hold_list; however, it
3774 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3775 * stripe with in flight i/o. The bypass_count will be reset when the
3776 * head of the hold_list has changed, i.e. the head was promoted to the
3779 static struct stripe_head
*__get_priority_stripe(raid5_conf_t
*conf
)
3781 struct stripe_head
*sh
;
3783 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3785 list_empty(&conf
->handle_list
) ? "empty" : "busy",
3786 list_empty(&conf
->hold_list
) ? "empty" : "busy",
3787 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
3789 if (!list_empty(&conf
->handle_list
)) {
3790 sh
= list_entry(conf
->handle_list
.next
, typeof(*sh
), lru
);
3792 if (list_empty(&conf
->hold_list
))
3793 conf
->bypass_count
= 0;
3794 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
3795 if (conf
->hold_list
.next
== conf
->last_hold
)
3796 conf
->bypass_count
++;
3798 conf
->last_hold
= conf
->hold_list
.next
;
3799 conf
->bypass_count
-= conf
->bypass_threshold
;
3800 if (conf
->bypass_count
< 0)
3801 conf
->bypass_count
= 0;
3804 } else if (!list_empty(&conf
->hold_list
) &&
3805 ((conf
->bypass_threshold
&&
3806 conf
->bypass_count
> conf
->bypass_threshold
) ||
3807 atomic_read(&conf
->pending_full_writes
) == 0)) {
3808 sh
= list_entry(conf
->hold_list
.next
,
3810 conf
->bypass_count
-= conf
->bypass_threshold
;
3811 if (conf
->bypass_count
< 0)
3812 conf
->bypass_count
= 0;
3816 list_del_init(&sh
->lru
);
3817 atomic_inc(&sh
->count
);
3818 BUG_ON(atomic_read(&sh
->count
) != 1);
3822 static int make_request(mddev_t
*mddev
, struct bio
* bi
)
3824 raid5_conf_t
*conf
= mddev
->private;
3826 sector_t new_sector
;
3827 sector_t logical_sector
, last_sector
;
3828 struct stripe_head
*sh
;
3829 const int rw
= bio_data_dir(bi
);
3833 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
3834 md_flush_request(mddev
, bi
);
3838 md_write_start(mddev
, bi
);
3841 mddev
->reshape_position
== MaxSector
&&
3842 chunk_aligned_read(mddev
,bi
))
3845 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
3846 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
3848 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
3850 plugged
= mddev_check_plugged(mddev
);
3851 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
3853 int disks
, data_disks
;
3858 disks
= conf
->raid_disks
;
3859 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
3860 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
3861 /* spinlock is needed as reshape_progress may be
3862 * 64bit on a 32bit platform, and so it might be
3863 * possible to see a half-updated value
3864 * Of course reshape_progress could change after
3865 * the lock is dropped, so once we get a reference
3866 * to the stripe that we think it is, we will have
3869 spin_lock_irq(&conf
->device_lock
);
3870 if (mddev
->delta_disks
< 0
3871 ? logical_sector
< conf
->reshape_progress
3872 : logical_sector
>= conf
->reshape_progress
) {
3873 disks
= conf
->previous_raid_disks
;
3876 if (mddev
->delta_disks
< 0
3877 ? logical_sector
< conf
->reshape_safe
3878 : logical_sector
>= conf
->reshape_safe
) {
3879 spin_unlock_irq(&conf
->device_lock
);
3884 spin_unlock_irq(&conf
->device_lock
);
3886 data_disks
= disks
- conf
->max_degraded
;
3888 new_sector
= raid5_compute_sector(conf
, logical_sector
,
3891 pr_debug("raid456: make_request, sector %llu logical %llu\n",
3892 (unsigned long long)new_sector
,
3893 (unsigned long long)logical_sector
);
3895 sh
= get_active_stripe(conf
, new_sector
, previous
,
3896 (bi
->bi_rw
&RWA_MASK
), 0);
3898 if (unlikely(previous
)) {
3899 /* expansion might have moved on while waiting for a
3900 * stripe, so we must do the range check again.
3901 * Expansion could still move past after this
3902 * test, but as we are holding a reference to
3903 * 'sh', we know that if that happens,
3904 * STRIPE_EXPANDING will get set and the expansion
3905 * won't proceed until we finish with the stripe.
3908 spin_lock_irq(&conf
->device_lock
);
3909 if (mddev
->delta_disks
< 0
3910 ? logical_sector
>= conf
->reshape_progress
3911 : logical_sector
< conf
->reshape_progress
)
3912 /* mismatch, need to try again */
3914 spin_unlock_irq(&conf
->device_lock
);
3923 logical_sector
>= mddev
->suspend_lo
&&
3924 logical_sector
< mddev
->suspend_hi
) {
3926 /* As the suspend_* range is controlled by
3927 * userspace, we want an interruptible
3930 flush_signals(current
);
3931 prepare_to_wait(&conf
->wait_for_overlap
,
3932 &w
, TASK_INTERRUPTIBLE
);
3933 if (logical_sector
>= mddev
->suspend_lo
&&
3934 logical_sector
< mddev
->suspend_hi
)
3939 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
3940 !add_stripe_bio(sh
, bi
, dd_idx
, rw
)) {
3941 /* Stripe is busy expanding or
3942 * add failed due to overlap. Flush everything
3945 md_wakeup_thread(mddev
->thread
);
3950 finish_wait(&conf
->wait_for_overlap
, &w
);
3951 set_bit(STRIPE_HANDLE
, &sh
->state
);
3952 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3953 if ((bi
->bi_rw
& REQ_SYNC
) &&
3954 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3955 atomic_inc(&conf
->preread_active_stripes
);
3958 /* cannot get stripe for read-ahead, just give-up */
3959 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3960 finish_wait(&conf
->wait_for_overlap
, &w
);
3966 md_wakeup_thread(mddev
->thread
);
3968 spin_lock_irq(&conf
->device_lock
);
3969 remaining
= raid5_dec_bi_phys_segments(bi
);
3970 spin_unlock_irq(&conf
->device_lock
);
3971 if (remaining
== 0) {
3974 md_write_end(mddev
);
3982 static sector_t
raid5_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
);
3984 static sector_t
reshape_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
)
3986 /* reshaping is quite different to recovery/resync so it is
3987 * handled quite separately ... here.
3989 * On each call to sync_request, we gather one chunk worth of
3990 * destination stripes and flag them as expanding.
3991 * Then we find all the source stripes and request reads.
3992 * As the reads complete, handle_stripe will copy the data
3993 * into the destination stripe and release that stripe.
3995 raid5_conf_t
*conf
= mddev
->private;
3996 struct stripe_head
*sh
;
3997 sector_t first_sector
, last_sector
;
3998 int raid_disks
= conf
->previous_raid_disks
;
3999 int data_disks
= raid_disks
- conf
->max_degraded
;
4000 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
4003 sector_t writepos
, readpos
, safepos
;
4004 sector_t stripe_addr
;
4005 int reshape_sectors
;
4006 struct list_head stripes
;
4008 if (sector_nr
== 0) {
4009 /* If restarting in the middle, skip the initial sectors */
4010 if (mddev
->delta_disks
< 0 &&
4011 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
4012 sector_nr
= raid5_size(mddev
, 0, 0)
4013 - conf
->reshape_progress
;
4014 } else if (mddev
->delta_disks
>= 0 &&
4015 conf
->reshape_progress
> 0)
4016 sector_nr
= conf
->reshape_progress
;
4017 sector_div(sector_nr
, new_data_disks
);
4019 mddev
->curr_resync_completed
= sector_nr
;
4020 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4026 /* We need to process a full chunk at a time.
4027 * If old and new chunk sizes differ, we need to process the
4030 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
)
4031 reshape_sectors
= mddev
->new_chunk_sectors
;
4033 reshape_sectors
= mddev
->chunk_sectors
;
4035 /* we update the metadata when there is more than 3Meg
4036 * in the block range (that is rather arbitrary, should
4037 * probably be time based) or when the data about to be
4038 * copied would over-write the source of the data at
4039 * the front of the range.
4040 * i.e. one new_stripe along from reshape_progress new_maps
4041 * to after where reshape_safe old_maps to
4043 writepos
= conf
->reshape_progress
;
4044 sector_div(writepos
, new_data_disks
);
4045 readpos
= conf
->reshape_progress
;
4046 sector_div(readpos
, data_disks
);
4047 safepos
= conf
->reshape_safe
;
4048 sector_div(safepos
, data_disks
);
4049 if (mddev
->delta_disks
< 0) {
4050 writepos
-= min_t(sector_t
, reshape_sectors
, writepos
);
4051 readpos
+= reshape_sectors
;
4052 safepos
+= reshape_sectors
;
4054 writepos
+= reshape_sectors
;
4055 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
4056 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
4059 /* 'writepos' is the most advanced device address we might write.
4060 * 'readpos' is the least advanced device address we might read.
4061 * 'safepos' is the least address recorded in the metadata as having
4063 * If 'readpos' is behind 'writepos', then there is no way that we can
4064 * ensure safety in the face of a crash - that must be done by userspace
4065 * making a backup of the data. So in that case there is no particular
4066 * rush to update metadata.
4067 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4068 * update the metadata to advance 'safepos' to match 'readpos' so that
4069 * we can be safe in the event of a crash.
4070 * So we insist on updating metadata if safepos is behind writepos and
4071 * readpos is beyond writepos.
4072 * In any case, update the metadata every 10 seconds.
4073 * Maybe that number should be configurable, but I'm not sure it is
4074 * worth it.... maybe it could be a multiple of safemode_delay???
4076 if ((mddev
->delta_disks
< 0
4077 ? (safepos
> writepos
&& readpos
< writepos
)
4078 : (safepos
< writepos
&& readpos
> writepos
)) ||
4079 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4080 /* Cannot proceed until we've updated the superblock... */
4081 wait_event(conf
->wait_for_overlap
,
4082 atomic_read(&conf
->reshape_stripes
)==0);
4083 mddev
->reshape_position
= conf
->reshape_progress
;
4084 mddev
->curr_resync_completed
= sector_nr
;
4085 conf
->reshape_checkpoint
= jiffies
;
4086 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4087 md_wakeup_thread(mddev
->thread
);
4088 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4089 kthread_should_stop());
4090 spin_lock_irq(&conf
->device_lock
);
4091 conf
->reshape_safe
= mddev
->reshape_position
;
4092 spin_unlock_irq(&conf
->device_lock
);
4093 wake_up(&conf
->wait_for_overlap
);
4094 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4097 if (mddev
->delta_disks
< 0) {
4098 BUG_ON(conf
->reshape_progress
== 0);
4099 stripe_addr
= writepos
;
4100 BUG_ON((mddev
->dev_sectors
&
4101 ~((sector_t
)reshape_sectors
- 1))
4102 - reshape_sectors
- stripe_addr
4105 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
4106 stripe_addr
= sector_nr
;
4108 INIT_LIST_HEAD(&stripes
);
4109 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
4111 int skipped_disk
= 0;
4112 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
4113 set_bit(STRIPE_EXPANDING
, &sh
->state
);
4114 atomic_inc(&conf
->reshape_stripes
);
4115 /* If any of this stripe is beyond the end of the old
4116 * array, then we need to zero those blocks
4118 for (j
=sh
->disks
; j
--;) {
4120 if (j
== sh
->pd_idx
)
4122 if (conf
->level
== 6 &&
4125 s
= compute_blocknr(sh
, j
, 0);
4126 if (s
< raid5_size(mddev
, 0, 0)) {
4130 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
4131 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
4132 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
4134 if (!skipped_disk
) {
4135 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4136 set_bit(STRIPE_HANDLE
, &sh
->state
);
4138 list_add(&sh
->lru
, &stripes
);
4140 spin_lock_irq(&conf
->device_lock
);
4141 if (mddev
->delta_disks
< 0)
4142 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
4144 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
4145 spin_unlock_irq(&conf
->device_lock
);
4146 /* Ok, those stripe are ready. We can start scheduling
4147 * reads on the source stripes.
4148 * The source stripes are determined by mapping the first and last
4149 * block on the destination stripes.
4152 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
4155 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
4156 * new_data_disks
- 1),
4158 if (last_sector
>= mddev
->dev_sectors
)
4159 last_sector
= mddev
->dev_sectors
- 1;
4160 while (first_sector
<= last_sector
) {
4161 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
4162 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
4163 set_bit(STRIPE_HANDLE
, &sh
->state
);
4165 first_sector
+= STRIPE_SECTORS
;
4167 /* Now that the sources are clearly marked, we can release
4168 * the destination stripes
4170 while (!list_empty(&stripes
)) {
4171 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
4172 list_del_init(&sh
->lru
);
4175 /* If this takes us to the resync_max point where we have to pause,
4176 * then we need to write out the superblock.
4178 sector_nr
+= reshape_sectors
;
4179 if ((sector_nr
- mddev
->curr_resync_completed
) * 2
4180 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
4181 /* Cannot proceed until we've updated the superblock... */
4182 wait_event(conf
->wait_for_overlap
,
4183 atomic_read(&conf
->reshape_stripes
) == 0);
4184 mddev
->reshape_position
= conf
->reshape_progress
;
4185 mddev
->curr_resync_completed
= sector_nr
;
4186 conf
->reshape_checkpoint
= jiffies
;
4187 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4188 md_wakeup_thread(mddev
->thread
);
4189 wait_event(mddev
->sb_wait
,
4190 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
4191 || kthread_should_stop());
4192 spin_lock_irq(&conf
->device_lock
);
4193 conf
->reshape_safe
= mddev
->reshape_position
;
4194 spin_unlock_irq(&conf
->device_lock
);
4195 wake_up(&conf
->wait_for_overlap
);
4196 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4198 return reshape_sectors
;
4201 /* FIXME go_faster isn't used */
4202 static inline sector_t
sync_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
4204 raid5_conf_t
*conf
= mddev
->private;
4205 struct stripe_head
*sh
;
4206 sector_t max_sector
= mddev
->dev_sectors
;
4207 sector_t sync_blocks
;
4208 int still_degraded
= 0;
4211 if (sector_nr
>= max_sector
) {
4212 /* just being told to finish up .. nothing much to do */
4214 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
4219 if (mddev
->curr_resync
< max_sector
) /* aborted */
4220 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
4222 else /* completed sync */
4224 bitmap_close_sync(mddev
->bitmap
);
4229 /* Allow raid5_quiesce to complete */
4230 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
4232 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
4233 return reshape_request(mddev
, sector_nr
, skipped
);
4235 /* No need to check resync_max as we never do more than one
4236 * stripe, and as resync_max will always be on a chunk boundary,
4237 * if the check in md_do_sync didn't fire, there is no chance
4238 * of overstepping resync_max here
4241 /* if there is too many failed drives and we are trying
4242 * to resync, then assert that we are finished, because there is
4243 * nothing we can do.
4245 if (mddev
->degraded
>= conf
->max_degraded
&&
4246 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
4247 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
4251 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
4252 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
4253 !conf
->fullsync
&& sync_blocks
>= STRIPE_SECTORS
) {
4254 /* we can skip this block, and probably more */
4255 sync_blocks
/= STRIPE_SECTORS
;
4257 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
4261 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
4263 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
4265 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
4266 /* make sure we don't swamp the stripe cache if someone else
4267 * is trying to get access
4269 schedule_timeout_uninterruptible(1);
4271 /* Need to check if array will still be degraded after recovery/resync
4272 * We don't need to check the 'failed' flag as when that gets set,
4275 for (i
= 0; i
< conf
->raid_disks
; i
++)
4276 if (conf
->disks
[i
].rdev
== NULL
)
4279 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
4281 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
4286 return STRIPE_SECTORS
;
4289 static int retry_aligned_read(raid5_conf_t
*conf
, struct bio
*raid_bio
)
4291 /* We may not be able to submit a whole bio at once as there
4292 * may not be enough stripe_heads available.
4293 * We cannot pre-allocate enough stripe_heads as we may need
4294 * more than exist in the cache (if we allow ever large chunks).
4295 * So we do one stripe head at a time and record in
4296 * ->bi_hw_segments how many have been done.
4298 * We *know* that this entire raid_bio is in one chunk, so
4299 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4301 struct stripe_head
*sh
;
4303 sector_t sector
, logical_sector
, last_sector
;
4308 logical_sector
= raid_bio
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4309 sector
= raid5_compute_sector(conf
, logical_sector
,
4311 last_sector
= raid_bio
->bi_sector
+ (raid_bio
->bi_size
>>9);
4313 for (; logical_sector
< last_sector
;
4314 logical_sector
+= STRIPE_SECTORS
,
4315 sector
+= STRIPE_SECTORS
,
4318 if (scnt
< raid5_bi_hw_segments(raid_bio
))
4319 /* already done this stripe */
4322 sh
= get_active_stripe(conf
, sector
, 0, 1, 0);
4325 /* failed to get a stripe - must wait */
4326 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4327 conf
->retry_read_aligned
= raid_bio
;
4331 set_bit(R5_ReadError
, &sh
->dev
[dd_idx
].flags
);
4332 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0)) {
4334 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4335 conf
->retry_read_aligned
= raid_bio
;
4343 spin_lock_irq(&conf
->device_lock
);
4344 remaining
= raid5_dec_bi_phys_segments(raid_bio
);
4345 spin_unlock_irq(&conf
->device_lock
);
4347 bio_endio(raid_bio
, 0);
4348 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4349 wake_up(&conf
->wait_for_stripe
);
4355 * This is our raid5 kernel thread.
4357 * We scan the hash table for stripes which can be handled now.
4358 * During the scan, completed stripes are saved for us by the interrupt
4359 * handler, so that they will not have to wait for our next wakeup.
4361 static void raid5d(mddev_t
*mddev
)
4363 struct stripe_head
*sh
;
4364 raid5_conf_t
*conf
= mddev
->private;
4366 struct blk_plug plug
;
4368 pr_debug("+++ raid5d active\n");
4370 md_check_recovery(mddev
);
4372 blk_start_plug(&plug
);
4374 spin_lock_irq(&conf
->device_lock
);
4378 if (atomic_read(&mddev
->plug_cnt
) == 0 &&
4379 !list_empty(&conf
->bitmap_list
)) {
4380 /* Now is a good time to flush some bitmap updates */
4382 spin_unlock_irq(&conf
->device_lock
);
4383 bitmap_unplug(mddev
->bitmap
);
4384 spin_lock_irq(&conf
->device_lock
);
4385 conf
->seq_write
= conf
->seq_flush
;
4386 activate_bit_delay(conf
);
4388 if (atomic_read(&mddev
->plug_cnt
) == 0)
4389 raid5_activate_delayed(conf
);
4391 while ((bio
= remove_bio_from_retry(conf
))) {
4393 spin_unlock_irq(&conf
->device_lock
);
4394 ok
= retry_aligned_read(conf
, bio
);
4395 spin_lock_irq(&conf
->device_lock
);
4401 sh
= __get_priority_stripe(conf
);
4405 spin_unlock_irq(&conf
->device_lock
);
4412 spin_lock_irq(&conf
->device_lock
);
4414 pr_debug("%d stripes handled\n", handled
);
4416 spin_unlock_irq(&conf
->device_lock
);
4418 async_tx_issue_pending_all();
4419 blk_finish_plug(&plug
);
4421 pr_debug("--- raid5d inactive\n");
4425 raid5_show_stripe_cache_size(mddev_t
*mddev
, char *page
)
4427 raid5_conf_t
*conf
= mddev
->private;
4429 return sprintf(page
, "%d\n", conf
->max_nr_stripes
);
4435 raid5_set_cache_size(mddev_t
*mddev
, int size
)
4437 raid5_conf_t
*conf
= mddev
->private;
4440 if (size
<= 16 || size
> 32768)
4442 while (size
< conf
->max_nr_stripes
) {
4443 if (drop_one_stripe(conf
))
4444 conf
->max_nr_stripes
--;
4448 err
= md_allow_write(mddev
);
4451 while (size
> conf
->max_nr_stripes
) {
4452 if (grow_one_stripe(conf
))
4453 conf
->max_nr_stripes
++;
4458 EXPORT_SYMBOL(raid5_set_cache_size
);
4461 raid5_store_stripe_cache_size(mddev_t
*mddev
, const char *page
, size_t len
)
4463 raid5_conf_t
*conf
= mddev
->private;
4467 if (len
>= PAGE_SIZE
)
4472 if (strict_strtoul(page
, 10, &new))
4474 err
= raid5_set_cache_size(mddev
, new);
4480 static struct md_sysfs_entry
4481 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
4482 raid5_show_stripe_cache_size
,
4483 raid5_store_stripe_cache_size
);
4486 raid5_show_preread_threshold(mddev_t
*mddev
, char *page
)
4488 raid5_conf_t
*conf
= mddev
->private;
4490 return sprintf(page
, "%d\n", conf
->bypass_threshold
);
4496 raid5_store_preread_threshold(mddev_t
*mddev
, const char *page
, size_t len
)
4498 raid5_conf_t
*conf
= mddev
->private;
4500 if (len
>= PAGE_SIZE
)
4505 if (strict_strtoul(page
, 10, &new))
4507 if (new > conf
->max_nr_stripes
)
4509 conf
->bypass_threshold
= new;
4513 static struct md_sysfs_entry
4514 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
4516 raid5_show_preread_threshold
,
4517 raid5_store_preread_threshold
);
4520 stripe_cache_active_show(mddev_t
*mddev
, char *page
)
4522 raid5_conf_t
*conf
= mddev
->private;
4524 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
4529 static struct md_sysfs_entry
4530 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
4532 static struct attribute
*raid5_attrs
[] = {
4533 &raid5_stripecache_size
.attr
,
4534 &raid5_stripecache_active
.attr
,
4535 &raid5_preread_bypass_threshold
.attr
,
4538 static struct attribute_group raid5_attrs_group
= {
4540 .attrs
= raid5_attrs
,
4544 raid5_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
)
4546 raid5_conf_t
*conf
= mddev
->private;
4549 sectors
= mddev
->dev_sectors
;
4551 /* size is defined by the smallest of previous and new size */
4552 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
4554 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
4555 sectors
&= ~((sector_t
)mddev
->new_chunk_sectors
- 1);
4556 return sectors
* (raid_disks
- conf
->max_degraded
);
4559 static void raid5_free_percpu(raid5_conf_t
*conf
)
4561 struct raid5_percpu
*percpu
;
4568 for_each_possible_cpu(cpu
) {
4569 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4570 safe_put_page(percpu
->spare_page
);
4571 kfree(percpu
->scribble
);
4573 #ifdef CONFIG_HOTPLUG_CPU
4574 unregister_cpu_notifier(&conf
->cpu_notify
);
4578 free_percpu(conf
->percpu
);
4581 static void free_conf(raid5_conf_t
*conf
)
4583 shrink_stripes(conf
);
4584 raid5_free_percpu(conf
);
4586 kfree(conf
->stripe_hashtbl
);
4590 #ifdef CONFIG_HOTPLUG_CPU
4591 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
4594 raid5_conf_t
*conf
= container_of(nfb
, raid5_conf_t
, cpu_notify
);
4595 long cpu
= (long)hcpu
;
4596 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4599 case CPU_UP_PREPARE
:
4600 case CPU_UP_PREPARE_FROZEN
:
4601 if (conf
->level
== 6 && !percpu
->spare_page
)
4602 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
4603 if (!percpu
->scribble
)
4604 percpu
->scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
4606 if (!percpu
->scribble
||
4607 (conf
->level
== 6 && !percpu
->spare_page
)) {
4608 safe_put_page(percpu
->spare_page
);
4609 kfree(percpu
->scribble
);
4610 pr_err("%s: failed memory allocation for cpu%ld\n",
4612 return notifier_from_errno(-ENOMEM
);
4616 case CPU_DEAD_FROZEN
:
4617 safe_put_page(percpu
->spare_page
);
4618 kfree(percpu
->scribble
);
4619 percpu
->spare_page
= NULL
;
4620 percpu
->scribble
= NULL
;
4629 static int raid5_alloc_percpu(raid5_conf_t
*conf
)
4632 struct page
*spare_page
;
4633 struct raid5_percpu __percpu
*allcpus
;
4637 allcpus
= alloc_percpu(struct raid5_percpu
);
4640 conf
->percpu
= allcpus
;
4644 for_each_present_cpu(cpu
) {
4645 if (conf
->level
== 6) {
4646 spare_page
= alloc_page(GFP_KERNEL
);
4651 per_cpu_ptr(conf
->percpu
, cpu
)->spare_page
= spare_page
;
4653 scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
4658 per_cpu_ptr(conf
->percpu
, cpu
)->scribble
= scribble
;
4660 #ifdef CONFIG_HOTPLUG_CPU
4661 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
4662 conf
->cpu_notify
.priority
= 0;
4664 err
= register_cpu_notifier(&conf
->cpu_notify
);
4671 static raid5_conf_t
*setup_conf(mddev_t
*mddev
)
4674 int raid_disk
, memory
, max_disks
;
4676 struct disk_info
*disk
;
4678 if (mddev
->new_level
!= 5
4679 && mddev
->new_level
!= 4
4680 && mddev
->new_level
!= 6) {
4681 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4682 mdname(mddev
), mddev
->new_level
);
4683 return ERR_PTR(-EIO
);
4685 if ((mddev
->new_level
== 5
4686 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
4687 (mddev
->new_level
== 6
4688 && !algorithm_valid_raid6(mddev
->new_layout
))) {
4689 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
4690 mdname(mddev
), mddev
->new_layout
);
4691 return ERR_PTR(-EIO
);
4693 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
4694 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4695 mdname(mddev
), mddev
->raid_disks
);
4696 return ERR_PTR(-EINVAL
);
4699 if (!mddev
->new_chunk_sectors
||
4700 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
4701 !is_power_of_2(mddev
->new_chunk_sectors
)) {
4702 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
4703 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
4704 return ERR_PTR(-EINVAL
);
4707 conf
= kzalloc(sizeof(raid5_conf_t
), GFP_KERNEL
);
4710 spin_lock_init(&conf
->device_lock
);
4711 init_waitqueue_head(&conf
->wait_for_stripe
);
4712 init_waitqueue_head(&conf
->wait_for_overlap
);
4713 INIT_LIST_HEAD(&conf
->handle_list
);
4714 INIT_LIST_HEAD(&conf
->hold_list
);
4715 INIT_LIST_HEAD(&conf
->delayed_list
);
4716 INIT_LIST_HEAD(&conf
->bitmap_list
);
4717 INIT_LIST_HEAD(&conf
->inactive_list
);
4718 atomic_set(&conf
->active_stripes
, 0);
4719 atomic_set(&conf
->preread_active_stripes
, 0);
4720 atomic_set(&conf
->active_aligned_reads
, 0);
4721 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
4723 conf
->raid_disks
= mddev
->raid_disks
;
4724 if (mddev
->reshape_position
== MaxSector
)
4725 conf
->previous_raid_disks
= mddev
->raid_disks
;
4727 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4728 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
4729 conf
->scribble_len
= scribble_len(max_disks
);
4731 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
4736 conf
->mddev
= mddev
;
4738 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
4741 conf
->level
= mddev
->new_level
;
4742 if (raid5_alloc_percpu(conf
) != 0)
4745 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
4747 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
4748 raid_disk
= rdev
->raid_disk
;
4749 if (raid_disk
>= max_disks
4752 disk
= conf
->disks
+ raid_disk
;
4756 if (test_bit(In_sync
, &rdev
->flags
)) {
4757 char b
[BDEVNAME_SIZE
];
4758 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
4760 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
4761 } else if (rdev
->saved_raid_disk
!= raid_disk
)
4762 /* Cannot rely on bitmap to complete recovery */
4766 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
4767 conf
->level
= mddev
->new_level
;
4768 if (conf
->level
== 6)
4769 conf
->max_degraded
= 2;
4771 conf
->max_degraded
= 1;
4772 conf
->algorithm
= mddev
->new_layout
;
4773 conf
->max_nr_stripes
= NR_STRIPES
;
4774 conf
->reshape_progress
= mddev
->reshape_position
;
4775 if (conf
->reshape_progress
!= MaxSector
) {
4776 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
4777 conf
->prev_algo
= mddev
->layout
;
4780 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
4781 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
4782 if (grow_stripes(conf
, conf
->max_nr_stripes
)) {
4784 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4785 mdname(mddev
), memory
);
4788 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
4789 mdname(mddev
), memory
);
4791 conf
->thread
= md_register_thread(raid5d
, mddev
, NULL
);
4792 if (!conf
->thread
) {
4794 "md/raid:%s: couldn't allocate thread.\n",
4804 return ERR_PTR(-EIO
);
4806 return ERR_PTR(-ENOMEM
);
4810 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
4813 case ALGORITHM_PARITY_0
:
4814 if (raid_disk
< max_degraded
)
4817 case ALGORITHM_PARITY_N
:
4818 if (raid_disk
>= raid_disks
- max_degraded
)
4821 case ALGORITHM_PARITY_0_6
:
4822 if (raid_disk
== 0 ||
4823 raid_disk
== raid_disks
- 1)
4826 case ALGORITHM_LEFT_ASYMMETRIC_6
:
4827 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
4828 case ALGORITHM_LEFT_SYMMETRIC_6
:
4829 case ALGORITHM_RIGHT_SYMMETRIC_6
:
4830 if (raid_disk
== raid_disks
- 1)
4836 static int run(mddev_t
*mddev
)
4839 int working_disks
= 0;
4840 int dirty_parity_disks
= 0;
4842 sector_t reshape_offset
= 0;
4844 if (mddev
->recovery_cp
!= MaxSector
)
4845 printk(KERN_NOTICE
"md/raid:%s: not clean"
4846 " -- starting background reconstruction\n",
4848 if (mddev
->reshape_position
!= MaxSector
) {
4849 /* Check that we can continue the reshape.
4850 * Currently only disks can change, it must
4851 * increase, and we must be past the point where
4852 * a stripe over-writes itself
4854 sector_t here_new
, here_old
;
4856 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
4858 if (mddev
->new_level
!= mddev
->level
) {
4859 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
4860 "required - aborting.\n",
4864 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4865 /* reshape_position must be on a new-stripe boundary, and one
4866 * further up in new geometry must map after here in old
4869 here_new
= mddev
->reshape_position
;
4870 if (sector_div(here_new
, mddev
->new_chunk_sectors
*
4871 (mddev
->raid_disks
- max_degraded
))) {
4872 printk(KERN_ERR
"md/raid:%s: reshape_position not "
4873 "on a stripe boundary\n", mdname(mddev
));
4876 reshape_offset
= here_new
* mddev
->new_chunk_sectors
;
4877 /* here_new is the stripe we will write to */
4878 here_old
= mddev
->reshape_position
;
4879 sector_div(here_old
, mddev
->chunk_sectors
*
4880 (old_disks
-max_degraded
));
4881 /* here_old is the first stripe that we might need to read
4883 if (mddev
->delta_disks
== 0) {
4884 /* We cannot be sure it is safe to start an in-place
4885 * reshape. It is only safe if user-space if monitoring
4886 * and taking constant backups.
4887 * mdadm always starts a situation like this in
4888 * readonly mode so it can take control before
4889 * allowing any writes. So just check for that.
4891 if ((here_new
* mddev
->new_chunk_sectors
!=
4892 here_old
* mddev
->chunk_sectors
) ||
4894 printk(KERN_ERR
"md/raid:%s: in-place reshape must be started"
4895 " in read-only mode - aborting\n",
4899 } else if (mddev
->delta_disks
< 0
4900 ? (here_new
* mddev
->new_chunk_sectors
<=
4901 here_old
* mddev
->chunk_sectors
)
4902 : (here_new
* mddev
->new_chunk_sectors
>=
4903 here_old
* mddev
->chunk_sectors
)) {
4904 /* Reading from the same stripe as writing to - bad */
4905 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
4906 "auto-recovery - aborting.\n",
4910 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
4912 /* OK, we should be able to continue; */
4914 BUG_ON(mddev
->level
!= mddev
->new_level
);
4915 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
4916 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
4917 BUG_ON(mddev
->delta_disks
!= 0);
4920 if (mddev
->private == NULL
)
4921 conf
= setup_conf(mddev
);
4923 conf
= mddev
->private;
4926 return PTR_ERR(conf
);
4928 mddev
->thread
= conf
->thread
;
4929 conf
->thread
= NULL
;
4930 mddev
->private = conf
;
4933 * 0 for a fully functional array, 1 or 2 for a degraded array.
4935 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
4936 if (rdev
->raid_disk
< 0)
4938 if (test_bit(In_sync
, &rdev
->flags
)) {
4942 /* This disc is not fully in-sync. However if it
4943 * just stored parity (beyond the recovery_offset),
4944 * when we don't need to be concerned about the
4945 * array being dirty.
4946 * When reshape goes 'backwards', we never have
4947 * partially completed devices, so we only need
4948 * to worry about reshape going forwards.
4950 /* Hack because v0.91 doesn't store recovery_offset properly. */
4951 if (mddev
->major_version
== 0 &&
4952 mddev
->minor_version
> 90)
4953 rdev
->recovery_offset
= reshape_offset
;
4955 if (rdev
->recovery_offset
< reshape_offset
) {
4956 /* We need to check old and new layout */
4957 if (!only_parity(rdev
->raid_disk
,
4960 conf
->max_degraded
))
4963 if (!only_parity(rdev
->raid_disk
,
4965 conf
->previous_raid_disks
,
4966 conf
->max_degraded
))
4968 dirty_parity_disks
++;
4971 mddev
->degraded
= (max(conf
->raid_disks
, conf
->previous_raid_disks
)
4974 if (has_failed(conf
)) {
4975 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
4976 " (%d/%d failed)\n",
4977 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
4981 /* device size must be a multiple of chunk size */
4982 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
4983 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
4985 if (mddev
->degraded
> dirty_parity_disks
&&
4986 mddev
->recovery_cp
!= MaxSector
) {
4987 if (mddev
->ok_start_degraded
)
4989 "md/raid:%s: starting dirty degraded array"
4990 " - data corruption possible.\n",
4994 "md/raid:%s: cannot start dirty degraded array.\n",
5000 if (mddev
->degraded
== 0)
5001 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
5002 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
5003 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
5006 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
5007 " out of %d devices, algorithm %d\n",
5008 mdname(mddev
), conf
->level
,
5009 mddev
->raid_disks
- mddev
->degraded
,
5010 mddev
->raid_disks
, mddev
->new_layout
);
5012 print_raid5_conf(conf
);
5014 if (conf
->reshape_progress
!= MaxSector
) {
5015 conf
->reshape_safe
= conf
->reshape_progress
;
5016 atomic_set(&conf
->reshape_stripes
, 0);
5017 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5018 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5019 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5020 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5021 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5026 /* Ok, everything is just fine now */
5027 if (mddev
->to_remove
== &raid5_attrs_group
)
5028 mddev
->to_remove
= NULL
;
5029 else if (mddev
->kobj
.sd
&&
5030 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
5032 "raid5: failed to create sysfs attributes for %s\n",
5034 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5038 /* read-ahead size must cover two whole stripes, which
5039 * is 2 * (datadisks) * chunksize where 'n' is the
5040 * number of raid devices
5042 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
5043 int stripe
= data_disks
*
5044 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
5045 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
5046 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
5048 blk_queue_merge_bvec(mddev
->queue
, raid5_mergeable_bvec
);
5050 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
5051 mddev
->queue
->backing_dev_info
.congested_fn
= raid5_congested
;
5053 chunk_size
= mddev
->chunk_sectors
<< 9;
5054 blk_queue_io_min(mddev
->queue
, chunk_size
);
5055 blk_queue_io_opt(mddev
->queue
, chunk_size
*
5056 (conf
->raid_disks
- conf
->max_degraded
));
5058 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
5059 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
5060 rdev
->data_offset
<< 9);
5065 md_unregister_thread(mddev
->thread
);
5066 mddev
->thread
= NULL
;
5068 print_raid5_conf(conf
);
5071 mddev
->private = NULL
;
5072 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
5076 static int stop(mddev_t
*mddev
)
5078 raid5_conf_t
*conf
= mddev
->private;
5080 md_unregister_thread(mddev
->thread
);
5081 mddev
->thread
= NULL
;
5083 mddev
->queue
->backing_dev_info
.congested_fn
= NULL
;
5085 mddev
->private = NULL
;
5086 mddev
->to_remove
= &raid5_attrs_group
;
5091 static void print_sh(struct seq_file
*seq
, struct stripe_head
*sh
)
5095 seq_printf(seq
, "sh %llu, pd_idx %d, state %ld.\n",
5096 (unsigned long long)sh
->sector
, sh
->pd_idx
, sh
->state
);
5097 seq_printf(seq
, "sh %llu, count %d.\n",
5098 (unsigned long long)sh
->sector
, atomic_read(&sh
->count
));
5099 seq_printf(seq
, "sh %llu, ", (unsigned long long)sh
->sector
);
5100 for (i
= 0; i
< sh
->disks
; i
++) {
5101 seq_printf(seq
, "(cache%d: %p %ld) ",
5102 i
, sh
->dev
[i
].page
, sh
->dev
[i
].flags
);
5104 seq_printf(seq
, "\n");
5107 static void printall(struct seq_file
*seq
, raid5_conf_t
*conf
)
5109 struct stripe_head
*sh
;
5110 struct hlist_node
*hn
;
5113 spin_lock_irq(&conf
->device_lock
);
5114 for (i
= 0; i
< NR_HASH
; i
++) {
5115 hlist_for_each_entry(sh
, hn
, &conf
->stripe_hashtbl
[i
], hash
) {
5116 if (sh
->raid_conf
!= conf
)
5121 spin_unlock_irq(&conf
->device_lock
);
5125 static void status(struct seq_file
*seq
, mddev_t
*mddev
)
5127 raid5_conf_t
*conf
= mddev
->private;
5130 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
5131 mddev
->chunk_sectors
/ 2, mddev
->layout
);
5132 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
5133 for (i
= 0; i
< conf
->raid_disks
; i
++)
5134 seq_printf (seq
, "%s",
5135 conf
->disks
[i
].rdev
&&
5136 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
5137 seq_printf (seq
, "]");
5139 seq_printf (seq
, "\n");
5140 printall(seq
, conf
);
5144 static void print_raid5_conf (raid5_conf_t
*conf
)
5147 struct disk_info
*tmp
;
5149 printk(KERN_DEBUG
"RAID conf printout:\n");
5151 printk("(conf==NULL)\n");
5154 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
5156 conf
->raid_disks
- conf
->mddev
->degraded
);
5158 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5159 char b
[BDEVNAME_SIZE
];
5160 tmp
= conf
->disks
+ i
;
5162 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
5163 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
5164 bdevname(tmp
->rdev
->bdev
, b
));
5168 static int raid5_spare_active(mddev_t
*mddev
)
5171 raid5_conf_t
*conf
= mddev
->private;
5172 struct disk_info
*tmp
;
5174 unsigned long flags
;
5176 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5177 tmp
= conf
->disks
+ i
;
5179 && tmp
->rdev
->recovery_offset
== MaxSector
5180 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
5181 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
5183 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
5186 spin_lock_irqsave(&conf
->device_lock
, flags
);
5187 mddev
->degraded
-= count
;
5188 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5189 print_raid5_conf(conf
);
5193 static int raid5_remove_disk(mddev_t
*mddev
, int number
)
5195 raid5_conf_t
*conf
= mddev
->private;
5198 struct disk_info
*p
= conf
->disks
+ number
;
5200 print_raid5_conf(conf
);
5203 if (number
>= conf
->raid_disks
&&
5204 conf
->reshape_progress
== MaxSector
)
5205 clear_bit(In_sync
, &rdev
->flags
);
5207 if (test_bit(In_sync
, &rdev
->flags
) ||
5208 atomic_read(&rdev
->nr_pending
)) {
5212 /* Only remove non-faulty devices if recovery
5215 if (!test_bit(Faulty
, &rdev
->flags
) &&
5216 !has_failed(conf
) &&
5217 number
< conf
->raid_disks
) {
5223 if (atomic_read(&rdev
->nr_pending
)) {
5224 /* lost the race, try later */
5231 print_raid5_conf(conf
);
5235 static int raid5_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
5237 raid5_conf_t
*conf
= mddev
->private;
5240 struct disk_info
*p
;
5242 int last
= conf
->raid_disks
- 1;
5244 if (has_failed(conf
))
5245 /* no point adding a device */
5248 if (rdev
->raid_disk
>= 0)
5249 first
= last
= rdev
->raid_disk
;
5252 * find the disk ... but prefer rdev->saved_raid_disk
5255 if (rdev
->saved_raid_disk
>= 0 &&
5256 rdev
->saved_raid_disk
>= first
&&
5257 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
5258 disk
= rdev
->saved_raid_disk
;
5261 for ( ; disk
<= last
; disk
++)
5262 if ((p
=conf
->disks
+ disk
)->rdev
== NULL
) {
5263 clear_bit(In_sync
, &rdev
->flags
);
5264 rdev
->raid_disk
= disk
;
5266 if (rdev
->saved_raid_disk
!= disk
)
5268 rcu_assign_pointer(p
->rdev
, rdev
);
5271 print_raid5_conf(conf
);
5275 static int raid5_resize(mddev_t
*mddev
, sector_t sectors
)
5277 /* no resync is happening, and there is enough space
5278 * on all devices, so we can resize.
5279 * We need to make sure resync covers any new space.
5280 * If the array is shrinking we should possibly wait until
5281 * any io in the removed space completes, but it hardly seems
5284 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
5285 md_set_array_sectors(mddev
, raid5_size(mddev
, sectors
,
5286 mddev
->raid_disks
));
5287 if (mddev
->array_sectors
>
5288 raid5_size(mddev
, sectors
, mddev
->raid_disks
))
5290 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5291 revalidate_disk(mddev
->gendisk
);
5292 if (sectors
> mddev
->dev_sectors
&&
5293 mddev
->recovery_cp
> mddev
->dev_sectors
) {
5294 mddev
->recovery_cp
= mddev
->dev_sectors
;
5295 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
5297 mddev
->dev_sectors
= sectors
;
5298 mddev
->resync_max_sectors
= sectors
;
5302 static int check_stripe_cache(mddev_t
*mddev
)
5304 /* Can only proceed if there are plenty of stripe_heads.
5305 * We need a minimum of one full stripe,, and for sensible progress
5306 * it is best to have about 4 times that.
5307 * If we require 4 times, then the default 256 4K stripe_heads will
5308 * allow for chunk sizes up to 256K, which is probably OK.
5309 * If the chunk size is greater, user-space should request more
5310 * stripe_heads first.
5312 raid5_conf_t
*conf
= mddev
->private;
5313 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5314 > conf
->max_nr_stripes
||
5315 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5316 > conf
->max_nr_stripes
) {
5317 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5319 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
5326 static int check_reshape(mddev_t
*mddev
)
5328 raid5_conf_t
*conf
= mddev
->private;
5330 if (mddev
->delta_disks
== 0 &&
5331 mddev
->new_layout
== mddev
->layout
&&
5332 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
5333 return 0; /* nothing to do */
5335 /* Cannot grow a bitmap yet */
5337 if (has_failed(conf
))
5339 if (mddev
->delta_disks
< 0) {
5340 /* We might be able to shrink, but the devices must
5341 * be made bigger first.
5342 * For raid6, 4 is the minimum size.
5343 * Otherwise 2 is the minimum
5346 if (mddev
->level
== 6)
5348 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
5352 if (!check_stripe_cache(mddev
))
5355 return resize_stripes(conf
, conf
->raid_disks
+ mddev
->delta_disks
);
5358 static int raid5_start_reshape(mddev_t
*mddev
)
5360 raid5_conf_t
*conf
= mddev
->private;
5363 unsigned long flags
;
5365 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
5368 if (!check_stripe_cache(mddev
))
5371 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
5372 if (!test_bit(In_sync
, &rdev
->flags
)
5373 && !test_bit(Faulty
, &rdev
->flags
))
5376 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
5377 /* Not enough devices even to make a degraded array
5382 /* Refuse to reduce size of the array. Any reductions in
5383 * array size must be through explicit setting of array_size
5386 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
5387 < mddev
->array_sectors
) {
5388 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
5389 "before number of disks\n", mdname(mddev
));
5393 atomic_set(&conf
->reshape_stripes
, 0);
5394 spin_lock_irq(&conf
->device_lock
);
5395 conf
->previous_raid_disks
= conf
->raid_disks
;
5396 conf
->raid_disks
+= mddev
->delta_disks
;
5397 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
5398 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
5399 conf
->prev_algo
= conf
->algorithm
;
5400 conf
->algorithm
= mddev
->new_layout
;
5401 if (mddev
->delta_disks
< 0)
5402 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
5404 conf
->reshape_progress
= 0;
5405 conf
->reshape_safe
= conf
->reshape_progress
;
5407 spin_unlock_irq(&conf
->device_lock
);
5409 /* Add some new drives, as many as will fit.
5410 * We know there are enough to make the newly sized array work.
5411 * Don't add devices if we are reducing the number of
5412 * devices in the array. This is because it is not possible
5413 * to correctly record the "partially reconstructed" state of
5414 * such devices during the reshape and confusion could result.
5416 if (mddev
->delta_disks
>= 0) {
5417 int added_devices
= 0;
5418 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
5419 if (rdev
->raid_disk
< 0 &&
5420 !test_bit(Faulty
, &rdev
->flags
)) {
5421 if (raid5_add_disk(mddev
, rdev
) == 0) {
5424 >= conf
->previous_raid_disks
) {
5425 set_bit(In_sync
, &rdev
->flags
);
5428 rdev
->recovery_offset
= 0;
5429 sprintf(nm
, "rd%d", rdev
->raid_disk
);
5430 if (sysfs_create_link(&mddev
->kobj
,
5432 /* Failure here is OK */;
5434 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
5435 && !test_bit(Faulty
, &rdev
->flags
)) {
5436 /* This is a spare that was manually added */
5437 set_bit(In_sync
, &rdev
->flags
);
5441 /* When a reshape changes the number of devices,
5442 * ->degraded is measured against the larger of the
5443 * pre and post number of devices.
5445 spin_lock_irqsave(&conf
->device_lock
, flags
);
5446 mddev
->degraded
+= (conf
->raid_disks
- conf
->previous_raid_disks
)
5448 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5450 mddev
->raid_disks
= conf
->raid_disks
;
5451 mddev
->reshape_position
= conf
->reshape_progress
;
5452 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5454 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5455 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5456 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5457 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5458 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5460 if (!mddev
->sync_thread
) {
5461 mddev
->recovery
= 0;
5462 spin_lock_irq(&conf
->device_lock
);
5463 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
5464 conf
->reshape_progress
= MaxSector
;
5465 spin_unlock_irq(&conf
->device_lock
);
5468 conf
->reshape_checkpoint
= jiffies
;
5469 md_wakeup_thread(mddev
->sync_thread
);
5470 md_new_event(mddev
);
5474 /* This is called from the reshape thread and should make any
5475 * changes needed in 'conf'
5477 static void end_reshape(raid5_conf_t
*conf
)
5480 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
5482 spin_lock_irq(&conf
->device_lock
);
5483 conf
->previous_raid_disks
= conf
->raid_disks
;
5484 conf
->reshape_progress
= MaxSector
;
5485 spin_unlock_irq(&conf
->device_lock
);
5486 wake_up(&conf
->wait_for_overlap
);
5488 /* read-ahead size must cover two whole stripes, which is
5489 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5491 if (conf
->mddev
->queue
) {
5492 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5493 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
5495 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
5496 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
5501 /* This is called from the raid5d thread with mddev_lock held.
5502 * It makes config changes to the device.
5504 static void raid5_finish_reshape(mddev_t
*mddev
)
5506 raid5_conf_t
*conf
= mddev
->private;
5508 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
5510 if (mddev
->delta_disks
> 0) {
5511 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5512 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5513 revalidate_disk(mddev
->gendisk
);
5516 mddev
->degraded
= conf
->raid_disks
;
5517 for (d
= 0; d
< conf
->raid_disks
; d
++)
5518 if (conf
->disks
[d
].rdev
&&
5520 &conf
->disks
[d
].rdev
->flags
))
5522 for (d
= conf
->raid_disks
;
5523 d
< conf
->raid_disks
- mddev
->delta_disks
;
5525 mdk_rdev_t
*rdev
= conf
->disks
[d
].rdev
;
5526 if (rdev
&& raid5_remove_disk(mddev
, d
) == 0) {
5528 sprintf(nm
, "rd%d", rdev
->raid_disk
);
5529 sysfs_remove_link(&mddev
->kobj
, nm
);
5530 rdev
->raid_disk
= -1;
5534 mddev
->layout
= conf
->algorithm
;
5535 mddev
->chunk_sectors
= conf
->chunk_sectors
;
5536 mddev
->reshape_position
= MaxSector
;
5537 mddev
->delta_disks
= 0;
5541 static void raid5_quiesce(mddev_t
*mddev
, int state
)
5543 raid5_conf_t
*conf
= mddev
->private;
5546 case 2: /* resume for a suspend */
5547 wake_up(&conf
->wait_for_overlap
);
5550 case 1: /* stop all writes */
5551 spin_lock_irq(&conf
->device_lock
);
5552 /* '2' tells resync/reshape to pause so that all
5553 * active stripes can drain
5556 wait_event_lock_irq(conf
->wait_for_stripe
,
5557 atomic_read(&conf
->active_stripes
) == 0 &&
5558 atomic_read(&conf
->active_aligned_reads
) == 0,
5559 conf
->device_lock
, /* nothing */);
5561 spin_unlock_irq(&conf
->device_lock
);
5562 /* allow reshape to continue */
5563 wake_up(&conf
->wait_for_overlap
);
5566 case 0: /* re-enable writes */
5567 spin_lock_irq(&conf
->device_lock
);
5569 wake_up(&conf
->wait_for_stripe
);
5570 wake_up(&conf
->wait_for_overlap
);
5571 spin_unlock_irq(&conf
->device_lock
);
5577 static void *raid45_takeover_raid0(mddev_t
*mddev
, int level
)
5579 struct raid0_private_data
*raid0_priv
= mddev
->private;
5582 /* for raid0 takeover only one zone is supported */
5583 if (raid0_priv
->nr_strip_zones
> 1) {
5584 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5586 return ERR_PTR(-EINVAL
);
5589 sectors
= raid0_priv
->strip_zone
[0].zone_end
;
5590 sector_div(sectors
, raid0_priv
->strip_zone
[0].nb_dev
);
5591 mddev
->dev_sectors
= sectors
;
5592 mddev
->new_level
= level
;
5593 mddev
->new_layout
= ALGORITHM_PARITY_N
;
5594 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
5595 mddev
->raid_disks
+= 1;
5596 mddev
->delta_disks
= 1;
5597 /* make sure it will be not marked as dirty */
5598 mddev
->recovery_cp
= MaxSector
;
5600 return setup_conf(mddev
);
5604 static void *raid5_takeover_raid1(mddev_t
*mddev
)
5608 if (mddev
->raid_disks
!= 2 ||
5609 mddev
->degraded
> 1)
5610 return ERR_PTR(-EINVAL
);
5612 /* Should check if there are write-behind devices? */
5614 chunksect
= 64*2; /* 64K by default */
5616 /* The array must be an exact multiple of chunksize */
5617 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
5620 if ((chunksect
<<9) < STRIPE_SIZE
)
5621 /* array size does not allow a suitable chunk size */
5622 return ERR_PTR(-EINVAL
);
5624 mddev
->new_level
= 5;
5625 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5626 mddev
->new_chunk_sectors
= chunksect
;
5628 return setup_conf(mddev
);
5631 static void *raid5_takeover_raid6(mddev_t
*mddev
)
5635 switch (mddev
->layout
) {
5636 case ALGORITHM_LEFT_ASYMMETRIC_6
:
5637 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
5639 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
5640 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
5642 case ALGORITHM_LEFT_SYMMETRIC_6
:
5643 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5645 case ALGORITHM_RIGHT_SYMMETRIC_6
:
5646 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
5648 case ALGORITHM_PARITY_0_6
:
5649 new_layout
= ALGORITHM_PARITY_0
;
5651 case ALGORITHM_PARITY_N
:
5652 new_layout
= ALGORITHM_PARITY_N
;
5655 return ERR_PTR(-EINVAL
);
5657 mddev
->new_level
= 5;
5658 mddev
->new_layout
= new_layout
;
5659 mddev
->delta_disks
= -1;
5660 mddev
->raid_disks
-= 1;
5661 return setup_conf(mddev
);
5665 static int raid5_check_reshape(mddev_t
*mddev
)
5667 /* For a 2-drive array, the layout and chunk size can be changed
5668 * immediately as not restriping is needed.
5669 * For larger arrays we record the new value - after validation
5670 * to be used by a reshape pass.
5672 raid5_conf_t
*conf
= mddev
->private;
5673 int new_chunk
= mddev
->new_chunk_sectors
;
5675 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
5677 if (new_chunk
> 0) {
5678 if (!is_power_of_2(new_chunk
))
5680 if (new_chunk
< (PAGE_SIZE
>>9))
5682 if (mddev
->array_sectors
& (new_chunk
-1))
5683 /* not factor of array size */
5687 /* They look valid */
5689 if (mddev
->raid_disks
== 2) {
5690 /* can make the change immediately */
5691 if (mddev
->new_layout
>= 0) {
5692 conf
->algorithm
= mddev
->new_layout
;
5693 mddev
->layout
= mddev
->new_layout
;
5695 if (new_chunk
> 0) {
5696 conf
->chunk_sectors
= new_chunk
;
5697 mddev
->chunk_sectors
= new_chunk
;
5699 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5700 md_wakeup_thread(mddev
->thread
);
5702 return check_reshape(mddev
);
5705 static int raid6_check_reshape(mddev_t
*mddev
)
5707 int new_chunk
= mddev
->new_chunk_sectors
;
5709 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
5711 if (new_chunk
> 0) {
5712 if (!is_power_of_2(new_chunk
))
5714 if (new_chunk
< (PAGE_SIZE
>> 9))
5716 if (mddev
->array_sectors
& (new_chunk
-1))
5717 /* not factor of array size */
5721 /* They look valid */
5722 return check_reshape(mddev
);
5725 static void *raid5_takeover(mddev_t
*mddev
)
5727 /* raid5 can take over:
5728 * raid0 - if there is only one strip zone - make it a raid4 layout
5729 * raid1 - if there are two drives. We need to know the chunk size
5730 * raid4 - trivial - just use a raid4 layout.
5731 * raid6 - Providing it is a *_6 layout
5733 if (mddev
->level
== 0)
5734 return raid45_takeover_raid0(mddev
, 5);
5735 if (mddev
->level
== 1)
5736 return raid5_takeover_raid1(mddev
);
5737 if (mddev
->level
== 4) {
5738 mddev
->new_layout
= ALGORITHM_PARITY_N
;
5739 mddev
->new_level
= 5;
5740 return setup_conf(mddev
);
5742 if (mddev
->level
== 6)
5743 return raid5_takeover_raid6(mddev
);
5745 return ERR_PTR(-EINVAL
);
5748 static void *raid4_takeover(mddev_t
*mddev
)
5750 /* raid4 can take over:
5751 * raid0 - if there is only one strip zone
5752 * raid5 - if layout is right
5754 if (mddev
->level
== 0)
5755 return raid45_takeover_raid0(mddev
, 4);
5756 if (mddev
->level
== 5 &&
5757 mddev
->layout
== ALGORITHM_PARITY_N
) {
5758 mddev
->new_layout
= 0;
5759 mddev
->new_level
= 4;
5760 return setup_conf(mddev
);
5762 return ERR_PTR(-EINVAL
);
5765 static struct mdk_personality raid5_personality
;
5767 static void *raid6_takeover(mddev_t
*mddev
)
5769 /* Currently can only take over a raid5. We map the
5770 * personality to an equivalent raid6 personality
5771 * with the Q block at the end.
5775 if (mddev
->pers
!= &raid5_personality
)
5776 return ERR_PTR(-EINVAL
);
5777 if (mddev
->degraded
> 1)
5778 return ERR_PTR(-EINVAL
);
5779 if (mddev
->raid_disks
> 253)
5780 return ERR_PTR(-EINVAL
);
5781 if (mddev
->raid_disks
< 3)
5782 return ERR_PTR(-EINVAL
);
5784 switch (mddev
->layout
) {
5785 case ALGORITHM_LEFT_ASYMMETRIC
:
5786 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
5788 case ALGORITHM_RIGHT_ASYMMETRIC
:
5789 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
5791 case ALGORITHM_LEFT_SYMMETRIC
:
5792 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
5794 case ALGORITHM_RIGHT_SYMMETRIC
:
5795 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
5797 case ALGORITHM_PARITY_0
:
5798 new_layout
= ALGORITHM_PARITY_0_6
;
5800 case ALGORITHM_PARITY_N
:
5801 new_layout
= ALGORITHM_PARITY_N
;
5804 return ERR_PTR(-EINVAL
);
5806 mddev
->new_level
= 6;
5807 mddev
->new_layout
= new_layout
;
5808 mddev
->delta_disks
= 1;
5809 mddev
->raid_disks
+= 1;
5810 return setup_conf(mddev
);
5814 static struct mdk_personality raid6_personality
=
5818 .owner
= THIS_MODULE
,
5819 .make_request
= make_request
,
5823 .error_handler
= error
,
5824 .hot_add_disk
= raid5_add_disk
,
5825 .hot_remove_disk
= raid5_remove_disk
,
5826 .spare_active
= raid5_spare_active
,
5827 .sync_request
= sync_request
,
5828 .resize
= raid5_resize
,
5830 .check_reshape
= raid6_check_reshape
,
5831 .start_reshape
= raid5_start_reshape
,
5832 .finish_reshape
= raid5_finish_reshape
,
5833 .quiesce
= raid5_quiesce
,
5834 .takeover
= raid6_takeover
,
5836 static struct mdk_personality raid5_personality
=
5840 .owner
= THIS_MODULE
,
5841 .make_request
= make_request
,
5845 .error_handler
= error
,
5846 .hot_add_disk
= raid5_add_disk
,
5847 .hot_remove_disk
= raid5_remove_disk
,
5848 .spare_active
= raid5_spare_active
,
5849 .sync_request
= sync_request
,
5850 .resize
= raid5_resize
,
5852 .check_reshape
= raid5_check_reshape
,
5853 .start_reshape
= raid5_start_reshape
,
5854 .finish_reshape
= raid5_finish_reshape
,
5855 .quiesce
= raid5_quiesce
,
5856 .takeover
= raid5_takeover
,
5859 static struct mdk_personality raid4_personality
=
5863 .owner
= THIS_MODULE
,
5864 .make_request
= make_request
,
5868 .error_handler
= error
,
5869 .hot_add_disk
= raid5_add_disk
,
5870 .hot_remove_disk
= raid5_remove_disk
,
5871 .spare_active
= raid5_spare_active
,
5872 .sync_request
= sync_request
,
5873 .resize
= raid5_resize
,
5875 .check_reshape
= raid5_check_reshape
,
5876 .start_reshape
= raid5_start_reshape
,
5877 .finish_reshape
= raid5_finish_reshape
,
5878 .quiesce
= raid5_quiesce
,
5879 .takeover
= raid4_takeover
,
5882 static int __init
raid5_init(void)
5884 register_md_personality(&raid6_personality
);
5885 register_md_personality(&raid5_personality
);
5886 register_md_personality(&raid4_personality
);
5890 static void raid5_exit(void)
5892 unregister_md_personality(&raid6_personality
);
5893 unregister_md_personality(&raid5_personality
);
5894 unregister_md_personality(&raid4_personality
);
5897 module_init(raid5_init
);
5898 module_exit(raid5_exit
);
5899 MODULE_LICENSE("GPL");
5900 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
5901 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5902 MODULE_ALIAS("md-raid5");
5903 MODULE_ALIAS("md-raid4");
5904 MODULE_ALIAS("md-level-5");
5905 MODULE_ALIAS("md-level-4");
5906 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5907 MODULE_ALIAS("md-raid6");
5908 MODULE_ALIAS("md-level-6");
5910 /* This used to be two separate modules, they were: */
5911 MODULE_ALIAS("raid5");
5912 MODULE_ALIAS("raid6");