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
);
437 static void raid5_unplug_device(raid5_conf_t
*conf
);
439 static struct stripe_head
*
440 get_active_stripe(raid5_conf_t
*conf
, sector_t sector
,
441 int previous
, int noblock
, int noquiesce
)
443 struct stripe_head
*sh
;
445 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
447 spin_lock_irq(&conf
->device_lock
);
450 wait_event_lock_irq(conf
->wait_for_stripe
,
451 conf
->quiesce
== 0 || noquiesce
,
452 conf
->device_lock
, /* nothing */);
453 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
455 if (!conf
->inactive_blocked
)
456 sh
= get_free_stripe(conf
);
457 if (noblock
&& sh
== NULL
)
460 conf
->inactive_blocked
= 1;
461 wait_event_lock_irq(conf
->wait_for_stripe
,
462 !list_empty(&conf
->inactive_list
) &&
463 (atomic_read(&conf
->active_stripes
)
464 < (conf
->max_nr_stripes
*3/4)
465 || !conf
->inactive_blocked
),
467 raid5_unplug_device(conf
)
469 conf
->inactive_blocked
= 0;
471 init_stripe(sh
, sector
, previous
);
473 if (atomic_read(&sh
->count
)) {
474 BUG_ON(!list_empty(&sh
->lru
)
475 && !test_bit(STRIPE_EXPANDING
, &sh
->state
));
477 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
478 atomic_inc(&conf
->active_stripes
);
479 if (list_empty(&sh
->lru
) &&
480 !test_bit(STRIPE_EXPANDING
, &sh
->state
))
482 list_del_init(&sh
->lru
);
485 } while (sh
== NULL
);
488 atomic_inc(&sh
->count
);
490 spin_unlock_irq(&conf
->device_lock
);
495 raid5_end_read_request(struct bio
*bi
, int error
);
497 raid5_end_write_request(struct bio
*bi
, int error
);
499 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
501 raid5_conf_t
*conf
= sh
->raid_conf
;
502 int i
, disks
= sh
->disks
;
506 for (i
= disks
; i
--; ) {
510 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
))
512 else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
517 bi
= &sh
->dev
[i
].req
;
521 bi
->bi_end_io
= raid5_end_write_request
;
523 bi
->bi_end_io
= raid5_end_read_request
;
526 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
527 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
530 atomic_inc(&rdev
->nr_pending
);
534 if (s
->syncing
|| s
->expanding
|| s
->expanded
)
535 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
537 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
539 bi
->bi_bdev
= rdev
->bdev
;
540 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
541 __func__
, (unsigned long long)sh
->sector
,
543 atomic_inc(&sh
->count
);
544 bi
->bi_sector
= sh
->sector
+ rdev
->data_offset
;
545 bi
->bi_flags
= 1 << BIO_UPTODATE
;
549 bi
->bi_io_vec
= &sh
->dev
[i
].vec
;
550 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
551 bi
->bi_io_vec
[0].bv_offset
= 0;
552 bi
->bi_size
= STRIPE_SIZE
;
555 test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
))
556 atomic_add(STRIPE_SECTORS
,
557 &rdev
->corrected_errors
);
558 generic_make_request(bi
);
561 set_bit(STRIPE_DEGRADED
, &sh
->state
);
562 pr_debug("skip op %ld on disc %d for sector %llu\n",
563 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
564 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
565 set_bit(STRIPE_HANDLE
, &sh
->state
);
570 static struct dma_async_tx_descriptor
*
571 async_copy_data(int frombio
, struct bio
*bio
, struct page
*page
,
572 sector_t sector
, struct dma_async_tx_descriptor
*tx
)
575 struct page
*bio_page
;
578 struct async_submit_ctl submit
;
579 enum async_tx_flags flags
= 0;
581 if (bio
->bi_sector
>= sector
)
582 page_offset
= (signed)(bio
->bi_sector
- sector
) * 512;
584 page_offset
= (signed)(sector
- bio
->bi_sector
) * -512;
587 flags
|= ASYNC_TX_FENCE
;
588 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
590 bio_for_each_segment(bvl
, bio
, i
) {
591 int len
= bio_iovec_idx(bio
, i
)->bv_len
;
595 if (page_offset
< 0) {
596 b_offset
= -page_offset
;
597 page_offset
+= b_offset
;
601 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
602 clen
= STRIPE_SIZE
- page_offset
;
607 b_offset
+= bio_iovec_idx(bio
, i
)->bv_offset
;
608 bio_page
= bio_iovec_idx(bio
, i
)->bv_page
;
610 tx
= async_memcpy(page
, bio_page
, page_offset
,
611 b_offset
, clen
, &submit
);
613 tx
= async_memcpy(bio_page
, page
, b_offset
,
614 page_offset
, clen
, &submit
);
616 /* chain the operations */
617 submit
.depend_tx
= tx
;
619 if (clen
< len
) /* hit end of page */
627 static void ops_complete_biofill(void *stripe_head_ref
)
629 struct stripe_head
*sh
= stripe_head_ref
;
630 struct bio
*return_bi
= NULL
;
631 raid5_conf_t
*conf
= sh
->raid_conf
;
634 pr_debug("%s: stripe %llu\n", __func__
,
635 (unsigned long long)sh
->sector
);
637 /* clear completed biofills */
638 spin_lock_irq(&conf
->device_lock
);
639 for (i
= sh
->disks
; i
--; ) {
640 struct r5dev
*dev
= &sh
->dev
[i
];
642 /* acknowledge completion of a biofill operation */
643 /* and check if we need to reply to a read request,
644 * new R5_Wantfill requests are held off until
645 * !STRIPE_BIOFILL_RUN
647 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
648 struct bio
*rbi
, *rbi2
;
653 while (rbi
&& rbi
->bi_sector
<
654 dev
->sector
+ STRIPE_SECTORS
) {
655 rbi2
= r5_next_bio(rbi
, dev
->sector
);
656 if (!raid5_dec_bi_phys_segments(rbi
)) {
657 rbi
->bi_next
= return_bi
;
664 spin_unlock_irq(&conf
->device_lock
);
665 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
667 return_io(return_bi
);
669 set_bit(STRIPE_HANDLE
, &sh
->state
);
673 static void ops_run_biofill(struct stripe_head
*sh
)
675 struct dma_async_tx_descriptor
*tx
= NULL
;
676 raid5_conf_t
*conf
= sh
->raid_conf
;
677 struct async_submit_ctl submit
;
680 pr_debug("%s: stripe %llu\n", __func__
,
681 (unsigned long long)sh
->sector
);
683 for (i
= sh
->disks
; i
--; ) {
684 struct r5dev
*dev
= &sh
->dev
[i
];
685 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
687 spin_lock_irq(&conf
->device_lock
);
688 dev
->read
= rbi
= dev
->toread
;
690 spin_unlock_irq(&conf
->device_lock
);
691 while (rbi
&& rbi
->bi_sector
<
692 dev
->sector
+ STRIPE_SECTORS
) {
693 tx
= async_copy_data(0, rbi
, dev
->page
,
695 rbi
= r5_next_bio(rbi
, dev
->sector
);
700 atomic_inc(&sh
->count
);
701 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
702 async_trigger_callback(&submit
);
705 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
712 tgt
= &sh
->dev
[target
];
713 set_bit(R5_UPTODATE
, &tgt
->flags
);
714 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
715 clear_bit(R5_Wantcompute
, &tgt
->flags
);
718 static void ops_complete_compute(void *stripe_head_ref
)
720 struct stripe_head
*sh
= stripe_head_ref
;
722 pr_debug("%s: stripe %llu\n", __func__
,
723 (unsigned long long)sh
->sector
);
725 /* mark the computed target(s) as uptodate */
726 mark_target_uptodate(sh
, sh
->ops
.target
);
727 mark_target_uptodate(sh
, sh
->ops
.target2
);
729 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
730 if (sh
->check_state
== check_state_compute_run
)
731 sh
->check_state
= check_state_compute_result
;
732 set_bit(STRIPE_HANDLE
, &sh
->state
);
736 /* return a pointer to the address conversion region of the scribble buffer */
737 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
738 struct raid5_percpu
*percpu
)
740 return percpu
->scribble
+ sizeof(struct page
*) * (sh
->disks
+ 2);
743 static struct dma_async_tx_descriptor
*
744 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
746 int disks
= sh
->disks
;
747 struct page
**xor_srcs
= percpu
->scribble
;
748 int target
= sh
->ops
.target
;
749 struct r5dev
*tgt
= &sh
->dev
[target
];
750 struct page
*xor_dest
= tgt
->page
;
752 struct dma_async_tx_descriptor
*tx
;
753 struct async_submit_ctl submit
;
756 pr_debug("%s: stripe %llu block: %d\n",
757 __func__
, (unsigned long long)sh
->sector
, target
);
758 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
760 for (i
= disks
; i
--; )
762 xor_srcs
[count
++] = sh
->dev
[i
].page
;
764 atomic_inc(&sh
->count
);
766 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
767 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
));
768 if (unlikely(count
== 1))
769 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
771 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
776 /* set_syndrome_sources - populate source buffers for gen_syndrome
777 * @srcs - (struct page *) array of size sh->disks
778 * @sh - stripe_head to parse
780 * Populates srcs in proper layout order for the stripe and returns the
781 * 'count' of sources to be used in a call to async_gen_syndrome. The P
782 * destination buffer is recorded in srcs[count] and the Q destination
783 * is recorded in srcs[count+1]].
785 static int set_syndrome_sources(struct page
**srcs
, struct stripe_head
*sh
)
787 int disks
= sh
->disks
;
788 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
789 int d0_idx
= raid6_d0(sh
);
793 for (i
= 0; i
< disks
; i
++)
799 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
801 srcs
[slot
] = sh
->dev
[i
].page
;
802 i
= raid6_next_disk(i
, disks
);
803 } while (i
!= d0_idx
);
805 return syndrome_disks
;
808 static struct dma_async_tx_descriptor
*
809 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
811 int disks
= sh
->disks
;
812 struct page
**blocks
= percpu
->scribble
;
814 int qd_idx
= sh
->qd_idx
;
815 struct dma_async_tx_descriptor
*tx
;
816 struct async_submit_ctl submit
;
822 if (sh
->ops
.target
< 0)
823 target
= sh
->ops
.target2
;
824 else if (sh
->ops
.target2
< 0)
825 target
= sh
->ops
.target
;
827 /* we should only have one valid target */
830 pr_debug("%s: stripe %llu block: %d\n",
831 __func__
, (unsigned long long)sh
->sector
, target
);
833 tgt
= &sh
->dev
[target
];
834 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
837 atomic_inc(&sh
->count
);
839 if (target
== qd_idx
) {
840 count
= set_syndrome_sources(blocks
, sh
);
841 blocks
[count
] = NULL
; /* regenerating p is not necessary */
842 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
843 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
844 ops_complete_compute
, sh
,
845 to_addr_conv(sh
, percpu
));
846 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
848 /* Compute any data- or p-drive using XOR */
850 for (i
= disks
; i
-- ; ) {
851 if (i
== target
|| i
== qd_idx
)
853 blocks
[count
++] = sh
->dev
[i
].page
;
856 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
857 NULL
, ops_complete_compute
, sh
,
858 to_addr_conv(sh
, percpu
));
859 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
865 static struct dma_async_tx_descriptor
*
866 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
868 int i
, count
, disks
= sh
->disks
;
869 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
870 int d0_idx
= raid6_d0(sh
);
871 int faila
= -1, failb
= -1;
872 int target
= sh
->ops
.target
;
873 int target2
= sh
->ops
.target2
;
874 struct r5dev
*tgt
= &sh
->dev
[target
];
875 struct r5dev
*tgt2
= &sh
->dev
[target2
];
876 struct dma_async_tx_descriptor
*tx
;
877 struct page
**blocks
= percpu
->scribble
;
878 struct async_submit_ctl submit
;
880 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
881 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
882 BUG_ON(target
< 0 || target2
< 0);
883 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
884 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
886 /* we need to open-code set_syndrome_sources to handle the
887 * slot number conversion for 'faila' and 'failb'
889 for (i
= 0; i
< disks
; i
++)
894 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
896 blocks
[slot
] = sh
->dev
[i
].page
;
902 i
= raid6_next_disk(i
, disks
);
903 } while (i
!= d0_idx
);
905 BUG_ON(faila
== failb
);
908 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
909 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
911 atomic_inc(&sh
->count
);
913 if (failb
== syndrome_disks
+1) {
914 /* Q disk is one of the missing disks */
915 if (faila
== syndrome_disks
) {
916 /* Missing P+Q, just recompute */
917 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
918 ops_complete_compute
, sh
,
919 to_addr_conv(sh
, percpu
));
920 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
921 STRIPE_SIZE
, &submit
);
925 int qd_idx
= sh
->qd_idx
;
927 /* Missing D+Q: recompute D from P, then recompute Q */
928 if (target
== qd_idx
)
929 data_target
= target2
;
931 data_target
= target
;
934 for (i
= disks
; i
-- ; ) {
935 if (i
== data_target
|| i
== qd_idx
)
937 blocks
[count
++] = sh
->dev
[i
].page
;
939 dest
= sh
->dev
[data_target
].page
;
940 init_async_submit(&submit
,
941 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
943 to_addr_conv(sh
, percpu
));
944 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
947 count
= set_syndrome_sources(blocks
, sh
);
948 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
949 ops_complete_compute
, sh
,
950 to_addr_conv(sh
, percpu
));
951 return async_gen_syndrome(blocks
, 0, count
+2,
952 STRIPE_SIZE
, &submit
);
955 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
956 ops_complete_compute
, sh
,
957 to_addr_conv(sh
, percpu
));
958 if (failb
== syndrome_disks
) {
959 /* We're missing D+P. */
960 return async_raid6_datap_recov(syndrome_disks
+2,
964 /* We're missing D+D. */
965 return async_raid6_2data_recov(syndrome_disks
+2,
966 STRIPE_SIZE
, faila
, failb
,
973 static void ops_complete_prexor(void *stripe_head_ref
)
975 struct stripe_head
*sh
= stripe_head_ref
;
977 pr_debug("%s: stripe %llu\n", __func__
,
978 (unsigned long long)sh
->sector
);
981 static struct dma_async_tx_descriptor
*
982 ops_run_prexor(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
983 struct dma_async_tx_descriptor
*tx
)
985 int disks
= sh
->disks
;
986 struct page
**xor_srcs
= percpu
->scribble
;
987 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
988 struct async_submit_ctl submit
;
990 /* existing parity data subtracted */
991 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
993 pr_debug("%s: stripe %llu\n", __func__
,
994 (unsigned long long)sh
->sector
);
996 for (i
= disks
; i
--; ) {
997 struct r5dev
*dev
= &sh
->dev
[i
];
998 /* Only process blocks that are known to be uptodate */
999 if (test_bit(R5_Wantdrain
, &dev
->flags
))
1000 xor_srcs
[count
++] = dev
->page
;
1003 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1004 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
));
1005 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1010 static struct dma_async_tx_descriptor
*
1011 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1013 int disks
= sh
->disks
;
1016 pr_debug("%s: stripe %llu\n", __func__
,
1017 (unsigned long long)sh
->sector
);
1019 for (i
= disks
; i
--; ) {
1020 struct r5dev
*dev
= &sh
->dev
[i
];
1023 if (test_and_clear_bit(R5_Wantdrain
, &dev
->flags
)) {
1026 spin_lock(&sh
->lock
);
1027 chosen
= dev
->towrite
;
1028 dev
->towrite
= NULL
;
1029 BUG_ON(dev
->written
);
1030 wbi
= dev
->written
= chosen
;
1031 spin_unlock(&sh
->lock
);
1033 while (wbi
&& wbi
->bi_sector
<
1034 dev
->sector
+ STRIPE_SECTORS
) {
1035 tx
= async_copy_data(1, wbi
, dev
->page
,
1037 wbi
= r5_next_bio(wbi
, dev
->sector
);
1045 static void ops_complete_reconstruct(void *stripe_head_ref
)
1047 struct stripe_head
*sh
= stripe_head_ref
;
1048 int disks
= sh
->disks
;
1049 int pd_idx
= sh
->pd_idx
;
1050 int qd_idx
= sh
->qd_idx
;
1053 pr_debug("%s: stripe %llu\n", __func__
,
1054 (unsigned long long)sh
->sector
);
1056 for (i
= disks
; i
--; ) {
1057 struct r5dev
*dev
= &sh
->dev
[i
];
1059 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
)
1060 set_bit(R5_UPTODATE
, &dev
->flags
);
1063 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1064 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1065 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1066 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1068 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1069 sh
->reconstruct_state
= reconstruct_state_result
;
1072 set_bit(STRIPE_HANDLE
, &sh
->state
);
1077 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1078 struct dma_async_tx_descriptor
*tx
)
1080 int disks
= sh
->disks
;
1081 struct page
**xor_srcs
= percpu
->scribble
;
1082 struct async_submit_ctl submit
;
1083 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1084 struct page
*xor_dest
;
1086 unsigned long flags
;
1088 pr_debug("%s: stripe %llu\n", __func__
,
1089 (unsigned long long)sh
->sector
);
1091 /* check if prexor is active which means only process blocks
1092 * that are part of a read-modify-write (written)
1094 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1096 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1097 for (i
= disks
; i
--; ) {
1098 struct r5dev
*dev
= &sh
->dev
[i
];
1100 xor_srcs
[count
++] = dev
->page
;
1103 xor_dest
= sh
->dev
[pd_idx
].page
;
1104 for (i
= disks
; i
--; ) {
1105 struct r5dev
*dev
= &sh
->dev
[i
];
1107 xor_srcs
[count
++] = dev
->page
;
1111 /* 1/ if we prexor'd then the dest is reused as a source
1112 * 2/ if we did not prexor then we are redoing the parity
1113 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1114 * for the synchronous xor case
1116 flags
= ASYNC_TX_ACK
|
1117 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1119 atomic_inc(&sh
->count
);
1121 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, sh
,
1122 to_addr_conv(sh
, percpu
));
1123 if (unlikely(count
== 1))
1124 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1126 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1130 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1131 struct dma_async_tx_descriptor
*tx
)
1133 struct async_submit_ctl submit
;
1134 struct page
**blocks
= percpu
->scribble
;
1137 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1139 count
= set_syndrome_sources(blocks
, sh
);
1141 atomic_inc(&sh
->count
);
1143 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_reconstruct
,
1144 sh
, to_addr_conv(sh
, percpu
));
1145 async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1148 static void ops_complete_check(void *stripe_head_ref
)
1150 struct stripe_head
*sh
= stripe_head_ref
;
1152 pr_debug("%s: stripe %llu\n", __func__
,
1153 (unsigned long long)sh
->sector
);
1155 sh
->check_state
= check_state_check_result
;
1156 set_bit(STRIPE_HANDLE
, &sh
->state
);
1160 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1162 int disks
= sh
->disks
;
1163 int pd_idx
= sh
->pd_idx
;
1164 int qd_idx
= sh
->qd_idx
;
1165 struct page
*xor_dest
;
1166 struct page
**xor_srcs
= percpu
->scribble
;
1167 struct dma_async_tx_descriptor
*tx
;
1168 struct async_submit_ctl submit
;
1172 pr_debug("%s: stripe %llu\n", __func__
,
1173 (unsigned long long)sh
->sector
);
1176 xor_dest
= sh
->dev
[pd_idx
].page
;
1177 xor_srcs
[count
++] = xor_dest
;
1178 for (i
= disks
; i
--; ) {
1179 if (i
== pd_idx
|| i
== qd_idx
)
1181 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1184 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1185 to_addr_conv(sh
, percpu
));
1186 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1187 &sh
->ops
.zero_sum_result
, &submit
);
1189 atomic_inc(&sh
->count
);
1190 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1191 tx
= async_trigger_callback(&submit
);
1194 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1196 struct page
**srcs
= percpu
->scribble
;
1197 struct async_submit_ctl submit
;
1200 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1201 (unsigned long long)sh
->sector
, checkp
);
1203 count
= set_syndrome_sources(srcs
, sh
);
1207 atomic_inc(&sh
->count
);
1208 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1209 sh
, to_addr_conv(sh
, percpu
));
1210 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1211 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1214 static void __raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1216 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1217 struct dma_async_tx_descriptor
*tx
= NULL
;
1218 raid5_conf_t
*conf
= sh
->raid_conf
;
1219 int level
= conf
->level
;
1220 struct raid5_percpu
*percpu
;
1224 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1225 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1226 ops_run_biofill(sh
);
1230 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1232 tx
= ops_run_compute5(sh
, percpu
);
1234 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
1235 tx
= ops_run_compute6_1(sh
, percpu
);
1237 tx
= ops_run_compute6_2(sh
, percpu
);
1239 /* terminate the chain if reconstruct is not set to be run */
1240 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
1244 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
))
1245 tx
= ops_run_prexor(sh
, percpu
, tx
);
1247 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
1248 tx
= ops_run_biodrain(sh
, tx
);
1252 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
1254 ops_run_reconstruct5(sh
, percpu
, tx
);
1256 ops_run_reconstruct6(sh
, percpu
, tx
);
1259 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
1260 if (sh
->check_state
== check_state_run
)
1261 ops_run_check_p(sh
, percpu
);
1262 else if (sh
->check_state
== check_state_run_q
)
1263 ops_run_check_pq(sh
, percpu
, 0);
1264 else if (sh
->check_state
== check_state_run_pq
)
1265 ops_run_check_pq(sh
, percpu
, 1);
1271 for (i
= disks
; i
--; ) {
1272 struct r5dev
*dev
= &sh
->dev
[i
];
1273 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
1274 wake_up(&sh
->raid_conf
->wait_for_overlap
);
1279 #ifdef CONFIG_MULTICORE_RAID456
1280 static void async_run_ops(void *param
, async_cookie_t cookie
)
1282 struct stripe_head
*sh
= param
;
1283 unsigned long ops_request
= sh
->ops
.request
;
1285 clear_bit_unlock(STRIPE_OPS_REQ_PENDING
, &sh
->state
);
1286 wake_up(&sh
->ops
.wait_for_ops
);
1288 __raid_run_ops(sh
, ops_request
);
1292 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1294 /* since handle_stripe can be called outside of raid5d context
1295 * we need to ensure sh->ops.request is de-staged before another
1298 wait_event(sh
->ops
.wait_for_ops
,
1299 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING
, &sh
->state
));
1300 sh
->ops
.request
= ops_request
;
1302 atomic_inc(&sh
->count
);
1303 async_schedule(async_run_ops
, sh
);
1306 #define raid_run_ops __raid_run_ops
1309 static int grow_one_stripe(raid5_conf_t
*conf
)
1311 struct stripe_head
*sh
;
1312 sh
= kmem_cache_alloc(conf
->slab_cache
, GFP_KERNEL
);
1315 memset(sh
, 0, sizeof(*sh
) + (conf
->pool_size
-1)*sizeof(struct r5dev
));
1316 sh
->raid_conf
= conf
;
1317 spin_lock_init(&sh
->lock
);
1318 #ifdef CONFIG_MULTICORE_RAID456
1319 init_waitqueue_head(&sh
->ops
.wait_for_ops
);
1322 if (grow_buffers(sh
)) {
1324 kmem_cache_free(conf
->slab_cache
, sh
);
1327 /* we just created an active stripe so... */
1328 atomic_set(&sh
->count
, 1);
1329 atomic_inc(&conf
->active_stripes
);
1330 INIT_LIST_HEAD(&sh
->lru
);
1335 static int grow_stripes(raid5_conf_t
*conf
, int num
)
1337 struct kmem_cache
*sc
;
1338 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
1340 if (conf
->mddev
->gendisk
)
1341 sprintf(conf
->cache_name
[0],
1342 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
1344 sprintf(conf
->cache_name
[0],
1345 "raid%d-%p", conf
->level
, conf
->mddev
);
1346 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
1348 conf
->active_name
= 0;
1349 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
1350 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
1354 conf
->slab_cache
= sc
;
1355 conf
->pool_size
= devs
;
1357 if (!grow_one_stripe(conf
))
1363 * scribble_len - return the required size of the scribble region
1364 * @num - total number of disks in the array
1366 * The size must be enough to contain:
1367 * 1/ a struct page pointer for each device in the array +2
1368 * 2/ room to convert each entry in (1) to its corresponding dma
1369 * (dma_map_page()) or page (page_address()) address.
1371 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1372 * calculate over all devices (not just the data blocks), using zeros in place
1373 * of the P and Q blocks.
1375 static size_t scribble_len(int num
)
1379 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
1384 static int resize_stripes(raid5_conf_t
*conf
, int newsize
)
1386 /* Make all the stripes able to hold 'newsize' devices.
1387 * New slots in each stripe get 'page' set to a new page.
1389 * This happens in stages:
1390 * 1/ create a new kmem_cache and allocate the required number of
1392 * 2/ gather all the old stripe_heads and tranfer the pages across
1393 * to the new stripe_heads. This will have the side effect of
1394 * freezing the array as once all stripe_heads have been collected,
1395 * no IO will be possible. Old stripe heads are freed once their
1396 * pages have been transferred over, and the old kmem_cache is
1397 * freed when all stripes are done.
1398 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1399 * we simple return a failre status - no need to clean anything up.
1400 * 4/ allocate new pages for the new slots in the new stripe_heads.
1401 * If this fails, we don't bother trying the shrink the
1402 * stripe_heads down again, we just leave them as they are.
1403 * As each stripe_head is processed the new one is released into
1406 * Once step2 is started, we cannot afford to wait for a write,
1407 * so we use GFP_NOIO allocations.
1409 struct stripe_head
*osh
, *nsh
;
1410 LIST_HEAD(newstripes
);
1411 struct disk_info
*ndisks
;
1414 struct kmem_cache
*sc
;
1417 if (newsize
<= conf
->pool_size
)
1418 return 0; /* never bother to shrink */
1420 err
= md_allow_write(conf
->mddev
);
1425 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
1426 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
1431 for (i
= conf
->max_nr_stripes
; i
; i
--) {
1432 nsh
= kmem_cache_alloc(sc
, GFP_KERNEL
);
1436 memset(nsh
, 0, sizeof(*nsh
) + (newsize
-1)*sizeof(struct r5dev
));
1438 nsh
->raid_conf
= conf
;
1439 spin_lock_init(&nsh
->lock
);
1440 #ifdef CONFIG_MULTICORE_RAID456
1441 init_waitqueue_head(&nsh
->ops
.wait_for_ops
);
1444 list_add(&nsh
->lru
, &newstripes
);
1447 /* didn't get enough, give up */
1448 while (!list_empty(&newstripes
)) {
1449 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1450 list_del(&nsh
->lru
);
1451 kmem_cache_free(sc
, nsh
);
1453 kmem_cache_destroy(sc
);
1456 /* Step 2 - Must use GFP_NOIO now.
1457 * OK, we have enough stripes, start collecting inactive
1458 * stripes and copying them over
1460 list_for_each_entry(nsh
, &newstripes
, lru
) {
1461 spin_lock_irq(&conf
->device_lock
);
1462 wait_event_lock_irq(conf
->wait_for_stripe
,
1463 !list_empty(&conf
->inactive_list
),
1465 unplug_slaves(conf
->mddev
)
1467 osh
= get_free_stripe(conf
);
1468 spin_unlock_irq(&conf
->device_lock
);
1469 atomic_set(&nsh
->count
, 1);
1470 for(i
=0; i
<conf
->pool_size
; i
++)
1471 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
1472 for( ; i
<newsize
; i
++)
1473 nsh
->dev
[i
].page
= NULL
;
1474 kmem_cache_free(conf
->slab_cache
, osh
);
1476 kmem_cache_destroy(conf
->slab_cache
);
1479 * At this point, we are holding all the stripes so the array
1480 * is completely stalled, so now is a good time to resize
1481 * conf->disks and the scribble region
1483 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
1485 for (i
=0; i
<conf
->raid_disks
; i
++)
1486 ndisks
[i
] = conf
->disks
[i
];
1488 conf
->disks
= ndisks
;
1493 conf
->scribble_len
= scribble_len(newsize
);
1494 for_each_present_cpu(cpu
) {
1495 struct raid5_percpu
*percpu
;
1498 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1499 scribble
= kmalloc(conf
->scribble_len
, GFP_NOIO
);
1502 kfree(percpu
->scribble
);
1503 percpu
->scribble
= scribble
;
1511 /* Step 4, return new stripes to service */
1512 while(!list_empty(&newstripes
)) {
1513 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1514 list_del_init(&nsh
->lru
);
1516 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
1517 if (nsh
->dev
[i
].page
== NULL
) {
1518 struct page
*p
= alloc_page(GFP_NOIO
);
1519 nsh
->dev
[i
].page
= p
;
1523 release_stripe(nsh
);
1525 /* critical section pass, GFP_NOIO no longer needed */
1527 conf
->slab_cache
= sc
;
1528 conf
->active_name
= 1-conf
->active_name
;
1529 conf
->pool_size
= newsize
;
1533 static int drop_one_stripe(raid5_conf_t
*conf
)
1535 struct stripe_head
*sh
;
1537 spin_lock_irq(&conf
->device_lock
);
1538 sh
= get_free_stripe(conf
);
1539 spin_unlock_irq(&conf
->device_lock
);
1542 BUG_ON(atomic_read(&sh
->count
));
1544 kmem_cache_free(conf
->slab_cache
, sh
);
1545 atomic_dec(&conf
->active_stripes
);
1549 static void shrink_stripes(raid5_conf_t
*conf
)
1551 while (drop_one_stripe(conf
))
1554 if (conf
->slab_cache
)
1555 kmem_cache_destroy(conf
->slab_cache
);
1556 conf
->slab_cache
= NULL
;
1559 static void raid5_end_read_request(struct bio
* bi
, int error
)
1561 struct stripe_head
*sh
= bi
->bi_private
;
1562 raid5_conf_t
*conf
= sh
->raid_conf
;
1563 int disks
= sh
->disks
, i
;
1564 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1565 char b
[BDEVNAME_SIZE
];
1569 for (i
=0 ; i
<disks
; i
++)
1570 if (bi
== &sh
->dev
[i
].req
)
1573 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1574 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1582 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1583 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
1584 rdev
= conf
->disks
[i
].rdev
;
1585 printk_rl(KERN_INFO
"md/raid:%s: read error corrected"
1586 " (%lu sectors at %llu on %s)\n",
1587 mdname(conf
->mddev
), STRIPE_SECTORS
,
1588 (unsigned long long)(sh
->sector
1589 + rdev
->data_offset
),
1590 bdevname(rdev
->bdev
, b
));
1591 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1592 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1594 if (atomic_read(&conf
->disks
[i
].rdev
->read_errors
))
1595 atomic_set(&conf
->disks
[i
].rdev
->read_errors
, 0);
1597 const char *bdn
= bdevname(conf
->disks
[i
].rdev
->bdev
, b
);
1599 rdev
= conf
->disks
[i
].rdev
;
1601 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1602 atomic_inc(&rdev
->read_errors
);
1603 if (conf
->mddev
->degraded
>= conf
->max_degraded
)
1604 printk_rl(KERN_WARNING
1605 "md/raid:%s: read error not correctable "
1606 "(sector %llu on %s).\n",
1607 mdname(conf
->mddev
),
1608 (unsigned long long)(sh
->sector
1609 + rdev
->data_offset
),
1611 else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
))
1613 printk_rl(KERN_WARNING
1614 "md/raid:%s: read error NOT corrected!! "
1615 "(sector %llu on %s).\n",
1616 mdname(conf
->mddev
),
1617 (unsigned long long)(sh
->sector
1618 + rdev
->data_offset
),
1620 else if (atomic_read(&rdev
->read_errors
)
1621 > conf
->max_nr_stripes
)
1623 "md/raid:%s: Too many read errors, failing device %s.\n",
1624 mdname(conf
->mddev
), bdn
);
1628 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1630 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1631 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1632 md_error(conf
->mddev
, rdev
);
1635 rdev_dec_pending(conf
->disks
[i
].rdev
, conf
->mddev
);
1636 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1637 set_bit(STRIPE_HANDLE
, &sh
->state
);
1641 static void raid5_end_write_request(struct bio
*bi
, int error
)
1643 struct stripe_head
*sh
= bi
->bi_private
;
1644 raid5_conf_t
*conf
= sh
->raid_conf
;
1645 int disks
= sh
->disks
, i
;
1646 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1648 for (i
=0 ; i
<disks
; i
++)
1649 if (bi
== &sh
->dev
[i
].req
)
1652 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1653 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1661 md_error(conf
->mddev
, conf
->disks
[i
].rdev
);
1663 rdev_dec_pending(conf
->disks
[i
].rdev
, conf
->mddev
);
1665 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1666 set_bit(STRIPE_HANDLE
, &sh
->state
);
1671 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
);
1673 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
1675 struct r5dev
*dev
= &sh
->dev
[i
];
1677 bio_init(&dev
->req
);
1678 dev
->req
.bi_io_vec
= &dev
->vec
;
1680 dev
->req
.bi_max_vecs
++;
1681 dev
->vec
.bv_page
= dev
->page
;
1682 dev
->vec
.bv_len
= STRIPE_SIZE
;
1683 dev
->vec
.bv_offset
= 0;
1685 dev
->req
.bi_sector
= sh
->sector
;
1686 dev
->req
.bi_private
= sh
;
1689 dev
->sector
= compute_blocknr(sh
, i
, previous
);
1692 static void error(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1694 char b
[BDEVNAME_SIZE
];
1695 raid5_conf_t
*conf
= mddev
->private;
1696 pr_debug("raid456: error called\n");
1698 if (!test_bit(Faulty
, &rdev
->flags
)) {
1699 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1700 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1701 unsigned long flags
;
1702 spin_lock_irqsave(&conf
->device_lock
, flags
);
1704 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1706 * if recovery was running, make sure it aborts.
1708 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1710 set_bit(Faulty
, &rdev
->flags
);
1712 "md/raid:%s: Disk failure on %s, disabling device.\n"
1714 "md/raid:%s: Operation continuing on %d devices.\n",
1716 bdevname(rdev
->bdev
, b
),
1718 conf
->raid_disks
- mddev
->degraded
);
1723 * Input: a 'big' sector number,
1724 * Output: index of the data and parity disk, and the sector # in them.
1726 static sector_t
raid5_compute_sector(raid5_conf_t
*conf
, sector_t r_sector
,
1727 int previous
, int *dd_idx
,
1728 struct stripe_head
*sh
)
1730 sector_t stripe
, stripe2
;
1731 sector_t chunk_number
;
1732 unsigned int chunk_offset
;
1735 sector_t new_sector
;
1736 int algorithm
= previous
? conf
->prev_algo
1738 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
1739 : conf
->chunk_sectors
;
1740 int raid_disks
= previous
? conf
->previous_raid_disks
1742 int data_disks
= raid_disks
- conf
->max_degraded
;
1744 /* First compute the information on this sector */
1747 * Compute the chunk number and the sector offset inside the chunk
1749 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
1750 chunk_number
= r_sector
;
1753 * Compute the stripe number
1755 stripe
= chunk_number
;
1756 *dd_idx
= sector_div(stripe
, data_disks
);
1759 * Select the parity disk based on the user selected algorithm.
1761 pd_idx
= qd_idx
= ~0;
1762 switch(conf
->level
) {
1764 pd_idx
= data_disks
;
1767 switch (algorithm
) {
1768 case ALGORITHM_LEFT_ASYMMETRIC
:
1769 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
1770 if (*dd_idx
>= pd_idx
)
1773 case ALGORITHM_RIGHT_ASYMMETRIC
:
1774 pd_idx
= sector_div(stripe2
, raid_disks
);
1775 if (*dd_idx
>= pd_idx
)
1778 case ALGORITHM_LEFT_SYMMETRIC
:
1779 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
1780 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1782 case ALGORITHM_RIGHT_SYMMETRIC
:
1783 pd_idx
= sector_div(stripe2
, raid_disks
);
1784 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1786 case ALGORITHM_PARITY_0
:
1790 case ALGORITHM_PARITY_N
:
1791 pd_idx
= data_disks
;
1799 switch (algorithm
) {
1800 case ALGORITHM_LEFT_ASYMMETRIC
:
1801 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1802 qd_idx
= pd_idx
+ 1;
1803 if (pd_idx
== raid_disks
-1) {
1804 (*dd_idx
)++; /* Q D D D P */
1806 } else if (*dd_idx
>= pd_idx
)
1807 (*dd_idx
) += 2; /* D D P Q D */
1809 case ALGORITHM_RIGHT_ASYMMETRIC
:
1810 pd_idx
= sector_div(stripe2
, raid_disks
);
1811 qd_idx
= pd_idx
+ 1;
1812 if (pd_idx
== raid_disks
-1) {
1813 (*dd_idx
)++; /* Q D D D P */
1815 } else if (*dd_idx
>= pd_idx
)
1816 (*dd_idx
) += 2; /* D D P Q D */
1818 case ALGORITHM_LEFT_SYMMETRIC
:
1819 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1820 qd_idx
= (pd_idx
+ 1) % raid_disks
;
1821 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
1823 case ALGORITHM_RIGHT_SYMMETRIC
:
1824 pd_idx
= sector_div(stripe2
, raid_disks
);
1825 qd_idx
= (pd_idx
+ 1) % raid_disks
;
1826 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
1829 case ALGORITHM_PARITY_0
:
1834 case ALGORITHM_PARITY_N
:
1835 pd_idx
= data_disks
;
1836 qd_idx
= data_disks
+ 1;
1839 case ALGORITHM_ROTATING_ZERO_RESTART
:
1840 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1841 * of blocks for computing Q is different.
1843 pd_idx
= sector_div(stripe2
, raid_disks
);
1844 qd_idx
= pd_idx
+ 1;
1845 if (pd_idx
== raid_disks
-1) {
1846 (*dd_idx
)++; /* Q D D D P */
1848 } else if (*dd_idx
>= pd_idx
)
1849 (*dd_idx
) += 2; /* D D P Q D */
1853 case ALGORITHM_ROTATING_N_RESTART
:
1854 /* Same a left_asymmetric, by first stripe is
1855 * D D D P Q rather than
1859 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1860 qd_idx
= pd_idx
+ 1;
1861 if (pd_idx
== raid_disks
-1) {
1862 (*dd_idx
)++; /* Q D D D P */
1864 } else if (*dd_idx
>= pd_idx
)
1865 (*dd_idx
) += 2; /* D D P Q D */
1869 case ALGORITHM_ROTATING_N_CONTINUE
:
1870 /* Same as left_symmetric but Q is before P */
1871 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
1872 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
1873 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1877 case ALGORITHM_LEFT_ASYMMETRIC_6
:
1878 /* RAID5 left_asymmetric, with Q on last device */
1879 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
1880 if (*dd_idx
>= pd_idx
)
1882 qd_idx
= raid_disks
- 1;
1885 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
1886 pd_idx
= sector_div(stripe2
, raid_disks
-1);
1887 if (*dd_idx
>= pd_idx
)
1889 qd_idx
= raid_disks
- 1;
1892 case ALGORITHM_LEFT_SYMMETRIC_6
:
1893 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
1894 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
1895 qd_idx
= raid_disks
- 1;
1898 case ALGORITHM_RIGHT_SYMMETRIC_6
:
1899 pd_idx
= sector_div(stripe2
, raid_disks
-1);
1900 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
1901 qd_idx
= raid_disks
- 1;
1904 case ALGORITHM_PARITY_0_6
:
1907 qd_idx
= raid_disks
- 1;
1917 sh
->pd_idx
= pd_idx
;
1918 sh
->qd_idx
= qd_idx
;
1919 sh
->ddf_layout
= ddf_layout
;
1922 * Finally, compute the new sector number
1924 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
1929 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
1931 raid5_conf_t
*conf
= sh
->raid_conf
;
1932 int raid_disks
= sh
->disks
;
1933 int data_disks
= raid_disks
- conf
->max_degraded
;
1934 sector_t new_sector
= sh
->sector
, check
;
1935 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
1936 : conf
->chunk_sectors
;
1937 int algorithm
= previous
? conf
->prev_algo
1941 sector_t chunk_number
;
1942 int dummy1
, dd_idx
= i
;
1944 struct stripe_head sh2
;
1947 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
1948 stripe
= new_sector
;
1950 if (i
== sh
->pd_idx
)
1952 switch(conf
->level
) {
1955 switch (algorithm
) {
1956 case ALGORITHM_LEFT_ASYMMETRIC
:
1957 case ALGORITHM_RIGHT_ASYMMETRIC
:
1961 case ALGORITHM_LEFT_SYMMETRIC
:
1962 case ALGORITHM_RIGHT_SYMMETRIC
:
1965 i
-= (sh
->pd_idx
+ 1);
1967 case ALGORITHM_PARITY_0
:
1970 case ALGORITHM_PARITY_N
:
1977 if (i
== sh
->qd_idx
)
1978 return 0; /* It is the Q disk */
1979 switch (algorithm
) {
1980 case ALGORITHM_LEFT_ASYMMETRIC
:
1981 case ALGORITHM_RIGHT_ASYMMETRIC
:
1982 case ALGORITHM_ROTATING_ZERO_RESTART
:
1983 case ALGORITHM_ROTATING_N_RESTART
:
1984 if (sh
->pd_idx
== raid_disks
-1)
1985 i
--; /* Q D D D P */
1986 else if (i
> sh
->pd_idx
)
1987 i
-= 2; /* D D P Q D */
1989 case ALGORITHM_LEFT_SYMMETRIC
:
1990 case ALGORITHM_RIGHT_SYMMETRIC
:
1991 if (sh
->pd_idx
== raid_disks
-1)
1992 i
--; /* Q D D D P */
1997 i
-= (sh
->pd_idx
+ 2);
2000 case ALGORITHM_PARITY_0
:
2003 case ALGORITHM_PARITY_N
:
2005 case ALGORITHM_ROTATING_N_CONTINUE
:
2006 /* Like left_symmetric, but P is before Q */
2007 if (sh
->pd_idx
== 0)
2008 i
--; /* P D D D Q */
2013 i
-= (sh
->pd_idx
+ 1);
2016 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2017 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2021 case ALGORITHM_LEFT_SYMMETRIC_6
:
2022 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2024 i
+= data_disks
+ 1;
2025 i
-= (sh
->pd_idx
+ 1);
2027 case ALGORITHM_PARITY_0_6
:
2036 chunk_number
= stripe
* data_disks
+ i
;
2037 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2039 check
= raid5_compute_sector(conf
, r_sector
,
2040 previous
, &dummy1
, &sh2
);
2041 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2042 || sh2
.qd_idx
!= sh
->qd_idx
) {
2043 printk(KERN_ERR
"md/raid:%s: compute_blocknr: map not correct\n",
2044 mdname(conf
->mddev
));
2052 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2053 int rcw
, int expand
)
2055 int i
, pd_idx
= sh
->pd_idx
, disks
= sh
->disks
;
2056 raid5_conf_t
*conf
= sh
->raid_conf
;
2057 int level
= conf
->level
;
2060 /* if we are not expanding this is a proper write request, and
2061 * there will be bios with new data to be drained into the
2065 sh
->reconstruct_state
= reconstruct_state_drain_run
;
2066 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2068 sh
->reconstruct_state
= reconstruct_state_run
;
2070 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2072 for (i
= disks
; i
--; ) {
2073 struct r5dev
*dev
= &sh
->dev
[i
];
2076 set_bit(R5_LOCKED
, &dev
->flags
);
2077 set_bit(R5_Wantdrain
, &dev
->flags
);
2079 clear_bit(R5_UPTODATE
, &dev
->flags
);
2083 if (s
->locked
+ conf
->max_degraded
== disks
)
2084 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2085 atomic_inc(&conf
->pending_full_writes
);
2088 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
2089 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
2091 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
2092 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
2093 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2094 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2096 for (i
= disks
; i
--; ) {
2097 struct r5dev
*dev
= &sh
->dev
[i
];
2102 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2103 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2104 set_bit(R5_Wantdrain
, &dev
->flags
);
2105 set_bit(R5_LOCKED
, &dev
->flags
);
2106 clear_bit(R5_UPTODATE
, &dev
->flags
);
2112 /* keep the parity disk(s) locked while asynchronous operations
2115 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
2116 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2120 int qd_idx
= sh
->qd_idx
;
2121 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
2123 set_bit(R5_LOCKED
, &dev
->flags
);
2124 clear_bit(R5_UPTODATE
, &dev
->flags
);
2128 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2129 __func__
, (unsigned long long)sh
->sector
,
2130 s
->locked
, s
->ops_request
);
2134 * Each stripe/dev can have one or more bion attached.
2135 * toread/towrite point to the first in a chain.
2136 * The bi_next chain must be in order.
2138 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
, int forwrite
)
2141 raid5_conf_t
*conf
= sh
->raid_conf
;
2144 pr_debug("adding bh b#%llu to stripe s#%llu\n",
2145 (unsigned long long)bi
->bi_sector
,
2146 (unsigned long long)sh
->sector
);
2149 spin_lock(&sh
->lock
);
2150 spin_lock_irq(&conf
->device_lock
);
2152 bip
= &sh
->dev
[dd_idx
].towrite
;
2153 if (*bip
== NULL
&& sh
->dev
[dd_idx
].written
== NULL
)
2156 bip
= &sh
->dev
[dd_idx
].toread
;
2157 while (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
) {
2158 if ((*bip
)->bi_sector
+ ((*bip
)->bi_size
>> 9) > bi
->bi_sector
)
2160 bip
= & (*bip
)->bi_next
;
2162 if (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
+ ((bi
->bi_size
)>>9))
2165 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
2169 bi
->bi_phys_segments
++;
2170 spin_unlock_irq(&conf
->device_lock
);
2171 spin_unlock(&sh
->lock
);
2173 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2174 (unsigned long long)bi
->bi_sector
,
2175 (unsigned long long)sh
->sector
, dd_idx
);
2177 if (conf
->mddev
->bitmap
&& firstwrite
) {
2178 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
2180 sh
->bm_seq
= conf
->seq_flush
+1;
2181 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
2185 /* check if page is covered */
2186 sector_t sector
= sh
->dev
[dd_idx
].sector
;
2187 for (bi
=sh
->dev
[dd_idx
].towrite
;
2188 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
2189 bi
&& bi
->bi_sector
<= sector
;
2190 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
2191 if (bi
->bi_sector
+ (bi
->bi_size
>>9) >= sector
)
2192 sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
2194 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
2195 set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
);
2200 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
2201 spin_unlock_irq(&conf
->device_lock
);
2202 spin_unlock(&sh
->lock
);
2206 static void end_reshape(raid5_conf_t
*conf
);
2208 static void stripe_set_idx(sector_t stripe
, raid5_conf_t
*conf
, int previous
,
2209 struct stripe_head
*sh
)
2211 int sectors_per_chunk
=
2212 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
2214 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
2215 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
2217 raid5_compute_sector(conf
,
2218 stripe
* (disks
- conf
->max_degraded
)
2219 *sectors_per_chunk
+ chunk_offset
,
2225 handle_failed_stripe(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2226 struct stripe_head_state
*s
, int disks
,
2227 struct bio
**return_bi
)
2230 for (i
= disks
; i
--; ) {
2234 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2237 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2238 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
2239 /* multiple read failures in one stripe */
2240 md_error(conf
->mddev
, rdev
);
2243 spin_lock_irq(&conf
->device_lock
);
2244 /* fail all writes first */
2245 bi
= sh
->dev
[i
].towrite
;
2246 sh
->dev
[i
].towrite
= NULL
;
2252 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2253 wake_up(&conf
->wait_for_overlap
);
2255 while (bi
&& bi
->bi_sector
<
2256 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2257 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2258 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2259 if (!raid5_dec_bi_phys_segments(bi
)) {
2260 md_write_end(conf
->mddev
);
2261 bi
->bi_next
= *return_bi
;
2266 /* and fail all 'written' */
2267 bi
= sh
->dev
[i
].written
;
2268 sh
->dev
[i
].written
= NULL
;
2269 if (bi
) bitmap_end
= 1;
2270 while (bi
&& bi
->bi_sector
<
2271 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2272 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2273 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2274 if (!raid5_dec_bi_phys_segments(bi
)) {
2275 md_write_end(conf
->mddev
);
2276 bi
->bi_next
= *return_bi
;
2282 /* fail any reads if this device is non-operational and
2283 * the data has not reached the cache yet.
2285 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
2286 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
2287 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
2288 bi
= sh
->dev
[i
].toread
;
2289 sh
->dev
[i
].toread
= NULL
;
2290 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2291 wake_up(&conf
->wait_for_overlap
);
2292 if (bi
) s
->to_read
--;
2293 while (bi
&& bi
->bi_sector
<
2294 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2295 struct bio
*nextbi
=
2296 r5_next_bio(bi
, sh
->dev
[i
].sector
);
2297 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2298 if (!raid5_dec_bi_phys_segments(bi
)) {
2299 bi
->bi_next
= *return_bi
;
2305 spin_unlock_irq(&conf
->device_lock
);
2307 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2308 STRIPE_SECTORS
, 0, 0);
2311 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2312 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2313 md_wakeup_thread(conf
->mddev
->thread
);
2316 /* fetch_block5 - checks the given member device to see if its data needs
2317 * to be read or computed to satisfy a request.
2319 * Returns 1 when no more member devices need to be checked, otherwise returns
2320 * 0 to tell the loop in handle_stripe_fill5 to continue
2322 static int fetch_block5(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2323 int disk_idx
, int disks
)
2325 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
2326 struct r5dev
*failed_dev
= &sh
->dev
[s
->failed_num
];
2328 /* is the data in this block needed, and can we get it? */
2329 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2330 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2332 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2333 s
->syncing
|| s
->expanding
||
2335 (failed_dev
->toread
||
2336 (failed_dev
->towrite
&&
2337 !test_bit(R5_OVERWRITE
, &failed_dev
->flags
)))))) {
2338 /* We would like to get this block, possibly by computing it,
2339 * otherwise read it if the backing disk is insync
2341 if ((s
->uptodate
== disks
- 1) &&
2342 (s
->failed
&& disk_idx
== s
->failed_num
)) {
2343 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2344 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2345 set_bit(R5_Wantcompute
, &dev
->flags
);
2346 sh
->ops
.target
= disk_idx
;
2347 sh
->ops
.target2
= -1;
2349 /* Careful: from this point on 'uptodate' is in the eye
2350 * of raid_run_ops which services 'compute' operations
2351 * before writes. R5_Wantcompute flags a block that will
2352 * be R5_UPTODATE by the time it is needed for a
2353 * subsequent operation.
2356 return 1; /* uptodate + compute == disks */
2357 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2358 set_bit(R5_LOCKED
, &dev
->flags
);
2359 set_bit(R5_Wantread
, &dev
->flags
);
2361 pr_debug("Reading block %d (sync=%d)\n", disk_idx
,
2370 * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2372 static void handle_stripe_fill5(struct stripe_head
*sh
,
2373 struct stripe_head_state
*s
, int disks
)
2377 /* look for blocks to read/compute, skip this if a compute
2378 * is already in flight, or if the stripe contents are in the
2379 * midst of changing due to a write
2381 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2382 !sh
->reconstruct_state
)
2383 for (i
= disks
; i
--; )
2384 if (fetch_block5(sh
, s
, i
, disks
))
2386 set_bit(STRIPE_HANDLE
, &sh
->state
);
2389 /* fetch_block6 - checks the given member device to see if its data needs
2390 * to be read or computed to satisfy a request.
2392 * Returns 1 when no more member devices need to be checked, otherwise returns
2393 * 0 to tell the loop in handle_stripe_fill6 to continue
2395 static int fetch_block6(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2396 struct r6_state
*r6s
, int disk_idx
, int disks
)
2398 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
2399 struct r5dev
*fdev
[2] = { &sh
->dev
[r6s
->failed_num
[0]],
2400 &sh
->dev
[r6s
->failed_num
[1]] };
2402 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2403 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2405 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2406 s
->syncing
|| s
->expanding
||
2408 (fdev
[0]->toread
|| s
->to_write
)) ||
2410 (fdev
[1]->toread
|| s
->to_write
)))) {
2411 /* we would like to get this block, possibly by computing it,
2412 * otherwise read it if the backing disk is insync
2414 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
2415 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
2416 if ((s
->uptodate
== disks
- 1) &&
2417 (s
->failed
&& (disk_idx
== r6s
->failed_num
[0] ||
2418 disk_idx
== r6s
->failed_num
[1]))) {
2419 /* have disk failed, and we're requested to fetch it;
2422 pr_debug("Computing stripe %llu block %d\n",
2423 (unsigned long long)sh
->sector
, disk_idx
);
2424 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2425 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2426 set_bit(R5_Wantcompute
, &dev
->flags
);
2427 sh
->ops
.target
= disk_idx
;
2428 sh
->ops
.target2
= -1; /* no 2nd target */
2432 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
2433 /* Computing 2-failure is *very* expensive; only
2434 * do it if failed >= 2
2437 for (other
= disks
; other
--; ) {
2438 if (other
== disk_idx
)
2440 if (!test_bit(R5_UPTODATE
,
2441 &sh
->dev
[other
].flags
))
2445 pr_debug("Computing stripe %llu blocks %d,%d\n",
2446 (unsigned long long)sh
->sector
,
2448 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2449 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2450 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
2451 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
2452 sh
->ops
.target
= disk_idx
;
2453 sh
->ops
.target2
= other
;
2457 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2458 set_bit(R5_LOCKED
, &dev
->flags
);
2459 set_bit(R5_Wantread
, &dev
->flags
);
2461 pr_debug("Reading block %d (sync=%d)\n",
2462 disk_idx
, s
->syncing
);
2470 * handle_stripe_fill6 - read or compute data to satisfy pending requests.
2472 static void handle_stripe_fill6(struct stripe_head
*sh
,
2473 struct stripe_head_state
*s
, struct r6_state
*r6s
,
2478 /* look for blocks to read/compute, skip this if a compute
2479 * is already in flight, or if the stripe contents are in the
2480 * midst of changing due to a write
2482 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2483 !sh
->reconstruct_state
)
2484 for (i
= disks
; i
--; )
2485 if (fetch_block6(sh
, s
, r6s
, i
, disks
))
2487 set_bit(STRIPE_HANDLE
, &sh
->state
);
2491 /* handle_stripe_clean_event
2492 * any written block on an uptodate or failed drive can be returned.
2493 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2494 * never LOCKED, so we don't need to test 'failed' directly.
2496 static void handle_stripe_clean_event(raid5_conf_t
*conf
,
2497 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
2502 for (i
= disks
; i
--; )
2503 if (sh
->dev
[i
].written
) {
2505 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2506 test_bit(R5_UPTODATE
, &dev
->flags
)) {
2507 /* We can return any write requests */
2508 struct bio
*wbi
, *wbi2
;
2510 pr_debug("Return write for disc %d\n", i
);
2511 spin_lock_irq(&conf
->device_lock
);
2513 dev
->written
= NULL
;
2514 while (wbi
&& wbi
->bi_sector
<
2515 dev
->sector
+ STRIPE_SECTORS
) {
2516 wbi2
= r5_next_bio(wbi
, dev
->sector
);
2517 if (!raid5_dec_bi_phys_segments(wbi
)) {
2518 md_write_end(conf
->mddev
);
2519 wbi
->bi_next
= *return_bi
;
2524 if (dev
->towrite
== NULL
)
2526 spin_unlock_irq(&conf
->device_lock
);
2528 bitmap_endwrite(conf
->mddev
->bitmap
,
2531 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
2536 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2537 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2538 md_wakeup_thread(conf
->mddev
->thread
);
2541 static void handle_stripe_dirtying5(raid5_conf_t
*conf
,
2542 struct stripe_head
*sh
, struct stripe_head_state
*s
, int disks
)
2544 int rmw
= 0, rcw
= 0, i
;
2545 for (i
= disks
; i
--; ) {
2546 /* would I have to read this buffer for read_modify_write */
2547 struct r5dev
*dev
= &sh
->dev
[i
];
2548 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2549 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2550 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2551 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2552 if (test_bit(R5_Insync
, &dev
->flags
))
2555 rmw
+= 2*disks
; /* cannot read it */
2557 /* Would I have to read this buffer for reconstruct_write */
2558 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) && 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
)) rcw
++;
2567 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2568 (unsigned long long)sh
->sector
, rmw
, rcw
);
2569 set_bit(STRIPE_HANDLE
, &sh
->state
);
2570 if (rmw
< rcw
&& rmw
> 0)
2571 /* prefer read-modify-write, but need to get some data */
2572 for (i
= disks
; i
--; ) {
2573 struct r5dev
*dev
= &sh
->dev
[i
];
2574 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2575 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2576 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2577 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2578 test_bit(R5_Insync
, &dev
->flags
)) {
2580 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2581 pr_debug("Read_old block "
2582 "%d for r-m-w\n", i
);
2583 set_bit(R5_LOCKED
, &dev
->flags
);
2584 set_bit(R5_Wantread
, &dev
->flags
);
2587 set_bit(STRIPE_DELAYED
, &sh
->state
);
2588 set_bit(STRIPE_HANDLE
, &sh
->state
);
2592 if (rcw
<= rmw
&& rcw
> 0)
2593 /* want reconstruct write, but need to get some data */
2594 for (i
= disks
; i
--; ) {
2595 struct r5dev
*dev
= &sh
->dev
[i
];
2596 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2598 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2599 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2600 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2601 test_bit(R5_Insync
, &dev
->flags
)) {
2603 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2604 pr_debug("Read_old block "
2605 "%d for Reconstruct\n", i
);
2606 set_bit(R5_LOCKED
, &dev
->flags
);
2607 set_bit(R5_Wantread
, &dev
->flags
);
2610 set_bit(STRIPE_DELAYED
, &sh
->state
);
2611 set_bit(STRIPE_HANDLE
, &sh
->state
);
2615 /* now if nothing is locked, and if we have enough data,
2616 * we can start a write request
2618 /* since handle_stripe can be called at any time we need to handle the
2619 * case where a compute block operation has been submitted and then a
2620 * subsequent call wants to start a write request. raid_run_ops only
2621 * handles the case where compute block and reconstruct are requested
2622 * simultaneously. If this is not the case then new writes need to be
2623 * held off until the compute completes.
2625 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
2626 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
2627 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
2628 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
2631 static void handle_stripe_dirtying6(raid5_conf_t
*conf
,
2632 struct stripe_head
*sh
, struct stripe_head_state
*s
,
2633 struct r6_state
*r6s
, int disks
)
2635 int rcw
= 0, pd_idx
= sh
->pd_idx
, i
;
2636 int qd_idx
= sh
->qd_idx
;
2638 set_bit(STRIPE_HANDLE
, &sh
->state
);
2639 for (i
= disks
; i
--; ) {
2640 struct r5dev
*dev
= &sh
->dev
[i
];
2641 /* check if we haven't enough data */
2642 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2643 i
!= pd_idx
&& i
!= qd_idx
&&
2644 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2645 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2646 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2648 if (!test_bit(R5_Insync
, &dev
->flags
))
2649 continue; /* it's a failed drive */
2652 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2653 pr_debug("Read_old stripe %llu "
2654 "block %d for Reconstruct\n",
2655 (unsigned long long)sh
->sector
, i
);
2656 set_bit(R5_LOCKED
, &dev
->flags
);
2657 set_bit(R5_Wantread
, &dev
->flags
);
2660 pr_debug("Request delayed stripe %llu "
2661 "block %d for Reconstruct\n",
2662 (unsigned long long)sh
->sector
, i
);
2663 set_bit(STRIPE_DELAYED
, &sh
->state
);
2664 set_bit(STRIPE_HANDLE
, &sh
->state
);
2668 /* now if nothing is locked, and if we have enough data, we can start a
2671 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
2672 s
->locked
== 0 && rcw
== 0 &&
2673 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)) {
2674 schedule_reconstruction(sh
, s
, 1, 0);
2678 static void handle_parity_checks5(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2679 struct stripe_head_state
*s
, int disks
)
2681 struct r5dev
*dev
= NULL
;
2683 set_bit(STRIPE_HANDLE
, &sh
->state
);
2685 switch (sh
->check_state
) {
2686 case check_state_idle
:
2687 /* start a new check operation if there are no failures */
2688 if (s
->failed
== 0) {
2689 BUG_ON(s
->uptodate
!= disks
);
2690 sh
->check_state
= check_state_run
;
2691 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2692 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2696 dev
= &sh
->dev
[s
->failed_num
];
2698 case check_state_compute_result
:
2699 sh
->check_state
= check_state_idle
;
2701 dev
= &sh
->dev
[sh
->pd_idx
];
2703 /* check that a write has not made the stripe insync */
2704 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2707 /* either failed parity check, or recovery is happening */
2708 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
2709 BUG_ON(s
->uptodate
!= disks
);
2711 set_bit(R5_LOCKED
, &dev
->flags
);
2713 set_bit(R5_Wantwrite
, &dev
->flags
);
2715 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2716 set_bit(STRIPE_INSYNC
, &sh
->state
);
2718 case check_state_run
:
2719 break; /* we will be called again upon completion */
2720 case check_state_check_result
:
2721 sh
->check_state
= check_state_idle
;
2723 /* if a failure occurred during the check operation, leave
2724 * STRIPE_INSYNC not set and let the stripe be handled again
2729 /* handle a successful check operation, if parity is correct
2730 * we are done. Otherwise update the mismatch count and repair
2731 * parity if !MD_RECOVERY_CHECK
2733 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
2734 /* parity is correct (on disc,
2735 * not in buffer any more)
2737 set_bit(STRIPE_INSYNC
, &sh
->state
);
2739 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2740 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2741 /* don't try to repair!! */
2742 set_bit(STRIPE_INSYNC
, &sh
->state
);
2744 sh
->check_state
= check_state_compute_run
;
2745 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2746 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2747 set_bit(R5_Wantcompute
,
2748 &sh
->dev
[sh
->pd_idx
].flags
);
2749 sh
->ops
.target
= sh
->pd_idx
;
2750 sh
->ops
.target2
= -1;
2755 case check_state_compute_run
:
2758 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
2759 __func__
, sh
->check_state
,
2760 (unsigned long long) sh
->sector
);
2766 static void handle_parity_checks6(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2767 struct stripe_head_state
*s
,
2768 struct r6_state
*r6s
, int disks
)
2770 int pd_idx
= sh
->pd_idx
;
2771 int qd_idx
= sh
->qd_idx
;
2774 set_bit(STRIPE_HANDLE
, &sh
->state
);
2776 BUG_ON(s
->failed
> 2);
2778 /* Want to check and possibly repair P and Q.
2779 * However there could be one 'failed' device, in which
2780 * case we can only check one of them, possibly using the
2781 * other to generate missing data
2784 switch (sh
->check_state
) {
2785 case check_state_idle
:
2786 /* start a new check operation if there are < 2 failures */
2787 if (s
->failed
== r6s
->q_failed
) {
2788 /* The only possible failed device holds Q, so it
2789 * makes sense to check P (If anything else were failed,
2790 * we would have used P to recreate it).
2792 sh
->check_state
= check_state_run
;
2794 if (!r6s
->q_failed
&& s
->failed
< 2) {
2795 /* Q is not failed, and we didn't use it to generate
2796 * anything, so it makes sense to check it
2798 if (sh
->check_state
== check_state_run
)
2799 sh
->check_state
= check_state_run_pq
;
2801 sh
->check_state
= check_state_run_q
;
2804 /* discard potentially stale zero_sum_result */
2805 sh
->ops
.zero_sum_result
= 0;
2807 if (sh
->check_state
== check_state_run
) {
2808 /* async_xor_zero_sum destroys the contents of P */
2809 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2812 if (sh
->check_state
>= check_state_run
&&
2813 sh
->check_state
<= check_state_run_pq
) {
2814 /* async_syndrome_zero_sum preserves P and Q, so
2815 * no need to mark them !uptodate here
2817 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2821 /* we have 2-disk failure */
2822 BUG_ON(s
->failed
!= 2);
2824 case check_state_compute_result
:
2825 sh
->check_state
= check_state_idle
;
2827 /* check that a write has not made the stripe insync */
2828 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2831 /* now write out any block on a failed drive,
2832 * or P or Q if they were recomputed
2834 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
2835 if (s
->failed
== 2) {
2836 dev
= &sh
->dev
[r6s
->failed_num
[1]];
2838 set_bit(R5_LOCKED
, &dev
->flags
);
2839 set_bit(R5_Wantwrite
, &dev
->flags
);
2841 if (s
->failed
>= 1) {
2842 dev
= &sh
->dev
[r6s
->failed_num
[0]];
2844 set_bit(R5_LOCKED
, &dev
->flags
);
2845 set_bit(R5_Wantwrite
, &dev
->flags
);
2847 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
2848 dev
= &sh
->dev
[pd_idx
];
2850 set_bit(R5_LOCKED
, &dev
->flags
);
2851 set_bit(R5_Wantwrite
, &dev
->flags
);
2853 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
2854 dev
= &sh
->dev
[qd_idx
];
2856 set_bit(R5_LOCKED
, &dev
->flags
);
2857 set_bit(R5_Wantwrite
, &dev
->flags
);
2859 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2861 set_bit(STRIPE_INSYNC
, &sh
->state
);
2863 case check_state_run
:
2864 case check_state_run_q
:
2865 case check_state_run_pq
:
2866 break; /* we will be called again upon completion */
2867 case check_state_check_result
:
2868 sh
->check_state
= check_state_idle
;
2870 /* handle a successful check operation, if parity is correct
2871 * we are done. Otherwise update the mismatch count and repair
2872 * parity if !MD_RECOVERY_CHECK
2874 if (sh
->ops
.zero_sum_result
== 0) {
2875 /* both parities are correct */
2877 set_bit(STRIPE_INSYNC
, &sh
->state
);
2879 /* in contrast to the raid5 case we can validate
2880 * parity, but still have a failure to write
2883 sh
->check_state
= check_state_compute_result
;
2884 /* Returning at this point means that we may go
2885 * off and bring p and/or q uptodate again so
2886 * we make sure to check zero_sum_result again
2887 * to verify if p or q need writeback
2891 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2892 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2893 /* don't try to repair!! */
2894 set_bit(STRIPE_INSYNC
, &sh
->state
);
2896 int *target
= &sh
->ops
.target
;
2898 sh
->ops
.target
= -1;
2899 sh
->ops
.target2
= -1;
2900 sh
->check_state
= check_state_compute_run
;
2901 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2902 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2903 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
2904 set_bit(R5_Wantcompute
,
2905 &sh
->dev
[pd_idx
].flags
);
2907 target
= &sh
->ops
.target2
;
2910 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
2911 set_bit(R5_Wantcompute
,
2912 &sh
->dev
[qd_idx
].flags
);
2919 case check_state_compute_run
:
2922 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
2923 __func__
, sh
->check_state
,
2924 (unsigned long long) sh
->sector
);
2929 static void handle_stripe_expansion(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2930 struct r6_state
*r6s
)
2934 /* We have read all the blocks in this stripe and now we need to
2935 * copy some of them into a target stripe for expand.
2937 struct dma_async_tx_descriptor
*tx
= NULL
;
2938 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
2939 for (i
= 0; i
< sh
->disks
; i
++)
2940 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
2942 struct stripe_head
*sh2
;
2943 struct async_submit_ctl submit
;
2945 sector_t bn
= compute_blocknr(sh
, i
, 1);
2946 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
2948 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
2950 /* so far only the early blocks of this stripe
2951 * have been requested. When later blocks
2952 * get requested, we will try again
2955 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
2956 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
2957 /* must have already done this block */
2958 release_stripe(sh2
);
2962 /* place all the copies on one channel */
2963 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
2964 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
2965 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
2968 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
2969 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
2970 for (j
= 0; j
< conf
->raid_disks
; j
++)
2971 if (j
!= sh2
->pd_idx
&&
2972 (!r6s
|| j
!= sh2
->qd_idx
) &&
2973 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
2975 if (j
== conf
->raid_disks
) {
2976 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
2977 set_bit(STRIPE_HANDLE
, &sh2
->state
);
2979 release_stripe(sh2
);
2982 /* done submitting copies, wait for them to complete */
2985 dma_wait_for_async_tx(tx
);
2991 * handle_stripe - do things to a stripe.
2993 * We lock the stripe and then examine the state of various bits
2994 * to see what needs to be done.
2996 * return some read request which now have data
2997 * return some write requests which are safely on disc
2998 * schedule a read on some buffers
2999 * schedule a write of some buffers
3000 * return confirmation of parity correctness
3002 * buffers are taken off read_list or write_list, and bh_cache buffers
3003 * get BH_Lock set before the stripe lock is released.
3007 static void handle_stripe5(struct stripe_head
*sh
)
3009 raid5_conf_t
*conf
= sh
->raid_conf
;
3010 int disks
= sh
->disks
, i
;
3011 struct bio
*return_bi
= NULL
;
3012 struct stripe_head_state s
;
3014 mdk_rdev_t
*blocked_rdev
= NULL
;
3016 int dec_preread_active
= 0;
3018 memset(&s
, 0, sizeof(s
));
3019 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
3020 "reconstruct:%d\n", (unsigned long long)sh
->sector
, sh
->state
,
3021 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->check_state
,
3022 sh
->reconstruct_state
);
3024 spin_lock(&sh
->lock
);
3025 clear_bit(STRIPE_HANDLE
, &sh
->state
);
3026 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3028 s
.syncing
= test_bit(STRIPE_SYNCING
, &sh
->state
);
3029 s
.expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3030 s
.expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3032 /* Now to look around and see what can be done */
3034 for (i
=disks
; i
--; ) {
3039 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
3040 "written %p\n", i
, dev
->flags
, dev
->toread
, dev
->read
,
3041 dev
->towrite
, dev
->written
);
3043 /* maybe we can request a biofill operation
3045 * new wantfill requests are only permitted while
3046 * ops_complete_biofill is guaranteed to be inactive
3048 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
3049 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
3050 set_bit(R5_Wantfill
, &dev
->flags
);
3052 /* now count some things */
3053 if (test_bit(R5_LOCKED
, &dev
->flags
)) s
.locked
++;
3054 if (test_bit(R5_UPTODATE
, &dev
->flags
)) s
.uptodate
++;
3055 if (test_bit(R5_Wantcompute
, &dev
->flags
)) s
.compute
++;
3057 if (test_bit(R5_Wantfill
, &dev
->flags
))
3059 else if (dev
->toread
)
3063 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3068 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3069 if (blocked_rdev
== NULL
&&
3070 rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
3071 blocked_rdev
= rdev
;
3072 atomic_inc(&rdev
->nr_pending
);
3074 clear_bit(R5_Insync
, &dev
->flags
);
3077 else if (test_bit(In_sync
, &rdev
->flags
))
3078 set_bit(R5_Insync
, &dev
->flags
);
3080 /* could be in-sync depending on recovery/reshape status */
3081 if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
3082 set_bit(R5_Insync
, &dev
->flags
);
3084 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3085 /* The ReadError flag will just be confusing now */
3086 clear_bit(R5_ReadError
, &dev
->flags
);
3087 clear_bit(R5_ReWrite
, &dev
->flags
);
3089 if (test_bit(R5_ReadError
, &dev
->flags
))
3090 clear_bit(R5_Insync
, &dev
->flags
);
3091 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3098 if (unlikely(blocked_rdev
)) {
3099 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
3100 s
.to_write
|| s
.written
) {
3101 set_bit(STRIPE_HANDLE
, &sh
->state
);
3104 /* There is nothing for the blocked_rdev to block */
3105 rdev_dec_pending(blocked_rdev
, conf
->mddev
);
3106 blocked_rdev
= NULL
;
3109 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
3110 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
3111 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
3114 pr_debug("locked=%d uptodate=%d to_read=%d"
3115 " to_write=%d failed=%d failed_num=%d\n",
3116 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
,
3117 s
.failed
, s
.failed_num
);
3118 /* check if the array has lost two devices and, if so, some requests might
3121 if (s
.failed
> 1 && s
.to_read
+s
.to_write
+s
.written
)
3122 handle_failed_stripe(conf
, sh
, &s
, disks
, &return_bi
);
3123 if (s
.failed
> 1 && s
.syncing
) {
3124 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,0);
3125 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3129 /* might be able to return some write requests if the parity block
3130 * is safe, or on a failed drive
3132 dev
= &sh
->dev
[sh
->pd_idx
];
3134 ((test_bit(R5_Insync
, &dev
->flags
) &&
3135 !test_bit(R5_LOCKED
, &dev
->flags
) &&
3136 test_bit(R5_UPTODATE
, &dev
->flags
)) ||
3137 (s
.failed
== 1 && s
.failed_num
== sh
->pd_idx
)))
3138 handle_stripe_clean_event(conf
, sh
, disks
, &return_bi
);
3140 /* Now we might consider reading some blocks, either to check/generate
3141 * parity, or to satisfy requests
3142 * or to load a block that is being partially written.
3144 if (s
.to_read
|| s
.non_overwrite
||
3145 (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
)) || s
.expanding
)
3146 handle_stripe_fill5(sh
, &s
, disks
);
3148 /* Now we check to see if any write operations have recently
3152 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
3154 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
3155 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
3156 sh
->reconstruct_state
= reconstruct_state_idle
;
3158 /* All the 'written' buffers and the parity block are ready to
3159 * be written back to disk
3161 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
));
3162 for (i
= disks
; i
--; ) {
3164 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
3165 (i
== sh
->pd_idx
|| dev
->written
)) {
3166 pr_debug("Writing block %d\n", i
);
3167 set_bit(R5_Wantwrite
, &dev
->flags
);
3170 if (!test_bit(R5_Insync
, &dev
->flags
) ||
3171 (i
== sh
->pd_idx
&& s
.failed
== 0))
3172 set_bit(STRIPE_INSYNC
, &sh
->state
);
3175 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3176 dec_preread_active
= 1;
3179 /* Now to consider new write requests and what else, if anything
3180 * should be read. We do not handle new writes when:
3181 * 1/ A 'write' operation (copy+xor) is already in flight.
3182 * 2/ A 'check' operation is in flight, as it may clobber the parity
3185 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
3186 handle_stripe_dirtying5(conf
, sh
, &s
, disks
);
3188 /* maybe we need to check and possibly fix the parity for this stripe
3189 * Any reads will already have been scheduled, so we just see if enough
3190 * data is available. The parity check is held off while parity
3191 * dependent operations are in flight.
3193 if (sh
->check_state
||
3194 (s
.syncing
&& s
.locked
== 0 &&
3195 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
3196 !test_bit(STRIPE_INSYNC
, &sh
->state
)))
3197 handle_parity_checks5(conf
, sh
, &s
, disks
);
3199 if (s
.syncing
&& s
.locked
== 0 && test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3200 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,1);
3201 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3204 /* If the failed drive is just a ReadError, then we might need to progress
3205 * the repair/check process
3207 if (s
.failed
== 1 && !conf
->mddev
->ro
&&
3208 test_bit(R5_ReadError
, &sh
->dev
[s
.failed_num
].flags
)
3209 && !test_bit(R5_LOCKED
, &sh
->dev
[s
.failed_num
].flags
)
3210 && test_bit(R5_UPTODATE
, &sh
->dev
[s
.failed_num
].flags
)
3212 dev
= &sh
->dev
[s
.failed_num
];
3213 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3214 set_bit(R5_Wantwrite
, &dev
->flags
);
3215 set_bit(R5_ReWrite
, &dev
->flags
);
3216 set_bit(R5_LOCKED
, &dev
->flags
);
3219 /* let's read it back */
3220 set_bit(R5_Wantread
, &dev
->flags
);
3221 set_bit(R5_LOCKED
, &dev
->flags
);
3226 /* Finish reconstruct operations initiated by the expansion process */
3227 if (sh
->reconstruct_state
== reconstruct_state_result
) {
3228 struct stripe_head
*sh2
3229 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
3230 if (sh2
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh2
->state
)) {
3231 /* sh cannot be written until sh2 has been read.
3232 * so arrange for sh to be delayed a little
3234 set_bit(STRIPE_DELAYED
, &sh
->state
);
3235 set_bit(STRIPE_HANDLE
, &sh
->state
);
3236 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
3238 atomic_inc(&conf
->preread_active_stripes
);
3239 release_stripe(sh2
);
3243 release_stripe(sh2
);
3245 sh
->reconstruct_state
= reconstruct_state_idle
;
3246 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3247 for (i
= conf
->raid_disks
; i
--; ) {
3248 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3249 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3254 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
3255 !sh
->reconstruct_state
) {
3256 /* Need to write out all blocks after computing parity */
3257 sh
->disks
= conf
->raid_disks
;
3258 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
3259 schedule_reconstruction(sh
, &s
, 1, 1);
3260 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
3261 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3262 atomic_dec(&conf
->reshape_stripes
);
3263 wake_up(&conf
->wait_for_overlap
);
3264 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3267 if (s
.expanding
&& s
.locked
== 0 &&
3268 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
3269 handle_stripe_expansion(conf
, sh
, NULL
);
3272 spin_unlock(&sh
->lock
);
3274 /* wait for this device to become unblocked */
3275 if (unlikely(blocked_rdev
))
3276 md_wait_for_blocked_rdev(blocked_rdev
, conf
->mddev
);
3279 raid_run_ops(sh
, s
.ops_request
);
3283 if (dec_preread_active
) {
3284 /* We delay this until after ops_run_io so that if make_request
3285 * is waiting on a barrier, it won't continue until the writes
3286 * have actually been submitted.
3288 atomic_dec(&conf
->preread_active_stripes
);
3289 if (atomic_read(&conf
->preread_active_stripes
) <
3291 md_wakeup_thread(conf
->mddev
->thread
);
3293 return_io(return_bi
);
3296 static void handle_stripe6(struct stripe_head
*sh
)
3298 raid5_conf_t
*conf
= sh
->raid_conf
;
3299 int disks
= sh
->disks
;
3300 struct bio
*return_bi
= NULL
;
3301 int i
, pd_idx
= sh
->pd_idx
, qd_idx
= sh
->qd_idx
;
3302 struct stripe_head_state s
;
3303 struct r6_state r6s
;
3304 struct r5dev
*dev
, *pdev
, *qdev
;
3305 mdk_rdev_t
*blocked_rdev
= NULL
;
3306 int dec_preread_active
= 0;
3308 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3309 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3310 (unsigned long long)sh
->sector
, sh
->state
,
3311 atomic_read(&sh
->count
), pd_idx
, qd_idx
,
3312 sh
->check_state
, sh
->reconstruct_state
);
3313 memset(&s
, 0, sizeof(s
));
3315 spin_lock(&sh
->lock
);
3316 clear_bit(STRIPE_HANDLE
, &sh
->state
);
3317 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3319 s
.syncing
= test_bit(STRIPE_SYNCING
, &sh
->state
);
3320 s
.expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3321 s
.expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3322 /* Now to look around and see what can be done */
3325 for (i
=disks
; i
--; ) {
3329 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3330 i
, dev
->flags
, dev
->toread
, dev
->towrite
, dev
->written
);
3331 /* maybe we can reply to a read
3333 * new wantfill requests are only permitted while
3334 * ops_complete_biofill is guaranteed to be inactive
3336 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
3337 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
3338 set_bit(R5_Wantfill
, &dev
->flags
);
3340 /* now count some things */
3341 if (test_bit(R5_LOCKED
, &dev
->flags
)) s
.locked
++;
3342 if (test_bit(R5_UPTODATE
, &dev
->flags
)) s
.uptodate
++;
3343 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
3345 BUG_ON(s
.compute
> 2);
3348 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
3350 } else if (dev
->toread
)
3354 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3359 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3360 if (blocked_rdev
== NULL
&&
3361 rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
3362 blocked_rdev
= rdev
;
3363 atomic_inc(&rdev
->nr_pending
);
3365 clear_bit(R5_Insync
, &dev
->flags
);
3368 else if (test_bit(In_sync
, &rdev
->flags
))
3369 set_bit(R5_Insync
, &dev
->flags
);
3371 /* in sync if before recovery_offset */
3372 if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
3373 set_bit(R5_Insync
, &dev
->flags
);
3375 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3376 /* The ReadError flag will just be confusing now */
3377 clear_bit(R5_ReadError
, &dev
->flags
);
3378 clear_bit(R5_ReWrite
, &dev
->flags
);
3380 if (test_bit(R5_ReadError
, &dev
->flags
))
3381 clear_bit(R5_Insync
, &dev
->flags
);
3382 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3384 r6s
.failed_num
[s
.failed
] = i
;
3390 if (unlikely(blocked_rdev
)) {
3391 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
3392 s
.to_write
|| s
.written
) {
3393 set_bit(STRIPE_HANDLE
, &sh
->state
);
3396 /* There is nothing for the blocked_rdev to block */
3397 rdev_dec_pending(blocked_rdev
, conf
->mddev
);
3398 blocked_rdev
= NULL
;
3401 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
3402 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
3403 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
3406 pr_debug("locked=%d uptodate=%d to_read=%d"
3407 " to_write=%d failed=%d failed_num=%d,%d\n",
3408 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
3409 r6s
.failed_num
[0], r6s
.failed_num
[1]);
3410 /* check if the array has lost >2 devices and, if so, some requests
3411 * might need to be failed
3413 if (s
.failed
> 2 && s
.to_read
+s
.to_write
+s
.written
)
3414 handle_failed_stripe(conf
, sh
, &s
, disks
, &return_bi
);
3415 if (s
.failed
> 2 && s
.syncing
) {
3416 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,0);
3417 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3422 * might be able to return some write requests if the parity blocks
3423 * are safe, or on a failed drive
3425 pdev
= &sh
->dev
[pd_idx
];
3426 r6s
.p_failed
= (s
.failed
>= 1 && r6s
.failed_num
[0] == pd_idx
)
3427 || (s
.failed
>= 2 && r6s
.failed_num
[1] == pd_idx
);
3428 qdev
= &sh
->dev
[qd_idx
];
3429 r6s
.q_failed
= (s
.failed
>= 1 && r6s
.failed_num
[0] == qd_idx
)
3430 || (s
.failed
>= 2 && r6s
.failed_num
[1] == qd_idx
);
3433 ( r6s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
3434 && !test_bit(R5_LOCKED
, &pdev
->flags
)
3435 && test_bit(R5_UPTODATE
, &pdev
->flags
)))) &&
3436 ( r6s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
3437 && !test_bit(R5_LOCKED
, &qdev
->flags
)
3438 && test_bit(R5_UPTODATE
, &qdev
->flags
)))))
3439 handle_stripe_clean_event(conf
, sh
, disks
, &return_bi
);
3441 /* Now we might consider reading some blocks, either to check/generate
3442 * parity, or to satisfy requests
3443 * or to load a block that is being partially written.
3445 if (s
.to_read
|| s
.non_overwrite
|| (s
.to_write
&& s
.failed
) ||
3446 (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
)) || s
.expanding
)
3447 handle_stripe_fill6(sh
, &s
, &r6s
, disks
);
3449 /* Now we check to see if any write operations have recently
3452 if (sh
->reconstruct_state
== reconstruct_state_drain_result
) {
3454 sh
->reconstruct_state
= reconstruct_state_idle
;
3455 /* All the 'written' buffers and the parity blocks are ready to
3456 * be written back to disk
3458 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
));
3459 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
));
3460 for (i
= disks
; i
--; ) {
3462 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
3463 (i
== sh
->pd_idx
|| i
== qd_idx
||
3465 pr_debug("Writing block %d\n", i
);
3466 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
3467 set_bit(R5_Wantwrite
, &dev
->flags
);
3468 if (!test_bit(R5_Insync
, &dev
->flags
) ||
3469 ((i
== sh
->pd_idx
|| i
== qd_idx
) &&
3471 set_bit(STRIPE_INSYNC
, &sh
->state
);
3474 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3475 dec_preread_active
= 1;
3478 /* Now to consider new write requests and what else, if anything
3479 * should be read. We do not handle new writes when:
3480 * 1/ A 'write' operation (copy+gen_syndrome) is already in flight.
3481 * 2/ A 'check' operation is in flight, as it may clobber the parity
3484 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
3485 handle_stripe_dirtying6(conf
, sh
, &s
, &r6s
, disks
);
3487 /* maybe we need to check and possibly fix the parity for this stripe
3488 * Any reads will already have been scheduled, so we just see if enough
3489 * data is available. The parity check is held off while parity
3490 * dependent operations are in flight.
3492 if (sh
->check_state
||
3493 (s
.syncing
&& s
.locked
== 0 &&
3494 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
3495 !test_bit(STRIPE_INSYNC
, &sh
->state
)))
3496 handle_parity_checks6(conf
, sh
, &s
, &r6s
, disks
);
3498 if (s
.syncing
&& s
.locked
== 0 && test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3499 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,1);
3500 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3503 /* If the failed drives are just a ReadError, then we might need
3504 * to progress the repair/check process
3506 if (s
.failed
<= 2 && !conf
->mddev
->ro
)
3507 for (i
= 0; i
< s
.failed
; i
++) {
3508 dev
= &sh
->dev
[r6s
.failed_num
[i
]];
3509 if (test_bit(R5_ReadError
, &dev
->flags
)
3510 && !test_bit(R5_LOCKED
, &dev
->flags
)
3511 && test_bit(R5_UPTODATE
, &dev
->flags
)
3513 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3514 set_bit(R5_Wantwrite
, &dev
->flags
);
3515 set_bit(R5_ReWrite
, &dev
->flags
);
3516 set_bit(R5_LOCKED
, &dev
->flags
);
3519 /* let's read it back */
3520 set_bit(R5_Wantread
, &dev
->flags
);
3521 set_bit(R5_LOCKED
, &dev
->flags
);
3527 /* Finish reconstruct operations initiated by the expansion process */
3528 if (sh
->reconstruct_state
== reconstruct_state_result
) {
3529 sh
->reconstruct_state
= reconstruct_state_idle
;
3530 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3531 for (i
= conf
->raid_disks
; i
--; ) {
3532 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3533 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3538 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
3539 !sh
->reconstruct_state
) {
3540 struct stripe_head
*sh2
3541 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
3542 if (sh2
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh2
->state
)) {
3543 /* sh cannot be written until sh2 has been read.
3544 * so arrange for sh to be delayed a little
3546 set_bit(STRIPE_DELAYED
, &sh
->state
);
3547 set_bit(STRIPE_HANDLE
, &sh
->state
);
3548 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
3550 atomic_inc(&conf
->preread_active_stripes
);
3551 release_stripe(sh2
);
3555 release_stripe(sh2
);
3557 /* Need to write out all blocks after computing P&Q */
3558 sh
->disks
= conf
->raid_disks
;
3559 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
3560 schedule_reconstruction(sh
, &s
, 1, 1);
3561 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
3562 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3563 atomic_dec(&conf
->reshape_stripes
);
3564 wake_up(&conf
->wait_for_overlap
);
3565 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3568 if (s
.expanding
&& s
.locked
== 0 &&
3569 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
3570 handle_stripe_expansion(conf
, sh
, &r6s
);
3573 spin_unlock(&sh
->lock
);
3575 /* wait for this device to become unblocked */
3576 if (unlikely(blocked_rdev
))
3577 md_wait_for_blocked_rdev(blocked_rdev
, conf
->mddev
);
3580 raid_run_ops(sh
, s
.ops_request
);
3585 if (dec_preread_active
) {
3586 /* We delay this until after ops_run_io so that if make_request
3587 * is waiting on a barrier, it won't continue until the writes
3588 * have actually been submitted.
3590 atomic_dec(&conf
->preread_active_stripes
);
3591 if (atomic_read(&conf
->preread_active_stripes
) <
3593 md_wakeup_thread(conf
->mddev
->thread
);
3596 return_io(return_bi
);
3599 static void handle_stripe(struct stripe_head
*sh
)
3601 if (sh
->raid_conf
->level
== 6)
3607 static void raid5_activate_delayed(raid5_conf_t
*conf
)
3609 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
3610 while (!list_empty(&conf
->delayed_list
)) {
3611 struct list_head
*l
= conf
->delayed_list
.next
;
3612 struct stripe_head
*sh
;
3613 sh
= list_entry(l
, struct stripe_head
, lru
);
3615 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3616 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3617 atomic_inc(&conf
->preread_active_stripes
);
3618 list_add_tail(&sh
->lru
, &conf
->hold_list
);
3621 plugger_set_plug(&conf
->plug
);
3624 static void activate_bit_delay(raid5_conf_t
*conf
)
3626 /* device_lock is held */
3627 struct list_head head
;
3628 list_add(&head
, &conf
->bitmap_list
);
3629 list_del_init(&conf
->bitmap_list
);
3630 while (!list_empty(&head
)) {
3631 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
3632 list_del_init(&sh
->lru
);
3633 atomic_inc(&sh
->count
);
3634 __release_stripe(conf
, sh
);
3638 static void unplug_slaves(mddev_t
*mddev
)
3640 raid5_conf_t
*conf
= mddev
->private;
3642 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
3645 for (i
= 0; i
< devs
; i
++) {
3646 mdk_rdev_t
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3647 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) && atomic_read(&rdev
->nr_pending
)) {
3648 struct request_queue
*r_queue
= bdev_get_queue(rdev
->bdev
);
3650 atomic_inc(&rdev
->nr_pending
);
3653 blk_unplug(r_queue
);
3655 rdev_dec_pending(rdev
, mddev
);
3662 static void raid5_unplug_device(raid5_conf_t
*conf
)
3664 unsigned long flags
;
3666 spin_lock_irqsave(&conf
->device_lock
, flags
);
3668 if (plugger_remove_plug(&conf
->plug
)) {
3670 raid5_activate_delayed(conf
);
3672 md_wakeup_thread(conf
->mddev
->thread
);
3674 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3676 unplug_slaves(conf
->mddev
);
3679 static void raid5_unplug(struct plug_handle
*plug
)
3681 raid5_conf_t
*conf
= container_of(plug
, raid5_conf_t
, plug
);
3682 raid5_unplug_device(conf
);
3685 static void raid5_unplug_queue(struct request_queue
*q
)
3687 mddev_t
*mddev
= q
->queuedata
;
3688 raid5_unplug_device(mddev
->private);
3691 int md_raid5_congested(mddev_t
*mddev
, int bits
)
3693 raid5_conf_t
*conf
= mddev
->private;
3695 /* No difference between reads and writes. Just check
3696 * how busy the stripe_cache is
3699 if (conf
->inactive_blocked
)
3703 if (list_empty_careful(&conf
->inactive_list
))
3708 EXPORT_SYMBOL_GPL(md_raid5_congested
);
3710 static int raid5_congested(void *data
, int bits
)
3712 mddev_t
*mddev
= data
;
3714 return mddev_congested(mddev
, bits
) ||
3715 md_raid5_congested(mddev
, bits
);
3718 /* We want read requests to align with chunks where possible,
3719 * but write requests don't need to.
3721 static int raid5_mergeable_bvec(struct request_queue
*q
,
3722 struct bvec_merge_data
*bvm
,
3723 struct bio_vec
*biovec
)
3725 mddev_t
*mddev
= q
->queuedata
;
3726 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
3728 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3729 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
3731 if ((bvm
->bi_rw
& 1) == WRITE
)
3732 return biovec
->bv_len
; /* always allow writes to be mergeable */
3734 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3735 chunk_sectors
= mddev
->new_chunk_sectors
;
3736 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
3737 if (max
< 0) max
= 0;
3738 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
3739 return biovec
->bv_len
;
3745 static int in_chunk_boundary(mddev_t
*mddev
, struct bio
*bio
)
3747 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
3748 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3749 unsigned int bio_sectors
= bio
->bi_size
>> 9;
3751 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3752 chunk_sectors
= mddev
->new_chunk_sectors
;
3753 return chunk_sectors
>=
3754 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
3758 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3759 * later sampled by raid5d.
3761 static void add_bio_to_retry(struct bio
*bi
,raid5_conf_t
*conf
)
3763 unsigned long flags
;
3765 spin_lock_irqsave(&conf
->device_lock
, flags
);
3767 bi
->bi_next
= conf
->retry_read_aligned_list
;
3768 conf
->retry_read_aligned_list
= bi
;
3770 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3771 md_wakeup_thread(conf
->mddev
->thread
);
3775 static struct bio
*remove_bio_from_retry(raid5_conf_t
*conf
)
3779 bi
= conf
->retry_read_aligned
;
3781 conf
->retry_read_aligned
= NULL
;
3784 bi
= conf
->retry_read_aligned_list
;
3786 conf
->retry_read_aligned_list
= bi
->bi_next
;
3789 * this sets the active strip count to 1 and the processed
3790 * strip count to zero (upper 8 bits)
3792 bi
->bi_phys_segments
= 1; /* biased count of active stripes */
3800 * The "raid5_align_endio" should check if the read succeeded and if it
3801 * did, call bio_endio on the original bio (having bio_put the new bio
3803 * If the read failed..
3805 static void raid5_align_endio(struct bio
*bi
, int error
)
3807 struct bio
* raid_bi
= bi
->bi_private
;
3810 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3815 rdev
= (void*)raid_bi
->bi_next
;
3816 raid_bi
->bi_next
= NULL
;
3817 mddev
= rdev
->mddev
;
3818 conf
= mddev
->private;
3820 rdev_dec_pending(rdev
, conf
->mddev
);
3822 if (!error
&& uptodate
) {
3823 bio_endio(raid_bi
, 0);
3824 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
3825 wake_up(&conf
->wait_for_stripe
);
3830 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3832 add_bio_to_retry(raid_bi
, conf
);
3835 static int bio_fits_rdev(struct bio
*bi
)
3837 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
3839 if ((bi
->bi_size
>>9) > queue_max_sectors(q
))
3841 blk_recount_segments(q
, bi
);
3842 if (bi
->bi_phys_segments
> queue_max_segments(q
))
3845 if (q
->merge_bvec_fn
)
3846 /* it's too hard to apply the merge_bvec_fn at this stage,
3855 static int chunk_aligned_read(mddev_t
*mddev
, struct bio
* raid_bio
)
3857 raid5_conf_t
*conf
= mddev
->private;
3859 struct bio
* align_bi
;
3862 if (!in_chunk_boundary(mddev
, raid_bio
)) {
3863 pr_debug("chunk_aligned_read : non aligned\n");
3867 * use bio_clone to make a copy of the bio
3869 align_bi
= bio_clone(raid_bio
, GFP_NOIO
);
3873 * set bi_end_io to a new function, and set bi_private to the
3876 align_bi
->bi_end_io
= raid5_align_endio
;
3877 align_bi
->bi_private
= raid_bio
;
3881 align_bi
->bi_sector
= raid5_compute_sector(conf
, raid_bio
->bi_sector
,
3886 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
3887 if (rdev
&& test_bit(In_sync
, &rdev
->flags
)) {
3888 atomic_inc(&rdev
->nr_pending
);
3890 raid_bio
->bi_next
= (void*)rdev
;
3891 align_bi
->bi_bdev
= rdev
->bdev
;
3892 align_bi
->bi_flags
&= ~(1 << BIO_SEG_VALID
);
3893 align_bi
->bi_sector
+= rdev
->data_offset
;
3895 if (!bio_fits_rdev(align_bi
)) {
3896 /* too big in some way */
3898 rdev_dec_pending(rdev
, mddev
);
3902 spin_lock_irq(&conf
->device_lock
);
3903 wait_event_lock_irq(conf
->wait_for_stripe
,
3905 conf
->device_lock
, /* nothing */);
3906 atomic_inc(&conf
->active_aligned_reads
);
3907 spin_unlock_irq(&conf
->device_lock
);
3909 generic_make_request(align_bi
);
3918 /* __get_priority_stripe - get the next stripe to process
3920 * Full stripe writes are allowed to pass preread active stripes up until
3921 * the bypass_threshold is exceeded. In general the bypass_count
3922 * increments when the handle_list is handled before the hold_list; however, it
3923 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3924 * stripe with in flight i/o. The bypass_count will be reset when the
3925 * head of the hold_list has changed, i.e. the head was promoted to the
3928 static struct stripe_head
*__get_priority_stripe(raid5_conf_t
*conf
)
3930 struct stripe_head
*sh
;
3932 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3934 list_empty(&conf
->handle_list
) ? "empty" : "busy",
3935 list_empty(&conf
->hold_list
) ? "empty" : "busy",
3936 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
3938 if (!list_empty(&conf
->handle_list
)) {
3939 sh
= list_entry(conf
->handle_list
.next
, typeof(*sh
), lru
);
3941 if (list_empty(&conf
->hold_list
))
3942 conf
->bypass_count
= 0;
3943 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
3944 if (conf
->hold_list
.next
== conf
->last_hold
)
3945 conf
->bypass_count
++;
3947 conf
->last_hold
= conf
->hold_list
.next
;
3948 conf
->bypass_count
-= conf
->bypass_threshold
;
3949 if (conf
->bypass_count
< 0)
3950 conf
->bypass_count
= 0;
3953 } else if (!list_empty(&conf
->hold_list
) &&
3954 ((conf
->bypass_threshold
&&
3955 conf
->bypass_count
> conf
->bypass_threshold
) ||
3956 atomic_read(&conf
->pending_full_writes
) == 0)) {
3957 sh
= list_entry(conf
->hold_list
.next
,
3959 conf
->bypass_count
-= conf
->bypass_threshold
;
3960 if (conf
->bypass_count
< 0)
3961 conf
->bypass_count
= 0;
3965 list_del_init(&sh
->lru
);
3966 atomic_inc(&sh
->count
);
3967 BUG_ON(atomic_read(&sh
->count
) != 1);
3971 static int make_request(mddev_t
*mddev
, struct bio
* bi
)
3973 raid5_conf_t
*conf
= mddev
->private;
3975 sector_t new_sector
;
3976 sector_t logical_sector
, last_sector
;
3977 struct stripe_head
*sh
;
3978 const int rw
= bio_data_dir(bi
);
3981 if (unlikely(bio_rw_flagged(bi
, BIO_RW_BARRIER
))) {
3982 /* Drain all pending writes. We only really need
3983 * to ensure they have been submitted, but this is
3986 mddev
->pers
->quiesce(mddev
, 1);
3987 mddev
->pers
->quiesce(mddev
, 0);
3988 md_barrier_request(mddev
, bi
);
3992 md_write_start(mddev
, bi
);
3995 mddev
->reshape_position
== MaxSector
&&
3996 chunk_aligned_read(mddev
,bi
))
3999 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4000 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
4002 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
4004 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
4006 int disks
, data_disks
;
4011 disks
= conf
->raid_disks
;
4012 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
4013 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
4014 /* spinlock is needed as reshape_progress may be
4015 * 64bit on a 32bit platform, and so it might be
4016 * possible to see a half-updated value
4017 * Ofcourse reshape_progress could change after
4018 * the lock is dropped, so once we get a reference
4019 * to the stripe that we think it is, we will have
4022 spin_lock_irq(&conf
->device_lock
);
4023 if (mddev
->delta_disks
< 0
4024 ? logical_sector
< conf
->reshape_progress
4025 : logical_sector
>= conf
->reshape_progress
) {
4026 disks
= conf
->previous_raid_disks
;
4029 if (mddev
->delta_disks
< 0
4030 ? logical_sector
< conf
->reshape_safe
4031 : logical_sector
>= conf
->reshape_safe
) {
4032 spin_unlock_irq(&conf
->device_lock
);
4037 spin_unlock_irq(&conf
->device_lock
);
4039 data_disks
= disks
- conf
->max_degraded
;
4041 new_sector
= raid5_compute_sector(conf
, logical_sector
,
4044 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4045 (unsigned long long)new_sector
,
4046 (unsigned long long)logical_sector
);
4048 sh
= get_active_stripe(conf
, new_sector
, previous
,
4049 (bi
->bi_rw
&RWA_MASK
), 0);
4051 if (unlikely(previous
)) {
4052 /* expansion might have moved on while waiting for a
4053 * stripe, so we must do the range check again.
4054 * Expansion could still move past after this
4055 * test, but as we are holding a reference to
4056 * 'sh', we know that if that happens,
4057 * STRIPE_EXPANDING will get set and the expansion
4058 * won't proceed until we finish with the stripe.
4061 spin_lock_irq(&conf
->device_lock
);
4062 if (mddev
->delta_disks
< 0
4063 ? logical_sector
>= conf
->reshape_progress
4064 : logical_sector
< conf
->reshape_progress
)
4065 /* mismatch, need to try again */
4067 spin_unlock_irq(&conf
->device_lock
);
4075 if (bio_data_dir(bi
) == WRITE
&&
4076 logical_sector
>= mddev
->suspend_lo
&&
4077 logical_sector
< mddev
->suspend_hi
) {
4079 /* As the suspend_* range is controlled by
4080 * userspace, we want an interruptible
4083 flush_signals(current
);
4084 prepare_to_wait(&conf
->wait_for_overlap
,
4085 &w
, TASK_INTERRUPTIBLE
);
4086 if (logical_sector
>= mddev
->suspend_lo
&&
4087 logical_sector
< mddev
->suspend_hi
)
4092 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
4093 !add_stripe_bio(sh
, bi
, dd_idx
, (bi
->bi_rw
&RW_MASK
))) {
4094 /* Stripe is busy expanding or
4095 * add failed due to overlap. Flush everything
4098 raid5_unplug_device(conf
);
4103 finish_wait(&conf
->wait_for_overlap
, &w
);
4104 set_bit(STRIPE_HANDLE
, &sh
->state
);
4105 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4106 if (mddev
->barrier
&&
4107 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4108 atomic_inc(&conf
->preread_active_stripes
);
4111 /* cannot get stripe for read-ahead, just give-up */
4112 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
4113 finish_wait(&conf
->wait_for_overlap
, &w
);
4118 spin_lock_irq(&conf
->device_lock
);
4119 remaining
= raid5_dec_bi_phys_segments(bi
);
4120 spin_unlock_irq(&conf
->device_lock
);
4121 if (remaining
== 0) {
4124 md_write_end(mddev
);
4129 if (mddev
->barrier
) {
4130 /* We need to wait for the stripes to all be handled.
4131 * So: wait for preread_active_stripes to drop to 0.
4133 wait_event(mddev
->thread
->wqueue
,
4134 atomic_read(&conf
->preread_active_stripes
) == 0);
4139 static sector_t
raid5_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
);
4141 static sector_t
reshape_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
)
4143 /* reshaping is quite different to recovery/resync so it is
4144 * handled quite separately ... here.
4146 * On each call to sync_request, we gather one chunk worth of
4147 * destination stripes and flag them as expanding.
4148 * Then we find all the source stripes and request reads.
4149 * As the reads complete, handle_stripe will copy the data
4150 * into the destination stripe and release that stripe.
4152 raid5_conf_t
*conf
= mddev
->private;
4153 struct stripe_head
*sh
;
4154 sector_t first_sector
, last_sector
;
4155 int raid_disks
= conf
->previous_raid_disks
;
4156 int data_disks
= raid_disks
- conf
->max_degraded
;
4157 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
4160 sector_t writepos
, readpos
, safepos
;
4161 sector_t stripe_addr
;
4162 int reshape_sectors
;
4163 struct list_head stripes
;
4165 if (sector_nr
== 0) {
4166 /* If restarting in the middle, skip the initial sectors */
4167 if (mddev
->delta_disks
< 0 &&
4168 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
4169 sector_nr
= raid5_size(mddev
, 0, 0)
4170 - conf
->reshape_progress
;
4171 } else if (mddev
->delta_disks
>= 0 &&
4172 conf
->reshape_progress
> 0)
4173 sector_nr
= conf
->reshape_progress
;
4174 sector_div(sector_nr
, new_data_disks
);
4176 mddev
->curr_resync_completed
= sector_nr
;
4177 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4183 /* We need to process a full chunk at a time.
4184 * If old and new chunk sizes differ, we need to process the
4187 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
)
4188 reshape_sectors
= mddev
->new_chunk_sectors
;
4190 reshape_sectors
= mddev
->chunk_sectors
;
4192 /* we update the metadata when there is more than 3Meg
4193 * in the block range (that is rather arbitrary, should
4194 * probably be time based) or when the data about to be
4195 * copied would over-write the source of the data at
4196 * the front of the range.
4197 * i.e. one new_stripe along from reshape_progress new_maps
4198 * to after where reshape_safe old_maps to
4200 writepos
= conf
->reshape_progress
;
4201 sector_div(writepos
, new_data_disks
);
4202 readpos
= conf
->reshape_progress
;
4203 sector_div(readpos
, data_disks
);
4204 safepos
= conf
->reshape_safe
;
4205 sector_div(safepos
, data_disks
);
4206 if (mddev
->delta_disks
< 0) {
4207 writepos
-= min_t(sector_t
, reshape_sectors
, writepos
);
4208 readpos
+= reshape_sectors
;
4209 safepos
+= reshape_sectors
;
4211 writepos
+= reshape_sectors
;
4212 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
4213 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
4216 /* 'writepos' is the most advanced device address we might write.
4217 * 'readpos' is the least advanced device address we might read.
4218 * 'safepos' is the least address recorded in the metadata as having
4220 * If 'readpos' is behind 'writepos', then there is no way that we can
4221 * ensure safety in the face of a crash - that must be done by userspace
4222 * making a backup of the data. So in that case there is no particular
4223 * rush to update metadata.
4224 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4225 * update the metadata to advance 'safepos' to match 'readpos' so that
4226 * we can be safe in the event of a crash.
4227 * So we insist on updating metadata if safepos is behind writepos and
4228 * readpos is beyond writepos.
4229 * In any case, update the metadata every 10 seconds.
4230 * Maybe that number should be configurable, but I'm not sure it is
4231 * worth it.... maybe it could be a multiple of safemode_delay???
4233 if ((mddev
->delta_disks
< 0
4234 ? (safepos
> writepos
&& readpos
< writepos
)
4235 : (safepos
< writepos
&& readpos
> writepos
)) ||
4236 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4237 /* Cannot proceed until we've updated the superblock... */
4238 wait_event(conf
->wait_for_overlap
,
4239 atomic_read(&conf
->reshape_stripes
)==0);
4240 mddev
->reshape_position
= conf
->reshape_progress
;
4241 mddev
->curr_resync_completed
= mddev
->curr_resync
;
4242 conf
->reshape_checkpoint
= jiffies
;
4243 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4244 md_wakeup_thread(mddev
->thread
);
4245 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4246 kthread_should_stop());
4247 spin_lock_irq(&conf
->device_lock
);
4248 conf
->reshape_safe
= mddev
->reshape_position
;
4249 spin_unlock_irq(&conf
->device_lock
);
4250 wake_up(&conf
->wait_for_overlap
);
4251 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4254 if (mddev
->delta_disks
< 0) {
4255 BUG_ON(conf
->reshape_progress
== 0);
4256 stripe_addr
= writepos
;
4257 BUG_ON((mddev
->dev_sectors
&
4258 ~((sector_t
)reshape_sectors
- 1))
4259 - reshape_sectors
- stripe_addr
4262 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
4263 stripe_addr
= sector_nr
;
4265 INIT_LIST_HEAD(&stripes
);
4266 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
4268 int skipped_disk
= 0;
4269 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
4270 set_bit(STRIPE_EXPANDING
, &sh
->state
);
4271 atomic_inc(&conf
->reshape_stripes
);
4272 /* If any of this stripe is beyond the end of the old
4273 * array, then we need to zero those blocks
4275 for (j
=sh
->disks
; j
--;) {
4277 if (j
== sh
->pd_idx
)
4279 if (conf
->level
== 6 &&
4282 s
= compute_blocknr(sh
, j
, 0);
4283 if (s
< raid5_size(mddev
, 0, 0)) {
4287 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
4288 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
4289 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
4291 if (!skipped_disk
) {
4292 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4293 set_bit(STRIPE_HANDLE
, &sh
->state
);
4295 list_add(&sh
->lru
, &stripes
);
4297 spin_lock_irq(&conf
->device_lock
);
4298 if (mddev
->delta_disks
< 0)
4299 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
4301 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
4302 spin_unlock_irq(&conf
->device_lock
);
4303 /* Ok, those stripe are ready. We can start scheduling
4304 * reads on the source stripes.
4305 * The source stripes are determined by mapping the first and last
4306 * block on the destination stripes.
4309 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
4312 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
4313 * new_data_disks
- 1),
4315 if (last_sector
>= mddev
->dev_sectors
)
4316 last_sector
= mddev
->dev_sectors
- 1;
4317 while (first_sector
<= last_sector
) {
4318 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
4319 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
4320 set_bit(STRIPE_HANDLE
, &sh
->state
);
4322 first_sector
+= STRIPE_SECTORS
;
4324 /* Now that the sources are clearly marked, we can release
4325 * the destination stripes
4327 while (!list_empty(&stripes
)) {
4328 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
4329 list_del_init(&sh
->lru
);
4332 /* If this takes us to the resync_max point where we have to pause,
4333 * then we need to write out the superblock.
4335 sector_nr
+= reshape_sectors
;
4336 if ((sector_nr
- mddev
->curr_resync_completed
) * 2
4337 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
4338 /* Cannot proceed until we've updated the superblock... */
4339 wait_event(conf
->wait_for_overlap
,
4340 atomic_read(&conf
->reshape_stripes
) == 0);
4341 mddev
->reshape_position
= conf
->reshape_progress
;
4342 mddev
->curr_resync_completed
= mddev
->curr_resync
+ reshape_sectors
;
4343 conf
->reshape_checkpoint
= jiffies
;
4344 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4345 md_wakeup_thread(mddev
->thread
);
4346 wait_event(mddev
->sb_wait
,
4347 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
4348 || kthread_should_stop());
4349 spin_lock_irq(&conf
->device_lock
);
4350 conf
->reshape_safe
= mddev
->reshape_position
;
4351 spin_unlock_irq(&conf
->device_lock
);
4352 wake_up(&conf
->wait_for_overlap
);
4353 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4355 return reshape_sectors
;
4358 /* FIXME go_faster isn't used */
4359 static inline sector_t
sync_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
4361 raid5_conf_t
*conf
= mddev
->private;
4362 struct stripe_head
*sh
;
4363 sector_t max_sector
= mddev
->dev_sectors
;
4365 int still_degraded
= 0;
4368 if (sector_nr
>= max_sector
) {
4369 /* just being told to finish up .. nothing much to do */
4370 unplug_slaves(mddev
);
4372 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
4377 if (mddev
->curr_resync
< max_sector
) /* aborted */
4378 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
4380 else /* completed sync */
4382 bitmap_close_sync(mddev
->bitmap
);
4387 /* Allow raid5_quiesce to complete */
4388 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
4390 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
4391 return reshape_request(mddev
, sector_nr
, skipped
);
4393 /* No need to check resync_max as we never do more than one
4394 * stripe, and as resync_max will always be on a chunk boundary,
4395 * if the check in md_do_sync didn't fire, there is no chance
4396 * of overstepping resync_max here
4399 /* if there is too many failed drives and we are trying
4400 * to resync, then assert that we are finished, because there is
4401 * nothing we can do.
4403 if (mddev
->degraded
>= conf
->max_degraded
&&
4404 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
4405 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
4409 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
4410 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
4411 !conf
->fullsync
&& sync_blocks
>= STRIPE_SECTORS
) {
4412 /* we can skip this block, and probably more */
4413 sync_blocks
/= STRIPE_SECTORS
;
4415 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
4419 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
4421 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
4423 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
4424 /* make sure we don't swamp the stripe cache if someone else
4425 * is trying to get access
4427 schedule_timeout_uninterruptible(1);
4429 /* Need to check if array will still be degraded after recovery/resync
4430 * We don't need to check the 'failed' flag as when that gets set,
4433 for (i
= 0; i
< conf
->raid_disks
; i
++)
4434 if (conf
->disks
[i
].rdev
== NULL
)
4437 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
4439 spin_lock(&sh
->lock
);
4440 set_bit(STRIPE_SYNCING
, &sh
->state
);
4441 clear_bit(STRIPE_INSYNC
, &sh
->state
);
4442 spin_unlock(&sh
->lock
);
4447 return STRIPE_SECTORS
;
4450 static int retry_aligned_read(raid5_conf_t
*conf
, struct bio
*raid_bio
)
4452 /* We may not be able to submit a whole bio at once as there
4453 * may not be enough stripe_heads available.
4454 * We cannot pre-allocate enough stripe_heads as we may need
4455 * more than exist in the cache (if we allow ever large chunks).
4456 * So we do one stripe head at a time and record in
4457 * ->bi_hw_segments how many have been done.
4459 * We *know* that this entire raid_bio is in one chunk, so
4460 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4462 struct stripe_head
*sh
;
4464 sector_t sector
, logical_sector
, last_sector
;
4469 logical_sector
= raid_bio
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4470 sector
= raid5_compute_sector(conf
, logical_sector
,
4472 last_sector
= raid_bio
->bi_sector
+ (raid_bio
->bi_size
>>9);
4474 for (; logical_sector
< last_sector
;
4475 logical_sector
+= STRIPE_SECTORS
,
4476 sector
+= STRIPE_SECTORS
,
4479 if (scnt
< raid5_bi_hw_segments(raid_bio
))
4480 /* already done this stripe */
4483 sh
= get_active_stripe(conf
, sector
, 0, 1, 0);
4486 /* failed to get a stripe - must wait */
4487 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4488 conf
->retry_read_aligned
= raid_bio
;
4492 set_bit(R5_ReadError
, &sh
->dev
[dd_idx
].flags
);
4493 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0)) {
4495 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4496 conf
->retry_read_aligned
= raid_bio
;
4504 spin_lock_irq(&conf
->device_lock
);
4505 remaining
= raid5_dec_bi_phys_segments(raid_bio
);
4506 spin_unlock_irq(&conf
->device_lock
);
4508 bio_endio(raid_bio
, 0);
4509 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4510 wake_up(&conf
->wait_for_stripe
);
4516 * This is our raid5 kernel thread.
4518 * We scan the hash table for stripes which can be handled now.
4519 * During the scan, completed stripes are saved for us by the interrupt
4520 * handler, so that they will not have to wait for our next wakeup.
4522 static void raid5d(mddev_t
*mddev
)
4524 struct stripe_head
*sh
;
4525 raid5_conf_t
*conf
= mddev
->private;
4528 pr_debug("+++ raid5d active\n");
4530 md_check_recovery(mddev
);
4533 spin_lock_irq(&conf
->device_lock
);
4537 if (conf
->seq_flush
!= conf
->seq_write
) {
4538 int seq
= conf
->seq_flush
;
4539 spin_unlock_irq(&conf
->device_lock
);
4540 bitmap_unplug(mddev
->bitmap
);
4541 spin_lock_irq(&conf
->device_lock
);
4542 conf
->seq_write
= seq
;
4543 activate_bit_delay(conf
);
4546 while ((bio
= remove_bio_from_retry(conf
))) {
4548 spin_unlock_irq(&conf
->device_lock
);
4549 ok
= retry_aligned_read(conf
, bio
);
4550 spin_lock_irq(&conf
->device_lock
);
4556 sh
= __get_priority_stripe(conf
);
4560 spin_unlock_irq(&conf
->device_lock
);
4567 spin_lock_irq(&conf
->device_lock
);
4569 pr_debug("%d stripes handled\n", handled
);
4571 spin_unlock_irq(&conf
->device_lock
);
4573 async_tx_issue_pending_all();
4574 unplug_slaves(mddev
);
4576 pr_debug("--- raid5d inactive\n");
4580 raid5_show_stripe_cache_size(mddev_t
*mddev
, char *page
)
4582 raid5_conf_t
*conf
= mddev
->private;
4584 return sprintf(page
, "%d\n", conf
->max_nr_stripes
);
4590 raid5_set_cache_size(mddev_t
*mddev
, int size
)
4592 raid5_conf_t
*conf
= mddev
->private;
4595 if (size
<= 16 || size
> 32768)
4597 while (size
< conf
->max_nr_stripes
) {
4598 if (drop_one_stripe(conf
))
4599 conf
->max_nr_stripes
--;
4603 err
= md_allow_write(mddev
);
4606 while (size
> conf
->max_nr_stripes
) {
4607 if (grow_one_stripe(conf
))
4608 conf
->max_nr_stripes
++;
4613 EXPORT_SYMBOL(raid5_set_cache_size
);
4616 raid5_store_stripe_cache_size(mddev_t
*mddev
, const char *page
, size_t len
)
4618 raid5_conf_t
*conf
= mddev
->private;
4622 if (len
>= PAGE_SIZE
)
4627 if (strict_strtoul(page
, 10, &new))
4629 err
= raid5_set_cache_size(mddev
, new);
4635 static struct md_sysfs_entry
4636 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
4637 raid5_show_stripe_cache_size
,
4638 raid5_store_stripe_cache_size
);
4641 raid5_show_preread_threshold(mddev_t
*mddev
, char *page
)
4643 raid5_conf_t
*conf
= mddev
->private;
4645 return sprintf(page
, "%d\n", conf
->bypass_threshold
);
4651 raid5_store_preread_threshold(mddev_t
*mddev
, const char *page
, size_t len
)
4653 raid5_conf_t
*conf
= mddev
->private;
4655 if (len
>= PAGE_SIZE
)
4660 if (strict_strtoul(page
, 10, &new))
4662 if (new > conf
->max_nr_stripes
)
4664 conf
->bypass_threshold
= new;
4668 static struct md_sysfs_entry
4669 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
4671 raid5_show_preread_threshold
,
4672 raid5_store_preread_threshold
);
4675 stripe_cache_active_show(mddev_t
*mddev
, char *page
)
4677 raid5_conf_t
*conf
= mddev
->private;
4679 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
4684 static struct md_sysfs_entry
4685 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
4687 static struct attribute
*raid5_attrs
[] = {
4688 &raid5_stripecache_size
.attr
,
4689 &raid5_stripecache_active
.attr
,
4690 &raid5_preread_bypass_threshold
.attr
,
4693 static struct attribute_group raid5_attrs_group
= {
4695 .attrs
= raid5_attrs
,
4699 raid5_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
)
4701 raid5_conf_t
*conf
= mddev
->private;
4704 sectors
= mddev
->dev_sectors
;
4706 /* size is defined by the smallest of previous and new size */
4707 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
4709 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
4710 sectors
&= ~((sector_t
)mddev
->new_chunk_sectors
- 1);
4711 return sectors
* (raid_disks
- conf
->max_degraded
);
4714 static void raid5_free_percpu(raid5_conf_t
*conf
)
4716 struct raid5_percpu
*percpu
;
4723 for_each_possible_cpu(cpu
) {
4724 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4725 safe_put_page(percpu
->spare_page
);
4726 kfree(percpu
->scribble
);
4728 #ifdef CONFIG_HOTPLUG_CPU
4729 unregister_cpu_notifier(&conf
->cpu_notify
);
4733 free_percpu(conf
->percpu
);
4736 static void free_conf(raid5_conf_t
*conf
)
4738 shrink_stripes(conf
);
4739 raid5_free_percpu(conf
);
4741 kfree(conf
->stripe_hashtbl
);
4745 #ifdef CONFIG_HOTPLUG_CPU
4746 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
4749 raid5_conf_t
*conf
= container_of(nfb
, raid5_conf_t
, cpu_notify
);
4750 long cpu
= (long)hcpu
;
4751 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4754 case CPU_UP_PREPARE
:
4755 case CPU_UP_PREPARE_FROZEN
:
4756 if (conf
->level
== 6 && !percpu
->spare_page
)
4757 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
4758 if (!percpu
->scribble
)
4759 percpu
->scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
4761 if (!percpu
->scribble
||
4762 (conf
->level
== 6 && !percpu
->spare_page
)) {
4763 safe_put_page(percpu
->spare_page
);
4764 kfree(percpu
->scribble
);
4765 pr_err("%s: failed memory allocation for cpu%ld\n",
4767 return notifier_from_errno(-ENOMEM
);
4771 case CPU_DEAD_FROZEN
:
4772 safe_put_page(percpu
->spare_page
);
4773 kfree(percpu
->scribble
);
4774 percpu
->spare_page
= NULL
;
4775 percpu
->scribble
= NULL
;
4784 static int raid5_alloc_percpu(raid5_conf_t
*conf
)
4787 struct page
*spare_page
;
4788 struct raid5_percpu __percpu
*allcpus
;
4792 allcpus
= alloc_percpu(struct raid5_percpu
);
4795 conf
->percpu
= allcpus
;
4799 for_each_present_cpu(cpu
) {
4800 if (conf
->level
== 6) {
4801 spare_page
= alloc_page(GFP_KERNEL
);
4806 per_cpu_ptr(conf
->percpu
, cpu
)->spare_page
= spare_page
;
4808 scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
4813 per_cpu_ptr(conf
->percpu
, cpu
)->scribble
= scribble
;
4815 #ifdef CONFIG_HOTPLUG_CPU
4816 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
4817 conf
->cpu_notify
.priority
= 0;
4819 err
= register_cpu_notifier(&conf
->cpu_notify
);
4826 static raid5_conf_t
*setup_conf(mddev_t
*mddev
)
4829 int raid_disk
, memory
, max_disks
;
4831 struct disk_info
*disk
;
4833 if (mddev
->new_level
!= 5
4834 && mddev
->new_level
!= 4
4835 && mddev
->new_level
!= 6) {
4836 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4837 mdname(mddev
), mddev
->new_level
);
4838 return ERR_PTR(-EIO
);
4840 if ((mddev
->new_level
== 5
4841 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
4842 (mddev
->new_level
== 6
4843 && !algorithm_valid_raid6(mddev
->new_layout
))) {
4844 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
4845 mdname(mddev
), mddev
->new_layout
);
4846 return ERR_PTR(-EIO
);
4848 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
4849 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4850 mdname(mddev
), mddev
->raid_disks
);
4851 return ERR_PTR(-EINVAL
);
4854 if (!mddev
->new_chunk_sectors
||
4855 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
4856 !is_power_of_2(mddev
->new_chunk_sectors
)) {
4857 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
4858 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
4859 return ERR_PTR(-EINVAL
);
4862 conf
= kzalloc(sizeof(raid5_conf_t
), GFP_KERNEL
);
4865 spin_lock_init(&conf
->device_lock
);
4866 init_waitqueue_head(&conf
->wait_for_stripe
);
4867 init_waitqueue_head(&conf
->wait_for_overlap
);
4868 INIT_LIST_HEAD(&conf
->handle_list
);
4869 INIT_LIST_HEAD(&conf
->hold_list
);
4870 INIT_LIST_HEAD(&conf
->delayed_list
);
4871 INIT_LIST_HEAD(&conf
->bitmap_list
);
4872 INIT_LIST_HEAD(&conf
->inactive_list
);
4873 atomic_set(&conf
->active_stripes
, 0);
4874 atomic_set(&conf
->preread_active_stripes
, 0);
4875 atomic_set(&conf
->active_aligned_reads
, 0);
4876 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
4878 conf
->raid_disks
= mddev
->raid_disks
;
4879 if (mddev
->reshape_position
== MaxSector
)
4880 conf
->previous_raid_disks
= mddev
->raid_disks
;
4882 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4883 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
4884 conf
->scribble_len
= scribble_len(max_disks
);
4886 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
4891 conf
->mddev
= mddev
;
4893 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
4896 conf
->level
= mddev
->new_level
;
4897 if (raid5_alloc_percpu(conf
) != 0)
4900 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
4902 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
4903 raid_disk
= rdev
->raid_disk
;
4904 if (raid_disk
>= max_disks
4907 disk
= conf
->disks
+ raid_disk
;
4911 if (test_bit(In_sync
, &rdev
->flags
)) {
4912 char b
[BDEVNAME_SIZE
];
4913 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
4915 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
4917 /* Cannot rely on bitmap to complete recovery */
4921 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
4922 conf
->level
= mddev
->new_level
;
4923 if (conf
->level
== 6)
4924 conf
->max_degraded
= 2;
4926 conf
->max_degraded
= 1;
4927 conf
->algorithm
= mddev
->new_layout
;
4928 conf
->max_nr_stripes
= NR_STRIPES
;
4929 conf
->reshape_progress
= mddev
->reshape_position
;
4930 if (conf
->reshape_progress
!= MaxSector
) {
4931 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
4932 conf
->prev_algo
= mddev
->layout
;
4935 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
4936 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
4937 if (grow_stripes(conf
, conf
->max_nr_stripes
)) {
4939 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4940 mdname(mddev
), memory
);
4943 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
4944 mdname(mddev
), memory
);
4946 conf
->thread
= md_register_thread(raid5d
, mddev
, NULL
);
4947 if (!conf
->thread
) {
4949 "md/raid:%s: couldn't allocate thread.\n",
4959 return ERR_PTR(-EIO
);
4961 return ERR_PTR(-ENOMEM
);
4965 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
4968 case ALGORITHM_PARITY_0
:
4969 if (raid_disk
< max_degraded
)
4972 case ALGORITHM_PARITY_N
:
4973 if (raid_disk
>= raid_disks
- max_degraded
)
4976 case ALGORITHM_PARITY_0_6
:
4977 if (raid_disk
== 0 ||
4978 raid_disk
== raid_disks
- 1)
4981 case ALGORITHM_LEFT_ASYMMETRIC_6
:
4982 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
4983 case ALGORITHM_LEFT_SYMMETRIC_6
:
4984 case ALGORITHM_RIGHT_SYMMETRIC_6
:
4985 if (raid_disk
== raid_disks
- 1)
4991 static int run(mddev_t
*mddev
)
4994 int working_disks
= 0, chunk_size
;
4995 int dirty_parity_disks
= 0;
4997 sector_t reshape_offset
= 0;
4999 if (mddev
->recovery_cp
!= MaxSector
)
5000 printk(KERN_NOTICE
"md/raid:%s: not clean"
5001 " -- starting background reconstruction\n",
5003 if (mddev
->reshape_position
!= MaxSector
) {
5004 /* Check that we can continue the reshape.
5005 * Currently only disks can change, it must
5006 * increase, and we must be past the point where
5007 * a stripe over-writes itself
5009 sector_t here_new
, here_old
;
5011 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
5013 if (mddev
->new_level
!= mddev
->level
) {
5014 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
5015 "required - aborting.\n",
5019 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
5020 /* reshape_position must be on a new-stripe boundary, and one
5021 * further up in new geometry must map after here in old
5024 here_new
= mddev
->reshape_position
;
5025 if (sector_div(here_new
, mddev
->new_chunk_sectors
*
5026 (mddev
->raid_disks
- max_degraded
))) {
5027 printk(KERN_ERR
"md/raid:%s: reshape_position not "
5028 "on a stripe boundary\n", mdname(mddev
));
5031 reshape_offset
= here_new
* mddev
->new_chunk_sectors
;
5032 /* here_new is the stripe we will write to */
5033 here_old
= mddev
->reshape_position
;
5034 sector_div(here_old
, mddev
->chunk_sectors
*
5035 (old_disks
-max_degraded
));
5036 /* here_old is the first stripe that we might need to read
5038 if (mddev
->delta_disks
== 0) {
5039 /* We cannot be sure it is safe to start an in-place
5040 * reshape. It is only safe if user-space if monitoring
5041 * and taking constant backups.
5042 * mdadm always starts a situation like this in
5043 * readonly mode so it can take control before
5044 * allowing any writes. So just check for that.
5046 if ((here_new
* mddev
->new_chunk_sectors
!=
5047 here_old
* mddev
->chunk_sectors
) ||
5049 printk(KERN_ERR
"md/raid:%s: in-place reshape must be started"
5050 " in read-only mode - aborting\n",
5054 } else if (mddev
->delta_disks
< 0
5055 ? (here_new
* mddev
->new_chunk_sectors
<=
5056 here_old
* mddev
->chunk_sectors
)
5057 : (here_new
* mddev
->new_chunk_sectors
>=
5058 here_old
* mddev
->chunk_sectors
)) {
5059 /* Reading from the same stripe as writing to - bad */
5060 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
5061 "auto-recovery - aborting.\n",
5065 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
5067 /* OK, we should be able to continue; */
5069 BUG_ON(mddev
->level
!= mddev
->new_level
);
5070 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
5071 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
5072 BUG_ON(mddev
->delta_disks
!= 0);
5075 if (mddev
->private == NULL
)
5076 conf
= setup_conf(mddev
);
5078 conf
= mddev
->private;
5081 return PTR_ERR(conf
);
5083 mddev
->thread
= conf
->thread
;
5084 conf
->thread
= NULL
;
5085 mddev
->private = conf
;
5088 * 0 for a fully functional array, 1 or 2 for a degraded array.
5090 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
5091 if (rdev
->raid_disk
< 0)
5093 if (test_bit(In_sync
, &rdev
->flags
)) {
5097 /* This disc is not fully in-sync. However if it
5098 * just stored parity (beyond the recovery_offset),
5099 * when we don't need to be concerned about the
5100 * array being dirty.
5101 * When reshape goes 'backwards', we never have
5102 * partially completed devices, so we only need
5103 * to worry about reshape going forwards.
5105 /* Hack because v0.91 doesn't store recovery_offset properly. */
5106 if (mddev
->major_version
== 0 &&
5107 mddev
->minor_version
> 90)
5108 rdev
->recovery_offset
= reshape_offset
;
5110 if (rdev
->recovery_offset
< reshape_offset
) {
5111 /* We need to check old and new layout */
5112 if (!only_parity(rdev
->raid_disk
,
5115 conf
->max_degraded
))
5118 if (!only_parity(rdev
->raid_disk
,
5120 conf
->previous_raid_disks
,
5121 conf
->max_degraded
))
5123 dirty_parity_disks
++;
5126 mddev
->degraded
= (max(conf
->raid_disks
, conf
->previous_raid_disks
)
5129 if (has_failed(conf
)) {
5130 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
5131 " (%d/%d failed)\n",
5132 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
5136 /* device size must be a multiple of chunk size */
5137 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
5138 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
5140 if (mddev
->degraded
> dirty_parity_disks
&&
5141 mddev
->recovery_cp
!= MaxSector
) {
5142 if (mddev
->ok_start_degraded
)
5144 "md/raid:%s: starting dirty degraded array"
5145 " - data corruption possible.\n",
5149 "md/raid:%s: cannot start dirty degraded array.\n",
5155 if (mddev
->degraded
== 0)
5156 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
5157 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
5158 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
5161 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
5162 " out of %d devices, algorithm %d\n",
5163 mdname(mddev
), conf
->level
,
5164 mddev
->raid_disks
- mddev
->degraded
,
5165 mddev
->raid_disks
, mddev
->new_layout
);
5167 print_raid5_conf(conf
);
5169 if (conf
->reshape_progress
!= MaxSector
) {
5170 conf
->reshape_safe
= conf
->reshape_progress
;
5171 atomic_set(&conf
->reshape_stripes
, 0);
5172 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5173 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5174 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5175 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5176 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5181 /* Ok, everything is just fine now */
5182 if (mddev
->to_remove
== &raid5_attrs_group
)
5183 mddev
->to_remove
= NULL
;
5184 else if (mddev
->kobj
.sd
&&
5185 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
5187 "raid5: failed to create sysfs attributes for %s\n",
5189 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5191 plugger_init(&conf
->plug
, raid5_unplug
);
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
;
5209 mddev
->queue
->queue_lock
= &conf
->device_lock
;
5211 mddev
->queue
->unplug_fn
= raid5_unplug_queue
;
5213 chunk_size
= mddev
->chunk_sectors
<< 9;
5214 blk_queue_io_min(mddev
->queue
, chunk_size
);
5215 blk_queue_io_opt(mddev
->queue
, chunk_size
*
5216 (conf
->raid_disks
- conf
->max_degraded
));
5218 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
5219 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
5220 rdev
->data_offset
<< 9);
5224 md_unregister_thread(mddev
->thread
);
5225 mddev
->thread
= NULL
;
5227 print_raid5_conf(conf
);
5230 mddev
->private = NULL
;
5231 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
5235 static int stop(mddev_t
*mddev
)
5237 raid5_conf_t
*conf
= mddev
->private;
5239 md_unregister_thread(mddev
->thread
);
5240 mddev
->thread
= NULL
;
5242 mddev
->queue
->backing_dev_info
.congested_fn
= NULL
;
5243 plugger_flush(&conf
->plug
); /* the unplug fn references 'conf'*/
5245 mddev
->private = NULL
;
5246 mddev
->to_remove
= &raid5_attrs_group
;
5251 static void print_sh(struct seq_file
*seq
, struct stripe_head
*sh
)
5255 seq_printf(seq
, "sh %llu, pd_idx %d, state %ld.\n",
5256 (unsigned long long)sh
->sector
, sh
->pd_idx
, sh
->state
);
5257 seq_printf(seq
, "sh %llu, count %d.\n",
5258 (unsigned long long)sh
->sector
, atomic_read(&sh
->count
));
5259 seq_printf(seq
, "sh %llu, ", (unsigned long long)sh
->sector
);
5260 for (i
= 0; i
< sh
->disks
; i
++) {
5261 seq_printf(seq
, "(cache%d: %p %ld) ",
5262 i
, sh
->dev
[i
].page
, sh
->dev
[i
].flags
);
5264 seq_printf(seq
, "\n");
5267 static void printall(struct seq_file
*seq
, raid5_conf_t
*conf
)
5269 struct stripe_head
*sh
;
5270 struct hlist_node
*hn
;
5273 spin_lock_irq(&conf
->device_lock
);
5274 for (i
= 0; i
< NR_HASH
; i
++) {
5275 hlist_for_each_entry(sh
, hn
, &conf
->stripe_hashtbl
[i
], hash
) {
5276 if (sh
->raid_conf
!= conf
)
5281 spin_unlock_irq(&conf
->device_lock
);
5285 static void status(struct seq_file
*seq
, mddev_t
*mddev
)
5287 raid5_conf_t
*conf
= mddev
->private;
5290 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
5291 mddev
->chunk_sectors
/ 2, mddev
->layout
);
5292 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
5293 for (i
= 0; i
< conf
->raid_disks
; i
++)
5294 seq_printf (seq
, "%s",
5295 conf
->disks
[i
].rdev
&&
5296 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
5297 seq_printf (seq
, "]");
5299 seq_printf (seq
, "\n");
5300 printall(seq
, conf
);
5304 static void print_raid5_conf (raid5_conf_t
*conf
)
5307 struct disk_info
*tmp
;
5309 printk(KERN_DEBUG
"RAID conf printout:\n");
5311 printk("(conf==NULL)\n");
5314 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
5316 conf
->raid_disks
- conf
->mddev
->degraded
);
5318 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5319 char b
[BDEVNAME_SIZE
];
5320 tmp
= conf
->disks
+ i
;
5322 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
5323 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
5324 bdevname(tmp
->rdev
->bdev
, b
));
5328 static int raid5_spare_active(mddev_t
*mddev
)
5331 raid5_conf_t
*conf
= mddev
->private;
5332 struct disk_info
*tmp
;
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
)) {
5340 unsigned long flags
;
5341 spin_lock_irqsave(&conf
->device_lock
, flags
);
5343 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5346 print_raid5_conf(conf
);
5350 static int raid5_remove_disk(mddev_t
*mddev
, int number
)
5352 raid5_conf_t
*conf
= mddev
->private;
5355 struct disk_info
*p
= conf
->disks
+ number
;
5357 print_raid5_conf(conf
);
5360 if (number
>= conf
->raid_disks
&&
5361 conf
->reshape_progress
== MaxSector
)
5362 clear_bit(In_sync
, &rdev
->flags
);
5364 if (test_bit(In_sync
, &rdev
->flags
) ||
5365 atomic_read(&rdev
->nr_pending
)) {
5369 /* Only remove non-faulty devices if recovery
5372 if (!test_bit(Faulty
, &rdev
->flags
) &&
5373 !has_failed(conf
) &&
5374 number
< conf
->raid_disks
) {
5380 if (atomic_read(&rdev
->nr_pending
)) {
5381 /* lost the race, try later */
5388 print_raid5_conf(conf
);
5392 static int raid5_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
5394 raid5_conf_t
*conf
= mddev
->private;
5397 struct disk_info
*p
;
5399 int last
= conf
->raid_disks
- 1;
5401 if (has_failed(conf
))
5402 /* no point adding a device */
5405 if (rdev
->raid_disk
>= 0)
5406 first
= last
= rdev
->raid_disk
;
5409 * find the disk ... but prefer rdev->saved_raid_disk
5412 if (rdev
->saved_raid_disk
>= 0 &&
5413 rdev
->saved_raid_disk
>= first
&&
5414 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
5415 disk
= rdev
->saved_raid_disk
;
5418 for ( ; disk
<= last
; disk
++)
5419 if ((p
=conf
->disks
+ disk
)->rdev
== NULL
) {
5420 clear_bit(In_sync
, &rdev
->flags
);
5421 rdev
->raid_disk
= disk
;
5423 if (rdev
->saved_raid_disk
!= disk
)
5425 rcu_assign_pointer(p
->rdev
, rdev
);
5428 print_raid5_conf(conf
);
5432 static int raid5_resize(mddev_t
*mddev
, sector_t sectors
)
5434 /* no resync is happening, and there is enough space
5435 * on all devices, so we can resize.
5436 * We need to make sure resync covers any new space.
5437 * If the array is shrinking we should possibly wait until
5438 * any io in the removed space completes, but it hardly seems
5441 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
5442 md_set_array_sectors(mddev
, raid5_size(mddev
, sectors
,
5443 mddev
->raid_disks
));
5444 if (mddev
->array_sectors
>
5445 raid5_size(mddev
, sectors
, mddev
->raid_disks
))
5447 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5448 revalidate_disk(mddev
->gendisk
);
5449 if (sectors
> mddev
->dev_sectors
&& mddev
->recovery_cp
== MaxSector
) {
5450 mddev
->recovery_cp
= mddev
->dev_sectors
;
5451 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
5453 mddev
->dev_sectors
= sectors
;
5454 mddev
->resync_max_sectors
= sectors
;
5458 static int check_stripe_cache(mddev_t
*mddev
)
5460 /* Can only proceed if there are plenty of stripe_heads.
5461 * We need a minimum of one full stripe,, and for sensible progress
5462 * it is best to have about 4 times that.
5463 * If we require 4 times, then the default 256 4K stripe_heads will
5464 * allow for chunk sizes up to 256K, which is probably OK.
5465 * If the chunk size is greater, user-space should request more
5466 * stripe_heads first.
5468 raid5_conf_t
*conf
= mddev
->private;
5469 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5470 > conf
->max_nr_stripes
||
5471 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5472 > conf
->max_nr_stripes
) {
5473 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5475 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
5482 static int check_reshape(mddev_t
*mddev
)
5484 raid5_conf_t
*conf
= mddev
->private;
5486 if (mddev
->delta_disks
== 0 &&
5487 mddev
->new_layout
== mddev
->layout
&&
5488 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
5489 return 0; /* nothing to do */
5491 /* Cannot grow a bitmap yet */
5493 if (has_failed(conf
))
5495 if (mddev
->delta_disks
< 0) {
5496 /* We might be able to shrink, but the devices must
5497 * be made bigger first.
5498 * For raid6, 4 is the minimum size.
5499 * Otherwise 2 is the minimum
5502 if (mddev
->level
== 6)
5504 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
5508 if (!check_stripe_cache(mddev
))
5511 return resize_stripes(conf
, conf
->raid_disks
+ mddev
->delta_disks
);
5514 static int raid5_start_reshape(mddev_t
*mddev
)
5516 raid5_conf_t
*conf
= mddev
->private;
5519 int added_devices
= 0;
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 (rdev
->raid_disk
< 0 &&
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 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
5575 if (rdev
->raid_disk
< 0 &&
5576 !test_bit(Faulty
, &rdev
->flags
)) {
5577 if (raid5_add_disk(mddev
, rdev
) == 0) {
5579 if (rdev
->raid_disk
>= conf
->previous_raid_disks
) {
5580 set_bit(In_sync
, &rdev
->flags
);
5583 rdev
->recovery_offset
= 0;
5584 sprintf(nm
, "rd%d", rdev
->raid_disk
);
5585 if (sysfs_create_link(&mddev
->kobj
,
5587 /* Failure here is OK */;
5592 /* When a reshape changes the number of devices, ->degraded
5593 * is measured against the larger of the pre and post number of
5595 if (mddev
->delta_disks
> 0) {
5596 spin_lock_irqsave(&conf
->device_lock
, flags
);
5597 mddev
->degraded
+= (conf
->raid_disks
- conf
->previous_raid_disks
)
5599 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5601 mddev
->raid_disks
= conf
->raid_disks
;
5602 mddev
->reshape_position
= conf
->reshape_progress
;
5603 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5605 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5606 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5607 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5608 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5609 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5611 if (!mddev
->sync_thread
) {
5612 mddev
->recovery
= 0;
5613 spin_lock_irq(&conf
->device_lock
);
5614 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
5615 conf
->reshape_progress
= MaxSector
;
5616 spin_unlock_irq(&conf
->device_lock
);
5619 conf
->reshape_checkpoint
= jiffies
;
5620 md_wakeup_thread(mddev
->sync_thread
);
5621 md_new_event(mddev
);
5625 /* This is called from the reshape thread and should make any
5626 * changes needed in 'conf'
5628 static void end_reshape(raid5_conf_t
*conf
)
5631 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
5633 spin_lock_irq(&conf
->device_lock
);
5634 conf
->previous_raid_disks
= conf
->raid_disks
;
5635 conf
->reshape_progress
= MaxSector
;
5636 spin_unlock_irq(&conf
->device_lock
);
5637 wake_up(&conf
->wait_for_overlap
);
5639 /* read-ahead size must cover two whole stripes, which is
5640 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5642 if (conf
->mddev
->queue
) {
5643 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5644 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
5646 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
5647 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
5652 /* This is called from the raid5d thread with mddev_lock held.
5653 * It makes config changes to the device.
5655 static void raid5_finish_reshape(mddev_t
*mddev
)
5657 raid5_conf_t
*conf
= mddev
->private;
5659 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
5661 if (mddev
->delta_disks
> 0) {
5662 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5663 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5664 revalidate_disk(mddev
->gendisk
);
5667 mddev
->degraded
= conf
->raid_disks
;
5668 for (d
= 0; d
< conf
->raid_disks
; d
++)
5669 if (conf
->disks
[d
].rdev
&&
5671 &conf
->disks
[d
].rdev
->flags
))
5673 for (d
= conf
->raid_disks
;
5674 d
< conf
->raid_disks
- mddev
->delta_disks
;
5676 mdk_rdev_t
*rdev
= conf
->disks
[d
].rdev
;
5677 if (rdev
&& raid5_remove_disk(mddev
, d
) == 0) {
5679 sprintf(nm
, "rd%d", rdev
->raid_disk
);
5680 sysfs_remove_link(&mddev
->kobj
, nm
);
5681 rdev
->raid_disk
= -1;
5685 mddev
->layout
= conf
->algorithm
;
5686 mddev
->chunk_sectors
= conf
->chunk_sectors
;
5687 mddev
->reshape_position
= MaxSector
;
5688 mddev
->delta_disks
= 0;
5692 static void raid5_quiesce(mddev_t
*mddev
, int state
)
5694 raid5_conf_t
*conf
= mddev
->private;
5697 case 2: /* resume for a suspend */
5698 wake_up(&conf
->wait_for_overlap
);
5701 case 1: /* stop all writes */
5702 spin_lock_irq(&conf
->device_lock
);
5703 /* '2' tells resync/reshape to pause so that all
5704 * active stripes can drain
5707 wait_event_lock_irq(conf
->wait_for_stripe
,
5708 atomic_read(&conf
->active_stripes
) == 0 &&
5709 atomic_read(&conf
->active_aligned_reads
) == 0,
5710 conf
->device_lock
, /* nothing */);
5712 spin_unlock_irq(&conf
->device_lock
);
5713 /* allow reshape to continue */
5714 wake_up(&conf
->wait_for_overlap
);
5717 case 0: /* re-enable writes */
5718 spin_lock_irq(&conf
->device_lock
);
5720 wake_up(&conf
->wait_for_stripe
);
5721 wake_up(&conf
->wait_for_overlap
);
5722 spin_unlock_irq(&conf
->device_lock
);
5728 static void *raid45_takeover_raid0(mddev_t
*mddev
, int level
)
5730 struct raid0_private_data
*raid0_priv
= mddev
->private;
5732 /* for raid0 takeover only one zone is supported */
5733 if (raid0_priv
->nr_strip_zones
> 1) {
5734 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5736 return ERR_PTR(-EINVAL
);
5739 mddev
->new_level
= level
;
5740 mddev
->new_layout
= ALGORITHM_PARITY_N
;
5741 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
5742 mddev
->raid_disks
+= 1;
5743 mddev
->delta_disks
= 1;
5744 /* make sure it will be not marked as dirty */
5745 mddev
->recovery_cp
= MaxSector
;
5747 return setup_conf(mddev
);
5751 static void *raid5_takeover_raid1(mddev_t
*mddev
)
5755 if (mddev
->raid_disks
!= 2 ||
5756 mddev
->degraded
> 1)
5757 return ERR_PTR(-EINVAL
);
5759 /* Should check if there are write-behind devices? */
5761 chunksect
= 64*2; /* 64K by default */
5763 /* The array must be an exact multiple of chunksize */
5764 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
5767 if ((chunksect
<<9) < STRIPE_SIZE
)
5768 /* array size does not allow a suitable chunk size */
5769 return ERR_PTR(-EINVAL
);
5771 mddev
->new_level
= 5;
5772 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5773 mddev
->new_chunk_sectors
= chunksect
;
5775 return setup_conf(mddev
);
5778 static void *raid5_takeover_raid6(mddev_t
*mddev
)
5782 switch (mddev
->layout
) {
5783 case ALGORITHM_LEFT_ASYMMETRIC_6
:
5784 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
5786 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
5787 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
5789 case ALGORITHM_LEFT_SYMMETRIC_6
:
5790 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5792 case ALGORITHM_RIGHT_SYMMETRIC_6
:
5793 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
5795 case ALGORITHM_PARITY_0_6
:
5796 new_layout
= ALGORITHM_PARITY_0
;
5798 case ALGORITHM_PARITY_N
:
5799 new_layout
= ALGORITHM_PARITY_N
;
5802 return ERR_PTR(-EINVAL
);
5804 mddev
->new_level
= 5;
5805 mddev
->new_layout
= new_layout
;
5806 mddev
->delta_disks
= -1;
5807 mddev
->raid_disks
-= 1;
5808 return setup_conf(mddev
);
5812 static int raid5_check_reshape(mddev_t
*mddev
)
5814 /* For a 2-drive array, the layout and chunk size can be changed
5815 * immediately as not restriping is needed.
5816 * For larger arrays we record the new value - after validation
5817 * to be used by a reshape pass.
5819 raid5_conf_t
*conf
= mddev
->private;
5820 int new_chunk
= mddev
->new_chunk_sectors
;
5822 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
5824 if (new_chunk
> 0) {
5825 if (!is_power_of_2(new_chunk
))
5827 if (new_chunk
< (PAGE_SIZE
>>9))
5829 if (mddev
->array_sectors
& (new_chunk
-1))
5830 /* not factor of array size */
5834 /* They look valid */
5836 if (mddev
->raid_disks
== 2) {
5837 /* can make the change immediately */
5838 if (mddev
->new_layout
>= 0) {
5839 conf
->algorithm
= mddev
->new_layout
;
5840 mddev
->layout
= mddev
->new_layout
;
5842 if (new_chunk
> 0) {
5843 conf
->chunk_sectors
= new_chunk
;
5844 mddev
->chunk_sectors
= new_chunk
;
5846 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5847 md_wakeup_thread(mddev
->thread
);
5849 return check_reshape(mddev
);
5852 static int raid6_check_reshape(mddev_t
*mddev
)
5854 int new_chunk
= mddev
->new_chunk_sectors
;
5856 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
5858 if (new_chunk
> 0) {
5859 if (!is_power_of_2(new_chunk
))
5861 if (new_chunk
< (PAGE_SIZE
>> 9))
5863 if (mddev
->array_sectors
& (new_chunk
-1))
5864 /* not factor of array size */
5868 /* They look valid */
5869 return check_reshape(mddev
);
5872 static void *raid5_takeover(mddev_t
*mddev
)
5874 /* raid5 can take over:
5875 * raid0 - if there is only one strip zone - make it a raid4 layout
5876 * raid1 - if there are two drives. We need to know the chunk size
5877 * raid4 - trivial - just use a raid4 layout.
5878 * raid6 - Providing it is a *_6 layout
5880 if (mddev
->level
== 0)
5881 return raid45_takeover_raid0(mddev
, 5);
5882 if (mddev
->level
== 1)
5883 return raid5_takeover_raid1(mddev
);
5884 if (mddev
->level
== 4) {
5885 mddev
->new_layout
= ALGORITHM_PARITY_N
;
5886 mddev
->new_level
= 5;
5887 return setup_conf(mddev
);
5889 if (mddev
->level
== 6)
5890 return raid5_takeover_raid6(mddev
);
5892 return ERR_PTR(-EINVAL
);
5895 static void *raid4_takeover(mddev_t
*mddev
)
5897 /* raid4 can take over:
5898 * raid0 - if there is only one strip zone
5899 * raid5 - if layout is right
5901 if (mddev
->level
== 0)
5902 return raid45_takeover_raid0(mddev
, 4);
5903 if (mddev
->level
== 5 &&
5904 mddev
->layout
== ALGORITHM_PARITY_N
) {
5905 mddev
->new_layout
= 0;
5906 mddev
->new_level
= 4;
5907 return setup_conf(mddev
);
5909 return ERR_PTR(-EINVAL
);
5912 static struct mdk_personality raid5_personality
;
5914 static void *raid6_takeover(mddev_t
*mddev
)
5916 /* Currently can only take over a raid5. We map the
5917 * personality to an equivalent raid6 personality
5918 * with the Q block at the end.
5922 if (mddev
->pers
!= &raid5_personality
)
5923 return ERR_PTR(-EINVAL
);
5924 if (mddev
->degraded
> 1)
5925 return ERR_PTR(-EINVAL
);
5926 if (mddev
->raid_disks
> 253)
5927 return ERR_PTR(-EINVAL
);
5928 if (mddev
->raid_disks
< 3)
5929 return ERR_PTR(-EINVAL
);
5931 switch (mddev
->layout
) {
5932 case ALGORITHM_LEFT_ASYMMETRIC
:
5933 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
5935 case ALGORITHM_RIGHT_ASYMMETRIC
:
5936 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
5938 case ALGORITHM_LEFT_SYMMETRIC
:
5939 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
5941 case ALGORITHM_RIGHT_SYMMETRIC
:
5942 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
5944 case ALGORITHM_PARITY_0
:
5945 new_layout
= ALGORITHM_PARITY_0_6
;
5947 case ALGORITHM_PARITY_N
:
5948 new_layout
= ALGORITHM_PARITY_N
;
5951 return ERR_PTR(-EINVAL
);
5953 mddev
->new_level
= 6;
5954 mddev
->new_layout
= new_layout
;
5955 mddev
->delta_disks
= 1;
5956 mddev
->raid_disks
+= 1;
5957 return setup_conf(mddev
);
5961 static struct mdk_personality raid6_personality
=
5965 .owner
= THIS_MODULE
,
5966 .make_request
= make_request
,
5970 .error_handler
= error
,
5971 .hot_add_disk
= raid5_add_disk
,
5972 .hot_remove_disk
= raid5_remove_disk
,
5973 .spare_active
= raid5_spare_active
,
5974 .sync_request
= sync_request
,
5975 .resize
= raid5_resize
,
5977 .check_reshape
= raid6_check_reshape
,
5978 .start_reshape
= raid5_start_reshape
,
5979 .finish_reshape
= raid5_finish_reshape
,
5980 .quiesce
= raid5_quiesce
,
5981 .takeover
= raid6_takeover
,
5983 static struct mdk_personality raid5_personality
=
5987 .owner
= THIS_MODULE
,
5988 .make_request
= make_request
,
5992 .error_handler
= error
,
5993 .hot_add_disk
= raid5_add_disk
,
5994 .hot_remove_disk
= raid5_remove_disk
,
5995 .spare_active
= raid5_spare_active
,
5996 .sync_request
= sync_request
,
5997 .resize
= raid5_resize
,
5999 .check_reshape
= raid5_check_reshape
,
6000 .start_reshape
= raid5_start_reshape
,
6001 .finish_reshape
= raid5_finish_reshape
,
6002 .quiesce
= raid5_quiesce
,
6003 .takeover
= raid5_takeover
,
6006 static struct mdk_personality raid4_personality
=
6010 .owner
= THIS_MODULE
,
6011 .make_request
= make_request
,
6015 .error_handler
= error
,
6016 .hot_add_disk
= raid5_add_disk
,
6017 .hot_remove_disk
= raid5_remove_disk
,
6018 .spare_active
= raid5_spare_active
,
6019 .sync_request
= sync_request
,
6020 .resize
= raid5_resize
,
6022 .check_reshape
= raid5_check_reshape
,
6023 .start_reshape
= raid5_start_reshape
,
6024 .finish_reshape
= raid5_finish_reshape
,
6025 .quiesce
= raid5_quiesce
,
6026 .takeover
= raid4_takeover
,
6029 static int __init
raid5_init(void)
6031 register_md_personality(&raid6_personality
);
6032 register_md_personality(&raid5_personality
);
6033 register_md_personality(&raid4_personality
);
6037 static void raid5_exit(void)
6039 unregister_md_personality(&raid6_personality
);
6040 unregister_md_personality(&raid5_personality
);
6041 unregister_md_personality(&raid4_personality
);
6044 module_init(raid5_init
);
6045 module_exit(raid5_exit
);
6046 MODULE_LICENSE("GPL");
6047 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6048 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6049 MODULE_ALIAS("md-raid5");
6050 MODULE_ALIAS("md-raid4");
6051 MODULE_ALIAS("md-level-5");
6052 MODULE_ALIAS("md-level-4");
6053 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6054 MODULE_ALIAS("md-raid6");
6055 MODULE_ALIAS("md-level-6");
6057 /* This used to be two separate modules, they were: */
6058 MODULE_ALIAS("raid5");
6059 MODULE_ALIAS("raid6");