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->bm_write is the number of the last batch successfully written.
31 * conf->bm_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 bm_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 plugger_set_plug(&conf
->plug
);
205 } else if (test_bit(STRIPE_BIT_DELAY
, &sh
->state
) &&
206 sh
->bm_seq
- conf
->seq_write
> 0) {
207 list_add_tail(&sh
->lru
, &conf
->bitmap_list
);
208 plugger_set_plug(&conf
->plug
);
210 clear_bit(STRIPE_BIT_DELAY
, &sh
->state
);
211 list_add_tail(&sh
->lru
, &conf
->handle_list
);
213 md_wakeup_thread(conf
->mddev
->thread
);
215 BUG_ON(stripe_operations_active(sh
));
216 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
217 atomic_dec(&conf
->preread_active_stripes
);
218 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
)
219 md_wakeup_thread(conf
->mddev
->thread
);
221 atomic_dec(&conf
->active_stripes
);
222 if (!test_bit(STRIPE_EXPANDING
, &sh
->state
)) {
223 list_add_tail(&sh
->lru
, &conf
->inactive_list
);
224 wake_up(&conf
->wait_for_stripe
);
225 if (conf
->retry_read_aligned
)
226 md_wakeup_thread(conf
->mddev
->thread
);
232 static void release_stripe(struct stripe_head
*sh
)
234 raid5_conf_t
*conf
= sh
->raid_conf
;
237 spin_lock_irqsave(&conf
->device_lock
, flags
);
238 __release_stripe(conf
, sh
);
239 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
242 static inline void remove_hash(struct stripe_head
*sh
)
244 pr_debug("remove_hash(), stripe %llu\n",
245 (unsigned long long)sh
->sector
);
247 hlist_del_init(&sh
->hash
);
250 static inline void insert_hash(raid5_conf_t
*conf
, struct stripe_head
*sh
)
252 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
254 pr_debug("insert_hash(), stripe %llu\n",
255 (unsigned long long)sh
->sector
);
258 hlist_add_head(&sh
->hash
, hp
);
262 /* find an idle stripe, make sure it is unhashed, and return it. */
263 static struct stripe_head
*get_free_stripe(raid5_conf_t
*conf
)
265 struct stripe_head
*sh
= NULL
;
266 struct list_head
*first
;
269 if (list_empty(&conf
->inactive_list
))
271 first
= conf
->inactive_list
.next
;
272 sh
= list_entry(first
, struct stripe_head
, lru
);
273 list_del_init(first
);
275 atomic_inc(&conf
->active_stripes
);
280 static void shrink_buffers(struct stripe_head
*sh
)
284 int num
= sh
->raid_conf
->pool_size
;
286 for (i
= 0; i
< num
; i
++) {
290 sh
->dev
[i
].page
= NULL
;
295 static int grow_buffers(struct stripe_head
*sh
)
298 int num
= sh
->raid_conf
->pool_size
;
300 for (i
= 0; i
< num
; i
++) {
303 if (!(page
= alloc_page(GFP_KERNEL
))) {
306 sh
->dev
[i
].page
= page
;
311 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
);
312 static void stripe_set_idx(sector_t stripe
, raid5_conf_t
*conf
, int previous
,
313 struct stripe_head
*sh
);
315 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int previous
)
317 raid5_conf_t
*conf
= sh
->raid_conf
;
320 BUG_ON(atomic_read(&sh
->count
) != 0);
321 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
322 BUG_ON(stripe_operations_active(sh
));
325 pr_debug("init_stripe called, stripe %llu\n",
326 (unsigned long long)sh
->sector
);
330 sh
->generation
= conf
->generation
- previous
;
331 sh
->disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
333 stripe_set_idx(sector
, conf
, previous
, sh
);
337 for (i
= sh
->disks
; i
--; ) {
338 struct r5dev
*dev
= &sh
->dev
[i
];
340 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
341 test_bit(R5_LOCKED
, &dev
->flags
)) {
342 printk(KERN_ERR
"sector=%llx i=%d %p %p %p %p %d\n",
343 (unsigned long long)sh
->sector
, i
, dev
->toread
,
344 dev
->read
, dev
->towrite
, dev
->written
,
345 test_bit(R5_LOCKED
, &dev
->flags
));
349 raid5_build_block(sh
, i
, previous
);
351 insert_hash(conf
, sh
);
354 static struct stripe_head
*__find_stripe(raid5_conf_t
*conf
, sector_t sector
,
357 struct stripe_head
*sh
;
358 struct hlist_node
*hn
;
361 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
362 hlist_for_each_entry(sh
, hn
, stripe_hash(conf
, sector
), hash
)
363 if (sh
->sector
== sector
&& sh
->generation
== generation
)
365 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
370 * Need to check if array has failed when deciding whether to:
372 * - remove non-faulty devices
375 * This determination is simple when no reshape is happening.
376 * However if there is a reshape, we need to carefully check
377 * both the before and after sections.
378 * This is because some failed devices may only affect one
379 * of the two sections, and some non-in_sync devices may
380 * be insync in the section most affected by failed devices.
382 static int has_failed(raid5_conf_t
*conf
)
386 if (conf
->mddev
->reshape_position
== MaxSector
)
387 return conf
->mddev
->degraded
> conf
->max_degraded
;
391 for (i
= 0; i
< conf
->previous_raid_disks
; i
++) {
392 mdk_rdev_t
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
393 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
395 else if (test_bit(In_sync
, &rdev
->flags
))
398 /* not in-sync or faulty.
399 * If the reshape increases the number of devices,
400 * this is being recovered by the reshape, so
401 * this 'previous' section is not in_sync.
402 * If the number of devices is being reduced however,
403 * the device can only be part of the array if
404 * we are reverting a reshape, so this section will
407 if (conf
->raid_disks
>= conf
->previous_raid_disks
)
411 if (degraded
> conf
->max_degraded
)
415 for (i
= 0; i
< conf
->raid_disks
; i
++) {
416 mdk_rdev_t
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
417 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
419 else if (test_bit(In_sync
, &rdev
->flags
))
422 /* not in-sync or faulty.
423 * If reshape increases the number of devices, this
424 * section has already been recovered, else it
425 * almost certainly hasn't.
427 if (conf
->raid_disks
<= conf
->previous_raid_disks
)
431 if (degraded
> conf
->max_degraded
)
436 static void unplug_slaves(mddev_t
*mddev
);
438 static struct stripe_head
*
439 get_active_stripe(raid5_conf_t
*conf
, sector_t sector
,
440 int previous
, int noblock
, int noquiesce
)
442 struct stripe_head
*sh
;
444 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
446 spin_lock_irq(&conf
->device_lock
);
449 wait_event_lock_irq(conf
->wait_for_stripe
,
450 conf
->quiesce
== 0 || noquiesce
,
451 conf
->device_lock
, /* nothing */);
452 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
454 if (!conf
->inactive_blocked
)
455 sh
= get_free_stripe(conf
);
456 if (noblock
&& sh
== NULL
)
459 conf
->inactive_blocked
= 1;
460 wait_event_lock_irq(conf
->wait_for_stripe
,
461 !list_empty(&conf
->inactive_list
) &&
462 (atomic_read(&conf
->active_stripes
)
463 < (conf
->max_nr_stripes
*3/4)
464 || !conf
->inactive_blocked
),
466 md_raid5_unplug_device(conf
)
468 conf
->inactive_blocked
= 0;
470 init_stripe(sh
, sector
, previous
);
472 if (atomic_read(&sh
->count
)) {
473 BUG_ON(!list_empty(&sh
->lru
)
474 && !test_bit(STRIPE_EXPANDING
, &sh
->state
));
476 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
477 atomic_inc(&conf
->active_stripes
);
478 if (list_empty(&sh
->lru
) &&
479 !test_bit(STRIPE_EXPANDING
, &sh
->state
))
481 list_del_init(&sh
->lru
);
484 } while (sh
== NULL
);
487 atomic_inc(&sh
->count
);
489 spin_unlock_irq(&conf
->device_lock
);
494 raid5_end_read_request(struct bio
*bi
, int error
);
496 raid5_end_write_request(struct bio
*bi
, int error
);
498 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
500 raid5_conf_t
*conf
= sh
->raid_conf
;
501 int i
, disks
= sh
->disks
;
505 for (i
= disks
; i
--; ) {
509 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
510 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
514 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
519 bi
= &sh
->dev
[i
].req
;
523 bi
->bi_end_io
= raid5_end_write_request
;
525 bi
->bi_end_io
= raid5_end_read_request
;
528 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
529 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
532 atomic_inc(&rdev
->nr_pending
);
536 if (s
->syncing
|| s
->expanding
|| s
->expanded
)
537 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
539 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
541 bi
->bi_bdev
= rdev
->bdev
;
542 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
543 __func__
, (unsigned long long)sh
->sector
,
545 atomic_inc(&sh
->count
);
546 bi
->bi_sector
= sh
->sector
+ rdev
->data_offset
;
547 bi
->bi_flags
= 1 << BIO_UPTODATE
;
551 bi
->bi_io_vec
= &sh
->dev
[i
].vec
;
552 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
553 bi
->bi_io_vec
[0].bv_offset
= 0;
554 bi
->bi_size
= STRIPE_SIZE
;
557 test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
))
558 atomic_add(STRIPE_SECTORS
,
559 &rdev
->corrected_errors
);
560 generic_make_request(bi
);
563 set_bit(STRIPE_DEGRADED
, &sh
->state
);
564 pr_debug("skip op %ld on disc %d for sector %llu\n",
565 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
566 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
567 set_bit(STRIPE_HANDLE
, &sh
->state
);
572 static struct dma_async_tx_descriptor
*
573 async_copy_data(int frombio
, struct bio
*bio
, struct page
*page
,
574 sector_t sector
, struct dma_async_tx_descriptor
*tx
)
577 struct page
*bio_page
;
580 struct async_submit_ctl submit
;
581 enum async_tx_flags flags
= 0;
583 if (bio
->bi_sector
>= sector
)
584 page_offset
= (signed)(bio
->bi_sector
- sector
) * 512;
586 page_offset
= (signed)(sector
- bio
->bi_sector
) * -512;
589 flags
|= ASYNC_TX_FENCE
;
590 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
592 bio_for_each_segment(bvl
, bio
, i
) {
593 int len
= bio_iovec_idx(bio
, i
)->bv_len
;
597 if (page_offset
< 0) {
598 b_offset
= -page_offset
;
599 page_offset
+= b_offset
;
603 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
604 clen
= STRIPE_SIZE
- page_offset
;
609 b_offset
+= bio_iovec_idx(bio
, i
)->bv_offset
;
610 bio_page
= bio_iovec_idx(bio
, i
)->bv_page
;
612 tx
= async_memcpy(page
, bio_page
, page_offset
,
613 b_offset
, clen
, &submit
);
615 tx
= async_memcpy(bio_page
, page
, b_offset
,
616 page_offset
, clen
, &submit
);
618 /* chain the operations */
619 submit
.depend_tx
= tx
;
621 if (clen
< len
) /* hit end of page */
629 static void ops_complete_biofill(void *stripe_head_ref
)
631 struct stripe_head
*sh
= stripe_head_ref
;
632 struct bio
*return_bi
= NULL
;
633 raid5_conf_t
*conf
= sh
->raid_conf
;
636 pr_debug("%s: stripe %llu\n", __func__
,
637 (unsigned long long)sh
->sector
);
639 /* clear completed biofills */
640 spin_lock_irq(&conf
->device_lock
);
641 for (i
= sh
->disks
; i
--; ) {
642 struct r5dev
*dev
= &sh
->dev
[i
];
644 /* acknowledge completion of a biofill operation */
645 /* and check if we need to reply to a read request,
646 * new R5_Wantfill requests are held off until
647 * !STRIPE_BIOFILL_RUN
649 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
650 struct bio
*rbi
, *rbi2
;
655 while (rbi
&& rbi
->bi_sector
<
656 dev
->sector
+ STRIPE_SECTORS
) {
657 rbi2
= r5_next_bio(rbi
, dev
->sector
);
658 if (!raid5_dec_bi_phys_segments(rbi
)) {
659 rbi
->bi_next
= return_bi
;
666 spin_unlock_irq(&conf
->device_lock
);
667 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
669 return_io(return_bi
);
671 set_bit(STRIPE_HANDLE
, &sh
->state
);
675 static void ops_run_biofill(struct stripe_head
*sh
)
677 struct dma_async_tx_descriptor
*tx
= NULL
;
678 raid5_conf_t
*conf
= sh
->raid_conf
;
679 struct async_submit_ctl submit
;
682 pr_debug("%s: stripe %llu\n", __func__
,
683 (unsigned long long)sh
->sector
);
685 for (i
= sh
->disks
; i
--; ) {
686 struct r5dev
*dev
= &sh
->dev
[i
];
687 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
689 spin_lock_irq(&conf
->device_lock
);
690 dev
->read
= rbi
= dev
->toread
;
692 spin_unlock_irq(&conf
->device_lock
);
693 while (rbi
&& rbi
->bi_sector
<
694 dev
->sector
+ STRIPE_SECTORS
) {
695 tx
= async_copy_data(0, rbi
, dev
->page
,
697 rbi
= r5_next_bio(rbi
, dev
->sector
);
702 atomic_inc(&sh
->count
);
703 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
704 async_trigger_callback(&submit
);
707 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
714 tgt
= &sh
->dev
[target
];
715 set_bit(R5_UPTODATE
, &tgt
->flags
);
716 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
717 clear_bit(R5_Wantcompute
, &tgt
->flags
);
720 static void ops_complete_compute(void *stripe_head_ref
)
722 struct stripe_head
*sh
= stripe_head_ref
;
724 pr_debug("%s: stripe %llu\n", __func__
,
725 (unsigned long long)sh
->sector
);
727 /* mark the computed target(s) as uptodate */
728 mark_target_uptodate(sh
, sh
->ops
.target
);
729 mark_target_uptodate(sh
, sh
->ops
.target2
);
731 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
732 if (sh
->check_state
== check_state_compute_run
)
733 sh
->check_state
= check_state_compute_result
;
734 set_bit(STRIPE_HANDLE
, &sh
->state
);
738 /* return a pointer to the address conversion region of the scribble buffer */
739 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
740 struct raid5_percpu
*percpu
)
742 return percpu
->scribble
+ sizeof(struct page
*) * (sh
->disks
+ 2);
745 static struct dma_async_tx_descriptor
*
746 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
748 int disks
= sh
->disks
;
749 struct page
**xor_srcs
= percpu
->scribble
;
750 int target
= sh
->ops
.target
;
751 struct r5dev
*tgt
= &sh
->dev
[target
];
752 struct page
*xor_dest
= tgt
->page
;
754 struct dma_async_tx_descriptor
*tx
;
755 struct async_submit_ctl submit
;
758 pr_debug("%s: stripe %llu block: %d\n",
759 __func__
, (unsigned long long)sh
->sector
, target
);
760 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
762 for (i
= disks
; i
--; )
764 xor_srcs
[count
++] = sh
->dev
[i
].page
;
766 atomic_inc(&sh
->count
);
768 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
769 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
));
770 if (unlikely(count
== 1))
771 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
773 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
778 /* set_syndrome_sources - populate source buffers for gen_syndrome
779 * @srcs - (struct page *) array of size sh->disks
780 * @sh - stripe_head to parse
782 * Populates srcs in proper layout order for the stripe and returns the
783 * 'count' of sources to be used in a call to async_gen_syndrome. The P
784 * destination buffer is recorded in srcs[count] and the Q destination
785 * is recorded in srcs[count+1]].
787 static int set_syndrome_sources(struct page
**srcs
, struct stripe_head
*sh
)
789 int disks
= sh
->disks
;
790 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
791 int d0_idx
= raid6_d0(sh
);
795 for (i
= 0; i
< disks
; i
++)
801 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
803 srcs
[slot
] = sh
->dev
[i
].page
;
804 i
= raid6_next_disk(i
, disks
);
805 } while (i
!= d0_idx
);
807 return syndrome_disks
;
810 static struct dma_async_tx_descriptor
*
811 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
813 int disks
= sh
->disks
;
814 struct page
**blocks
= percpu
->scribble
;
816 int qd_idx
= sh
->qd_idx
;
817 struct dma_async_tx_descriptor
*tx
;
818 struct async_submit_ctl submit
;
824 if (sh
->ops
.target
< 0)
825 target
= sh
->ops
.target2
;
826 else if (sh
->ops
.target2
< 0)
827 target
= sh
->ops
.target
;
829 /* we should only have one valid target */
832 pr_debug("%s: stripe %llu block: %d\n",
833 __func__
, (unsigned long long)sh
->sector
, target
);
835 tgt
= &sh
->dev
[target
];
836 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
839 atomic_inc(&sh
->count
);
841 if (target
== qd_idx
) {
842 count
= set_syndrome_sources(blocks
, sh
);
843 blocks
[count
] = NULL
; /* regenerating p is not necessary */
844 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
845 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
846 ops_complete_compute
, sh
,
847 to_addr_conv(sh
, percpu
));
848 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
850 /* Compute any data- or p-drive using XOR */
852 for (i
= disks
; i
-- ; ) {
853 if (i
== target
|| i
== qd_idx
)
855 blocks
[count
++] = sh
->dev
[i
].page
;
858 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
859 NULL
, ops_complete_compute
, sh
,
860 to_addr_conv(sh
, percpu
));
861 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
867 static struct dma_async_tx_descriptor
*
868 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
870 int i
, count
, disks
= sh
->disks
;
871 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
872 int d0_idx
= raid6_d0(sh
);
873 int faila
= -1, failb
= -1;
874 int target
= sh
->ops
.target
;
875 int target2
= sh
->ops
.target2
;
876 struct r5dev
*tgt
= &sh
->dev
[target
];
877 struct r5dev
*tgt2
= &sh
->dev
[target2
];
878 struct dma_async_tx_descriptor
*tx
;
879 struct page
**blocks
= percpu
->scribble
;
880 struct async_submit_ctl submit
;
882 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
883 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
884 BUG_ON(target
< 0 || target2
< 0);
885 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
886 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
888 /* we need to open-code set_syndrome_sources to handle the
889 * slot number conversion for 'faila' and 'failb'
891 for (i
= 0; i
< disks
; i
++)
896 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
898 blocks
[slot
] = sh
->dev
[i
].page
;
904 i
= raid6_next_disk(i
, disks
);
905 } while (i
!= d0_idx
);
907 BUG_ON(faila
== failb
);
910 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
911 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
913 atomic_inc(&sh
->count
);
915 if (failb
== syndrome_disks
+1) {
916 /* Q disk is one of the missing disks */
917 if (faila
== syndrome_disks
) {
918 /* Missing P+Q, just recompute */
919 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
920 ops_complete_compute
, sh
,
921 to_addr_conv(sh
, percpu
));
922 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
923 STRIPE_SIZE
, &submit
);
927 int qd_idx
= sh
->qd_idx
;
929 /* Missing D+Q: recompute D from P, then recompute Q */
930 if (target
== qd_idx
)
931 data_target
= target2
;
933 data_target
= target
;
936 for (i
= disks
; i
-- ; ) {
937 if (i
== data_target
|| i
== qd_idx
)
939 blocks
[count
++] = sh
->dev
[i
].page
;
941 dest
= sh
->dev
[data_target
].page
;
942 init_async_submit(&submit
,
943 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
945 to_addr_conv(sh
, percpu
));
946 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
949 count
= set_syndrome_sources(blocks
, sh
);
950 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
951 ops_complete_compute
, sh
,
952 to_addr_conv(sh
, percpu
));
953 return async_gen_syndrome(blocks
, 0, count
+2,
954 STRIPE_SIZE
, &submit
);
957 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
958 ops_complete_compute
, sh
,
959 to_addr_conv(sh
, percpu
));
960 if (failb
== syndrome_disks
) {
961 /* We're missing D+P. */
962 return async_raid6_datap_recov(syndrome_disks
+2,
966 /* We're missing D+D. */
967 return async_raid6_2data_recov(syndrome_disks
+2,
968 STRIPE_SIZE
, faila
, failb
,
975 static void ops_complete_prexor(void *stripe_head_ref
)
977 struct stripe_head
*sh
= stripe_head_ref
;
979 pr_debug("%s: stripe %llu\n", __func__
,
980 (unsigned long long)sh
->sector
);
983 static struct dma_async_tx_descriptor
*
984 ops_run_prexor(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
985 struct dma_async_tx_descriptor
*tx
)
987 int disks
= sh
->disks
;
988 struct page
**xor_srcs
= percpu
->scribble
;
989 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
990 struct async_submit_ctl submit
;
992 /* existing parity data subtracted */
993 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
995 pr_debug("%s: stripe %llu\n", __func__
,
996 (unsigned long long)sh
->sector
);
998 for (i
= disks
; i
--; ) {
999 struct r5dev
*dev
= &sh
->dev
[i
];
1000 /* Only process blocks that are known to be uptodate */
1001 if (test_bit(R5_Wantdrain
, &dev
->flags
))
1002 xor_srcs
[count
++] = dev
->page
;
1005 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1006 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
));
1007 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1012 static struct dma_async_tx_descriptor
*
1013 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1015 int disks
= sh
->disks
;
1018 pr_debug("%s: stripe %llu\n", __func__
,
1019 (unsigned long long)sh
->sector
);
1021 for (i
= disks
; i
--; ) {
1022 struct r5dev
*dev
= &sh
->dev
[i
];
1025 if (test_and_clear_bit(R5_Wantdrain
, &dev
->flags
)) {
1028 spin_lock(&sh
->lock
);
1029 chosen
= dev
->towrite
;
1030 dev
->towrite
= NULL
;
1031 BUG_ON(dev
->written
);
1032 wbi
= dev
->written
= chosen
;
1033 spin_unlock(&sh
->lock
);
1035 while (wbi
&& wbi
->bi_sector
<
1036 dev
->sector
+ STRIPE_SECTORS
) {
1037 if (wbi
->bi_rw
& REQ_FUA
)
1038 set_bit(R5_WantFUA
, &dev
->flags
);
1039 tx
= async_copy_data(1, wbi
, dev
->page
,
1041 wbi
= r5_next_bio(wbi
, dev
->sector
);
1049 static void ops_complete_reconstruct(void *stripe_head_ref
)
1051 struct stripe_head
*sh
= stripe_head_ref
;
1052 int disks
= sh
->disks
;
1053 int pd_idx
= sh
->pd_idx
;
1054 int qd_idx
= sh
->qd_idx
;
1058 pr_debug("%s: stripe %llu\n", __func__
,
1059 (unsigned long long)sh
->sector
);
1061 for (i
= disks
; i
--; )
1062 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1064 for (i
= disks
; i
--; ) {
1065 struct r5dev
*dev
= &sh
->dev
[i
];
1067 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1068 set_bit(R5_UPTODATE
, &dev
->flags
);
1070 set_bit(R5_WantFUA
, &dev
->flags
);
1074 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1075 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1076 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1077 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1079 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1080 sh
->reconstruct_state
= reconstruct_state_result
;
1083 set_bit(STRIPE_HANDLE
, &sh
->state
);
1088 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1089 struct dma_async_tx_descriptor
*tx
)
1091 int disks
= sh
->disks
;
1092 struct page
**xor_srcs
= percpu
->scribble
;
1093 struct async_submit_ctl submit
;
1094 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1095 struct page
*xor_dest
;
1097 unsigned long flags
;
1099 pr_debug("%s: stripe %llu\n", __func__
,
1100 (unsigned long long)sh
->sector
);
1102 /* check if prexor is active which means only process blocks
1103 * that are part of a read-modify-write (written)
1105 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1107 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1108 for (i
= disks
; i
--; ) {
1109 struct r5dev
*dev
= &sh
->dev
[i
];
1111 xor_srcs
[count
++] = dev
->page
;
1114 xor_dest
= sh
->dev
[pd_idx
].page
;
1115 for (i
= disks
; i
--; ) {
1116 struct r5dev
*dev
= &sh
->dev
[i
];
1118 xor_srcs
[count
++] = dev
->page
;
1122 /* 1/ if we prexor'd then the dest is reused as a source
1123 * 2/ if we did not prexor then we are redoing the parity
1124 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1125 * for the synchronous xor case
1127 flags
= ASYNC_TX_ACK
|
1128 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1130 atomic_inc(&sh
->count
);
1132 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, sh
,
1133 to_addr_conv(sh
, percpu
));
1134 if (unlikely(count
== 1))
1135 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1137 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1141 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1142 struct dma_async_tx_descriptor
*tx
)
1144 struct async_submit_ctl submit
;
1145 struct page
**blocks
= percpu
->scribble
;
1148 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1150 count
= set_syndrome_sources(blocks
, sh
);
1152 atomic_inc(&sh
->count
);
1154 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_reconstruct
,
1155 sh
, to_addr_conv(sh
, percpu
));
1156 async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1159 static void ops_complete_check(void *stripe_head_ref
)
1161 struct stripe_head
*sh
= stripe_head_ref
;
1163 pr_debug("%s: stripe %llu\n", __func__
,
1164 (unsigned long long)sh
->sector
);
1166 sh
->check_state
= check_state_check_result
;
1167 set_bit(STRIPE_HANDLE
, &sh
->state
);
1171 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1173 int disks
= sh
->disks
;
1174 int pd_idx
= sh
->pd_idx
;
1175 int qd_idx
= sh
->qd_idx
;
1176 struct page
*xor_dest
;
1177 struct page
**xor_srcs
= percpu
->scribble
;
1178 struct dma_async_tx_descriptor
*tx
;
1179 struct async_submit_ctl submit
;
1183 pr_debug("%s: stripe %llu\n", __func__
,
1184 (unsigned long long)sh
->sector
);
1187 xor_dest
= sh
->dev
[pd_idx
].page
;
1188 xor_srcs
[count
++] = xor_dest
;
1189 for (i
= disks
; i
--; ) {
1190 if (i
== pd_idx
|| i
== qd_idx
)
1192 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1195 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1196 to_addr_conv(sh
, percpu
));
1197 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1198 &sh
->ops
.zero_sum_result
, &submit
);
1200 atomic_inc(&sh
->count
);
1201 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1202 tx
= async_trigger_callback(&submit
);
1205 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1207 struct page
**srcs
= percpu
->scribble
;
1208 struct async_submit_ctl submit
;
1211 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1212 (unsigned long long)sh
->sector
, checkp
);
1214 count
= set_syndrome_sources(srcs
, sh
);
1218 atomic_inc(&sh
->count
);
1219 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1220 sh
, to_addr_conv(sh
, percpu
));
1221 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1222 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1225 static void __raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1227 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1228 struct dma_async_tx_descriptor
*tx
= NULL
;
1229 raid5_conf_t
*conf
= sh
->raid_conf
;
1230 int level
= conf
->level
;
1231 struct raid5_percpu
*percpu
;
1235 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1236 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1237 ops_run_biofill(sh
);
1241 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1243 tx
= ops_run_compute5(sh
, percpu
);
1245 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
1246 tx
= ops_run_compute6_1(sh
, percpu
);
1248 tx
= ops_run_compute6_2(sh
, percpu
);
1250 /* terminate the chain if reconstruct is not set to be run */
1251 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
1255 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
))
1256 tx
= ops_run_prexor(sh
, percpu
, tx
);
1258 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
1259 tx
= ops_run_biodrain(sh
, tx
);
1263 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
1265 ops_run_reconstruct5(sh
, percpu
, tx
);
1267 ops_run_reconstruct6(sh
, percpu
, tx
);
1270 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
1271 if (sh
->check_state
== check_state_run
)
1272 ops_run_check_p(sh
, percpu
);
1273 else if (sh
->check_state
== check_state_run_q
)
1274 ops_run_check_pq(sh
, percpu
, 0);
1275 else if (sh
->check_state
== check_state_run_pq
)
1276 ops_run_check_pq(sh
, percpu
, 1);
1282 for (i
= disks
; i
--; ) {
1283 struct r5dev
*dev
= &sh
->dev
[i
];
1284 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
1285 wake_up(&sh
->raid_conf
->wait_for_overlap
);
1290 #ifdef CONFIG_MULTICORE_RAID456
1291 static void async_run_ops(void *param
, async_cookie_t cookie
)
1293 struct stripe_head
*sh
= param
;
1294 unsigned long ops_request
= sh
->ops
.request
;
1296 clear_bit_unlock(STRIPE_OPS_REQ_PENDING
, &sh
->state
);
1297 wake_up(&sh
->ops
.wait_for_ops
);
1299 __raid_run_ops(sh
, ops_request
);
1303 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1305 /* since handle_stripe can be called outside of raid5d context
1306 * we need to ensure sh->ops.request is de-staged before another
1309 wait_event(sh
->ops
.wait_for_ops
,
1310 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING
, &sh
->state
));
1311 sh
->ops
.request
= ops_request
;
1313 atomic_inc(&sh
->count
);
1314 async_schedule(async_run_ops
, sh
);
1317 #define raid_run_ops __raid_run_ops
1320 static int grow_one_stripe(raid5_conf_t
*conf
)
1322 struct stripe_head
*sh
;
1323 sh
= kmem_cache_alloc(conf
->slab_cache
, GFP_KERNEL
);
1326 memset(sh
, 0, sizeof(*sh
) + (conf
->pool_size
-1)*sizeof(struct r5dev
));
1327 sh
->raid_conf
= conf
;
1328 spin_lock_init(&sh
->lock
);
1329 #ifdef CONFIG_MULTICORE_RAID456
1330 init_waitqueue_head(&sh
->ops
.wait_for_ops
);
1333 if (grow_buffers(sh
)) {
1335 kmem_cache_free(conf
->slab_cache
, sh
);
1338 /* we just created an active stripe so... */
1339 atomic_set(&sh
->count
, 1);
1340 atomic_inc(&conf
->active_stripes
);
1341 INIT_LIST_HEAD(&sh
->lru
);
1346 static int grow_stripes(raid5_conf_t
*conf
, int num
)
1348 struct kmem_cache
*sc
;
1349 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
1351 if (conf
->mddev
->gendisk
)
1352 sprintf(conf
->cache_name
[0],
1353 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
1355 sprintf(conf
->cache_name
[0],
1356 "raid%d-%p", conf
->level
, conf
->mddev
);
1357 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
1359 conf
->active_name
= 0;
1360 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
1361 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
1365 conf
->slab_cache
= sc
;
1366 conf
->pool_size
= devs
;
1368 if (!grow_one_stripe(conf
))
1374 * scribble_len - return the required size of the scribble region
1375 * @num - total number of disks in the array
1377 * The size must be enough to contain:
1378 * 1/ a struct page pointer for each device in the array +2
1379 * 2/ room to convert each entry in (1) to its corresponding dma
1380 * (dma_map_page()) or page (page_address()) address.
1382 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1383 * calculate over all devices (not just the data blocks), using zeros in place
1384 * of the P and Q blocks.
1386 static size_t scribble_len(int num
)
1390 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
1395 static int resize_stripes(raid5_conf_t
*conf
, int newsize
)
1397 /* Make all the stripes able to hold 'newsize' devices.
1398 * New slots in each stripe get 'page' set to a new page.
1400 * This happens in stages:
1401 * 1/ create a new kmem_cache and allocate the required number of
1403 * 2/ gather all the old stripe_heads and tranfer the pages across
1404 * to the new stripe_heads. This will have the side effect of
1405 * freezing the array as once all stripe_heads have been collected,
1406 * no IO will be possible. Old stripe heads are freed once their
1407 * pages have been transferred over, and the old kmem_cache is
1408 * freed when all stripes are done.
1409 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1410 * we simple return a failre status - no need to clean anything up.
1411 * 4/ allocate new pages for the new slots in the new stripe_heads.
1412 * If this fails, we don't bother trying the shrink the
1413 * stripe_heads down again, we just leave them as they are.
1414 * As each stripe_head is processed the new one is released into
1417 * Once step2 is started, we cannot afford to wait for a write,
1418 * so we use GFP_NOIO allocations.
1420 struct stripe_head
*osh
, *nsh
;
1421 LIST_HEAD(newstripes
);
1422 struct disk_info
*ndisks
;
1425 struct kmem_cache
*sc
;
1428 if (newsize
<= conf
->pool_size
)
1429 return 0; /* never bother to shrink */
1431 err
= md_allow_write(conf
->mddev
);
1436 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
1437 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
1442 for (i
= conf
->max_nr_stripes
; i
; i
--) {
1443 nsh
= kmem_cache_alloc(sc
, GFP_KERNEL
);
1447 memset(nsh
, 0, sizeof(*nsh
) + (newsize
-1)*sizeof(struct r5dev
));
1449 nsh
->raid_conf
= conf
;
1450 spin_lock_init(&nsh
->lock
);
1451 #ifdef CONFIG_MULTICORE_RAID456
1452 init_waitqueue_head(&nsh
->ops
.wait_for_ops
);
1455 list_add(&nsh
->lru
, &newstripes
);
1458 /* didn't get enough, give up */
1459 while (!list_empty(&newstripes
)) {
1460 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1461 list_del(&nsh
->lru
);
1462 kmem_cache_free(sc
, nsh
);
1464 kmem_cache_destroy(sc
);
1467 /* Step 2 - Must use GFP_NOIO now.
1468 * OK, we have enough stripes, start collecting inactive
1469 * stripes and copying them over
1471 list_for_each_entry(nsh
, &newstripes
, lru
) {
1472 spin_lock_irq(&conf
->device_lock
);
1473 wait_event_lock_irq(conf
->wait_for_stripe
,
1474 !list_empty(&conf
->inactive_list
),
1476 unplug_slaves(conf
->mddev
)
1478 osh
= get_free_stripe(conf
);
1479 spin_unlock_irq(&conf
->device_lock
);
1480 atomic_set(&nsh
->count
, 1);
1481 for(i
=0; i
<conf
->pool_size
; i
++)
1482 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
1483 for( ; i
<newsize
; i
++)
1484 nsh
->dev
[i
].page
= NULL
;
1485 kmem_cache_free(conf
->slab_cache
, osh
);
1487 kmem_cache_destroy(conf
->slab_cache
);
1490 * At this point, we are holding all the stripes so the array
1491 * is completely stalled, so now is a good time to resize
1492 * conf->disks and the scribble region
1494 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
1496 for (i
=0; i
<conf
->raid_disks
; i
++)
1497 ndisks
[i
] = conf
->disks
[i
];
1499 conf
->disks
= ndisks
;
1504 conf
->scribble_len
= scribble_len(newsize
);
1505 for_each_present_cpu(cpu
) {
1506 struct raid5_percpu
*percpu
;
1509 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1510 scribble
= kmalloc(conf
->scribble_len
, GFP_NOIO
);
1513 kfree(percpu
->scribble
);
1514 percpu
->scribble
= scribble
;
1522 /* Step 4, return new stripes to service */
1523 while(!list_empty(&newstripes
)) {
1524 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1525 list_del_init(&nsh
->lru
);
1527 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
1528 if (nsh
->dev
[i
].page
== NULL
) {
1529 struct page
*p
= alloc_page(GFP_NOIO
);
1530 nsh
->dev
[i
].page
= p
;
1534 release_stripe(nsh
);
1536 /* critical section pass, GFP_NOIO no longer needed */
1538 conf
->slab_cache
= sc
;
1539 conf
->active_name
= 1-conf
->active_name
;
1540 conf
->pool_size
= newsize
;
1544 static int drop_one_stripe(raid5_conf_t
*conf
)
1546 struct stripe_head
*sh
;
1548 spin_lock_irq(&conf
->device_lock
);
1549 sh
= get_free_stripe(conf
);
1550 spin_unlock_irq(&conf
->device_lock
);
1553 BUG_ON(atomic_read(&sh
->count
));
1555 kmem_cache_free(conf
->slab_cache
, sh
);
1556 atomic_dec(&conf
->active_stripes
);
1560 static void shrink_stripes(raid5_conf_t
*conf
)
1562 while (drop_one_stripe(conf
))
1565 if (conf
->slab_cache
)
1566 kmem_cache_destroy(conf
->slab_cache
);
1567 conf
->slab_cache
= NULL
;
1570 static void raid5_end_read_request(struct bio
* bi
, int error
)
1572 struct stripe_head
*sh
= bi
->bi_private
;
1573 raid5_conf_t
*conf
= sh
->raid_conf
;
1574 int disks
= sh
->disks
, i
;
1575 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1576 char b
[BDEVNAME_SIZE
];
1580 for (i
=0 ; i
<disks
; i
++)
1581 if (bi
== &sh
->dev
[i
].req
)
1584 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1585 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1593 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1594 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
1595 rdev
= conf
->disks
[i
].rdev
;
1596 printk_rl(KERN_INFO
"md/raid:%s: read error corrected"
1597 " (%lu sectors at %llu on %s)\n",
1598 mdname(conf
->mddev
), STRIPE_SECTORS
,
1599 (unsigned long long)(sh
->sector
1600 + rdev
->data_offset
),
1601 bdevname(rdev
->bdev
, b
));
1602 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1603 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1605 if (atomic_read(&conf
->disks
[i
].rdev
->read_errors
))
1606 atomic_set(&conf
->disks
[i
].rdev
->read_errors
, 0);
1608 const char *bdn
= bdevname(conf
->disks
[i
].rdev
->bdev
, b
);
1610 rdev
= conf
->disks
[i
].rdev
;
1612 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1613 atomic_inc(&rdev
->read_errors
);
1614 if (conf
->mddev
->degraded
>= conf
->max_degraded
)
1615 printk_rl(KERN_WARNING
1616 "md/raid:%s: read error not correctable "
1617 "(sector %llu on %s).\n",
1618 mdname(conf
->mddev
),
1619 (unsigned long long)(sh
->sector
1620 + rdev
->data_offset
),
1622 else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
))
1624 printk_rl(KERN_WARNING
1625 "md/raid:%s: read error NOT corrected!! "
1626 "(sector %llu on %s).\n",
1627 mdname(conf
->mddev
),
1628 (unsigned long long)(sh
->sector
1629 + rdev
->data_offset
),
1631 else if (atomic_read(&rdev
->read_errors
)
1632 > conf
->max_nr_stripes
)
1634 "md/raid:%s: Too many read errors, failing device %s.\n",
1635 mdname(conf
->mddev
), bdn
);
1639 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1641 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1642 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1643 md_error(conf
->mddev
, rdev
);
1646 rdev_dec_pending(conf
->disks
[i
].rdev
, conf
->mddev
);
1647 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1648 set_bit(STRIPE_HANDLE
, &sh
->state
);
1652 static void raid5_end_write_request(struct bio
*bi
, int error
)
1654 struct stripe_head
*sh
= bi
->bi_private
;
1655 raid5_conf_t
*conf
= sh
->raid_conf
;
1656 int disks
= sh
->disks
, i
;
1657 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1659 for (i
=0 ; i
<disks
; i
++)
1660 if (bi
== &sh
->dev
[i
].req
)
1663 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1664 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1672 md_error(conf
->mddev
, conf
->disks
[i
].rdev
);
1674 rdev_dec_pending(conf
->disks
[i
].rdev
, conf
->mddev
);
1676 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1677 set_bit(STRIPE_HANDLE
, &sh
->state
);
1682 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
);
1684 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
1686 struct r5dev
*dev
= &sh
->dev
[i
];
1688 bio_init(&dev
->req
);
1689 dev
->req
.bi_io_vec
= &dev
->vec
;
1691 dev
->req
.bi_max_vecs
++;
1692 dev
->vec
.bv_page
= dev
->page
;
1693 dev
->vec
.bv_len
= STRIPE_SIZE
;
1694 dev
->vec
.bv_offset
= 0;
1696 dev
->req
.bi_sector
= sh
->sector
;
1697 dev
->req
.bi_private
= sh
;
1700 dev
->sector
= compute_blocknr(sh
, i
, previous
);
1703 static void error(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1705 char b
[BDEVNAME_SIZE
];
1706 raid5_conf_t
*conf
= mddev
->private;
1707 pr_debug("raid456: error called\n");
1709 if (!test_bit(Faulty
, &rdev
->flags
)) {
1710 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1711 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1712 unsigned long flags
;
1713 spin_lock_irqsave(&conf
->device_lock
, flags
);
1715 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1717 * if recovery was running, make sure it aborts.
1719 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1721 set_bit(Faulty
, &rdev
->flags
);
1723 "md/raid:%s: Disk failure on %s, disabling device.\n"
1724 "md/raid:%s: Operation continuing on %d devices.\n",
1726 bdevname(rdev
->bdev
, b
),
1728 conf
->raid_disks
- mddev
->degraded
);
1733 * Input: a 'big' sector number,
1734 * Output: index of the data and parity disk, and the sector # in them.
1736 static sector_t
raid5_compute_sector(raid5_conf_t
*conf
, sector_t r_sector
,
1737 int previous
, int *dd_idx
,
1738 struct stripe_head
*sh
)
1740 sector_t stripe
, stripe2
;
1741 sector_t chunk_number
;
1742 unsigned int chunk_offset
;
1745 sector_t new_sector
;
1746 int algorithm
= previous
? conf
->prev_algo
1748 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
1749 : conf
->chunk_sectors
;
1750 int raid_disks
= previous
? conf
->previous_raid_disks
1752 int data_disks
= raid_disks
- conf
->max_degraded
;
1754 /* First compute the information on this sector */
1757 * Compute the chunk number and the sector offset inside the chunk
1759 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
1760 chunk_number
= r_sector
;
1763 * Compute the stripe number
1765 stripe
= chunk_number
;
1766 *dd_idx
= sector_div(stripe
, data_disks
);
1769 * Select the parity disk based on the user selected algorithm.
1771 pd_idx
= qd_idx
= ~0;
1772 switch(conf
->level
) {
1774 pd_idx
= data_disks
;
1777 switch (algorithm
) {
1778 case ALGORITHM_LEFT_ASYMMETRIC
:
1779 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
1780 if (*dd_idx
>= pd_idx
)
1783 case ALGORITHM_RIGHT_ASYMMETRIC
:
1784 pd_idx
= sector_div(stripe2
, raid_disks
);
1785 if (*dd_idx
>= pd_idx
)
1788 case ALGORITHM_LEFT_SYMMETRIC
:
1789 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
1790 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1792 case ALGORITHM_RIGHT_SYMMETRIC
:
1793 pd_idx
= sector_div(stripe2
, raid_disks
);
1794 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1796 case ALGORITHM_PARITY_0
:
1800 case ALGORITHM_PARITY_N
:
1801 pd_idx
= data_disks
;
1809 switch (algorithm
) {
1810 case ALGORITHM_LEFT_ASYMMETRIC
:
1811 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1812 qd_idx
= pd_idx
+ 1;
1813 if (pd_idx
== raid_disks
-1) {
1814 (*dd_idx
)++; /* Q D D D P */
1816 } else if (*dd_idx
>= pd_idx
)
1817 (*dd_idx
) += 2; /* D D P Q D */
1819 case ALGORITHM_RIGHT_ASYMMETRIC
:
1820 pd_idx
= sector_div(stripe2
, raid_disks
);
1821 qd_idx
= pd_idx
+ 1;
1822 if (pd_idx
== raid_disks
-1) {
1823 (*dd_idx
)++; /* Q D D D P */
1825 } else if (*dd_idx
>= pd_idx
)
1826 (*dd_idx
) += 2; /* D D P Q D */
1828 case ALGORITHM_LEFT_SYMMETRIC
:
1829 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1830 qd_idx
= (pd_idx
+ 1) % raid_disks
;
1831 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
1833 case ALGORITHM_RIGHT_SYMMETRIC
:
1834 pd_idx
= sector_div(stripe2
, raid_disks
);
1835 qd_idx
= (pd_idx
+ 1) % raid_disks
;
1836 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
1839 case ALGORITHM_PARITY_0
:
1844 case ALGORITHM_PARITY_N
:
1845 pd_idx
= data_disks
;
1846 qd_idx
= data_disks
+ 1;
1849 case ALGORITHM_ROTATING_ZERO_RESTART
:
1850 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1851 * of blocks for computing Q is different.
1853 pd_idx
= sector_div(stripe2
, raid_disks
);
1854 qd_idx
= pd_idx
+ 1;
1855 if (pd_idx
== raid_disks
-1) {
1856 (*dd_idx
)++; /* Q D D D P */
1858 } else if (*dd_idx
>= pd_idx
)
1859 (*dd_idx
) += 2; /* D D P Q D */
1863 case ALGORITHM_ROTATING_N_RESTART
:
1864 /* Same a left_asymmetric, by first stripe is
1865 * D D D P Q rather than
1869 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1870 qd_idx
= pd_idx
+ 1;
1871 if (pd_idx
== raid_disks
-1) {
1872 (*dd_idx
)++; /* Q D D D P */
1874 } else if (*dd_idx
>= pd_idx
)
1875 (*dd_idx
) += 2; /* D D P Q D */
1879 case ALGORITHM_ROTATING_N_CONTINUE
:
1880 /* Same as left_symmetric but Q is before P */
1881 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1882 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
1883 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1887 case ALGORITHM_LEFT_ASYMMETRIC_6
:
1888 /* RAID5 left_asymmetric, with Q on last device */
1889 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
1890 if (*dd_idx
>= pd_idx
)
1892 qd_idx
= raid_disks
- 1;
1895 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
1896 pd_idx
= sector_div(stripe2
, raid_disks
-1);
1897 if (*dd_idx
>= pd_idx
)
1899 qd_idx
= raid_disks
- 1;
1902 case ALGORITHM_LEFT_SYMMETRIC_6
:
1903 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
1904 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
1905 qd_idx
= raid_disks
- 1;
1908 case ALGORITHM_RIGHT_SYMMETRIC_6
:
1909 pd_idx
= sector_div(stripe2
, raid_disks
-1);
1910 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
1911 qd_idx
= raid_disks
- 1;
1914 case ALGORITHM_PARITY_0_6
:
1917 qd_idx
= raid_disks
- 1;
1927 sh
->pd_idx
= pd_idx
;
1928 sh
->qd_idx
= qd_idx
;
1929 sh
->ddf_layout
= ddf_layout
;
1932 * Finally, compute the new sector number
1934 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
1939 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
1941 raid5_conf_t
*conf
= sh
->raid_conf
;
1942 int raid_disks
= sh
->disks
;
1943 int data_disks
= raid_disks
- conf
->max_degraded
;
1944 sector_t new_sector
= sh
->sector
, check
;
1945 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
1946 : conf
->chunk_sectors
;
1947 int algorithm
= previous
? conf
->prev_algo
1951 sector_t chunk_number
;
1952 int dummy1
, dd_idx
= i
;
1954 struct stripe_head sh2
;
1957 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
1958 stripe
= new_sector
;
1960 if (i
== sh
->pd_idx
)
1962 switch(conf
->level
) {
1965 switch (algorithm
) {
1966 case ALGORITHM_LEFT_ASYMMETRIC
:
1967 case ALGORITHM_RIGHT_ASYMMETRIC
:
1971 case ALGORITHM_LEFT_SYMMETRIC
:
1972 case ALGORITHM_RIGHT_SYMMETRIC
:
1975 i
-= (sh
->pd_idx
+ 1);
1977 case ALGORITHM_PARITY_0
:
1980 case ALGORITHM_PARITY_N
:
1987 if (i
== sh
->qd_idx
)
1988 return 0; /* It is the Q disk */
1989 switch (algorithm
) {
1990 case ALGORITHM_LEFT_ASYMMETRIC
:
1991 case ALGORITHM_RIGHT_ASYMMETRIC
:
1992 case ALGORITHM_ROTATING_ZERO_RESTART
:
1993 case ALGORITHM_ROTATING_N_RESTART
:
1994 if (sh
->pd_idx
== raid_disks
-1)
1995 i
--; /* Q D D D P */
1996 else if (i
> sh
->pd_idx
)
1997 i
-= 2; /* D D P Q D */
1999 case ALGORITHM_LEFT_SYMMETRIC
:
2000 case ALGORITHM_RIGHT_SYMMETRIC
:
2001 if (sh
->pd_idx
== raid_disks
-1)
2002 i
--; /* Q D D D P */
2007 i
-= (sh
->pd_idx
+ 2);
2010 case ALGORITHM_PARITY_0
:
2013 case ALGORITHM_PARITY_N
:
2015 case ALGORITHM_ROTATING_N_CONTINUE
:
2016 /* Like left_symmetric, but P is before Q */
2017 if (sh
->pd_idx
== 0)
2018 i
--; /* P D D D Q */
2023 i
-= (sh
->pd_idx
+ 1);
2026 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2027 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2031 case ALGORITHM_LEFT_SYMMETRIC_6
:
2032 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2034 i
+= data_disks
+ 1;
2035 i
-= (sh
->pd_idx
+ 1);
2037 case ALGORITHM_PARITY_0_6
:
2046 chunk_number
= stripe
* data_disks
+ i
;
2047 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2049 check
= raid5_compute_sector(conf
, r_sector
,
2050 previous
, &dummy1
, &sh2
);
2051 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2052 || sh2
.qd_idx
!= sh
->qd_idx
) {
2053 printk(KERN_ERR
"md/raid:%s: compute_blocknr: map not correct\n",
2054 mdname(conf
->mddev
));
2062 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2063 int rcw
, int expand
)
2065 int i
, pd_idx
= sh
->pd_idx
, disks
= sh
->disks
;
2066 raid5_conf_t
*conf
= sh
->raid_conf
;
2067 int level
= conf
->level
;
2070 /* if we are not expanding this is a proper write request, and
2071 * there will be bios with new data to be drained into the
2075 sh
->reconstruct_state
= reconstruct_state_drain_run
;
2076 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2078 sh
->reconstruct_state
= reconstruct_state_run
;
2080 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2082 for (i
= disks
; i
--; ) {
2083 struct r5dev
*dev
= &sh
->dev
[i
];
2086 set_bit(R5_LOCKED
, &dev
->flags
);
2087 set_bit(R5_Wantdrain
, &dev
->flags
);
2089 clear_bit(R5_UPTODATE
, &dev
->flags
);
2093 if (s
->locked
+ conf
->max_degraded
== disks
)
2094 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2095 atomic_inc(&conf
->pending_full_writes
);
2098 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
2099 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
2101 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
2102 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
2103 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2104 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2106 for (i
= disks
; i
--; ) {
2107 struct r5dev
*dev
= &sh
->dev
[i
];
2112 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2113 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2114 set_bit(R5_Wantdrain
, &dev
->flags
);
2115 set_bit(R5_LOCKED
, &dev
->flags
);
2116 clear_bit(R5_UPTODATE
, &dev
->flags
);
2122 /* keep the parity disk(s) locked while asynchronous operations
2125 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
2126 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2130 int qd_idx
= sh
->qd_idx
;
2131 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
2133 set_bit(R5_LOCKED
, &dev
->flags
);
2134 clear_bit(R5_UPTODATE
, &dev
->flags
);
2138 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2139 __func__
, (unsigned long long)sh
->sector
,
2140 s
->locked
, s
->ops_request
);
2144 * Each stripe/dev can have one or more bion attached.
2145 * toread/towrite point to the first in a chain.
2146 * The bi_next chain must be in order.
2148 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
, int forwrite
)
2151 raid5_conf_t
*conf
= sh
->raid_conf
;
2154 pr_debug("adding bh b#%llu to stripe s#%llu\n",
2155 (unsigned long long)bi
->bi_sector
,
2156 (unsigned long long)sh
->sector
);
2159 spin_lock(&sh
->lock
);
2160 spin_lock_irq(&conf
->device_lock
);
2162 bip
= &sh
->dev
[dd_idx
].towrite
;
2163 if (*bip
== NULL
&& sh
->dev
[dd_idx
].written
== NULL
)
2166 bip
= &sh
->dev
[dd_idx
].toread
;
2167 while (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
) {
2168 if ((*bip
)->bi_sector
+ ((*bip
)->bi_size
>> 9) > bi
->bi_sector
)
2170 bip
= & (*bip
)->bi_next
;
2172 if (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
+ ((bi
->bi_size
)>>9))
2175 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
2179 bi
->bi_phys_segments
++;
2180 spin_unlock_irq(&conf
->device_lock
);
2181 spin_unlock(&sh
->lock
);
2183 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2184 (unsigned long long)bi
->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
);
2195 /* check if page is covered */
2196 sector_t sector
= sh
->dev
[dd_idx
].sector
;
2197 for (bi
=sh
->dev
[dd_idx
].towrite
;
2198 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
2199 bi
&& bi
->bi_sector
<= sector
;
2200 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
2201 if (bi
->bi_sector
+ (bi
->bi_size
>>9) >= sector
)
2202 sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
2204 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
2205 set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
);
2210 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
2211 spin_unlock_irq(&conf
->device_lock
);
2212 spin_unlock(&sh
->lock
);
2216 static void end_reshape(raid5_conf_t
*conf
);
2218 static void stripe_set_idx(sector_t stripe
, raid5_conf_t
*conf
, int previous
,
2219 struct stripe_head
*sh
)
2221 int sectors_per_chunk
=
2222 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
2224 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
2225 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
2227 raid5_compute_sector(conf
,
2228 stripe
* (disks
- conf
->max_degraded
)
2229 *sectors_per_chunk
+ chunk_offset
,
2235 handle_failed_stripe(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2236 struct stripe_head_state
*s
, int disks
,
2237 struct bio
**return_bi
)
2240 for (i
= disks
; i
--; ) {
2244 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2247 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2248 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
2249 /* multiple read failures in one stripe */
2250 md_error(conf
->mddev
, rdev
);
2253 spin_lock_irq(&conf
->device_lock
);
2254 /* fail all writes first */
2255 bi
= sh
->dev
[i
].towrite
;
2256 sh
->dev
[i
].towrite
= NULL
;
2262 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2263 wake_up(&conf
->wait_for_overlap
);
2265 while (bi
&& bi
->bi_sector
<
2266 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2267 struct bio
*nextbi
= 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
;
2276 /* and fail all 'written' */
2277 bi
= sh
->dev
[i
].written
;
2278 sh
->dev
[i
].written
= NULL
;
2279 if (bi
) bitmap_end
= 1;
2280 while (bi
&& bi
->bi_sector
<
2281 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2282 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2283 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2284 if (!raid5_dec_bi_phys_segments(bi
)) {
2285 md_write_end(conf
->mddev
);
2286 bi
->bi_next
= *return_bi
;
2292 /* fail any reads if this device is non-operational and
2293 * the data has not reached the cache yet.
2295 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
2296 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
2297 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
2298 bi
= sh
->dev
[i
].toread
;
2299 sh
->dev
[i
].toread
= NULL
;
2300 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2301 wake_up(&conf
->wait_for_overlap
);
2302 if (bi
) s
->to_read
--;
2303 while (bi
&& bi
->bi_sector
<
2304 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2305 struct bio
*nextbi
=
2306 r5_next_bio(bi
, sh
->dev
[i
].sector
);
2307 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2308 if (!raid5_dec_bi_phys_segments(bi
)) {
2309 bi
->bi_next
= *return_bi
;
2315 spin_unlock_irq(&conf
->device_lock
);
2317 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2318 STRIPE_SECTORS
, 0, 0);
2321 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2322 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2323 md_wakeup_thread(conf
->mddev
->thread
);
2326 /* fetch_block5 - checks the given member device to see if its data needs
2327 * to be read or computed to satisfy a request.
2329 * Returns 1 when no more member devices need to be checked, otherwise returns
2330 * 0 to tell the loop in handle_stripe_fill5 to continue
2332 static int fetch_block5(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2333 int disk_idx
, int disks
)
2335 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
2336 struct r5dev
*failed_dev
= &sh
->dev
[s
->failed_num
];
2338 /* is the data in this block needed, and can we get it? */
2339 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2340 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2342 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2343 s
->syncing
|| s
->expanding
||
2345 (failed_dev
->toread
||
2346 (failed_dev
->towrite
&&
2347 !test_bit(R5_OVERWRITE
, &failed_dev
->flags
)))))) {
2348 /* We would like to get this block, possibly by computing it,
2349 * otherwise read it if the backing disk is insync
2351 if ((s
->uptodate
== disks
- 1) &&
2352 (s
->failed
&& disk_idx
== s
->failed_num
)) {
2353 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2354 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2355 set_bit(R5_Wantcompute
, &dev
->flags
);
2356 sh
->ops
.target
= disk_idx
;
2357 sh
->ops
.target2
= -1;
2359 /* Careful: from this point on 'uptodate' is in the eye
2360 * of raid_run_ops which services 'compute' operations
2361 * before writes. R5_Wantcompute flags a block that will
2362 * be R5_UPTODATE by the time it is needed for a
2363 * subsequent operation.
2366 return 1; /* uptodate + compute == disks */
2367 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2368 set_bit(R5_LOCKED
, &dev
->flags
);
2369 set_bit(R5_Wantread
, &dev
->flags
);
2371 pr_debug("Reading block %d (sync=%d)\n", disk_idx
,
2380 * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2382 static void handle_stripe_fill5(struct stripe_head
*sh
,
2383 struct stripe_head_state
*s
, int disks
)
2387 /* look for blocks to read/compute, skip this if a compute
2388 * is already in flight, or if the stripe contents are in the
2389 * midst of changing due to a write
2391 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2392 !sh
->reconstruct_state
)
2393 for (i
= disks
; i
--; )
2394 if (fetch_block5(sh
, s
, i
, disks
))
2396 set_bit(STRIPE_HANDLE
, &sh
->state
);
2399 /* fetch_block6 - checks the given member device to see if its data needs
2400 * to be read or computed to satisfy a request.
2402 * Returns 1 when no more member devices need to be checked, otherwise returns
2403 * 0 to tell the loop in handle_stripe_fill6 to continue
2405 static int fetch_block6(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2406 struct r6_state
*r6s
, int disk_idx
, int disks
)
2408 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
2409 struct r5dev
*fdev
[2] = { &sh
->dev
[r6s
->failed_num
[0]],
2410 &sh
->dev
[r6s
->failed_num
[1]] };
2412 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2413 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2415 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2416 s
->syncing
|| s
->expanding
||
2418 (fdev
[0]->toread
|| s
->to_write
)) ||
2420 (fdev
[1]->toread
|| s
->to_write
)))) {
2421 /* we would like to get this block, possibly by computing it,
2422 * otherwise read it if the backing disk is insync
2424 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
2425 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
2426 if ((s
->uptodate
== disks
- 1) &&
2427 (s
->failed
&& (disk_idx
== r6s
->failed_num
[0] ||
2428 disk_idx
== r6s
->failed_num
[1]))) {
2429 /* have disk failed, and we're requested to fetch it;
2432 pr_debug("Computing stripe %llu block %d\n",
2433 (unsigned long long)sh
->sector
, disk_idx
);
2434 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2435 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2436 set_bit(R5_Wantcompute
, &dev
->flags
);
2437 sh
->ops
.target
= disk_idx
;
2438 sh
->ops
.target2
= -1; /* no 2nd target */
2442 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
2443 /* Computing 2-failure is *very* expensive; only
2444 * do it if failed >= 2
2447 for (other
= disks
; other
--; ) {
2448 if (other
== disk_idx
)
2450 if (!test_bit(R5_UPTODATE
,
2451 &sh
->dev
[other
].flags
))
2455 pr_debug("Computing stripe %llu blocks %d,%d\n",
2456 (unsigned long long)sh
->sector
,
2458 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2459 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2460 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
2461 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
2462 sh
->ops
.target
= disk_idx
;
2463 sh
->ops
.target2
= other
;
2467 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2468 set_bit(R5_LOCKED
, &dev
->flags
);
2469 set_bit(R5_Wantread
, &dev
->flags
);
2471 pr_debug("Reading block %d (sync=%d)\n",
2472 disk_idx
, s
->syncing
);
2480 * handle_stripe_fill6 - read or compute data to satisfy pending requests.
2482 static void handle_stripe_fill6(struct stripe_head
*sh
,
2483 struct stripe_head_state
*s
, struct r6_state
*r6s
,
2488 /* look for blocks to read/compute, skip this if a compute
2489 * is already in flight, or if the stripe contents are in the
2490 * midst of changing due to a write
2492 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2493 !sh
->reconstruct_state
)
2494 for (i
= disks
; i
--; )
2495 if (fetch_block6(sh
, s
, r6s
, i
, disks
))
2497 set_bit(STRIPE_HANDLE
, &sh
->state
);
2501 /* handle_stripe_clean_event
2502 * any written block on an uptodate or failed drive can be returned.
2503 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2504 * never LOCKED, so we don't need to test 'failed' directly.
2506 static void handle_stripe_clean_event(raid5_conf_t
*conf
,
2507 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
2512 for (i
= disks
; i
--; )
2513 if (sh
->dev
[i
].written
) {
2515 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2516 test_bit(R5_UPTODATE
, &dev
->flags
)) {
2517 /* We can return any write requests */
2518 struct bio
*wbi
, *wbi2
;
2520 pr_debug("Return write for disc %d\n", i
);
2521 spin_lock_irq(&conf
->device_lock
);
2523 dev
->written
= NULL
;
2524 while (wbi
&& wbi
->bi_sector
<
2525 dev
->sector
+ STRIPE_SECTORS
) {
2526 wbi2
= r5_next_bio(wbi
, dev
->sector
);
2527 if (!raid5_dec_bi_phys_segments(wbi
)) {
2528 md_write_end(conf
->mddev
);
2529 wbi
->bi_next
= *return_bi
;
2534 if (dev
->towrite
== NULL
)
2536 spin_unlock_irq(&conf
->device_lock
);
2538 bitmap_endwrite(conf
->mddev
->bitmap
,
2541 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
2546 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2547 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2548 md_wakeup_thread(conf
->mddev
->thread
);
2551 static void handle_stripe_dirtying5(raid5_conf_t
*conf
,
2552 struct stripe_head
*sh
, struct stripe_head_state
*s
, int disks
)
2554 int rmw
= 0, rcw
= 0, i
;
2555 for (i
= disks
; i
--; ) {
2556 /* would I have to read this buffer for read_modify_write */
2557 struct r5dev
*dev
= &sh
->dev
[i
];
2558 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2559 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2560 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2561 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2562 if (test_bit(R5_Insync
, &dev
->flags
))
2565 rmw
+= 2*disks
; /* cannot read it */
2567 /* Would I have to read this buffer for reconstruct_write */
2568 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) && i
!= sh
->pd_idx
&&
2569 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2570 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2571 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2572 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2577 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2578 (unsigned long long)sh
->sector
, rmw
, rcw
);
2579 set_bit(STRIPE_HANDLE
, &sh
->state
);
2580 if (rmw
< rcw
&& rmw
> 0)
2581 /* prefer read-modify-write, but need to get some data */
2582 for (i
= disks
; i
--; ) {
2583 struct r5dev
*dev
= &sh
->dev
[i
];
2584 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2585 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2586 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2587 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2588 test_bit(R5_Insync
, &dev
->flags
)) {
2590 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2591 pr_debug("Read_old block "
2592 "%d for r-m-w\n", i
);
2593 set_bit(R5_LOCKED
, &dev
->flags
);
2594 set_bit(R5_Wantread
, &dev
->flags
);
2597 set_bit(STRIPE_DELAYED
, &sh
->state
);
2598 set_bit(STRIPE_HANDLE
, &sh
->state
);
2602 if (rcw
<= rmw
&& rcw
> 0)
2603 /* want reconstruct write, but need to get some data */
2604 for (i
= disks
; i
--; ) {
2605 struct r5dev
*dev
= &sh
->dev
[i
];
2606 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2608 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2609 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2610 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2611 test_bit(R5_Insync
, &dev
->flags
)) {
2613 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2614 pr_debug("Read_old block "
2615 "%d for Reconstruct\n", i
);
2616 set_bit(R5_LOCKED
, &dev
->flags
);
2617 set_bit(R5_Wantread
, &dev
->flags
);
2620 set_bit(STRIPE_DELAYED
, &sh
->state
);
2621 set_bit(STRIPE_HANDLE
, &sh
->state
);
2625 /* now if nothing is locked, and if we have enough data,
2626 * we can start a write request
2628 /* since handle_stripe can be called at any time we need to handle the
2629 * case where a compute block operation has been submitted and then a
2630 * subsequent call wants to start a write request. raid_run_ops only
2631 * handles the case where compute block and reconstruct are requested
2632 * simultaneously. If this is not the case then new writes need to be
2633 * held off until the compute completes.
2635 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
2636 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
2637 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
2638 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
2641 static void handle_stripe_dirtying6(raid5_conf_t
*conf
,
2642 struct stripe_head
*sh
, struct stripe_head_state
*s
,
2643 struct r6_state
*r6s
, int disks
)
2645 int rcw
= 0, pd_idx
= sh
->pd_idx
, i
;
2646 int qd_idx
= sh
->qd_idx
;
2648 set_bit(STRIPE_HANDLE
, &sh
->state
);
2649 for (i
= disks
; i
--; ) {
2650 struct r5dev
*dev
= &sh
->dev
[i
];
2651 /* check if we haven't enough data */
2652 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2653 i
!= pd_idx
&& i
!= qd_idx
&&
2654 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2655 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2656 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2658 if (!test_bit(R5_Insync
, &dev
->flags
))
2659 continue; /* it's a failed drive */
2662 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2663 pr_debug("Read_old stripe %llu "
2664 "block %d for Reconstruct\n",
2665 (unsigned long long)sh
->sector
, i
);
2666 set_bit(R5_LOCKED
, &dev
->flags
);
2667 set_bit(R5_Wantread
, &dev
->flags
);
2670 pr_debug("Request delayed stripe %llu "
2671 "block %d for Reconstruct\n",
2672 (unsigned long long)sh
->sector
, i
);
2673 set_bit(STRIPE_DELAYED
, &sh
->state
);
2674 set_bit(STRIPE_HANDLE
, &sh
->state
);
2678 /* now if nothing is locked, and if we have enough data, we can start a
2681 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
2682 s
->locked
== 0 && rcw
== 0 &&
2683 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)) {
2684 schedule_reconstruction(sh
, s
, 1, 0);
2688 static void handle_parity_checks5(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2689 struct stripe_head_state
*s
, int disks
)
2691 struct r5dev
*dev
= NULL
;
2693 set_bit(STRIPE_HANDLE
, &sh
->state
);
2695 switch (sh
->check_state
) {
2696 case check_state_idle
:
2697 /* start a new check operation if there are no failures */
2698 if (s
->failed
== 0) {
2699 BUG_ON(s
->uptodate
!= disks
);
2700 sh
->check_state
= check_state_run
;
2701 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2702 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2706 dev
= &sh
->dev
[s
->failed_num
];
2708 case check_state_compute_result
:
2709 sh
->check_state
= check_state_idle
;
2711 dev
= &sh
->dev
[sh
->pd_idx
];
2713 /* check that a write has not made the stripe insync */
2714 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2717 /* either failed parity check, or recovery is happening */
2718 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
2719 BUG_ON(s
->uptodate
!= disks
);
2721 set_bit(R5_LOCKED
, &dev
->flags
);
2723 set_bit(R5_Wantwrite
, &dev
->flags
);
2725 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2726 set_bit(STRIPE_INSYNC
, &sh
->state
);
2728 case check_state_run
:
2729 break; /* we will be called again upon completion */
2730 case check_state_check_result
:
2731 sh
->check_state
= check_state_idle
;
2733 /* if a failure occurred during the check operation, leave
2734 * STRIPE_INSYNC not set and let the stripe be handled again
2739 /* handle a successful check operation, if parity is correct
2740 * we are done. Otherwise update the mismatch count and repair
2741 * parity if !MD_RECOVERY_CHECK
2743 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
2744 /* parity is correct (on disc,
2745 * not in buffer any more)
2747 set_bit(STRIPE_INSYNC
, &sh
->state
);
2749 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2750 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2751 /* don't try to repair!! */
2752 set_bit(STRIPE_INSYNC
, &sh
->state
);
2754 sh
->check_state
= check_state_compute_run
;
2755 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2756 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2757 set_bit(R5_Wantcompute
,
2758 &sh
->dev
[sh
->pd_idx
].flags
);
2759 sh
->ops
.target
= sh
->pd_idx
;
2760 sh
->ops
.target2
= -1;
2765 case check_state_compute_run
:
2768 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
2769 __func__
, sh
->check_state
,
2770 (unsigned long long) sh
->sector
);
2776 static void handle_parity_checks6(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2777 struct stripe_head_state
*s
,
2778 struct r6_state
*r6s
, int disks
)
2780 int pd_idx
= sh
->pd_idx
;
2781 int qd_idx
= sh
->qd_idx
;
2784 set_bit(STRIPE_HANDLE
, &sh
->state
);
2786 BUG_ON(s
->failed
> 2);
2788 /* Want to check and possibly repair P and Q.
2789 * However there could be one 'failed' device, in which
2790 * case we can only check one of them, possibly using the
2791 * other to generate missing data
2794 switch (sh
->check_state
) {
2795 case check_state_idle
:
2796 /* start a new check operation if there are < 2 failures */
2797 if (s
->failed
== r6s
->q_failed
) {
2798 /* The only possible failed device holds Q, so it
2799 * makes sense to check P (If anything else were failed,
2800 * we would have used P to recreate it).
2802 sh
->check_state
= check_state_run
;
2804 if (!r6s
->q_failed
&& s
->failed
< 2) {
2805 /* Q is not failed, and we didn't use it to generate
2806 * anything, so it makes sense to check it
2808 if (sh
->check_state
== check_state_run
)
2809 sh
->check_state
= check_state_run_pq
;
2811 sh
->check_state
= check_state_run_q
;
2814 /* discard potentially stale zero_sum_result */
2815 sh
->ops
.zero_sum_result
= 0;
2817 if (sh
->check_state
== check_state_run
) {
2818 /* async_xor_zero_sum destroys the contents of P */
2819 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2822 if (sh
->check_state
>= check_state_run
&&
2823 sh
->check_state
<= check_state_run_pq
) {
2824 /* async_syndrome_zero_sum preserves P and Q, so
2825 * no need to mark them !uptodate here
2827 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2831 /* we have 2-disk failure */
2832 BUG_ON(s
->failed
!= 2);
2834 case check_state_compute_result
:
2835 sh
->check_state
= check_state_idle
;
2837 /* check that a write has not made the stripe insync */
2838 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2841 /* now write out any block on a failed drive,
2842 * or P or Q if they were recomputed
2844 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
2845 if (s
->failed
== 2) {
2846 dev
= &sh
->dev
[r6s
->failed_num
[1]];
2848 set_bit(R5_LOCKED
, &dev
->flags
);
2849 set_bit(R5_Wantwrite
, &dev
->flags
);
2851 if (s
->failed
>= 1) {
2852 dev
= &sh
->dev
[r6s
->failed_num
[0]];
2854 set_bit(R5_LOCKED
, &dev
->flags
);
2855 set_bit(R5_Wantwrite
, &dev
->flags
);
2857 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
2858 dev
= &sh
->dev
[pd_idx
];
2860 set_bit(R5_LOCKED
, &dev
->flags
);
2861 set_bit(R5_Wantwrite
, &dev
->flags
);
2863 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
2864 dev
= &sh
->dev
[qd_idx
];
2866 set_bit(R5_LOCKED
, &dev
->flags
);
2867 set_bit(R5_Wantwrite
, &dev
->flags
);
2869 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2871 set_bit(STRIPE_INSYNC
, &sh
->state
);
2873 case check_state_run
:
2874 case check_state_run_q
:
2875 case check_state_run_pq
:
2876 break; /* we will be called again upon completion */
2877 case check_state_check_result
:
2878 sh
->check_state
= check_state_idle
;
2880 /* handle a successful check operation, if parity is correct
2881 * we are done. Otherwise update the mismatch count and repair
2882 * parity if !MD_RECOVERY_CHECK
2884 if (sh
->ops
.zero_sum_result
== 0) {
2885 /* both parities are correct */
2887 set_bit(STRIPE_INSYNC
, &sh
->state
);
2889 /* in contrast to the raid5 case we can validate
2890 * parity, but still have a failure to write
2893 sh
->check_state
= check_state_compute_result
;
2894 /* Returning at this point means that we may go
2895 * off and bring p and/or q uptodate again so
2896 * we make sure to check zero_sum_result again
2897 * to verify if p or q need writeback
2901 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2902 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2903 /* don't try to repair!! */
2904 set_bit(STRIPE_INSYNC
, &sh
->state
);
2906 int *target
= &sh
->ops
.target
;
2908 sh
->ops
.target
= -1;
2909 sh
->ops
.target2
= -1;
2910 sh
->check_state
= check_state_compute_run
;
2911 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2912 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2913 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
2914 set_bit(R5_Wantcompute
,
2915 &sh
->dev
[pd_idx
].flags
);
2917 target
= &sh
->ops
.target2
;
2920 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
2921 set_bit(R5_Wantcompute
,
2922 &sh
->dev
[qd_idx
].flags
);
2929 case check_state_compute_run
:
2932 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
2933 __func__
, sh
->check_state
,
2934 (unsigned long long) sh
->sector
);
2939 static void handle_stripe_expansion(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2940 struct r6_state
*r6s
)
2944 /* We have read all the blocks in this stripe and now we need to
2945 * copy some of them into a target stripe for expand.
2947 struct dma_async_tx_descriptor
*tx
= NULL
;
2948 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
2949 for (i
= 0; i
< sh
->disks
; i
++)
2950 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
2952 struct stripe_head
*sh2
;
2953 struct async_submit_ctl submit
;
2955 sector_t bn
= compute_blocknr(sh
, i
, 1);
2956 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
2958 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
2960 /* so far only the early blocks of this stripe
2961 * have been requested. When later blocks
2962 * get requested, we will try again
2965 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
2966 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
2967 /* must have already done this block */
2968 release_stripe(sh2
);
2972 /* place all the copies on one channel */
2973 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
2974 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
2975 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
2978 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
2979 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
2980 for (j
= 0; j
< conf
->raid_disks
; j
++)
2981 if (j
!= sh2
->pd_idx
&&
2982 (!r6s
|| j
!= sh2
->qd_idx
) &&
2983 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
2985 if (j
== conf
->raid_disks
) {
2986 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
2987 set_bit(STRIPE_HANDLE
, &sh2
->state
);
2989 release_stripe(sh2
);
2992 /* done submitting copies, wait for them to complete */
2995 dma_wait_for_async_tx(tx
);
3001 * handle_stripe - do things to a stripe.
3003 * We lock the stripe and then examine the state of various bits
3004 * to see what needs to be done.
3006 * return some read request which now have data
3007 * return some write requests which are safely on disc
3008 * schedule a read on some buffers
3009 * schedule a write of some buffers
3010 * return confirmation of parity correctness
3012 * buffers are taken off read_list or write_list, and bh_cache buffers
3013 * get BH_Lock set before the stripe lock is released.
3017 static void handle_stripe5(struct stripe_head
*sh
)
3019 raid5_conf_t
*conf
= sh
->raid_conf
;
3020 int disks
= sh
->disks
, i
;
3021 struct bio
*return_bi
= NULL
;
3022 struct stripe_head_state s
;
3024 mdk_rdev_t
*blocked_rdev
= NULL
;
3026 int dec_preread_active
= 0;
3028 memset(&s
, 0, sizeof(s
));
3029 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
3030 "reconstruct:%d\n", (unsigned long long)sh
->sector
, sh
->state
,
3031 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->check_state
,
3032 sh
->reconstruct_state
);
3034 spin_lock(&sh
->lock
);
3035 clear_bit(STRIPE_HANDLE
, &sh
->state
);
3036 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3038 s
.syncing
= test_bit(STRIPE_SYNCING
, &sh
->state
);
3039 s
.expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3040 s
.expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3042 /* Now to look around and see what can be done */
3044 for (i
=disks
; i
--; ) {
3049 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
3050 "written %p\n", i
, dev
->flags
, dev
->toread
, dev
->read
,
3051 dev
->towrite
, dev
->written
);
3053 /* maybe we can request a biofill operation
3055 * new wantfill requests are only permitted while
3056 * ops_complete_biofill is guaranteed to be inactive
3058 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
3059 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
3060 set_bit(R5_Wantfill
, &dev
->flags
);
3062 /* now count some things */
3063 if (test_bit(R5_LOCKED
, &dev
->flags
)) s
.locked
++;
3064 if (test_bit(R5_UPTODATE
, &dev
->flags
)) s
.uptodate
++;
3065 if (test_bit(R5_Wantcompute
, &dev
->flags
)) s
.compute
++;
3067 if (test_bit(R5_Wantfill
, &dev
->flags
))
3069 else if (dev
->toread
)
3073 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3078 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3079 if (blocked_rdev
== NULL
&&
3080 rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
3081 blocked_rdev
= rdev
;
3082 atomic_inc(&rdev
->nr_pending
);
3084 clear_bit(R5_Insync
, &dev
->flags
);
3087 else if (test_bit(In_sync
, &rdev
->flags
))
3088 set_bit(R5_Insync
, &dev
->flags
);
3090 /* could be in-sync depending on recovery/reshape status */
3091 if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
3092 set_bit(R5_Insync
, &dev
->flags
);
3094 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3095 /* The ReadError flag will just be confusing now */
3096 clear_bit(R5_ReadError
, &dev
->flags
);
3097 clear_bit(R5_ReWrite
, &dev
->flags
);
3099 if (test_bit(R5_ReadError
, &dev
->flags
))
3100 clear_bit(R5_Insync
, &dev
->flags
);
3101 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3108 if (unlikely(blocked_rdev
)) {
3109 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
3110 s
.to_write
|| s
.written
) {
3111 set_bit(STRIPE_HANDLE
, &sh
->state
);
3114 /* There is nothing for the blocked_rdev to block */
3115 rdev_dec_pending(blocked_rdev
, conf
->mddev
);
3116 blocked_rdev
= NULL
;
3119 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
3120 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
3121 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
3124 pr_debug("locked=%d uptodate=%d to_read=%d"
3125 " to_write=%d failed=%d failed_num=%d\n",
3126 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
,
3127 s
.failed
, s
.failed_num
);
3128 /* check if the array has lost two devices and, if so, some requests might
3131 if (s
.failed
> 1 && s
.to_read
+s
.to_write
+s
.written
)
3132 handle_failed_stripe(conf
, sh
, &s
, disks
, &return_bi
);
3133 if (s
.failed
> 1 && s
.syncing
) {
3134 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,0);
3135 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3139 /* might be able to return some write requests if the parity block
3140 * is safe, or on a failed drive
3142 dev
= &sh
->dev
[sh
->pd_idx
];
3144 ((test_bit(R5_Insync
, &dev
->flags
) &&
3145 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3146 test_bit(R5_UPTODATE
, &dev
->flags
)) ||
3147 (s
.failed
== 1 && s
.failed_num
== sh
->pd_idx
)))
3148 handle_stripe_clean_event(conf
, sh
, disks
, &return_bi
);
3150 /* Now we might consider reading some blocks, either to check/generate
3151 * parity, or to satisfy requests
3152 * or to load a block that is being partially written.
3154 if (s
.to_read
|| s
.non_overwrite
||
3155 (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
)) || s
.expanding
)
3156 handle_stripe_fill5(sh
, &s
, disks
);
3158 /* Now we check to see if any write operations have recently
3162 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
3164 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
3165 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
3166 sh
->reconstruct_state
= reconstruct_state_idle
;
3168 /* All the 'written' buffers and the parity block are ready to
3169 * be written back to disk
3171 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
));
3172 for (i
= disks
; i
--; ) {
3174 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
3175 (i
== sh
->pd_idx
|| dev
->written
)) {
3176 pr_debug("Writing block %d\n", i
);
3177 set_bit(R5_Wantwrite
, &dev
->flags
);
3180 if (!test_bit(R5_Insync
, &dev
->flags
) ||
3181 (i
== sh
->pd_idx
&& s
.failed
== 0))
3182 set_bit(STRIPE_INSYNC
, &sh
->state
);
3185 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3186 dec_preread_active
= 1;
3189 /* Now to consider new write requests and what else, if anything
3190 * should be read. We do not handle new writes when:
3191 * 1/ A 'write' operation (copy+xor) is already in flight.
3192 * 2/ A 'check' operation is in flight, as it may clobber the parity
3195 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
3196 handle_stripe_dirtying5(conf
, sh
, &s
, disks
);
3198 /* maybe we need to check and possibly fix the parity for this stripe
3199 * Any reads will already have been scheduled, so we just see if enough
3200 * data is available. The parity check is held off while parity
3201 * dependent operations are in flight.
3203 if (sh
->check_state
||
3204 (s
.syncing
&& s
.locked
== 0 &&
3205 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
3206 !test_bit(STRIPE_INSYNC
, &sh
->state
)))
3207 handle_parity_checks5(conf
, sh
, &s
, disks
);
3209 if (s
.syncing
&& s
.locked
== 0 && test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3210 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,1);
3211 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3214 /* If the failed drive is just a ReadError, then we might need to progress
3215 * the repair/check process
3217 if (s
.failed
== 1 && !conf
->mddev
->ro
&&
3218 test_bit(R5_ReadError
, &sh
->dev
[s
.failed_num
].flags
)
3219 && !test_bit(R5_LOCKED
, &sh
->dev
[s
.failed_num
].flags
)
3220 && test_bit(R5_UPTODATE
, &sh
->dev
[s
.failed_num
].flags
)
3222 dev
= &sh
->dev
[s
.failed_num
];
3223 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3224 set_bit(R5_Wantwrite
, &dev
->flags
);
3225 set_bit(R5_ReWrite
, &dev
->flags
);
3226 set_bit(R5_LOCKED
, &dev
->flags
);
3229 /* let's read it back */
3230 set_bit(R5_Wantread
, &dev
->flags
);
3231 set_bit(R5_LOCKED
, &dev
->flags
);
3236 /* Finish reconstruct operations initiated by the expansion process */
3237 if (sh
->reconstruct_state
== reconstruct_state_result
) {
3238 struct stripe_head
*sh2
3239 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
3240 if (sh2
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh2
->state
)) {
3241 /* sh cannot be written until sh2 has been read.
3242 * so arrange for sh to be delayed a little
3244 set_bit(STRIPE_DELAYED
, &sh
->state
);
3245 set_bit(STRIPE_HANDLE
, &sh
->state
);
3246 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
3248 atomic_inc(&conf
->preread_active_stripes
);
3249 release_stripe(sh2
);
3253 release_stripe(sh2
);
3255 sh
->reconstruct_state
= reconstruct_state_idle
;
3256 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3257 for (i
= conf
->raid_disks
; i
--; ) {
3258 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3259 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3264 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
3265 !sh
->reconstruct_state
) {
3266 /* Need to write out all blocks after computing parity */
3267 sh
->disks
= conf
->raid_disks
;
3268 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
3269 schedule_reconstruction(sh
, &s
, 1, 1);
3270 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
3271 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3272 atomic_dec(&conf
->reshape_stripes
);
3273 wake_up(&conf
->wait_for_overlap
);
3274 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3277 if (s
.expanding
&& s
.locked
== 0 &&
3278 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
3279 handle_stripe_expansion(conf
, sh
, NULL
);
3282 spin_unlock(&sh
->lock
);
3284 /* wait for this device to become unblocked */
3285 if (unlikely(blocked_rdev
))
3286 md_wait_for_blocked_rdev(blocked_rdev
, conf
->mddev
);
3289 raid_run_ops(sh
, s
.ops_request
);
3293 if (dec_preread_active
) {
3294 /* We delay this until after ops_run_io so that if make_request
3295 * is waiting on a flush, it won't continue until the writes
3296 * have actually been submitted.
3298 atomic_dec(&conf
->preread_active_stripes
);
3299 if (atomic_read(&conf
->preread_active_stripes
) <
3301 md_wakeup_thread(conf
->mddev
->thread
);
3303 return_io(return_bi
);
3306 static void handle_stripe6(struct stripe_head
*sh
)
3308 raid5_conf_t
*conf
= sh
->raid_conf
;
3309 int disks
= sh
->disks
;
3310 struct bio
*return_bi
= NULL
;
3311 int i
, pd_idx
= sh
->pd_idx
, qd_idx
= sh
->qd_idx
;
3312 struct stripe_head_state s
;
3313 struct r6_state r6s
;
3314 struct r5dev
*dev
, *pdev
, *qdev
;
3315 mdk_rdev_t
*blocked_rdev
= NULL
;
3316 int dec_preread_active
= 0;
3318 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3319 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3320 (unsigned long long)sh
->sector
, sh
->state
,
3321 atomic_read(&sh
->count
), pd_idx
, qd_idx
,
3322 sh
->check_state
, sh
->reconstruct_state
);
3323 memset(&s
, 0, sizeof(s
));
3325 spin_lock(&sh
->lock
);
3326 clear_bit(STRIPE_HANDLE
, &sh
->state
);
3327 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3329 s
.syncing
= test_bit(STRIPE_SYNCING
, &sh
->state
);
3330 s
.expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3331 s
.expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3332 /* Now to look around and see what can be done */
3335 for (i
=disks
; i
--; ) {
3339 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3340 i
, dev
->flags
, dev
->toread
, dev
->towrite
, dev
->written
);
3341 /* maybe we can reply to a read
3343 * new wantfill requests are only permitted while
3344 * ops_complete_biofill is guaranteed to be inactive
3346 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
3347 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
3348 set_bit(R5_Wantfill
, &dev
->flags
);
3350 /* now count some things */
3351 if (test_bit(R5_LOCKED
, &dev
->flags
)) s
.locked
++;
3352 if (test_bit(R5_UPTODATE
, &dev
->flags
)) s
.uptodate
++;
3353 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
3355 BUG_ON(s
.compute
> 2);
3358 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
3360 } else if (dev
->toread
)
3364 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3369 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3370 if (blocked_rdev
== NULL
&&
3371 rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
3372 blocked_rdev
= rdev
;
3373 atomic_inc(&rdev
->nr_pending
);
3375 clear_bit(R5_Insync
, &dev
->flags
);
3378 else if (test_bit(In_sync
, &rdev
->flags
))
3379 set_bit(R5_Insync
, &dev
->flags
);
3381 /* in sync if before recovery_offset */
3382 if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
3383 set_bit(R5_Insync
, &dev
->flags
);
3385 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3386 /* The ReadError flag will just be confusing now */
3387 clear_bit(R5_ReadError
, &dev
->flags
);
3388 clear_bit(R5_ReWrite
, &dev
->flags
);
3390 if (test_bit(R5_ReadError
, &dev
->flags
))
3391 clear_bit(R5_Insync
, &dev
->flags
);
3392 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3394 r6s
.failed_num
[s
.failed
] = i
;
3400 if (unlikely(blocked_rdev
)) {
3401 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
3402 s
.to_write
|| s
.written
) {
3403 set_bit(STRIPE_HANDLE
, &sh
->state
);
3406 /* There is nothing for the blocked_rdev to block */
3407 rdev_dec_pending(blocked_rdev
, conf
->mddev
);
3408 blocked_rdev
= NULL
;
3411 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
3412 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
3413 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
3416 pr_debug("locked=%d uptodate=%d to_read=%d"
3417 " to_write=%d failed=%d failed_num=%d,%d\n",
3418 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
3419 r6s
.failed_num
[0], r6s
.failed_num
[1]);
3420 /* check if the array has lost >2 devices and, if so, some requests
3421 * might need to be failed
3423 if (s
.failed
> 2 && s
.to_read
+s
.to_write
+s
.written
)
3424 handle_failed_stripe(conf
, sh
, &s
, disks
, &return_bi
);
3425 if (s
.failed
> 2 && s
.syncing
) {
3426 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,0);
3427 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3432 * might be able to return some write requests if the parity blocks
3433 * are safe, or on a failed drive
3435 pdev
= &sh
->dev
[pd_idx
];
3436 r6s
.p_failed
= (s
.failed
>= 1 && r6s
.failed_num
[0] == pd_idx
)
3437 || (s
.failed
>= 2 && r6s
.failed_num
[1] == pd_idx
);
3438 qdev
= &sh
->dev
[qd_idx
];
3439 r6s
.q_failed
= (s
.failed
>= 1 && r6s
.failed_num
[0] == qd_idx
)
3440 || (s
.failed
>= 2 && r6s
.failed_num
[1] == qd_idx
);
3443 ( r6s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
3444 && !test_bit(R5_LOCKED
, &pdev
->flags
)
3445 && test_bit(R5_UPTODATE
, &pdev
->flags
)))) &&
3446 ( r6s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
3447 && !test_bit(R5_LOCKED
, &qdev
->flags
)
3448 && test_bit(R5_UPTODATE
, &qdev
->flags
)))))
3449 handle_stripe_clean_event(conf
, sh
, disks
, &return_bi
);
3451 /* Now we might consider reading some blocks, either to check/generate
3452 * parity, or to satisfy requests
3453 * or to load a block that is being partially written.
3455 if (s
.to_read
|| s
.non_overwrite
|| (s
.to_write
&& s
.failed
) ||
3456 (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
)) || s
.expanding
)
3457 handle_stripe_fill6(sh
, &s
, &r6s
, disks
);
3459 /* Now we check to see if any write operations have recently
3462 if (sh
->reconstruct_state
== reconstruct_state_drain_result
) {
3464 sh
->reconstruct_state
= reconstruct_state_idle
;
3465 /* All the 'written' buffers and the parity blocks are ready to
3466 * be written back to disk
3468 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
));
3469 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
));
3470 for (i
= disks
; i
--; ) {
3472 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
3473 (i
== sh
->pd_idx
|| i
== qd_idx
||
3475 pr_debug("Writing block %d\n", i
);
3476 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
3477 set_bit(R5_Wantwrite
, &dev
->flags
);
3478 if (!test_bit(R5_Insync
, &dev
->flags
) ||
3479 ((i
== sh
->pd_idx
|| i
== qd_idx
) &&
3481 set_bit(STRIPE_INSYNC
, &sh
->state
);
3484 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3485 dec_preread_active
= 1;
3488 /* Now to consider new write requests and what else, if anything
3489 * should be read. We do not handle new writes when:
3490 * 1/ A 'write' operation (copy+gen_syndrome) is already in flight.
3491 * 2/ A 'check' operation is in flight, as it may clobber the parity
3494 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
3495 handle_stripe_dirtying6(conf
, sh
, &s
, &r6s
, disks
);
3497 /* maybe we need to check and possibly fix the parity for this stripe
3498 * Any reads will already have been scheduled, so we just see if enough
3499 * data is available. The parity check is held off while parity
3500 * dependent operations are in flight.
3502 if (sh
->check_state
||
3503 (s
.syncing
&& s
.locked
== 0 &&
3504 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
3505 !test_bit(STRIPE_INSYNC
, &sh
->state
)))
3506 handle_parity_checks6(conf
, sh
, &s
, &r6s
, disks
);
3508 if (s
.syncing
&& s
.locked
== 0 && test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3509 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,1);
3510 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3513 /* If the failed drives are just a ReadError, then we might need
3514 * to progress the repair/check process
3516 if (s
.failed
<= 2 && !conf
->mddev
->ro
)
3517 for (i
= 0; i
< s
.failed
; i
++) {
3518 dev
= &sh
->dev
[r6s
.failed_num
[i
]];
3519 if (test_bit(R5_ReadError
, &dev
->flags
)
3520 && !test_bit(R5_LOCKED
, &dev
->flags
)
3521 && test_bit(R5_UPTODATE
, &dev
->flags
)
3523 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3524 set_bit(R5_Wantwrite
, &dev
->flags
);
3525 set_bit(R5_ReWrite
, &dev
->flags
);
3526 set_bit(R5_LOCKED
, &dev
->flags
);
3529 /* let's read it back */
3530 set_bit(R5_Wantread
, &dev
->flags
);
3531 set_bit(R5_LOCKED
, &dev
->flags
);
3537 /* Finish reconstruct operations initiated by the expansion process */
3538 if (sh
->reconstruct_state
== reconstruct_state_result
) {
3539 sh
->reconstruct_state
= reconstruct_state_idle
;
3540 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3541 for (i
= conf
->raid_disks
; i
--; ) {
3542 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3543 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3548 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
3549 !sh
->reconstruct_state
) {
3550 struct stripe_head
*sh2
3551 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
3552 if (sh2
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh2
->state
)) {
3553 /* sh cannot be written until sh2 has been read.
3554 * so arrange for sh to be delayed a little
3556 set_bit(STRIPE_DELAYED
, &sh
->state
);
3557 set_bit(STRIPE_HANDLE
, &sh
->state
);
3558 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
3560 atomic_inc(&conf
->preread_active_stripes
);
3561 release_stripe(sh2
);
3565 release_stripe(sh2
);
3567 /* Need to write out all blocks after computing P&Q */
3568 sh
->disks
= conf
->raid_disks
;
3569 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
3570 schedule_reconstruction(sh
, &s
, 1, 1);
3571 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
3572 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3573 atomic_dec(&conf
->reshape_stripes
);
3574 wake_up(&conf
->wait_for_overlap
);
3575 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3578 if (s
.expanding
&& s
.locked
== 0 &&
3579 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
3580 handle_stripe_expansion(conf
, sh
, &r6s
);
3583 spin_unlock(&sh
->lock
);
3585 /* wait for this device to become unblocked */
3586 if (unlikely(blocked_rdev
))
3587 md_wait_for_blocked_rdev(blocked_rdev
, conf
->mddev
);
3590 raid_run_ops(sh
, s
.ops_request
);
3595 if (dec_preread_active
) {
3596 /* We delay this until after ops_run_io so that if make_request
3597 * is waiting on a flush, it won't continue until the writes
3598 * have actually been submitted.
3600 atomic_dec(&conf
->preread_active_stripes
);
3601 if (atomic_read(&conf
->preread_active_stripes
) <
3603 md_wakeup_thread(conf
->mddev
->thread
);
3606 return_io(return_bi
);
3609 static void handle_stripe(struct stripe_head
*sh
)
3611 if (sh
->raid_conf
->level
== 6)
3617 static void raid5_activate_delayed(raid5_conf_t
*conf
)
3619 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
3620 while (!list_empty(&conf
->delayed_list
)) {
3621 struct list_head
*l
= conf
->delayed_list
.next
;
3622 struct stripe_head
*sh
;
3623 sh
= list_entry(l
, struct stripe_head
, lru
);
3625 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3626 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3627 atomic_inc(&conf
->preread_active_stripes
);
3628 list_add_tail(&sh
->lru
, &conf
->hold_list
);
3631 plugger_set_plug(&conf
->plug
);
3634 static void activate_bit_delay(raid5_conf_t
*conf
)
3636 /* device_lock is held */
3637 struct list_head head
;
3638 list_add(&head
, &conf
->bitmap_list
);
3639 list_del_init(&conf
->bitmap_list
);
3640 while (!list_empty(&head
)) {
3641 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
3642 list_del_init(&sh
->lru
);
3643 atomic_inc(&sh
->count
);
3644 __release_stripe(conf
, sh
);
3648 static void unplug_slaves(mddev_t
*mddev
)
3650 raid5_conf_t
*conf
= mddev
->private;
3652 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
3655 for (i
= 0; i
< devs
; i
++) {
3656 mdk_rdev_t
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3657 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) && atomic_read(&rdev
->nr_pending
)) {
3658 struct request_queue
*r_queue
= bdev_get_queue(rdev
->bdev
);
3660 atomic_inc(&rdev
->nr_pending
);
3663 blk_unplug(r_queue
);
3665 rdev_dec_pending(rdev
, mddev
);
3672 void md_raid5_unplug_device(raid5_conf_t
*conf
)
3674 unsigned long flags
;
3676 spin_lock_irqsave(&conf
->device_lock
, flags
);
3678 if (plugger_remove_plug(&conf
->plug
)) {
3680 raid5_activate_delayed(conf
);
3682 md_wakeup_thread(conf
->mddev
->thread
);
3684 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3686 unplug_slaves(conf
->mddev
);
3688 EXPORT_SYMBOL_GPL(md_raid5_unplug_device
);
3690 static void raid5_unplug(struct plug_handle
*plug
)
3692 raid5_conf_t
*conf
= container_of(plug
, raid5_conf_t
, plug
);
3693 md_raid5_unplug_device(conf
);
3696 static void raid5_unplug_queue(struct request_queue
*q
)
3698 mddev_t
*mddev
= q
->queuedata
;
3699 md_raid5_unplug_device(mddev
->private);
3702 int md_raid5_congested(mddev_t
*mddev
, int bits
)
3704 raid5_conf_t
*conf
= mddev
->private;
3706 /* No difference between reads and writes. Just check
3707 * how busy the stripe_cache is
3710 if (conf
->inactive_blocked
)
3714 if (list_empty_careful(&conf
->inactive_list
))
3719 EXPORT_SYMBOL_GPL(md_raid5_congested
);
3721 static int raid5_congested(void *data
, int bits
)
3723 mddev_t
*mddev
= data
;
3725 return mddev_congested(mddev
, bits
) ||
3726 md_raid5_congested(mddev
, bits
);
3729 /* We want read requests to align with chunks where possible,
3730 * but write requests don't need to.
3732 static int raid5_mergeable_bvec(struct request_queue
*q
,
3733 struct bvec_merge_data
*bvm
,
3734 struct bio_vec
*biovec
)
3736 mddev_t
*mddev
= q
->queuedata
;
3737 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
3739 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3740 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
3742 if ((bvm
->bi_rw
& 1) == WRITE
)
3743 return biovec
->bv_len
; /* always allow writes to be mergeable */
3745 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3746 chunk_sectors
= mddev
->new_chunk_sectors
;
3747 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
3748 if (max
< 0) max
= 0;
3749 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
3750 return biovec
->bv_len
;
3756 static int in_chunk_boundary(mddev_t
*mddev
, struct bio
*bio
)
3758 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
3759 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3760 unsigned int bio_sectors
= bio
->bi_size
>> 9;
3762 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3763 chunk_sectors
= mddev
->new_chunk_sectors
;
3764 return chunk_sectors
>=
3765 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
3769 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3770 * later sampled by raid5d.
3772 static void add_bio_to_retry(struct bio
*bi
,raid5_conf_t
*conf
)
3774 unsigned long flags
;
3776 spin_lock_irqsave(&conf
->device_lock
, flags
);
3778 bi
->bi_next
= conf
->retry_read_aligned_list
;
3779 conf
->retry_read_aligned_list
= bi
;
3781 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3782 md_wakeup_thread(conf
->mddev
->thread
);
3786 static struct bio
*remove_bio_from_retry(raid5_conf_t
*conf
)
3790 bi
= conf
->retry_read_aligned
;
3792 conf
->retry_read_aligned
= NULL
;
3795 bi
= conf
->retry_read_aligned_list
;
3797 conf
->retry_read_aligned_list
= bi
->bi_next
;
3800 * this sets the active strip count to 1 and the processed
3801 * strip count to zero (upper 8 bits)
3803 bi
->bi_phys_segments
= 1; /* biased count of active stripes */
3811 * The "raid5_align_endio" should check if the read succeeded and if it
3812 * did, call bio_endio on the original bio (having bio_put the new bio
3814 * If the read failed..
3816 static void raid5_align_endio(struct bio
*bi
, int error
)
3818 struct bio
* raid_bi
= bi
->bi_private
;
3821 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3826 rdev
= (void*)raid_bi
->bi_next
;
3827 raid_bi
->bi_next
= NULL
;
3828 mddev
= rdev
->mddev
;
3829 conf
= mddev
->private;
3831 rdev_dec_pending(rdev
, conf
->mddev
);
3833 if (!error
&& uptodate
) {
3834 bio_endio(raid_bi
, 0);
3835 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
3836 wake_up(&conf
->wait_for_stripe
);
3841 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3843 add_bio_to_retry(raid_bi
, conf
);
3846 static int bio_fits_rdev(struct bio
*bi
)
3848 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
3850 if ((bi
->bi_size
>>9) > queue_max_sectors(q
))
3852 blk_recount_segments(q
, bi
);
3853 if (bi
->bi_phys_segments
> queue_max_segments(q
))
3856 if (q
->merge_bvec_fn
)
3857 /* it's too hard to apply the merge_bvec_fn at this stage,
3866 static int chunk_aligned_read(mddev_t
*mddev
, struct bio
* raid_bio
)
3868 raid5_conf_t
*conf
= mddev
->private;
3870 struct bio
* align_bi
;
3873 if (!in_chunk_boundary(mddev
, raid_bio
)) {
3874 pr_debug("chunk_aligned_read : non aligned\n");
3878 * use bio_clone_mddev to make a copy of the bio
3880 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
3884 * set bi_end_io to a new function, and set bi_private to the
3887 align_bi
->bi_end_io
= raid5_align_endio
;
3888 align_bi
->bi_private
= raid_bio
;
3892 align_bi
->bi_sector
= raid5_compute_sector(conf
, raid_bio
->bi_sector
,
3897 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
3898 if (rdev
&& test_bit(In_sync
, &rdev
->flags
)) {
3899 atomic_inc(&rdev
->nr_pending
);
3901 raid_bio
->bi_next
= (void*)rdev
;
3902 align_bi
->bi_bdev
= rdev
->bdev
;
3903 align_bi
->bi_flags
&= ~(1 << BIO_SEG_VALID
);
3904 align_bi
->bi_sector
+= rdev
->data_offset
;
3906 if (!bio_fits_rdev(align_bi
)) {
3907 /* too big in some way */
3909 rdev_dec_pending(rdev
, mddev
);
3913 spin_lock_irq(&conf
->device_lock
);
3914 wait_event_lock_irq(conf
->wait_for_stripe
,
3916 conf
->device_lock
, /* nothing */);
3917 atomic_inc(&conf
->active_aligned_reads
);
3918 spin_unlock_irq(&conf
->device_lock
);
3920 generic_make_request(align_bi
);
3929 /* __get_priority_stripe - get the next stripe to process
3931 * Full stripe writes are allowed to pass preread active stripes up until
3932 * the bypass_threshold is exceeded. In general the bypass_count
3933 * increments when the handle_list is handled before the hold_list; however, it
3934 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3935 * stripe with in flight i/o. The bypass_count will be reset when the
3936 * head of the hold_list has changed, i.e. the head was promoted to the
3939 static struct stripe_head
*__get_priority_stripe(raid5_conf_t
*conf
)
3941 struct stripe_head
*sh
;
3943 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3945 list_empty(&conf
->handle_list
) ? "empty" : "busy",
3946 list_empty(&conf
->hold_list
) ? "empty" : "busy",
3947 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
3949 if (!list_empty(&conf
->handle_list
)) {
3950 sh
= list_entry(conf
->handle_list
.next
, typeof(*sh
), lru
);
3952 if (list_empty(&conf
->hold_list
))
3953 conf
->bypass_count
= 0;
3954 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
3955 if (conf
->hold_list
.next
== conf
->last_hold
)
3956 conf
->bypass_count
++;
3958 conf
->last_hold
= conf
->hold_list
.next
;
3959 conf
->bypass_count
-= conf
->bypass_threshold
;
3960 if (conf
->bypass_count
< 0)
3961 conf
->bypass_count
= 0;
3964 } else if (!list_empty(&conf
->hold_list
) &&
3965 ((conf
->bypass_threshold
&&
3966 conf
->bypass_count
> conf
->bypass_threshold
) ||
3967 atomic_read(&conf
->pending_full_writes
) == 0)) {
3968 sh
= list_entry(conf
->hold_list
.next
,
3970 conf
->bypass_count
-= conf
->bypass_threshold
;
3971 if (conf
->bypass_count
< 0)
3972 conf
->bypass_count
= 0;
3976 list_del_init(&sh
->lru
);
3977 atomic_inc(&sh
->count
);
3978 BUG_ON(atomic_read(&sh
->count
) != 1);
3982 static int make_request(mddev_t
*mddev
, struct bio
* bi
)
3984 raid5_conf_t
*conf
= mddev
->private;
3986 sector_t new_sector
;
3987 sector_t logical_sector
, last_sector
;
3988 struct stripe_head
*sh
;
3989 const int rw
= bio_data_dir(bi
);
3992 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
3993 md_flush_request(mddev
, bi
);
3997 md_write_start(mddev
, bi
);
4000 mddev
->reshape_position
== MaxSector
&&
4001 chunk_aligned_read(mddev
,bi
))
4004 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4005 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
4007 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
4009 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
4011 int disks
, data_disks
;
4016 disks
= conf
->raid_disks
;
4017 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
4018 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
4019 /* spinlock is needed as reshape_progress may be
4020 * 64bit on a 32bit platform, and so it might be
4021 * possible to see a half-updated value
4022 * Ofcourse reshape_progress could change after
4023 * the lock is dropped, so once we get a reference
4024 * to the stripe that we think it is, we will have
4027 spin_lock_irq(&conf
->device_lock
);
4028 if (mddev
->delta_disks
< 0
4029 ? logical_sector
< conf
->reshape_progress
4030 : logical_sector
>= conf
->reshape_progress
) {
4031 disks
= conf
->previous_raid_disks
;
4034 if (mddev
->delta_disks
< 0
4035 ? logical_sector
< conf
->reshape_safe
4036 : logical_sector
>= conf
->reshape_safe
) {
4037 spin_unlock_irq(&conf
->device_lock
);
4042 spin_unlock_irq(&conf
->device_lock
);
4044 data_disks
= disks
- conf
->max_degraded
;
4046 new_sector
= raid5_compute_sector(conf
, logical_sector
,
4049 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4050 (unsigned long long)new_sector
,
4051 (unsigned long long)logical_sector
);
4053 sh
= get_active_stripe(conf
, new_sector
, previous
,
4054 (bi
->bi_rw
&RWA_MASK
), 0);
4056 if (unlikely(previous
)) {
4057 /* expansion might have moved on while waiting for a
4058 * stripe, so we must do the range check again.
4059 * Expansion could still move past after this
4060 * test, but as we are holding a reference to
4061 * 'sh', we know that if that happens,
4062 * STRIPE_EXPANDING will get set and the expansion
4063 * won't proceed until we finish with the stripe.
4066 spin_lock_irq(&conf
->device_lock
);
4067 if (mddev
->delta_disks
< 0
4068 ? logical_sector
>= conf
->reshape_progress
4069 : logical_sector
< conf
->reshape_progress
)
4070 /* mismatch, need to try again */
4072 spin_unlock_irq(&conf
->device_lock
);
4080 if (bio_data_dir(bi
) == WRITE
&&
4081 logical_sector
>= mddev
->suspend_lo
&&
4082 logical_sector
< mddev
->suspend_hi
) {
4084 /* As the suspend_* range is controlled by
4085 * userspace, we want an interruptible
4088 flush_signals(current
);
4089 prepare_to_wait(&conf
->wait_for_overlap
,
4090 &w
, TASK_INTERRUPTIBLE
);
4091 if (logical_sector
>= mddev
->suspend_lo
&&
4092 logical_sector
< mddev
->suspend_hi
)
4097 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
4098 !add_stripe_bio(sh
, bi
, dd_idx
, (bi
->bi_rw
&RW_MASK
))) {
4099 /* Stripe is busy expanding or
4100 * add failed due to overlap. Flush everything
4103 md_raid5_unplug_device(conf
);
4108 finish_wait(&conf
->wait_for_overlap
, &w
);
4109 set_bit(STRIPE_HANDLE
, &sh
->state
);
4110 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4111 if ((bi
->bi_rw
& REQ_SYNC
) &&
4112 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4113 atomic_inc(&conf
->preread_active_stripes
);
4116 /* cannot get stripe for read-ahead, just give-up */
4117 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
4118 finish_wait(&conf
->wait_for_overlap
, &w
);
4123 spin_lock_irq(&conf
->device_lock
);
4124 remaining
= raid5_dec_bi_phys_segments(bi
);
4125 spin_unlock_irq(&conf
->device_lock
);
4126 if (remaining
== 0) {
4129 md_write_end(mddev
);
4137 static sector_t
raid5_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
);
4139 static sector_t
reshape_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
)
4141 /* reshaping is quite different to recovery/resync so it is
4142 * handled quite separately ... here.
4144 * On each call to sync_request, we gather one chunk worth of
4145 * destination stripes and flag them as expanding.
4146 * Then we find all the source stripes and request reads.
4147 * As the reads complete, handle_stripe will copy the data
4148 * into the destination stripe and release that stripe.
4150 raid5_conf_t
*conf
= mddev
->private;
4151 struct stripe_head
*sh
;
4152 sector_t first_sector
, last_sector
;
4153 int raid_disks
= conf
->previous_raid_disks
;
4154 int data_disks
= raid_disks
- conf
->max_degraded
;
4155 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
4158 sector_t writepos
, readpos
, safepos
;
4159 sector_t stripe_addr
;
4160 int reshape_sectors
;
4161 struct list_head stripes
;
4163 if (sector_nr
== 0) {
4164 /* If restarting in the middle, skip the initial sectors */
4165 if (mddev
->delta_disks
< 0 &&
4166 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
4167 sector_nr
= raid5_size(mddev
, 0, 0)
4168 - conf
->reshape_progress
;
4169 } else if (mddev
->delta_disks
>= 0 &&
4170 conf
->reshape_progress
> 0)
4171 sector_nr
= conf
->reshape_progress
;
4172 sector_div(sector_nr
, new_data_disks
);
4174 mddev
->curr_resync_completed
= sector_nr
;
4175 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4181 /* We need to process a full chunk at a time.
4182 * If old and new chunk sizes differ, we need to process the
4185 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
)
4186 reshape_sectors
= mddev
->new_chunk_sectors
;
4188 reshape_sectors
= mddev
->chunk_sectors
;
4190 /* we update the metadata when there is more than 3Meg
4191 * in the block range (that is rather arbitrary, should
4192 * probably be time based) or when the data about to be
4193 * copied would over-write the source of the data at
4194 * the front of the range.
4195 * i.e. one new_stripe along from reshape_progress new_maps
4196 * to after where reshape_safe old_maps to
4198 writepos
= conf
->reshape_progress
;
4199 sector_div(writepos
, new_data_disks
);
4200 readpos
= conf
->reshape_progress
;
4201 sector_div(readpos
, data_disks
);
4202 safepos
= conf
->reshape_safe
;
4203 sector_div(safepos
, data_disks
);
4204 if (mddev
->delta_disks
< 0) {
4205 writepos
-= min_t(sector_t
, reshape_sectors
, writepos
);
4206 readpos
+= reshape_sectors
;
4207 safepos
+= reshape_sectors
;
4209 writepos
+= reshape_sectors
;
4210 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
4211 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
4214 /* 'writepos' is the most advanced device address we might write.
4215 * 'readpos' is the least advanced device address we might read.
4216 * 'safepos' is the least address recorded in the metadata as having
4218 * If 'readpos' is behind 'writepos', then there is no way that we can
4219 * ensure safety in the face of a crash - that must be done by userspace
4220 * making a backup of the data. So in that case there is no particular
4221 * rush to update metadata.
4222 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4223 * update the metadata to advance 'safepos' to match 'readpos' so that
4224 * we can be safe in the event of a crash.
4225 * So we insist on updating metadata if safepos is behind writepos and
4226 * readpos is beyond writepos.
4227 * In any case, update the metadata every 10 seconds.
4228 * Maybe that number should be configurable, but I'm not sure it is
4229 * worth it.... maybe it could be a multiple of safemode_delay???
4231 if ((mddev
->delta_disks
< 0
4232 ? (safepos
> writepos
&& readpos
< writepos
)
4233 : (safepos
< writepos
&& readpos
> writepos
)) ||
4234 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4235 /* Cannot proceed until we've updated the superblock... */
4236 wait_event(conf
->wait_for_overlap
,
4237 atomic_read(&conf
->reshape_stripes
)==0);
4238 mddev
->reshape_position
= conf
->reshape_progress
;
4239 mddev
->curr_resync_completed
= sector_nr
;
4240 conf
->reshape_checkpoint
= jiffies
;
4241 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4242 md_wakeup_thread(mddev
->thread
);
4243 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4244 kthread_should_stop());
4245 spin_lock_irq(&conf
->device_lock
);
4246 conf
->reshape_safe
= mddev
->reshape_position
;
4247 spin_unlock_irq(&conf
->device_lock
);
4248 wake_up(&conf
->wait_for_overlap
);
4249 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4252 if (mddev
->delta_disks
< 0) {
4253 BUG_ON(conf
->reshape_progress
== 0);
4254 stripe_addr
= writepos
;
4255 BUG_ON((mddev
->dev_sectors
&
4256 ~((sector_t
)reshape_sectors
- 1))
4257 - reshape_sectors
- stripe_addr
4260 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
4261 stripe_addr
= sector_nr
;
4263 INIT_LIST_HEAD(&stripes
);
4264 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
4266 int skipped_disk
= 0;
4267 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
4268 set_bit(STRIPE_EXPANDING
, &sh
->state
);
4269 atomic_inc(&conf
->reshape_stripes
);
4270 /* If any of this stripe is beyond the end of the old
4271 * array, then we need to zero those blocks
4273 for (j
=sh
->disks
; j
--;) {
4275 if (j
== sh
->pd_idx
)
4277 if (conf
->level
== 6 &&
4280 s
= compute_blocknr(sh
, j
, 0);
4281 if (s
< raid5_size(mddev
, 0, 0)) {
4285 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
4286 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
4287 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
4289 if (!skipped_disk
) {
4290 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4291 set_bit(STRIPE_HANDLE
, &sh
->state
);
4293 list_add(&sh
->lru
, &stripes
);
4295 spin_lock_irq(&conf
->device_lock
);
4296 if (mddev
->delta_disks
< 0)
4297 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
4299 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
4300 spin_unlock_irq(&conf
->device_lock
);
4301 /* Ok, those stripe are ready. We can start scheduling
4302 * reads on the source stripes.
4303 * The source stripes are determined by mapping the first and last
4304 * block on the destination stripes.
4307 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
4310 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
4311 * new_data_disks
- 1),
4313 if (last_sector
>= mddev
->dev_sectors
)
4314 last_sector
= mddev
->dev_sectors
- 1;
4315 while (first_sector
<= last_sector
) {
4316 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
4317 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
4318 set_bit(STRIPE_HANDLE
, &sh
->state
);
4320 first_sector
+= STRIPE_SECTORS
;
4322 /* Now that the sources are clearly marked, we can release
4323 * the destination stripes
4325 while (!list_empty(&stripes
)) {
4326 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
4327 list_del_init(&sh
->lru
);
4330 /* If this takes us to the resync_max point where we have to pause,
4331 * then we need to write out the superblock.
4333 sector_nr
+= reshape_sectors
;
4334 if ((sector_nr
- mddev
->curr_resync_completed
) * 2
4335 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
4336 /* Cannot proceed until we've updated the superblock... */
4337 wait_event(conf
->wait_for_overlap
,
4338 atomic_read(&conf
->reshape_stripes
) == 0);
4339 mddev
->reshape_position
= conf
->reshape_progress
;
4340 mddev
->curr_resync_completed
= sector_nr
;
4341 conf
->reshape_checkpoint
= jiffies
;
4342 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4343 md_wakeup_thread(mddev
->thread
);
4344 wait_event(mddev
->sb_wait
,
4345 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
4346 || kthread_should_stop());
4347 spin_lock_irq(&conf
->device_lock
);
4348 conf
->reshape_safe
= mddev
->reshape_position
;
4349 spin_unlock_irq(&conf
->device_lock
);
4350 wake_up(&conf
->wait_for_overlap
);
4351 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4353 return reshape_sectors
;
4356 /* FIXME go_faster isn't used */
4357 static inline sector_t
sync_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
4359 raid5_conf_t
*conf
= mddev
->private;
4360 struct stripe_head
*sh
;
4361 sector_t max_sector
= mddev
->dev_sectors
;
4362 sector_t sync_blocks
;
4363 int still_degraded
= 0;
4366 if (sector_nr
>= max_sector
) {
4367 /* just being told to finish up .. nothing much to do */
4368 unplug_slaves(mddev
);
4370 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
4375 if (mddev
->curr_resync
< max_sector
) /* aborted */
4376 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
4378 else /* completed sync */
4380 bitmap_close_sync(mddev
->bitmap
);
4385 /* Allow raid5_quiesce to complete */
4386 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
4388 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
4389 return reshape_request(mddev
, sector_nr
, skipped
);
4391 /* No need to check resync_max as we never do more than one
4392 * stripe, and as resync_max will always be on a chunk boundary,
4393 * if the check in md_do_sync didn't fire, there is no chance
4394 * of overstepping resync_max here
4397 /* if there is too many failed drives and we are trying
4398 * to resync, then assert that we are finished, because there is
4399 * nothing we can do.
4401 if (mddev
->degraded
>= conf
->max_degraded
&&
4402 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
4403 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
4407 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
4408 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
4409 !conf
->fullsync
&& sync_blocks
>= STRIPE_SECTORS
) {
4410 /* we can skip this block, and probably more */
4411 sync_blocks
/= STRIPE_SECTORS
;
4413 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
4417 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
4419 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
4421 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
4422 /* make sure we don't swamp the stripe cache if someone else
4423 * is trying to get access
4425 schedule_timeout_uninterruptible(1);
4427 /* Need to check if array will still be degraded after recovery/resync
4428 * We don't need to check the 'failed' flag as when that gets set,
4431 for (i
= 0; i
< conf
->raid_disks
; i
++)
4432 if (conf
->disks
[i
].rdev
== NULL
)
4435 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
4437 spin_lock(&sh
->lock
);
4438 set_bit(STRIPE_SYNCING
, &sh
->state
);
4439 clear_bit(STRIPE_INSYNC
, &sh
->state
);
4440 spin_unlock(&sh
->lock
);
4445 return STRIPE_SECTORS
;
4448 static int retry_aligned_read(raid5_conf_t
*conf
, struct bio
*raid_bio
)
4450 /* We may not be able to submit a whole bio at once as there
4451 * may not be enough stripe_heads available.
4452 * We cannot pre-allocate enough stripe_heads as we may need
4453 * more than exist in the cache (if we allow ever large chunks).
4454 * So we do one stripe head at a time and record in
4455 * ->bi_hw_segments how many have been done.
4457 * We *know* that this entire raid_bio is in one chunk, so
4458 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4460 struct stripe_head
*sh
;
4462 sector_t sector
, logical_sector
, last_sector
;
4467 logical_sector
= raid_bio
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4468 sector
= raid5_compute_sector(conf
, logical_sector
,
4470 last_sector
= raid_bio
->bi_sector
+ (raid_bio
->bi_size
>>9);
4472 for (; logical_sector
< last_sector
;
4473 logical_sector
+= STRIPE_SECTORS
,
4474 sector
+= STRIPE_SECTORS
,
4477 if (scnt
< raid5_bi_hw_segments(raid_bio
))
4478 /* already done this stripe */
4481 sh
= get_active_stripe(conf
, sector
, 0, 1, 0);
4484 /* failed to get a stripe - must wait */
4485 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4486 conf
->retry_read_aligned
= raid_bio
;
4490 set_bit(R5_ReadError
, &sh
->dev
[dd_idx
].flags
);
4491 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0)) {
4493 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4494 conf
->retry_read_aligned
= raid_bio
;
4502 spin_lock_irq(&conf
->device_lock
);
4503 remaining
= raid5_dec_bi_phys_segments(raid_bio
);
4504 spin_unlock_irq(&conf
->device_lock
);
4506 bio_endio(raid_bio
, 0);
4507 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4508 wake_up(&conf
->wait_for_stripe
);
4514 * This is our raid5 kernel thread.
4516 * We scan the hash table for stripes which can be handled now.
4517 * During the scan, completed stripes are saved for us by the interrupt
4518 * handler, so that they will not have to wait for our next wakeup.
4520 static void raid5d(mddev_t
*mddev
)
4522 struct stripe_head
*sh
;
4523 raid5_conf_t
*conf
= mddev
->private;
4526 pr_debug("+++ raid5d active\n");
4528 md_check_recovery(mddev
);
4531 spin_lock_irq(&conf
->device_lock
);
4535 if (conf
->seq_flush
!= conf
->seq_write
) {
4536 int seq
= conf
->seq_flush
;
4537 spin_unlock_irq(&conf
->device_lock
);
4538 bitmap_unplug(mddev
->bitmap
);
4539 spin_lock_irq(&conf
->device_lock
);
4540 conf
->seq_write
= seq
;
4541 activate_bit_delay(conf
);
4544 while ((bio
= remove_bio_from_retry(conf
))) {
4546 spin_unlock_irq(&conf
->device_lock
);
4547 ok
= retry_aligned_read(conf
, bio
);
4548 spin_lock_irq(&conf
->device_lock
);
4554 sh
= __get_priority_stripe(conf
);
4558 spin_unlock_irq(&conf
->device_lock
);
4565 spin_lock_irq(&conf
->device_lock
);
4567 pr_debug("%d stripes handled\n", handled
);
4569 spin_unlock_irq(&conf
->device_lock
);
4571 async_tx_issue_pending_all();
4572 unplug_slaves(mddev
);
4574 pr_debug("--- raid5d inactive\n");
4578 raid5_show_stripe_cache_size(mddev_t
*mddev
, char *page
)
4580 raid5_conf_t
*conf
= mddev
->private;
4582 return sprintf(page
, "%d\n", conf
->max_nr_stripes
);
4588 raid5_set_cache_size(mddev_t
*mddev
, int size
)
4590 raid5_conf_t
*conf
= mddev
->private;
4593 if (size
<= 16 || size
> 32768)
4595 while (size
< conf
->max_nr_stripes
) {
4596 if (drop_one_stripe(conf
))
4597 conf
->max_nr_stripes
--;
4601 err
= md_allow_write(mddev
);
4604 while (size
> conf
->max_nr_stripes
) {
4605 if (grow_one_stripe(conf
))
4606 conf
->max_nr_stripes
++;
4611 EXPORT_SYMBOL(raid5_set_cache_size
);
4614 raid5_store_stripe_cache_size(mddev_t
*mddev
, const char *page
, size_t len
)
4616 raid5_conf_t
*conf
= mddev
->private;
4620 if (len
>= PAGE_SIZE
)
4625 if (strict_strtoul(page
, 10, &new))
4627 err
= raid5_set_cache_size(mddev
, new);
4633 static struct md_sysfs_entry
4634 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
4635 raid5_show_stripe_cache_size
,
4636 raid5_store_stripe_cache_size
);
4639 raid5_show_preread_threshold(mddev_t
*mddev
, char *page
)
4641 raid5_conf_t
*conf
= mddev
->private;
4643 return sprintf(page
, "%d\n", conf
->bypass_threshold
);
4649 raid5_store_preread_threshold(mddev_t
*mddev
, const char *page
, size_t len
)
4651 raid5_conf_t
*conf
= mddev
->private;
4653 if (len
>= PAGE_SIZE
)
4658 if (strict_strtoul(page
, 10, &new))
4660 if (new > conf
->max_nr_stripes
)
4662 conf
->bypass_threshold
= new;
4666 static struct md_sysfs_entry
4667 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
4669 raid5_show_preread_threshold
,
4670 raid5_store_preread_threshold
);
4673 stripe_cache_active_show(mddev_t
*mddev
, char *page
)
4675 raid5_conf_t
*conf
= mddev
->private;
4677 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
4682 static struct md_sysfs_entry
4683 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
4685 static struct attribute
*raid5_attrs
[] = {
4686 &raid5_stripecache_size
.attr
,
4687 &raid5_stripecache_active
.attr
,
4688 &raid5_preread_bypass_threshold
.attr
,
4691 static struct attribute_group raid5_attrs_group
= {
4693 .attrs
= raid5_attrs
,
4697 raid5_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
)
4699 raid5_conf_t
*conf
= mddev
->private;
4702 sectors
= mddev
->dev_sectors
;
4704 /* size is defined by the smallest of previous and new size */
4705 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
4707 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
4708 sectors
&= ~((sector_t
)mddev
->new_chunk_sectors
- 1);
4709 return sectors
* (raid_disks
- conf
->max_degraded
);
4712 static void raid5_free_percpu(raid5_conf_t
*conf
)
4714 struct raid5_percpu
*percpu
;
4721 for_each_possible_cpu(cpu
) {
4722 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4723 safe_put_page(percpu
->spare_page
);
4724 kfree(percpu
->scribble
);
4726 #ifdef CONFIG_HOTPLUG_CPU
4727 unregister_cpu_notifier(&conf
->cpu_notify
);
4731 free_percpu(conf
->percpu
);
4734 static void free_conf(raid5_conf_t
*conf
)
4736 shrink_stripes(conf
);
4737 raid5_free_percpu(conf
);
4739 kfree(conf
->stripe_hashtbl
);
4743 #ifdef CONFIG_HOTPLUG_CPU
4744 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
4747 raid5_conf_t
*conf
= container_of(nfb
, raid5_conf_t
, cpu_notify
);
4748 long cpu
= (long)hcpu
;
4749 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4752 case CPU_UP_PREPARE
:
4753 case CPU_UP_PREPARE_FROZEN
:
4754 if (conf
->level
== 6 && !percpu
->spare_page
)
4755 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
4756 if (!percpu
->scribble
)
4757 percpu
->scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
4759 if (!percpu
->scribble
||
4760 (conf
->level
== 6 && !percpu
->spare_page
)) {
4761 safe_put_page(percpu
->spare_page
);
4762 kfree(percpu
->scribble
);
4763 pr_err("%s: failed memory allocation for cpu%ld\n",
4765 return notifier_from_errno(-ENOMEM
);
4769 case CPU_DEAD_FROZEN
:
4770 safe_put_page(percpu
->spare_page
);
4771 kfree(percpu
->scribble
);
4772 percpu
->spare_page
= NULL
;
4773 percpu
->scribble
= NULL
;
4782 static int raid5_alloc_percpu(raid5_conf_t
*conf
)
4785 struct page
*spare_page
;
4786 struct raid5_percpu __percpu
*allcpus
;
4790 allcpus
= alloc_percpu(struct raid5_percpu
);
4793 conf
->percpu
= allcpus
;
4797 for_each_present_cpu(cpu
) {
4798 if (conf
->level
== 6) {
4799 spare_page
= alloc_page(GFP_KERNEL
);
4804 per_cpu_ptr(conf
->percpu
, cpu
)->spare_page
= spare_page
;
4806 scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
4811 per_cpu_ptr(conf
->percpu
, cpu
)->scribble
= scribble
;
4813 #ifdef CONFIG_HOTPLUG_CPU
4814 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
4815 conf
->cpu_notify
.priority
= 0;
4817 err
= register_cpu_notifier(&conf
->cpu_notify
);
4824 static raid5_conf_t
*setup_conf(mddev_t
*mddev
)
4827 int raid_disk
, memory
, max_disks
;
4829 struct disk_info
*disk
;
4831 if (mddev
->new_level
!= 5
4832 && mddev
->new_level
!= 4
4833 && mddev
->new_level
!= 6) {
4834 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4835 mdname(mddev
), mddev
->new_level
);
4836 return ERR_PTR(-EIO
);
4838 if ((mddev
->new_level
== 5
4839 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
4840 (mddev
->new_level
== 6
4841 && !algorithm_valid_raid6(mddev
->new_layout
))) {
4842 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
4843 mdname(mddev
), mddev
->new_layout
);
4844 return ERR_PTR(-EIO
);
4846 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
4847 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4848 mdname(mddev
), mddev
->raid_disks
);
4849 return ERR_PTR(-EINVAL
);
4852 if (!mddev
->new_chunk_sectors
||
4853 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
4854 !is_power_of_2(mddev
->new_chunk_sectors
)) {
4855 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
4856 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
4857 return ERR_PTR(-EINVAL
);
4860 conf
= kzalloc(sizeof(raid5_conf_t
), GFP_KERNEL
);
4863 spin_lock_init(&conf
->device_lock
);
4864 init_waitqueue_head(&conf
->wait_for_stripe
);
4865 init_waitqueue_head(&conf
->wait_for_overlap
);
4866 INIT_LIST_HEAD(&conf
->handle_list
);
4867 INIT_LIST_HEAD(&conf
->hold_list
);
4868 INIT_LIST_HEAD(&conf
->delayed_list
);
4869 INIT_LIST_HEAD(&conf
->bitmap_list
);
4870 INIT_LIST_HEAD(&conf
->inactive_list
);
4871 atomic_set(&conf
->active_stripes
, 0);
4872 atomic_set(&conf
->preread_active_stripes
, 0);
4873 atomic_set(&conf
->active_aligned_reads
, 0);
4874 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
4876 conf
->raid_disks
= mddev
->raid_disks
;
4877 if (mddev
->reshape_position
== MaxSector
)
4878 conf
->previous_raid_disks
= mddev
->raid_disks
;
4880 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4881 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
4882 conf
->scribble_len
= scribble_len(max_disks
);
4884 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
4889 conf
->mddev
= mddev
;
4891 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
4894 conf
->level
= mddev
->new_level
;
4895 if (raid5_alloc_percpu(conf
) != 0)
4898 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
4900 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
4901 raid_disk
= rdev
->raid_disk
;
4902 if (raid_disk
>= max_disks
4905 disk
= conf
->disks
+ raid_disk
;
4909 if (test_bit(In_sync
, &rdev
->flags
)) {
4910 char b
[BDEVNAME_SIZE
];
4911 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
4913 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
4915 /* Cannot rely on bitmap to complete recovery */
4919 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
4920 conf
->level
= mddev
->new_level
;
4921 if (conf
->level
== 6)
4922 conf
->max_degraded
= 2;
4924 conf
->max_degraded
= 1;
4925 conf
->algorithm
= mddev
->new_layout
;
4926 conf
->max_nr_stripes
= NR_STRIPES
;
4927 conf
->reshape_progress
= mddev
->reshape_position
;
4928 if (conf
->reshape_progress
!= MaxSector
) {
4929 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
4930 conf
->prev_algo
= mddev
->layout
;
4933 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
4934 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
4935 if (grow_stripes(conf
, conf
->max_nr_stripes
)) {
4937 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4938 mdname(mddev
), memory
);
4941 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
4942 mdname(mddev
), memory
);
4944 conf
->thread
= md_register_thread(raid5d
, mddev
, NULL
);
4945 if (!conf
->thread
) {
4947 "md/raid:%s: couldn't allocate thread.\n",
4957 return ERR_PTR(-EIO
);
4959 return ERR_PTR(-ENOMEM
);
4963 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
4966 case ALGORITHM_PARITY_0
:
4967 if (raid_disk
< max_degraded
)
4970 case ALGORITHM_PARITY_N
:
4971 if (raid_disk
>= raid_disks
- max_degraded
)
4974 case ALGORITHM_PARITY_0_6
:
4975 if (raid_disk
== 0 ||
4976 raid_disk
== raid_disks
- 1)
4979 case ALGORITHM_LEFT_ASYMMETRIC_6
:
4980 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
4981 case ALGORITHM_LEFT_SYMMETRIC_6
:
4982 case ALGORITHM_RIGHT_SYMMETRIC_6
:
4983 if (raid_disk
== raid_disks
- 1)
4989 static int run(mddev_t
*mddev
)
4992 int working_disks
= 0;
4993 int dirty_parity_disks
= 0;
4995 sector_t reshape_offset
= 0;
4997 if (mddev
->recovery_cp
!= MaxSector
)
4998 printk(KERN_NOTICE
"md/raid:%s: not clean"
4999 " -- starting background reconstruction\n",
5001 if (mddev
->reshape_position
!= MaxSector
) {
5002 /* Check that we can continue the reshape.
5003 * Currently only disks can change, it must
5004 * increase, and we must be past the point where
5005 * a stripe over-writes itself
5007 sector_t here_new
, here_old
;
5009 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
5011 if (mddev
->new_level
!= mddev
->level
) {
5012 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
5013 "required - aborting.\n",
5017 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
5018 /* reshape_position must be on a new-stripe boundary, and one
5019 * further up in new geometry must map after here in old
5022 here_new
= mddev
->reshape_position
;
5023 if (sector_div(here_new
, mddev
->new_chunk_sectors
*
5024 (mddev
->raid_disks
- max_degraded
))) {
5025 printk(KERN_ERR
"md/raid:%s: reshape_position not "
5026 "on a stripe boundary\n", mdname(mddev
));
5029 reshape_offset
= here_new
* mddev
->new_chunk_sectors
;
5030 /* here_new is the stripe we will write to */
5031 here_old
= mddev
->reshape_position
;
5032 sector_div(here_old
, mddev
->chunk_sectors
*
5033 (old_disks
-max_degraded
));
5034 /* here_old is the first stripe that we might need to read
5036 if (mddev
->delta_disks
== 0) {
5037 /* We cannot be sure it is safe to start an in-place
5038 * reshape. It is only safe if user-space if monitoring
5039 * and taking constant backups.
5040 * mdadm always starts a situation like this in
5041 * readonly mode so it can take control before
5042 * allowing any writes. So just check for that.
5044 if ((here_new
* mddev
->new_chunk_sectors
!=
5045 here_old
* mddev
->chunk_sectors
) ||
5047 printk(KERN_ERR
"md/raid:%s: in-place reshape must be started"
5048 " in read-only mode - aborting\n",
5052 } else if (mddev
->delta_disks
< 0
5053 ? (here_new
* mddev
->new_chunk_sectors
<=
5054 here_old
* mddev
->chunk_sectors
)
5055 : (here_new
* mddev
->new_chunk_sectors
>=
5056 here_old
* mddev
->chunk_sectors
)) {
5057 /* Reading from the same stripe as writing to - bad */
5058 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
5059 "auto-recovery - aborting.\n",
5063 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
5065 /* OK, we should be able to continue; */
5067 BUG_ON(mddev
->level
!= mddev
->new_level
);
5068 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
5069 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
5070 BUG_ON(mddev
->delta_disks
!= 0);
5073 if (mddev
->private == NULL
)
5074 conf
= setup_conf(mddev
);
5076 conf
= mddev
->private;
5079 return PTR_ERR(conf
);
5081 mddev
->thread
= conf
->thread
;
5082 conf
->thread
= NULL
;
5083 mddev
->private = conf
;
5086 * 0 for a fully functional array, 1 or 2 for a degraded array.
5088 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
5089 if (rdev
->raid_disk
< 0)
5091 if (test_bit(In_sync
, &rdev
->flags
)) {
5095 /* This disc is not fully in-sync. However if it
5096 * just stored parity (beyond the recovery_offset),
5097 * when we don't need to be concerned about the
5098 * array being dirty.
5099 * When reshape goes 'backwards', we never have
5100 * partially completed devices, so we only need
5101 * to worry about reshape going forwards.
5103 /* Hack because v0.91 doesn't store recovery_offset properly. */
5104 if (mddev
->major_version
== 0 &&
5105 mddev
->minor_version
> 90)
5106 rdev
->recovery_offset
= reshape_offset
;
5108 if (rdev
->recovery_offset
< reshape_offset
) {
5109 /* We need to check old and new layout */
5110 if (!only_parity(rdev
->raid_disk
,
5113 conf
->max_degraded
))
5116 if (!only_parity(rdev
->raid_disk
,
5118 conf
->previous_raid_disks
,
5119 conf
->max_degraded
))
5121 dirty_parity_disks
++;
5124 mddev
->degraded
= (max(conf
->raid_disks
, conf
->previous_raid_disks
)
5127 if (has_failed(conf
)) {
5128 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
5129 " (%d/%d failed)\n",
5130 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
5134 /* device size must be a multiple of chunk size */
5135 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
5136 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
5138 if (mddev
->degraded
> dirty_parity_disks
&&
5139 mddev
->recovery_cp
!= MaxSector
) {
5140 if (mddev
->ok_start_degraded
)
5142 "md/raid:%s: starting dirty degraded array"
5143 " - data corruption possible.\n",
5147 "md/raid:%s: cannot start dirty degraded array.\n",
5153 if (mddev
->degraded
== 0)
5154 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
5155 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
5156 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
5159 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
5160 " out of %d devices, algorithm %d\n",
5161 mdname(mddev
), conf
->level
,
5162 mddev
->raid_disks
- mddev
->degraded
,
5163 mddev
->raid_disks
, mddev
->new_layout
);
5165 print_raid5_conf(conf
);
5167 if (conf
->reshape_progress
!= MaxSector
) {
5168 conf
->reshape_safe
= conf
->reshape_progress
;
5169 atomic_set(&conf
->reshape_stripes
, 0);
5170 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5171 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5172 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5173 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5174 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5179 /* Ok, everything is just fine now */
5180 if (mddev
->to_remove
== &raid5_attrs_group
)
5181 mddev
->to_remove
= NULL
;
5182 else if (mddev
->kobj
.sd
&&
5183 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
5185 "raid5: failed to create sysfs attributes for %s\n",
5187 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5189 plugger_init(&conf
->plug
, raid5_unplug
);
5190 mddev
->plug
= &conf
->plug
;
5193 /* read-ahead size must cover two whole stripes, which
5194 * is 2 * (datadisks) * chunksize where 'n' is the
5195 * number of raid devices
5197 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
5198 int stripe
= data_disks
*
5199 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
5200 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
5201 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
5203 blk_queue_merge_bvec(mddev
->queue
, raid5_mergeable_bvec
);
5205 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
5206 mddev
->queue
->backing_dev_info
.congested_fn
= raid5_congested
;
5207 mddev
->queue
->queue_lock
= &conf
->device_lock
;
5208 mddev
->queue
->unplug_fn
= raid5_unplug_queue
;
5210 chunk_size
= mddev
->chunk_sectors
<< 9;
5211 blk_queue_io_min(mddev
->queue
, chunk_size
);
5212 blk_queue_io_opt(mddev
->queue
, chunk_size
*
5213 (conf
->raid_disks
- conf
->max_degraded
));
5215 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
5216 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
5217 rdev
->data_offset
<< 9);
5222 md_unregister_thread(mddev
->thread
);
5223 mddev
->thread
= NULL
;
5225 print_raid5_conf(conf
);
5228 mddev
->private = NULL
;
5229 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
5233 static int stop(mddev_t
*mddev
)
5235 raid5_conf_t
*conf
= mddev
->private;
5237 md_unregister_thread(mddev
->thread
);
5238 mddev
->thread
= NULL
;
5240 mddev
->queue
->backing_dev_info
.congested_fn
= NULL
;
5241 plugger_flush(&conf
->plug
); /* the unplug fn references 'conf'*/
5243 mddev
->private = NULL
;
5244 mddev
->to_remove
= &raid5_attrs_group
;
5249 static void print_sh(struct seq_file
*seq
, struct stripe_head
*sh
)
5253 seq_printf(seq
, "sh %llu, pd_idx %d, state %ld.\n",
5254 (unsigned long long)sh
->sector
, sh
->pd_idx
, sh
->state
);
5255 seq_printf(seq
, "sh %llu, count %d.\n",
5256 (unsigned long long)sh
->sector
, atomic_read(&sh
->count
));
5257 seq_printf(seq
, "sh %llu, ", (unsigned long long)sh
->sector
);
5258 for (i
= 0; i
< sh
->disks
; i
++) {
5259 seq_printf(seq
, "(cache%d: %p %ld) ",
5260 i
, sh
->dev
[i
].page
, sh
->dev
[i
].flags
);
5262 seq_printf(seq
, "\n");
5265 static void printall(struct seq_file
*seq
, raid5_conf_t
*conf
)
5267 struct stripe_head
*sh
;
5268 struct hlist_node
*hn
;
5271 spin_lock_irq(&conf
->device_lock
);
5272 for (i
= 0; i
< NR_HASH
; i
++) {
5273 hlist_for_each_entry(sh
, hn
, &conf
->stripe_hashtbl
[i
], hash
) {
5274 if (sh
->raid_conf
!= conf
)
5279 spin_unlock_irq(&conf
->device_lock
);
5283 static void status(struct seq_file
*seq
, mddev_t
*mddev
)
5285 raid5_conf_t
*conf
= mddev
->private;
5288 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
5289 mddev
->chunk_sectors
/ 2, mddev
->layout
);
5290 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
5291 for (i
= 0; i
< conf
->raid_disks
; i
++)
5292 seq_printf (seq
, "%s",
5293 conf
->disks
[i
].rdev
&&
5294 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
5295 seq_printf (seq
, "]");
5297 seq_printf (seq
, "\n");
5298 printall(seq
, conf
);
5302 static void print_raid5_conf (raid5_conf_t
*conf
)
5305 struct disk_info
*tmp
;
5307 printk(KERN_DEBUG
"RAID conf printout:\n");
5309 printk("(conf==NULL)\n");
5312 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
5314 conf
->raid_disks
- conf
->mddev
->degraded
);
5316 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5317 char b
[BDEVNAME_SIZE
];
5318 tmp
= conf
->disks
+ i
;
5320 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
5321 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
5322 bdevname(tmp
->rdev
->bdev
, b
));
5326 static int raid5_spare_active(mddev_t
*mddev
)
5329 raid5_conf_t
*conf
= mddev
->private;
5330 struct disk_info
*tmp
;
5332 unsigned long flags
;
5334 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5335 tmp
= conf
->disks
+ i
;
5337 && tmp
->rdev
->recovery_offset
== MaxSector
5338 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
5339 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
5341 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
5344 spin_lock_irqsave(&conf
->device_lock
, flags
);
5345 mddev
->degraded
-= count
;
5346 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5347 print_raid5_conf(conf
);
5351 static int raid5_remove_disk(mddev_t
*mddev
, int number
)
5353 raid5_conf_t
*conf
= mddev
->private;
5356 struct disk_info
*p
= conf
->disks
+ number
;
5358 print_raid5_conf(conf
);
5361 if (number
>= conf
->raid_disks
&&
5362 conf
->reshape_progress
== MaxSector
)
5363 clear_bit(In_sync
, &rdev
->flags
);
5365 if (test_bit(In_sync
, &rdev
->flags
) ||
5366 atomic_read(&rdev
->nr_pending
)) {
5370 /* Only remove non-faulty devices if recovery
5373 if (!test_bit(Faulty
, &rdev
->flags
) &&
5374 !has_failed(conf
) &&
5375 number
< conf
->raid_disks
) {
5381 if (atomic_read(&rdev
->nr_pending
)) {
5382 /* lost the race, try later */
5389 print_raid5_conf(conf
);
5393 static int raid5_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
5395 raid5_conf_t
*conf
= mddev
->private;
5398 struct disk_info
*p
;
5400 int last
= conf
->raid_disks
- 1;
5402 if (has_failed(conf
))
5403 /* no point adding a device */
5406 if (rdev
->raid_disk
>= 0)
5407 first
= last
= rdev
->raid_disk
;
5410 * find the disk ... but prefer rdev->saved_raid_disk
5413 if (rdev
->saved_raid_disk
>= 0 &&
5414 rdev
->saved_raid_disk
>= first
&&
5415 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
5416 disk
= rdev
->saved_raid_disk
;
5419 for ( ; disk
<= last
; disk
++)
5420 if ((p
=conf
->disks
+ disk
)->rdev
== NULL
) {
5421 clear_bit(In_sync
, &rdev
->flags
);
5422 rdev
->raid_disk
= disk
;
5424 if (rdev
->saved_raid_disk
!= disk
)
5426 rcu_assign_pointer(p
->rdev
, rdev
);
5429 print_raid5_conf(conf
);
5433 static int raid5_resize(mddev_t
*mddev
, sector_t sectors
)
5435 /* no resync is happening, and there is enough space
5436 * on all devices, so we can resize.
5437 * We need to make sure resync covers any new space.
5438 * If the array is shrinking we should possibly wait until
5439 * any io in the removed space completes, but it hardly seems
5442 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
5443 md_set_array_sectors(mddev
, raid5_size(mddev
, sectors
,
5444 mddev
->raid_disks
));
5445 if (mddev
->array_sectors
>
5446 raid5_size(mddev
, sectors
, mddev
->raid_disks
))
5448 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5449 revalidate_disk(mddev
->gendisk
);
5450 if (sectors
> mddev
->dev_sectors
&& mddev
->recovery_cp
== MaxSector
) {
5451 mddev
->recovery_cp
= mddev
->dev_sectors
;
5452 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
5454 mddev
->dev_sectors
= sectors
;
5455 mddev
->resync_max_sectors
= sectors
;
5459 static int check_stripe_cache(mddev_t
*mddev
)
5461 /* Can only proceed if there are plenty of stripe_heads.
5462 * We need a minimum of one full stripe,, and for sensible progress
5463 * it is best to have about 4 times that.
5464 * If we require 4 times, then the default 256 4K stripe_heads will
5465 * allow for chunk sizes up to 256K, which is probably OK.
5466 * If the chunk size is greater, user-space should request more
5467 * stripe_heads first.
5469 raid5_conf_t
*conf
= mddev
->private;
5470 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5471 > conf
->max_nr_stripes
||
5472 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5473 > conf
->max_nr_stripes
) {
5474 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5476 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
5483 static int check_reshape(mddev_t
*mddev
)
5485 raid5_conf_t
*conf
= mddev
->private;
5487 if (mddev
->delta_disks
== 0 &&
5488 mddev
->new_layout
== mddev
->layout
&&
5489 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
5490 return 0; /* nothing to do */
5492 /* Cannot grow a bitmap yet */
5494 if (has_failed(conf
))
5496 if (mddev
->delta_disks
< 0) {
5497 /* We might be able to shrink, but the devices must
5498 * be made bigger first.
5499 * For raid6, 4 is the minimum size.
5500 * Otherwise 2 is the minimum
5503 if (mddev
->level
== 6)
5505 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
5509 if (!check_stripe_cache(mddev
))
5512 return resize_stripes(conf
, conf
->raid_disks
+ mddev
->delta_disks
);
5515 static int raid5_start_reshape(mddev_t
*mddev
)
5517 raid5_conf_t
*conf
= mddev
->private;
5520 unsigned long flags
;
5522 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
5525 if (!check_stripe_cache(mddev
))
5528 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
5529 if (!test_bit(In_sync
, &rdev
->flags
)
5530 && !test_bit(Faulty
, &rdev
->flags
))
5533 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
5534 /* Not enough devices even to make a degraded array
5539 /* Refuse to reduce size of the array. Any reductions in
5540 * array size must be through explicit setting of array_size
5543 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
5544 < mddev
->array_sectors
) {
5545 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
5546 "before number of disks\n", mdname(mddev
));
5550 atomic_set(&conf
->reshape_stripes
, 0);
5551 spin_lock_irq(&conf
->device_lock
);
5552 conf
->previous_raid_disks
= conf
->raid_disks
;
5553 conf
->raid_disks
+= mddev
->delta_disks
;
5554 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
5555 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
5556 conf
->prev_algo
= conf
->algorithm
;
5557 conf
->algorithm
= mddev
->new_layout
;
5558 if (mddev
->delta_disks
< 0)
5559 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
5561 conf
->reshape_progress
= 0;
5562 conf
->reshape_safe
= conf
->reshape_progress
;
5564 spin_unlock_irq(&conf
->device_lock
);
5566 /* Add some new drives, as many as will fit.
5567 * We know there are enough to make the newly sized array work.
5568 * Don't add devices if we are reducing the number of
5569 * devices in the array. This is because it is not possible
5570 * to correctly record the "partially reconstructed" state of
5571 * such devices during the reshape and confusion could result.
5573 if (mddev
->delta_disks
>= 0) {
5574 int added_devices
= 0;
5575 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
5576 if (rdev
->raid_disk
< 0 &&
5577 !test_bit(Faulty
, &rdev
->flags
)) {
5578 if (raid5_add_disk(mddev
, rdev
) == 0) {
5581 >= conf
->previous_raid_disks
) {
5582 set_bit(In_sync
, &rdev
->flags
);
5585 rdev
->recovery_offset
= 0;
5586 sprintf(nm
, "rd%d", rdev
->raid_disk
);
5587 if (sysfs_create_link(&mddev
->kobj
,
5589 /* Failure here is OK */;
5591 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
5592 && !test_bit(Faulty
, &rdev
->flags
)) {
5593 /* This is a spare that was manually added */
5594 set_bit(In_sync
, &rdev
->flags
);
5598 /* When a reshape changes the number of devices,
5599 * ->degraded is measured against the larger of the
5600 * pre and post number of devices.
5602 spin_lock_irqsave(&conf
->device_lock
, flags
);
5603 mddev
->degraded
+= (conf
->raid_disks
- conf
->previous_raid_disks
)
5605 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5607 mddev
->raid_disks
= conf
->raid_disks
;
5608 mddev
->reshape_position
= conf
->reshape_progress
;
5609 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5611 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5612 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5613 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5614 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5615 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5617 if (!mddev
->sync_thread
) {
5618 mddev
->recovery
= 0;
5619 spin_lock_irq(&conf
->device_lock
);
5620 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
5621 conf
->reshape_progress
= MaxSector
;
5622 spin_unlock_irq(&conf
->device_lock
);
5625 conf
->reshape_checkpoint
= jiffies
;
5626 md_wakeup_thread(mddev
->sync_thread
);
5627 md_new_event(mddev
);
5631 /* This is called from the reshape thread and should make any
5632 * changes needed in 'conf'
5634 static void end_reshape(raid5_conf_t
*conf
)
5637 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
5639 spin_lock_irq(&conf
->device_lock
);
5640 conf
->previous_raid_disks
= conf
->raid_disks
;
5641 conf
->reshape_progress
= MaxSector
;
5642 spin_unlock_irq(&conf
->device_lock
);
5643 wake_up(&conf
->wait_for_overlap
);
5645 /* read-ahead size must cover two whole stripes, which is
5646 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5648 if (conf
->mddev
->queue
) {
5649 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5650 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
5652 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
5653 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
5658 /* This is called from the raid5d thread with mddev_lock held.
5659 * It makes config changes to the device.
5661 static void raid5_finish_reshape(mddev_t
*mddev
)
5663 raid5_conf_t
*conf
= mddev
->private;
5665 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
5667 if (mddev
->delta_disks
> 0) {
5668 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5669 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5670 revalidate_disk(mddev
->gendisk
);
5673 mddev
->degraded
= conf
->raid_disks
;
5674 for (d
= 0; d
< conf
->raid_disks
; d
++)
5675 if (conf
->disks
[d
].rdev
&&
5677 &conf
->disks
[d
].rdev
->flags
))
5679 for (d
= conf
->raid_disks
;
5680 d
< conf
->raid_disks
- mddev
->delta_disks
;
5682 mdk_rdev_t
*rdev
= conf
->disks
[d
].rdev
;
5683 if (rdev
&& raid5_remove_disk(mddev
, d
) == 0) {
5685 sprintf(nm
, "rd%d", rdev
->raid_disk
);
5686 sysfs_remove_link(&mddev
->kobj
, nm
);
5687 rdev
->raid_disk
= -1;
5691 mddev
->layout
= conf
->algorithm
;
5692 mddev
->chunk_sectors
= conf
->chunk_sectors
;
5693 mddev
->reshape_position
= MaxSector
;
5694 mddev
->delta_disks
= 0;
5698 static void raid5_quiesce(mddev_t
*mddev
, int state
)
5700 raid5_conf_t
*conf
= mddev
->private;
5703 case 2: /* resume for a suspend */
5704 wake_up(&conf
->wait_for_overlap
);
5707 case 1: /* stop all writes */
5708 spin_lock_irq(&conf
->device_lock
);
5709 /* '2' tells resync/reshape to pause so that all
5710 * active stripes can drain
5713 wait_event_lock_irq(conf
->wait_for_stripe
,
5714 atomic_read(&conf
->active_stripes
) == 0 &&
5715 atomic_read(&conf
->active_aligned_reads
) == 0,
5716 conf
->device_lock
, /* nothing */);
5718 spin_unlock_irq(&conf
->device_lock
);
5719 /* allow reshape to continue */
5720 wake_up(&conf
->wait_for_overlap
);
5723 case 0: /* re-enable writes */
5724 spin_lock_irq(&conf
->device_lock
);
5726 wake_up(&conf
->wait_for_stripe
);
5727 wake_up(&conf
->wait_for_overlap
);
5728 spin_unlock_irq(&conf
->device_lock
);
5734 static void *raid45_takeover_raid0(mddev_t
*mddev
, int level
)
5736 struct raid0_private_data
*raid0_priv
= mddev
->private;
5738 /* for raid0 takeover only one zone is supported */
5739 if (raid0_priv
->nr_strip_zones
> 1) {
5740 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5742 return ERR_PTR(-EINVAL
);
5745 mddev
->new_level
= level
;
5746 mddev
->new_layout
= ALGORITHM_PARITY_N
;
5747 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
5748 mddev
->raid_disks
+= 1;
5749 mddev
->delta_disks
= 1;
5750 /* make sure it will be not marked as dirty */
5751 mddev
->recovery_cp
= MaxSector
;
5753 return setup_conf(mddev
);
5757 static void *raid5_takeover_raid1(mddev_t
*mddev
)
5761 if (mddev
->raid_disks
!= 2 ||
5762 mddev
->degraded
> 1)
5763 return ERR_PTR(-EINVAL
);
5765 /* Should check if there are write-behind devices? */
5767 chunksect
= 64*2; /* 64K by default */
5769 /* The array must be an exact multiple of chunksize */
5770 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
5773 if ((chunksect
<<9) < STRIPE_SIZE
)
5774 /* array size does not allow a suitable chunk size */
5775 return ERR_PTR(-EINVAL
);
5777 mddev
->new_level
= 5;
5778 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5779 mddev
->new_chunk_sectors
= chunksect
;
5781 return setup_conf(mddev
);
5784 static void *raid5_takeover_raid6(mddev_t
*mddev
)
5788 switch (mddev
->layout
) {
5789 case ALGORITHM_LEFT_ASYMMETRIC_6
:
5790 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
5792 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
5793 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
5795 case ALGORITHM_LEFT_SYMMETRIC_6
:
5796 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5798 case ALGORITHM_RIGHT_SYMMETRIC_6
:
5799 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
5801 case ALGORITHM_PARITY_0_6
:
5802 new_layout
= ALGORITHM_PARITY_0
;
5804 case ALGORITHM_PARITY_N
:
5805 new_layout
= ALGORITHM_PARITY_N
;
5808 return ERR_PTR(-EINVAL
);
5810 mddev
->new_level
= 5;
5811 mddev
->new_layout
= new_layout
;
5812 mddev
->delta_disks
= -1;
5813 mddev
->raid_disks
-= 1;
5814 return setup_conf(mddev
);
5818 static int raid5_check_reshape(mddev_t
*mddev
)
5820 /* For a 2-drive array, the layout and chunk size can be changed
5821 * immediately as not restriping is needed.
5822 * For larger arrays we record the new value - after validation
5823 * to be used by a reshape pass.
5825 raid5_conf_t
*conf
= mddev
->private;
5826 int new_chunk
= mddev
->new_chunk_sectors
;
5828 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
5830 if (new_chunk
> 0) {
5831 if (!is_power_of_2(new_chunk
))
5833 if (new_chunk
< (PAGE_SIZE
>>9))
5835 if (mddev
->array_sectors
& (new_chunk
-1))
5836 /* not factor of array size */
5840 /* They look valid */
5842 if (mddev
->raid_disks
== 2) {
5843 /* can make the change immediately */
5844 if (mddev
->new_layout
>= 0) {
5845 conf
->algorithm
= mddev
->new_layout
;
5846 mddev
->layout
= mddev
->new_layout
;
5848 if (new_chunk
> 0) {
5849 conf
->chunk_sectors
= new_chunk
;
5850 mddev
->chunk_sectors
= new_chunk
;
5852 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5853 md_wakeup_thread(mddev
->thread
);
5855 return check_reshape(mddev
);
5858 static int raid6_check_reshape(mddev_t
*mddev
)
5860 int new_chunk
= mddev
->new_chunk_sectors
;
5862 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
5864 if (new_chunk
> 0) {
5865 if (!is_power_of_2(new_chunk
))
5867 if (new_chunk
< (PAGE_SIZE
>> 9))
5869 if (mddev
->array_sectors
& (new_chunk
-1))
5870 /* not factor of array size */
5874 /* They look valid */
5875 return check_reshape(mddev
);
5878 static void *raid5_takeover(mddev_t
*mddev
)
5880 /* raid5 can take over:
5881 * raid0 - if there is only one strip zone - make it a raid4 layout
5882 * raid1 - if there are two drives. We need to know the chunk size
5883 * raid4 - trivial - just use a raid4 layout.
5884 * raid6 - Providing it is a *_6 layout
5886 if (mddev
->level
== 0)
5887 return raid45_takeover_raid0(mddev
, 5);
5888 if (mddev
->level
== 1)
5889 return raid5_takeover_raid1(mddev
);
5890 if (mddev
->level
== 4) {
5891 mddev
->new_layout
= ALGORITHM_PARITY_N
;
5892 mddev
->new_level
= 5;
5893 return setup_conf(mddev
);
5895 if (mddev
->level
== 6)
5896 return raid5_takeover_raid6(mddev
);
5898 return ERR_PTR(-EINVAL
);
5901 static void *raid4_takeover(mddev_t
*mddev
)
5903 /* raid4 can take over:
5904 * raid0 - if there is only one strip zone
5905 * raid5 - if layout is right
5907 if (mddev
->level
== 0)
5908 return raid45_takeover_raid0(mddev
, 4);
5909 if (mddev
->level
== 5 &&
5910 mddev
->layout
== ALGORITHM_PARITY_N
) {
5911 mddev
->new_layout
= 0;
5912 mddev
->new_level
= 4;
5913 return setup_conf(mddev
);
5915 return ERR_PTR(-EINVAL
);
5918 static struct mdk_personality raid5_personality
;
5920 static void *raid6_takeover(mddev_t
*mddev
)
5922 /* Currently can only take over a raid5. We map the
5923 * personality to an equivalent raid6 personality
5924 * with the Q block at the end.
5928 if (mddev
->pers
!= &raid5_personality
)
5929 return ERR_PTR(-EINVAL
);
5930 if (mddev
->degraded
> 1)
5931 return ERR_PTR(-EINVAL
);
5932 if (mddev
->raid_disks
> 253)
5933 return ERR_PTR(-EINVAL
);
5934 if (mddev
->raid_disks
< 3)
5935 return ERR_PTR(-EINVAL
);
5937 switch (mddev
->layout
) {
5938 case ALGORITHM_LEFT_ASYMMETRIC
:
5939 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
5941 case ALGORITHM_RIGHT_ASYMMETRIC
:
5942 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
5944 case ALGORITHM_LEFT_SYMMETRIC
:
5945 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
5947 case ALGORITHM_RIGHT_SYMMETRIC
:
5948 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
5950 case ALGORITHM_PARITY_0
:
5951 new_layout
= ALGORITHM_PARITY_0_6
;
5953 case ALGORITHM_PARITY_N
:
5954 new_layout
= ALGORITHM_PARITY_N
;
5957 return ERR_PTR(-EINVAL
);
5959 mddev
->new_level
= 6;
5960 mddev
->new_layout
= new_layout
;
5961 mddev
->delta_disks
= 1;
5962 mddev
->raid_disks
+= 1;
5963 return setup_conf(mddev
);
5967 static struct mdk_personality raid6_personality
=
5971 .owner
= THIS_MODULE
,
5972 .make_request
= make_request
,
5976 .error_handler
= error
,
5977 .hot_add_disk
= raid5_add_disk
,
5978 .hot_remove_disk
= raid5_remove_disk
,
5979 .spare_active
= raid5_spare_active
,
5980 .sync_request
= sync_request
,
5981 .resize
= raid5_resize
,
5983 .check_reshape
= raid6_check_reshape
,
5984 .start_reshape
= raid5_start_reshape
,
5985 .finish_reshape
= raid5_finish_reshape
,
5986 .quiesce
= raid5_quiesce
,
5987 .takeover
= raid6_takeover
,
5989 static struct mdk_personality raid5_personality
=
5993 .owner
= THIS_MODULE
,
5994 .make_request
= make_request
,
5998 .error_handler
= error
,
5999 .hot_add_disk
= raid5_add_disk
,
6000 .hot_remove_disk
= raid5_remove_disk
,
6001 .spare_active
= raid5_spare_active
,
6002 .sync_request
= sync_request
,
6003 .resize
= raid5_resize
,
6005 .check_reshape
= raid5_check_reshape
,
6006 .start_reshape
= raid5_start_reshape
,
6007 .finish_reshape
= raid5_finish_reshape
,
6008 .quiesce
= raid5_quiesce
,
6009 .takeover
= raid5_takeover
,
6012 static struct mdk_personality raid4_personality
=
6016 .owner
= THIS_MODULE
,
6017 .make_request
= make_request
,
6021 .error_handler
= error
,
6022 .hot_add_disk
= raid5_add_disk
,
6023 .hot_remove_disk
= raid5_remove_disk
,
6024 .spare_active
= raid5_spare_active
,
6025 .sync_request
= sync_request
,
6026 .resize
= raid5_resize
,
6028 .check_reshape
= raid5_check_reshape
,
6029 .start_reshape
= raid5_start_reshape
,
6030 .finish_reshape
= raid5_finish_reshape
,
6031 .quiesce
= raid5_quiesce
,
6032 .takeover
= raid4_takeover
,
6035 static int __init
raid5_init(void)
6037 register_md_personality(&raid6_personality
);
6038 register_md_personality(&raid5_personality
);
6039 register_md_personality(&raid4_personality
);
6043 static void raid5_exit(void)
6045 unregister_md_personality(&raid6_personality
);
6046 unregister_md_personality(&raid5_personality
);
6047 unregister_md_personality(&raid4_personality
);
6050 module_init(raid5_init
);
6051 module_exit(raid5_exit
);
6052 MODULE_LICENSE("GPL");
6053 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6054 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6055 MODULE_ALIAS("md-raid5");
6056 MODULE_ALIAS("md-raid4");
6057 MODULE_ALIAS("md-level-5");
6058 MODULE_ALIAS("md-level-4");
6059 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6060 MODULE_ALIAS("md-raid6");
6061 MODULE_ALIAS("md-level-6");
6063 /* This used to be two separate modules, they were: */
6064 MODULE_ALIAS("raid5");
6065 MODULE_ALIAS("raid6");